text
stringlengths 47
469k
| meta
dict | domain
stringclasses 1
value |
|---|---|---|
---
abstract: 'Two-mode squeezing and entanglement is obtained in a atom-cavity system cosisting a three-level atom and a two-mode cavity with driving laser fields. Here non-resonatn dressed-state transitions between the cavity modes and atom are used to get the effective Hamiltonian.'
author:
- 'Jinhua Zou,$^{1,2}$ [^1] Dahai Xu $^{2}$ and Huafeng Zhang,$^{2}$'
title: 'Generation of two-mode entanglement via atomic coherence created by non-resonant dressed-state transitions'
---
Introduction
============
Quantum entanglement is regarded as the magic part of mechanics and has became one central part in quantum information and quantum computation. Recently, continuous-variable (CV) entanglement has attracted much great interest in the application of entanglement-based quantum information processing [@1]. Thus generation continuous-variable entanglement have gained much attention since they can provide useful entanglement sources for quantum teleportation \[2\], quantum dense coding \[3\] and quantum swapping \[4\]. There are many interaction system candidates to generate CV entanglement such as cavity QED system \[5-9\], nondegenerate parametric down-conversion optical crystals \[10,11\], NV centers system \[12,13\], and so on.
As for cavity QED system, it has been a good candidate to investigate the quantum interaction in physics and quantum mechanics. Many schemes have been presented to generate entangled states. For example, the scheme for generating CV entangled light via correlated emission laser has been proposed \[14\]. They obtain a two-mode entangled light with large photon number. The scheme to generate CV entanglement via a single atom laser has also been presented \[15,16\]. Among the entanglement generating schemes, some schemes use atomic coherence to build the entanglement between various cavity modes \[17,18\]. The atomic coherence can be input or created by driving related transitions. In these schemes, atomic coherence induces entanglement somewhat. One scheme is proposed in Ref. \[19\], a two-step operation to get highly entangled two-mode light by using the atoms as a reservior is proposed. They use the dressed states of the driving field and choose resonant transition for the two cavity modes and the squeeze-transformed modes successively interact with the dressed atoms. One of the limitations in this scheme are as following. First, the two-step procedure and different initial state preparation make the experimental setup complicated. It is useful to investigate the scheme that has one-step procedure and simple initial state preparation.
Motivated by this, we propose a scheme to extend the atomic system to a three-level atom, which has an auxiliary driving transition to generate the two-mode entanglement light. By introducing the auxiliary driving transition and choosing non-resonant atom-cavity interaction, the dressed atomic-cavity interaction can be an effective two-mode squeezing interaction, which can be used to generate two-mode entangled light.
Model and equations
===================
The system we consider is a three-level $\Lambda $ atomic system locates in a two-mode optical cavity with drivings. Two classical laser fields with Rabi frequency $\Omega _{j}$ and frequency $\omega _{d,j\text{ }}$resonantly drive the two atomic transitions $|j\rangle \leftrightarrow |3\rangle $, respectively for $j=1$, $2$. The two cavity modes with coupling constant $%
g_{n}$ and frequency $\omega _{c,n}$ coupling transitions $|2\rangle
\leftrightarrow |3\rangle $ ( $n=1,2$). The whole system can be described by the following Hamiltonian ($\hbar =1$) $$\begin{aligned}
H &=&H_{\Omega }+H_{c} \nonumber \\
H_{\Omega } &=&-\sum_{j=1,2}\Omega _{j}(|3\rangle \langle j|+|j\rangle
\langle 3|) \nonumber \\
H_{c} &=&\sum_{n=1,2}\delta _{n}a_{n}^{\dagger
}a_{n}+\sum_{n=1,2}[(g_{n}a_{n}\sigma _{32}+g_{n}^{*}a_{n}^{\dagger }\sigma
_{23}) \label{1}\end{aligned}$$ where $\Delta _{j}=\omega _{3j}-\omega _{d,j\text{ }}$( $j=1,2$) and $\delta
_{n}=\omega _{c,n}-\omega _{d,2\text{ }}$with $\omega _{3j}$ are atomic resonant frequency for transition $|j\rangle \leftrightarrow |3\rangle $ ( $%
j=1,2$). Atomic operator $\sigma _{jk}$ are flip operators when $j\neq k$ and are population operators when $j=k$. $a_{n}$ ($a_{n}^{\dagger }$) is the creation (annihilation) operator for the $n$th cavity mode.
Assume that the laser fields are much strong, we first calculate the dressed states created by part Hamiltonian $H_{\Omega }$. The eigenvalues are $%
\lambda _{0}=0$, $\lambda _{\pm }=\pm \sqrt{\Omega _{2}^{2}+\Omega _{1}^{2}}%
=\pm \Omega _{e}$, and related eigenstates are
$$\begin{aligned}
|0\rangle &=&\cos \theta |2\rangle -\sin \theta |1\rangle \nonumber \\
|+\rangle &=&\frac{1}{\sqrt{2}}|3\rangle +\frac{\sin \theta }{\sqrt{2}}%
|2\rangle +\frac{\sin \theta }{\sqrt{2}}|1\rangle \label{2} \\
|-\rangle &=&-\frac{1}{\sqrt{2}}|3\rangle +\frac{\sin \theta }{\sqrt{2}}%
|2\rangle +\frac{\sin \theta }{\sqrt{2}}|1\rangle \nonumber\end{aligned}$$
in which $\sin \theta =\frac{\Omega _{1}}{\Omega _{e}}$. Thus in the eigenstates basis, the ralations hold $H_{\Omega }=$ $\Omega _{e}(\sigma
_{++}-\sigma _{--})$ and $\sigma _{32}=\frac{1}{\sqrt{2}}\cos \theta (\sigma
_{+0}-\sigma _{-0})+\frac{\sin \theta }{2}(\sigma _{++}-\sigma _{-+}+\sigma
_{+-}-\sigma _{--})$. Then make a unitary transformation $%
e^{iH_{0}t}H_{c}e^{-iH_{0}t}$ with $H_{0}=H_{\Omega }+\sum_{n=1,2}\delta
_{n}a_{n}^{\dagger }a_{n}$ and assume that the detunings $\delta
_{1}=2\Omega _{e}-d$ and $\delta _{2}=2\Omega _{e}+d$, where $d>0$ is a small quantity compared with $2\Omega _{e}$ but is still large when compared with the effective coupling strength (which will show later). Because of the strong driving condition, the unitary transformation will contain fast osicilating terms $e^{\pm i\Omega _{e}t}$ and $e^{\pm 2i\Omega _{e}t}$, by neglecting theses terms, i.e., when secular appoximation is made, the cavity related Hamiltonian now has the form
$$H_{c1}=(G_{1}a_{1}-G_{2}^{*}a_{2}^{\dagger })\sigma
_{+-}e^{idt}+(G_{1}^{*}a_{1}-G_{2}a_{2})\sigma _{-+}e^{-idt} \label{3}$$
in which $G_{1}=\frac{g_{1}}{2}\sin \theta $, and $G_{2}=\frac{g_{2}}{2}\sin
\theta $. If $d$ $\gg |G_{1}|,|G_{2}|$ is fulfilled, an effective Hamiltonian $H_{e}$ can be obtained as $$\begin{aligned}
H_{e} &=&\lambda (a_{1}a_{2}+a_{1}^{\dagger }a_{2}^{\dagger })(\sigma
_{++}-\sigma _{--}) \nonumber \\
&&+(\frac{|G_{1}|^{2}}{d}a_{1}^{\dagger }a_{1}+\frac{|G_{2}|^{2}}{d}%
a_{2}^{\dagger }a_{2})(\sigma _{++}-\sigma _{--}) \label{4} \\
&&+\lambda _{++}\sigma _{++}-\lambda _{--}\sigma _{--} \nonumber\end{aligned}$$ where $\lambda =\frac{G_{1}G2}{d}$ is the effective coupling between the two cavity modes, and $\lambda _{++}=$ $\frac{|G_{1}|^{2}}{d}$, $\lambda _{--}=%
\frac{|G_{2}|^{2}}{d}$. Assume that the two cavity modes are initially in their vacuum states $|0_{1}\rangle |0_{2}\rangle $, then the terms related to $a_{j}^{\dagger }a_{j}$ ($j=1,2$) have no contribution to the evolution and can be omitted. Thus the final effective Hamiltonian is
$$\begin{aligned}
H_{eff} &=&\lambda (a_{1}a_{2}+a_{1}^{\dagger }a_{2}^{\dagger })(\sigma
_{++}-\sigma _{--}) \nonumber \\
&&+\lambda _{++}\sigma _{++}-\lambda _{--}\sigma _{--} \label{5}\end{aligned}$$
Next we will show how to obtain squeezing and entanglement from this effective Hamiltonian for the two cavity modes. If the initial state of the whole system is $|\psi \left( 0\right) \rangle =|+\rangle |0_{1}\rangle
|0_{2}\rangle $, then the state will evolve as $|\psi \left( t\right)
\rangle =e^{-iH_{e}t}|\psi \left( 0\right) \rangle $ to the following state $$|\psi \left( t\right) \rangle =e^{-i\lambda _{++}t}|+\rangle e^{-i\lambda
(a_{1}a_{2}+a_{1}^{\dagger }a_{2}^{\dagger })t}|0_{1}\rangle |0_{2}\rangle
\label{6}$$ Thus if the whole atom is intially in state $|+\rangle |0_{1}\rangle
|0_{2}\rangle $, then after time $t$, the atom will still in its state $%
|+\rangle $, but the two cavity modes will evolve into the following state
$$|\varphi \left( t\right) \rangle _{c}=e^{-i\lambda
t(a_{1}a_{2}+a_{1}^{\dagger }a_{2}^{\dagger })}|0_{1}\rangle |0_{2}\rangle
\label{7}$$
It is obvious that $|\varphi \left( t\right) \rangle _{c}$ is a two-mode squeezed vacuum state. Let us define the quadrature operators of the two cavity modes which will be used to judge squeezing and entanglement as $%
x_{j}=\frac{a_{j}+a_{j}^{\dagger }}{\sqrt{2}}$, $p_{j}=-i\frac{%
a_{j}-a_{j}^{\dagger }}{\sqrt{2}}$, and $u=x_{1}+x_{2}$, $\upsilon
=p_{1}-p_{2}$. The quanties $\langle (\Delta u)^{2}\rangle =\langle
u^{2}\rangle -\langle u\rangle ^{2}$ and $\langle (\Delta \upsilon
)^{2}\rangle =\langle \upsilon ^{2}\rangle -\langle \upsilon \rangle ^{2}$ are needed to be calculated. According to the entantanglement criterion \[20\], if $M=\langle (\Delta u)^{2}+(\Delta \upsilon )^{2}\rangle <2$, then the two cavity modes are entangled. Using the state $|\varphi \left(
t\right) \rangle _{c}$ in Eq. (7), the following relations hold $\langle
u\rangle =\langle \upsilon \rangle =0$, $\langle a_{1}^{\dagger
}a_{1}\rangle =\langle a_{2}^{\dagger }a_{2}\rangle =\sinh (\lambda t)$ and $%
\langle a_{1}a_{2}\rangle =-i\cosh (\lambda t)\sinh (\lambda t)$. Thus $%
M=2+(e^{\lambda t}-e^{-\lambda t})^{2}=2\left( e^{2\lambda t}+e^{-2\lambda
t}\right) $. If $\lambda =i|\lambda |$ ($|\lambda |\neq 0$), then $M=2\cos
(2\lambda t)<2$, that is to say if we choose $g_{1}=|g|$, $g_{2}=i|g|$, then $\lambda =\frac{G_{1}G2}{d}=i\frac{|g|^{2}}{d}$, and entanglement criterion $%
M<2$ always holds. Thus two cavity modes are entangled states. Actually the state $|\varphi \left( t\right) \rangle _{c}$ is a two-mode entangled state and is an ideally entangled state when $\lambda =i|\lambda |$ ($|\lambda
|\neq 0$) holds.
Conclusion
==========
In conclusion, we have presented a scheme to generate two-mode entangled state in a cavity QED system by using the atomic coherence created by the two strong classical driving fields. The result shows that when we appropriately adjusting the coupling strengths of the two cavity modes, a two-mode squeezed vacuum state which is a two-mode entangled state is generated.\
**Acknowledgments**
This work is supported by the Scientific Research Plan of the 281 Provincial Education Department in Hubei (Grant No. Q20101304 ) and NSFC under Grant No. 11147153.
[99]{} S. L. Braunstein and P. van Look, Rev. Mod. Phys. **77**, 513 (2005).
D.Bouwmeester, J. W. Pan, K. Mattle, M. Eibl, H. weinfurter and A. Zeilinger, Nature (London) 390, **575** (1997).
X. Li, Q. Pan, J. Jing, J. Zhang, C. Xie and K. Peng, Phys. Rev. Lett. 88, 047904 (2002).
X. Jia, X. Su, Q. Pan, J. Gao, C. Xie and K. Peng, Phys. Rev. Lett. 93, 250503 (2004).
H. T. Tan and G. X. Li, Phys. Rev. A 84, 024301 (2011).
M. Abdi,Sh. Barzanjeh, P. Tombesi and D. Vitali1, Phys. Rev. A 84, 032325 (2011).
Eyob A. Sete and C. H. Raymond Ooi, Phys. Rev. A 85, 063819 (2012).
Q. Xu and X. M. Hu, Phys. Rev. A 86, 032337 (2012).
Q. X. Mu, Y. H. Ma and L. Zhou, Phys. Rev. A 81, 024301 (2010).
H. S. Eisenberg, G. Khoury, G. Durkin, C. Simon and D. Bouwmeester, Phys. Rev. Lett. 93, 193901 (2004).
G. X. Li, Y. P. Yang, K. Allaart and D. Lenstra, Phys. Rev. A 69, 014301 (2004).
Q. Cheng, W. L. Yang and M. Feng, Phys. Rev. A 86, 022327 (2012).
P. B. Li, S. Y. Gao, H. R. Li, S. L. Ma and F. L. Li, Phys. Rev. A 85, 042306 (2012).
H. Xiong, M. O. Scully and M. S. Zubairy, Phys. Rev. Lett. 94, 023601 (2005).
L. Zhou, H. Xiong and M. S. Zubairy, Phys. Rev. A 74, 022321 (2006).
X. B. Zou, Y. L. Dong and G. C. Guo, Phys. Rev. A 73, 025802 (2006).
G. L. Cheng, X. M. Hu, W. X. Zhong and Q. Li, Phys. Rev. A 78, 033811 (2009).
X. M. Hu and J. H. Zou, Phys. Rev. A 78, 045801 (2009).
S. Pielawa, G. Morigi, D. Vitali and L. Davidovich, Phys. Rev. Lett. 98, 240401 (2007)
L. M. Duan, G. Giedke, J. I. Cirac and P. Zoller, Rev. Lett. 84, 2722 (2000).
[^1]: E-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A famous result by Milner is that the $\l$-calculus can be simulated inside the $\pi$-calculus. This simulation, however, holds only modulo strong bisimilarity on processes, i.e. there is a slight mismatch between $\beta$-reduction and how it is simulated in the $\pi$-calculus. The idea is that evaluating a $\l$-term in the $\pi$-calculus is like running an environment-based abstract machine, rather than applying ordinary $\beta$-reduction. In this paper we show that such an abstract-machine evaluation corresponds to linear weak head reduction, a strategy arising from the representation of $\l$-terms as linear logic proof nets, and that the relation between the two is as tight as it can be. The study is also smoothly rephrased in the call-by-value case, introducing a call-by-value analogous of linear weak head reduction.'
author:
- Beniamino Accattoli
title: Evaluating functions as processes
---
Conclusions {#conclusions .unnumbered}
===========
We have shown how to refine the relation between the $\l$-calculus and the $\pi$-calculus, getting a perfect match of reductions steps in both call-by-name and call-by-value. The refinements crucially exploits rewriting rules at a distance, and unveil that the $\pi$-calculus evaluates $\l$-terms exactly as linear logic proof nets. A natural continuation would be to extend these relations to calculi with multiplicities [@DBLP:journals/iandc/BoudolL96], which are related to the study of observational equivalence. It would also be interesting to investigate linear weak applicative reduction, in particular in relation with complexity [@DBLP:conf/rta/AccattoliL12] or with Taylor-Ehrhard expansion [@DBLP:conf/csl/Ehrhard12]. Finally, given the compactness of the results and the involved reasoning about bound, free, and fresh variables, it would be interesting to try to formalize this work in Abella [@DBLP:conf/cade/Gacek08], which is a proof assistant provided with a nominal quantifier precisely developed to cope with the $\pi$-calculus [@DBLP:journals/tocl/TiuM10] and where reasoning about untyped calculi with binders is very close to pen-and-paper reasoning [@DBLP:conf/cpp/Accattoli12].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Leo P. Singer'
bibliography:
- 'ms.bib'
- 'iptf-gbm-paper/gcn.bib'
date: 'November 24, 2014'
title: |
The needle in the 100 deg$^2$ haystack:\
The hunt for binary neutron star mergers with and
---
To the love of my life, my wife Kristin, and our precious son Isaac.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: '[0cm]{}[8ex]{} Hybrid Monte Carlo (HMC) generates samples from a prescribed probability distribution in a configuration space by simulating Hamiltonian dynamics, followed by the Metropolis (-Hastings) acceptance/rejection step. Compressible HMC (CHMC) generalizes HMC to a situation in which the dynamics is reversible but not necessarily Hamiltonian. This article presents a framework to further extend the algorithm. Within the existing framework, each trajectory of the dynamics must be integrated for the same amount of (random) time to generate a valid Metropolis proposal. Our generalized acceptance/rejection mechanism allows a more deliberate choice of the integration time for each trajectory. The proposed algorithm in particular enables an effective application of variable step size integrators to HMC-type sampling algorithms based on reversible dynamics. The potential of our framework is further demonstrated by another extension of HMC which reduces the wasted computations due to unstable numerical approximations and corresponding rejected proposals.'
author:
- Akihiko Nishimura
- David Dunson
bibliography:
- 'VLT\_CHMC.bib'
title: Variable length trajectory compressible hybrid Monte Carlo
---
\[sec:intro\]Introduction
=========================
A study of molecular systems often relies on generating random variables from a prescribed (unnormalized) probability distribution $\rho({\bm{q}}) \propto \exp(-U({\bm{q}}))$ on the configuration space. Markov chain Monte Carlo (MCMC) generates samples from a target distribution by constructing a Markov chain whose stationary distribution coincides with the target distribution. Such a Markov chain can be realized by building a transition rule that satisfies the detailed balance condition. MCMC based on the Metropolis (-Hastings) algorithm [@metropolis53] is a general sampling approach widely used in computational physical science as well as in Bayesian statistics and machine learning. Many such algorithms are inefficient, producing highly correlated samples, and require a large number of iterations to adequately characterize the target distribution [@roberts01; @mattingly12; @pillai12].
Hybrid Monte Carlo [@duane87] (HMC) constructs the proposal distribution for the Metropolis algorithm by simulating molecular dynamics (MD), a procedure that can greatly reduce the correlation among successive MCMC samples. More precisely, HMC augments the state space by introducing a *momentum* variable ${\bm{p}}$; the original variable ${\bm{q}}$ is often referred to as *position* variable in the HMC framework. In this augmented state space $({\bm{q}},{\bm{p}})$, the proposal distribution for Metropolis is constructed by solving an ordinary differential equation (ODE) corresponding to Newton’s equations of motion with respect to the potential energy $U({\bm{q}})$. There are applications in which (partial) analytical solutions to an ODE can be exploited [@shahbaba13; @pakman13; @pakman14], but in general ODEs are discretized and integrated numerically.
Within the original HMC framework, an integrator for simulating MD must be reversible and volume-preserving to produce a valid Metropolis proposal [@neal10]. In fact, the volume-preserving property can be relaxed by including a Jacobian factor in the calculation of the Metropolis acceptance probability. [@leimkuhler09; @lan15] Under this generalization of HMC, any reversible (discrete) dynamics / bijective map can be applied to generate a proposal state. This algorithm is formalized as *compressible HMC* (CHMC) in Ref. . A generalization of HMC known as *Riemannian manifold HMC* (RMHMC) [@girolami11] also falls within the framework of CHMC.
This article presents an algorithm to relax another condition required by (compressible) HMC. Given a reversible map $\bm{F}$ and state $({\bm{q}},{\bm{p}})$, CHMC proposes the next state by applying the map $n$ times, where the number of steps $n$ can be drawn randomly at each iteration. Though often not stated explicitly, the detailed balance requires the number of steps to be determined independently of the trajectory $\{({\bm{q}},{\bm{p}}), \bm{F}({\bm{q}},{\bm{p}}), \bm{F}^2({\bm{q}},{\bm{p}}), \ldots \}$. As we will show in Section \[sec:vlt\_chmc\], this constraint can prevent realizing the full potential of MCMC algorithms based on reversible dynamics. Our algorithm generalizes the acceptance-rejection mechanism behind CHMC to allow the number of steps to depend on each trajectory of the dynamics while preserving the detailed balance. The number of numerical integration steps taken in simulating a trajectory of HMC is commonly referred to as the “path length” of a trajectory in the statistics literature. We therefore call our algorithm *variable length trajectory CHMC* (VLT-CHMC). It should be mentioned that the No-U-Turn-Sampler (NUTS) is another variant of HMC that allows the path lengths to vary from one trajectory to another. [@hoffman14] However, the motivation behind NUTS is to spare a user the trouble of manually tuning the number of steps, and NUTS in general performs no better than HMC with well-chosen path lengths. [@hoffman14; @wang13] On the other hand VLT-CHMC can improve the performance of CHMC in a more fundamental and significant way. In particular, VLT-CHMC enables an effective application of reversible variable step size integrators to HMC-type sampling algorithms based on reversible dynamics. The rest of the paper is organized as follows. Section \[sec:chmc\] reviews the main ideas behind CHMC and provides an example in which the compressible dynamics arises from the use of non-traditional integrators in HMC settings. Such integrators have proven to be more efficient than the commonly used volume-preserving integrators in various applications. The example also serves to introduce the notations and concepts needed in the next section, where VLT-CHMC is motivated as a method to effectively apply variable step size integrators in HMC settings. The presentation is self-contained, but some familiarity with HMC is assumed. VLT-CHMC is developed in Section \[sec:vlt\_chmc\]. Section \[sec:vlt\_chmc\_special\] explains how the existing framework limits the utility of variable step size integrators to sampling algorithms. The key observation in addressing this issue leads to a special case of VLT-CHMC. More general construction of VLT-CHMC is provided in Section \[sec:vlt\_chmc\_general\]. Section \[sec:rahmc\] presents another use case of VLT-CHMC, where HMC is modified to reduce the wasted computation due to unstable numerical approximations and corresponding rejected proposals. The simulation results are shown in Section \[sec:simulation\] to demonstrate the potential gains from the framework of VLT-CHMC.
Review of compressible HMC {#sec:chmc}
==========================
Basic Theory
------------
To keep the description of CHMC and the subsequent development of VLT-CHMC more intuitive, the version of CHMC described here is slightly less general than the one in Ref. . It is straightforward to extend the variable length trajectory algorithm of Section \[sec:vlt\_chmc\] to the general settings.
A bijective map ${\bm{F}}$ is said to be *reversible* if $$\label{eq:reversibility}
{\bm{F}}^{-1} = {\bm{R}}\circ {\bm{F}}\circ {\bm{R}}$$ or equivalently $\left( {\bm{R}}\circ {\bm{F}}\right)^{-1} = {\bm{R}}\circ {\bm{F}}$ for an involution ${\bm{R}}$ (i.e. ${\bm{R}}\circ {\bm{R}}= \text{id}$). Note that the reversiblity of ${\bm{F}}$ implies that of ${\bm{F}}^n$ for any $n$. Let ${\bm{D}}{\bm{F}}^n$ denote the Jacobian matrix of ${\bm{F}}^n$ and $|{\bm{D}}{\bm{F}}^n|$ its determinant. Given a state ${\bm{z}}$ and integer $n$, CHMC proposes the state ${\bm{z}}^* = {\bm{R}}\circ {\bm{F}}^n ({\bm{z}})$ and accepts or rejects the proposal with probability $$\label{eq:acceptance_prob}
\min \left\{ 1,
\frac{ \rho({\bm{z}}^*) | {\bm{D}}{\bm{F}}^n({\bm{z}}) | }{ \rho({\bm{z}})} \right\}$$ To see that this transition rule satisfies the detailed balance with respect to $\rho(\cdot)$, consider a small neighborhood $B$ around ${\bm{z}}$ and $B^* = {\bm{R}}\circ {\bm{F}}^n(B)$ around ${\bm{z}}^*$, so that ${\bm{R}}\circ {\bm{F}}^n(B^*) = B$. The proposal move sends the probability mass $$\int_B \rho({\bm{z}}') {\rm d}{\bm{z}}' \approx \rho({\bm{z}}) \text{vol}(B)$$ from $B$ to $B^*$. On the other hand, the mass sent from $B^*$ to $B$ by the proposal move can be seen to be $$\begin{aligned}
\int_{B^*} \rho({\bm{z}}') {\rm d}{\bm{z}}'
&= \int_{B} \rho({\bm{z}}') |{\bm{D}}({\bm{R}}\circ {\bm{F}}^n)({\bm{z}}')|{\rm d}{\bm{z}}' \\
&\approx \rho({\bm{z}}^*) |{\bm{D}}{\bm{F}}^n({\bm{z}})| \, \text{vol}(B)
\end{aligned}$$ by the change of variable formula and the fact $|{\bm{R}}| = 1$. The acceptance and rejection step of CHMC amounts to rejecting the fraction of move by the ratio of the probability fluxes and thus imposes the detailed balance.
The above transition rule preserves the target density $\rho(\cdot)$ for any $n$, so in practice the number of steps can be drawn randomly at each iteration of CHMC. The steps of CHMC are summarized in Algorithm \[alg:chmc\] below, where the distribution $p(\cdot)$ for the number of steps is a tuning parameter a user must specify. The use of a deterministic map as a proposal distribution does not yield an ergodic Markov chain, and therefore such a transition rule must be alternated with another transition rule that preserves the target density $\rho(\cdot)$, as done in Step 1 of the algorithm. We do not concern ourselves here with how to choose such a random move since the choice depends critically on the particular form of $\rho(\cdot)$.
\[alg:chmc\] With a prespecified probability mass function $p(\cdot)$ on $\mathbb{Z}^+$, CHMC generates a Markov chain $\{ {\bm{z}}^{(m)} \}_m$ with the following transition rule ${\bm{z}}^{(m)} \to {\bm{z}}^{(m+1)}$:
1. Make a random change ${\bm{z}}^{(m)} \to {\bm{z}}$ that preserves the target density $\rho(\cdot)$.
2. Sample $n \sim p(\cdot)$ and propose the state ${\bm{z}}^* = {\bm{R}}\circ {\bm{F}}^n({\bm{z}})$.
3. Let ${\bm{z}}^{(m+1)} = {\bm{z}}^*$ with probability $$\min \left\{1,
\frac{ \rho({\bm{z}}^*) | {\bm{D}}{\bm{F}}^n({\bm{z}}) |
}{
\rho({\bm{z}})} \right\}$$ Otherwise, let ${\bm{z}}^{(m+1)} = {\bm{z}}$.
Example: (Riemann manifold) HMC with non-volume-preserving integrators {#sec:hmc_wo_volume_preservation}
----------------------------------------------------------------------
HMC and its extension Riemann manifold HMC (RMHMC) construct a reversible and volume-preserving bijective map by numerically approximating Hamiltonian dynamics. To this end, they require a geometric integrator that preserves the reversibility and volume-preservation property of Hamiltonian dynamics. Under the CHMC framework, however, Hamiltonian dynamics can be approximated using a wider range of integration techniques.
In order to sample from a probability density of interest $\rho_0({\bm{q}}) \propto \exp(-U({\bm{q}}))$ in $\mathbb{R}^d$, RMHMC introduces an auxiliary variable ${\bm{p}}\in \mathbb{R}^d$ whose distribution is defined conditionally as ${\bm{p}}| {\bm{q}}\sim \mathcal{N}(\bm{0}, {\bm{M}}({\bm{q}}))$ for a family of positive definite matrices known as *mass tensors* $\{{\bm{M}}({\bm{q}})\}_{{\bm{q}}}$. [@bennett75; @neal10; @girolami11] The joint density $\rho({\bm{q}},{\bm{p}})$ in the phase space then is given as $\rho({\bm{q}},{\bm{p}}) \propto \exp(-H({\bm{q}},{\bm{p}}))$ where the *Hamiltonian* $H({\bm{q}},{\bm{p}})$ is given by $$\label{eq:rmhmc_hamiltonian}
H({\bm{q}},{\bm{p}})
= U({\bm{q}}) + \frac{1}{2} {\bm{p}}^T {\bm{M}}({\bm{q}})^{-1} {\bm{p}}+ \frac{1}{2} \log | {\bm{M}}({\bm{q}}) |$$ The proposal is generated by approximating the solution to *Hamilton’s equations*: $$\label{eq:hamilton's}
\frac{\text{d} {\bm{q}}}{\text{d} t}
= \nabla_{{\bm{p}}} H({\bm{q}}, {\bm{p}}), \
\frac{\text{d} {\bm{p}}}{\text{d} t}
= - \nabla_{{\bm{q}}} H({\bm{q}}, {\bm{p}})$$ For the Hamiltonian , the solution operator of is reversible with respect to a momentum flip operator ${\bm{R}}({\bm{q}},{\bm{p}}) = ({\bm{q}},-{\bm{p}})$. Solving using a reversible integrator with a constant step size ${{\Delta t}}$ yields a reversible map ${\bm{F}}_{{{\Delta t}}}$ so that $$\label{eq:approximate_map}
{\bm{F}}_{{{\Delta t}}}^n({\bm{q}}_0,{\bm{p}}_0) \approx ({\bm{q}}(n {{\Delta t}}), {\bm{p}}(n {{\Delta t}}))$$ where $\{({\bm{q}}(t), {\bm{p}}(t))\}_t$ denotes the exact solution with the initial condition $({\bm{q}}_0,{\bm{p}}_0)$. In other words, ${\bm{F}}_{{\Delta t}}$ approximates the solution operator ${\bm{\Phi}}_{{\Delta t}}$ of defined through the relation $$\label{eq:solution_operator}
\frac{\text{d} {\bm{\Phi}}_t}{\text{d} t}
= \big( (\nabla_{{\bm{p}}} H) \circ {\bm{\Phi}}_t, - (\nabla_{{\bm{q}}} H) \circ {\bm{\Phi}}_t \big)$$ for all $t$. If the reversible map ${\bm{F}}_{{{\Delta t}}}$ is further required to be volume preserving, then we have $|{\bm{D}}{\bm{F}}_{{{\Delta t}}}^n| = 1$ and the Jacobian factor drops from , recovering HMC and RMHMC algorithms of Ref. . In some applications however, non-volume-preserving approximations of have been shown to offer substantial gains in computational efficiency. [@lan15; @fang14]
For example, Lan et. al. [@lan15] considers the ODE corresponding to in terms of reparametrization $({\bm{q}}, \mathbf{v}) = ({\bm{q}}, {\bm{M}}({\bm{q}})^{-1} {\bm{p}})$. The reparametrized ODE admits semi-explicit and explicit reversible approximations, requiring fewer or no fixed point iterations compared to the Störmer-Verlet integrator typically employed in RMHMC. The proposal move using a simulated trajectory is alternated with sampling $\mathbf{v}$ from its conditional density $\mathbf{v} | {\bm{q}}\sim \mathcal{N}(\bm{0}, {\bm{M}}({\bm{q}})^{-1})$, a random move corresponding to Step 1 in Algorithm \[alg:chmc\]. The CHMC algorithm based on the semi-explicit and explicit integrator are found to significantly outperform RMHMC based on the Störmer-Verlet integrator over a range of examples.
Variable length trajectory CHMC {#sec:vlt_chmc}
===============================
Variable length trajectory CHMC (VLT-CHMC) is most naturally motivated as a method to effectively apply variable step size integrators in RMHMC settings. For this reason, we first develop this special case of VLT-CHMC in Section \[sec:vlt\_chmc\_special\]. A more general theory is developed in Section \[sec:vlt\_chmc\_general\]. Section \[sec:rahmc\] illustrates the use and potential benefits of the general VLT-CHMC algorithm through another example.
Special case of VLT-CHMC {#sec:vlt_chmc_special}
------------------------
### Motivation: RMHMC with variable step size integrators and limitations of CHMC
In Section \[sec:hmc\_wo\_volume\_preservation\], we discussed how CHMC allows us to approximate Hamiltonian dynamics with non-volume-preserving integrators and still generate a valid Metropolis proposal. We in particular considered the use of a reversible integrator with a constant step size. A wider range of reversible integration techniques for Hamiltonian systems are available in the literature, however, including a number of variable step size integrators. [@calvo98; @blanes12; @leimkuhler04; @hairer06] In theory, a variable step size integrator similarly produces a valid CHMC proposal as long as the integrator is reversible. However, the use of such an integrator under the existing CHMC framework generally leads to an algorithm with suboptimal sampling efficiency, for the reasons we describe now.
Each step of a variable step size integrator approximates the evolution $({\bm{q}}(t_n), {\bm{p}}(t_n)) \to ({\bm{q}}(t_n + {{\Delta t}}_n), {\bm{p}}(t_n + {{\Delta t}}_n))$ where the step size ${{\Delta t}}_n$ depends on the current state $({\bm{q}}(t_n), {\bm{p}}(t_n))$ through a *step size controller* $g({\bm{q}},{\bm{p}})$. The simplest choice of step size would be ${{\Delta t}}_n = g({\bm{q}}(t_n),{\bm{p}}(t_n)) {{\Delta s}}$, but the reversibility requires a slightly more sophisticated relationship and the condition $g({\bm{q}},{\bm{p}}) = g({\bm{q}},-{\bm{p}})$ (see Section \[sec:vlt\_chmc\_algorithm\_in\_special\_case\]). Most importantly for our discussion, a variable step size scheme is equivalent to approximating the following *time-rescaled* Hamiltonian dynamics in a new time scale $\text{d} s = g({\bm{q}},{\bm{p}})^{-1} \text{d} t$ with a constant step size ${{\Delta s}}$: $$\label{eq:time_rescaled_hamilton}
\frac{\text{d} {\bm{q}}}{\text{d} s}
= g({\bm{q}}, {\bm{p}}) \nabla_{{\bm{p}}} H({\bm{q}}, {\bm{p}}), \
\frac{\text{d} {\bm{p}}}{\text{d} s}
= - g({\bm{q}}, {\bm{p}}) \nabla_{{\bm{q}}} H({\bm{q}}, {\bm{p}})$$ In other words, a reversible variable step size approximation of yields a reversible map ${\bm{F}}_{{{\Delta s}}}$ such that $$\label{eq:approximate_time_rescaled_map}
{\bm{F}}_{{{\Delta s}}}^n({\bm{q}}_0,{\bm{p}}_0) \approx ({\bm{q}}(n {{\Delta s}}), {\bm{p}}(n {{\Delta s}}))$$ where $\{{\bm{q}}(s),{\bm{p}}(s)\}_s$ is the solution to the time-rescaled dynamics with the initial condition $({\bm{q}}_0,{\bm{p}}_0)$.
This implicit time-rescaling behind variable step size integration causes trouble for CHMC. The utility of Hamiltonian dynamics as a proposal generation mechanism stems from the fact that $\rho({\bm{q}},{\bm{p}}) \propto \exp(-H({\bm{q}},{\bm{p}}))$ is the *invariant distribution* of the dynamics i.e. if $({\bm{q}}_0,{\bm{p}}_0)$ has the distribution $\rho({\bm{q}},{\bm{p}}) \propto \exp(-H({\bm{q}},{\bm{p}}))$, then ${\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)$ also has the same distribution $\rho(\cdot)$ for all $t$. As a consequence, the proposal generated by an approximate solution $({\bm{q}}^*,{\bm{p}}^*) = {\bm{F}}_{{{\Delta t}}}^n({\bm{q}}_0, {\bm{p}}_0)$ as in can be accepted with probability 1 in the limit ${{\Delta t}}\to 0$ and $n {{\Delta t}}\to t'$. On the other hand, the time-rescaled dynamics in general does not preserve the target density $\rho({\bm{q}},{\bm{p}})$, and the proposal generated by the approximate solution $({\bm{q}}^*, {\bm{p}}^*) = {\bm{F}}_{{{\Delta s}}}^n({\bm{q}}_0,{\bm{p}}_0)$ may not be accepted with high probability even in the limit ${{\Delta s}}\to 0$ and $n {{\Delta s}}\to s'$. In fact, the acceptance probability of the CHMC proposal in the limit is given by: $$\label{eq:variable_stepsize_chmc_acceptance_prob}
\min \left\{ 1, \frac{ g({\bm{q}}(s'),{\bm{p}}(s')) }{ g({\bm{q}}_0,{\bm{p}}_0) } \right\}$$ where $\{{\bm{q}}(s),{\bm{p}}(s)\}_s$ denotes the solution to with the initial condition $({\bm{q}}_0,{\bm{p}}_0)$. The derivation is given in Appendix \[app:acceptance\_prob\_derivation\].
### Algorithm: variable length trajectory scheme for time-rescaled dynamics {#sec:vlt_chmc_algorithm_in_special_case}
In order to address the issue caused by the implicit time-rescaling associated with variable step size integrators, VLT-CHMC approximates the dynamics in the original time scale as follows. Fix the initial condition $({\bm{q}}_0, {\bm{p}}_0)$ and denote $({\bm{q}}_i,{\bm{p}}_i) = {\bm{F}}_{{\Delta s}}^i ({\bm{q}}_0,{\bm{p}}_0)$ where ${\bm{F}}_{{\Delta s}}$ approximates the dynamics in the time scale $s$ as in . The evolution $({\bm{q}}_0, {\bm{p}}_0) \to ({\bm{q}}(t),{\bm{p}}(t))$ in the original time scale can be approximated by taking the trajectory dependent number of steps $N({\bm{q}}_0, {\bm{p}}_0) = N(t, {\bm{q}}_0, {\bm{p}}_0)$ defined as $$\label{eq:traj_len_function}
\begin{aligned}
N(& {\bm{q}}_0, {\bm{p}}_0) = \\
&\min \left\{ n :
\sum_{i=1}^n \frac{{{\Delta s}}}{2} \left( g({\bm{q}}_{i - 1}, {\bm{p}}_{i - 1}) + g({\bm{q}}_{i}, {\bm{p}}_{i}) \right) > t \right\}
\end{aligned}$$ Now we consider the map ${\bm{F}}_{{\Delta s}}^N$ defined as $${\bm{F}}_{{\Delta s}}^N ({\bm{q}}, {\bm{p}}) = {\bm{F}}_{{\Delta s}}^{N({\bm{q}},{\bm{p}})} ({\bm{q}},{\bm{p}})$$ which approximates the solution operator ${\bm{\Phi}}_t$ as defined in . The map however cannot be used directly to generate a proposal because in general it is neither reversible or even bijective. The map would be reversible if $N({\bm{q}}_0^*, {\bm{p}}_0^*) = N({\bm{q}}_0, {\bm{p}}_0)$ where $({\bm{q}}_0^*, {\bm{p}}_0^*) = {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N ({\bm{q}}_0, {\bm{p}}_0)$, but only implies $N({\bm{q}}_0^*, {\bm{p}}_0^*) \leq N({\bm{q}}_0, {\bm{p}}_0)$. For example when $g({\bm{q}}_0^*, {\bm{p}}_0^*) \gg g({\bm{q}}_0, {\bm{p}}_0)$, the simulated time along the reverse trajectory $\left\{ ({\bm{q}}_i^*, {\bm{p}}_i^*) = {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0^*, {\bm{p}}_0^*) \right\}_{i=0}^n$ $$\sum_{i=1}^n \frac{{{\Delta s}}}{2} \big( g({\bm{q}}^*_{i - 1}, {\bm{p}}^*_{i - 1}) + g({\bm{q}}^*_{i}, {\bm{p}}^*_{i}) \big)$$ will likely reach the threshold $t$ before $n = N({\bm{q}}_0, {\bm{p}}_0)$ steps.
The key observation behind VLT-CHMC is that we can nonetheless construct collections of states $S$ and $S^*$ containing $({\bm{q}}_0, {\bm{p}}_0)$ and $({\bm{q}}_0^*, {\bm{p}}_0^*)$ such that $$\label{eq:generalized_reversibility}
\begin{aligned}
{\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N(S) \subset S^*
&\ \text{ and } \
{\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N(S^*) \subset S \\
{\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N(S^c) \subset (S^{*})^c
&\ \text{ and } \
{\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N \left((S^{*})^c \right) \subset S^c
\end{aligned}$$ The existence of such sets $S$ and $S^*$ is a property of the map ${\bm{F}}_{{\Delta s}}^N$ and generalizes the notion of reversibility . The set $S$ is essentially the pre-image of $\{({\bm{q}}_0^*, {\bm{p}}_0^*)\}$ under ${\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N$ and can be constructed by defining $S = \left\{({\bm{q}}_{- \ell}, {\bm{p}}_{- \ell}), \thinspace \ldots, ({\bm{q}}_r, {\bm{p}}_r) \right\}$ by choosing $\ell, r \geq 0$ such that $$\label{eq:nstep_forward_backward}
\begin{aligned}
\ell
&= \max \left\{ j \geq 0: {\bm{F}}_{{\Delta s}}^N({\bm{q}}_{- j}, {\bm{p}}_{- j}) = {\bm{F}}_{{\Delta s}}^N({\bm{q}}_0, {\bm{p}}_0) \right\} \\
r
&= \max \left\{ j \geq 0: {\bm{F}}_{{\Delta s}}^N({\bm{q}}_{j}, {\bm{p}}_{j}) = {\bm{F}}_{{\Delta s}}^N({\bm{q}}_0, {\bm{p}}_0) \right\}
\end{aligned}$$ Algorithmically, $\ell$ and $r$ can be found by solving the dynamics backward and forward from $({\bm{q}}_0, {\bm{p}}_0)$ using the equivalent definitions below: $$\label{eq:nstep_special_case}
\begin{aligned}
\ell
&=\max \left\{ j \geq 0: \sum_{i=-j}^{N(t,{\bm{q}}_0,{\bm{p}}_0) - 1} {{\Delta t}}_i < t \right\} \\
r
&= \max \left\{ j \geq 0: \sum_{i=j}^{N(t,{\bm{q}}_0,{\bm{p}}_0)} {{\Delta t}}_i > t \right\} \\
&\hspace{3ex} \text{ where } {{\Delta t}}_i = \frac{{{\Delta s}}}{2} \left( g({\bm{q}}_{i - 1}, {\bm{p}}_{i - 1}) + g({\bm{q}}_{i}, {\bm{p}}_{i}) \right)
\end{aligned}$$ The set $S^*$ is the pre-image of $\{({\bm{q}}_r, {\bm{p}}_r)\}$ under ${\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N$ and can analogously be constructed. Denoting $({\bm{q}}^*_i, {\bm{p}}^*_i) = {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0^*, {\bm{p}}_0^*)$, let $S^* = \left\{({\bm{q}}^*_{- \ell^*}, {\bm{p}}^*_{- \ell^*}), \thinspace \ldots, ({\bm{q}}^*_{r^*}, {\bm{p}}^*_{r^*}) \right\}$ where $\ell^*, r^* \geq 0$ is defined as $$\label{eq:nstep_forward_backward_star}
\begin{aligned}
\ell^*
&= \max \left\{ j \geq 0: {\bm{F}}_{{\Delta s}}^N({\bm{q}}^*_{- j}, {\bm{p}}^*_{- j}) = {\bm{F}}_{{\Delta s}}^N({\bm{q}}^*_0, {\bm{p}}^*_0) \right\} \\
r^*
&= \max \left\{ j \geq 0: {\bm{F}}_{{\Delta s}}^N({\bm{q}}^*_{j}, {\bm{p}}^*_{j}) = {\bm{F}}_{{\Delta s}}^N({\bm{q}}^*_0, {\bm{p}}^*_0) \right\}
\end{aligned}$$ It is shown in Appendix \[app:vlt\_chmc\_justification\] that the above definition actually implies $r^* = 0$. The proof of and of other facts regarding $S$ and $S^*$ are also given in Appendix \[app:vlt\_chmc\_justification\].
Having constructed the sets $S$ and $S^*$ with the property , VLT-CHMC imposes the detailed balance by rejecting a fraction of moves between $S$ and $S^*$ as described in Algorithm \[alg:vlt\_chmc\] below.
\[alg:vlt\_chmc\] Given a reversible map ${\bm{F}}_{{\Delta s}}$ as in and a trajectory length function $N$ as in , VLT-CHMC generates a Markov chain $\{({\bm{q}}^{(m)}, {\bm{p}}^{(m)})\}_m$ with the following transition rule $({\bm{q}}^{(m)}, {\bm{p}}^{(m)}) \to ({\bm{q}}^{(m+1)}, {\bm{p}}^{(m+1)})$:
1. Sample ${\bm{p}}_0$ from the conditional density ${\bm{p}}| {\bm{q}}^{(m)}$ and set ${\bm{q}}_0 = {\bm{q}}^{(m)}$.
2. Find the indices $\ell, r, \ell^*, r^*$ as in and by simulating the dynamics forward and backward from $({\bm{q}}_0, {\bm{p}}_0)$ and ${\bm{R}}\circ {\bm{F}}_{{\Delta s}}^N({\bm{q}}_0,{\bm{p}}_0)$. Then set $$\begin{aligned}
S &= \left\{ {\bm{F}}_{{\Delta s}}^{- \ell}({\bm{q}}_0, {\bm{p}}_0), \thinspace \ldots, {\bm{F}}_{{\Delta s}}^r({\bm{q}}_0, {\bm{p}}_0) \right\} \\
S^* &= \Big\{ {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0 - r^*}({\bm{q}}_0, {\bm{p}}_0), \\
&\hspace{15ex} \thinspace \ldots, {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0 + \ell^*}({\bm{q}}_0, {\bm{p}}_0) \Big\}
\end{aligned}$$ where $N_0 = N({\bm{q}}_0, {\bm{p}}_0)$.
3. Propose the transition from $S$ to $S^*$ with the acceptance probability which is the smaller of 1 and $$\label{eq:valet_accept_prob}
\frac{
\sum\limits_{j=-r*}^{\ell^*} \rho \left( {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0, {\bm{p}}_0) \right) \left|{\bm{D}}{\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_{0},{\bm{p}}_{0}) \right|
}{
\sum\limits_{i=-\ell}^r \rho \left( {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0, {\bm{p}}_0) \right) \left| {\bm{D}}{\bm{F}}_{{\Delta s}}^{i}({\bm{q}}_{0},{\bm{p}}_{0}) \right|
}$$
4. If the transition in Step 3 is accepted, choose a state ${\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0, {\bm{p}}_0)$ from $S^*$ with the probability proportional to $$\rho \left( {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0, {\bm{p}}_0) \right) \left|{\bm{D}}{\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_{0},{\bm{p}}_{0}) \right|$$ and set $({\bm{q}}^{(m+1)}, {\bm{p}}^{(m+1)}) = {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0, {\bm{p}}_0)$. Otherwise, choose a state ${\bm{F}}_{{\Delta s}}^i({\bm{q}}_0, {\bm{p}}_0)$ from $S$ with the probability proportional to $$\rho \left( {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0, {\bm{p}}_0) \right) \left| {\bm{D}}{\bm{F}}_{{\Delta s}}^{i}({\bm{q}}_{0},{\bm{p}}_{0}) \right|$$ and set $({\bm{q}}^{(m+1)}, {\bm{p}}^{(m+1)}) = {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0, {\bm{p}}_0)$.
### Theory: VLT-CHMC and detailed-balance condition {#sec:detailed_balance_for_vlt_chmc}
Too see how VLT-CHMC achieves the detailed balance, consider a small neighborhood $B_0$ around $({\bm{q}}_0, {\bm{p}}_0)$. The total probability in the neighborhood $B = \cup_{i=- \ell}^r {\bm{F}}_{{\Delta s}}^i(B_0)$ of $S$ is $$\label{eq:mass_around_S}
\begin{aligned}
\int_{B}
&\rho({\bm{q}},{\bm{p}}) \, {\rm d} {\bm{q}}\, {\rm d} {\bm{p}}\\
&\approx \sum_{i=- \ell}^r \rho \left( {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0, {\bm{p}}_0) \right) \left| {\bm{D}}{\bm{F}}_{{\Delta s}}^{i}({\bm{q}}_{0},{\bm{p}}_{0}) \right| \big|B_0 \big|
\end{aligned}$$ assuming that $B_0$ is small enough that ${\bm{F}}_{{\Delta s}}^i(B_0)$’s are disjoint. Similarly, the total probability in the neighborhood $B^* = \cup_{j=- r^*}^{\ell^*} {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}(B_0)$ of $S^*$ is $$\label{eq:mass_around_S_star}
\begin{aligned}
\int_{B^*}
&\rho({\bm{q}},{\bm{p}}) \, {\rm d} {\bm{q}}\, {\rm d} {\bm{p}}\\
&\hspace{-2ex} \approx \sum_{j=- r^*}^{\ell^*} \rho \left( {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0, {\bm{p}}_0) \right) \left|{\bm{D}}{\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_{0},{\bm{p}}_{0}) \right| \big|B_0 \big|
\end{aligned}$$ Comparing the acceptance probability with the probability fluxes and , one can see that the acceptance-rejection procedure of Step 3 controls the probability fluxes appropriately to achieve the detailed balance between the neighborhoods $B$ and $B^*$. Step 4 then imposes the detailed balance within $B$ and $B^*$ by sampling a state according to the relative amount of probability in the individual components $\{ {\bm{F}}_{{\Delta s}}^i(B_0) \}_{i=- \ell}^r$ of $B$ and $\{ {\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}(B_0) \}_{j=- r^*}^{\ell^*}$ of $B^*$.
### Theoretical efficiency: improvement over CHMC
Throughout Section \[sec:vlt\_chmc\_special\] we considered the compressible dynamics arising from a variable step size integration of Hamiltonian dynamics. In this specific setting with the trajectory length function $N$ as defined in , VLT-CHMC is guaranteed to have a high average acceptance probability. In fact, in the limit ${{\Delta s}}\to 0$ with $t$ fixed, the acceptance probability of a VLT-CHMC proposal from $({\bm{q}}_0,{\bm{p}}_0)$ converges to a value bounded below by $$\label{eq:vlt_accept_prob_lower_bound}
\frac{ g({\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) }{ g({\bm{q}}_0,{\bm{p}}_0) }
\left \lfloor \frac{ g({\bm{q}}_0,{\bm{p}}_0) }{ g({\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) } \right \rfloor$$ when $g({\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) < g({\bm{q}}_0,{\bm{p}}_0)$. In case $g({\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) > g({\bm{q}}_0,{\bm{p}}_0)$, a similar lower bound holds for the proposal from ${\bm{R}}\circ {\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)$. Note that the quantity is always larger than $1/2$ and it tends to 1 as the ratio $g({\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) / g({\bm{q}}_0,{\bm{p}}_0)$ increases, in contrast with the acceptance probability of CHMC. More precise results on the acceptance probability of a VLT-CHMC proposal are derived in Appendix \[app:acceptance\_prob\_derivation\].
Of course, the acceptance rate of a proposal distribution is not the only factor determining the efficiency of an MCMC algorithm. Nonetheless, the theoretical result above highlights an advantage VLT-CHMC has over the usual CHMC. The bottom line is that VLT-CHMC proposals approximate the original dynamic while CHMC proposals approximate the time-rescaled dynamics . Therefore, VLT-CHMC will generally outperform CHMC whenever the exact solution of the original dynamics constitutes an efficient Markov chain propagator as is typically the case in RMHMC applications. [@girolami11; @nishimura16] This is substantiated by our simulation study in Section \[sec:simulation\].
General VLT-CHMC {#sec:vlt_chmc_general}
----------------
The key step in Algorithm \[alg:vlt\_chmc\] is the construction of the sets $S$ and $S^*$ with the property . More generally, the detailed balance can be imposed by the same type of acceptance-rejection mechanism whenever the phase space can be partitioned into a collection of pairs $S$ and $S^*$ such that the set $S \cup S^*$ and $(S \cup S^*)^c$ is closed under a (deterministic) transition rule. Conceivably, a wide range of algorithms can be devised under this general condition. In this section we present one systematic way to generalize the framework of Section \[sec:vlt\_chmc\_special\].
Consider a generic reversible map ${\bm{F}}$ on a state space ${\bm{z}}$ and associated involution ${\bm{R}}$. Fix ${\bm{z}}_0$ and denote ${\bm{z}}_i = {\bm{F}}^i({\bm{z}}_0)$. Choose a *trajectory termination criteria*, or more precisely boolean valued functions $b_n({\bm{z}}_0, \ldots, {\bm{z}}_n) \in \{0, 1\}$, with the following property $$\label{eq:criteria_symmetry}
b_n({\bm{z}}_0, \ldots, {\bm{z}}_n) = b_n({\bm{R}}({\bm{z}}_n), \ldots, {\bm{R}}({\bm{z}}_0))$$ as well as the property $$\label{eq:criteria_monotonicity}
b_n({\bm{z}}_0, \ldots, {\bm{z}}_n) = 1 \ \text{ only if } \
b_{n-i}({\bm{z}}_i, \ldots, {\bm{z}}_n) = 1$$ for any $i > 0$. These properties are satisfied, for example, by a termination criteria $\sum_{i=1}^n a({\bm{z}}_{i}) + a({\bm{z}}_{i-1}) > c$ for a scalar function $a({\bm{z}}) \geq 0$. Define a corresponding trajectory length function $N({\bm{z}}_0)$ as $$\label{eq:traj_length_function_general}
\begin{aligned}
N({\bm{z}}_0)
&= \min \{ N'({\bm{z}}_0), N_{\rm max} \} \\
&\hspace{-2ex} \text{for } \ N'({\bm{z}}_0) = \min \big\{ n : b_n({\bm{z}}_0, \ldots, {\bm{z}}_n) = 1
\big\}
\end{aligned}$$ With the reversible map ${\bm{F}}_{{\Delta s}}$ and trajectory length function $N$ of replaced by the generic ones as above, Algorithm \[alg:vlt\_chmc\] remains a valid MCMC scheme. This is because the justification of the algorithm (in Appendix \[app:vlt\_chmc\_justification\]) only require a trajectory length function $N$ to satisfy the *short return* condition $$\label{eq:short_return}
N({\bm{z}}^*) \leq N({\bm{z}}) \ \text{ where } {\bm{z}}^* = {\bm{R}}\circ {\bm{F}}^{N({\bm{z}})} ({\bm{z}})$$ and *order preserving* condition $$\label{eq:order_perserving}
N({\bm{z}}) - n \leq N({\bm{F}}^n({\bm{z}})) \ \text{ for any } n$$ The intuition behind the terminologies are explained in Appendix \[app:vlt\_chmc\_justification\] along with the proof of the general VLT-CHMC algorithm.
Example: Rejection Avoiding HMC {#sec:rahmc}
-------------------------------
Here we illustrate a use of the general VLT-CHMC framework through an algorithm of very different flavor from the special case presented in Section \[sec:vlt\_chmc\_special\].
A step size required for stable numerical integration of Hamilton’s equation can vary significantly at different regions of a phase space in some application areas of HMC. [@neal10] In such situations, the Hamiltonian may be approximately preserved along a simulated trajectory for a while until it suddenly starts to deviate wildly, leading to a proposal with little chance of acceptance. VLT-CHMC provides a way to “detect” when the trajectory becomes unstable and select an alternate state along the trajectory to transition to. Let ${\bm{F}}_{{\Delta t}}$ be a volume-preserving and reversible map as in , approximating Hamiltonian dynamics. Consider a trajectory $\left\{({\bm{q}}_i, {\bm{p}}_i) = {\bm{F}}_{{\Delta t}}^i ({\bm{q}}_0, {\bm{p}}_0) \right\}_{i = 0, 1, 2, \ldots}$. When the trajectory becomes unstable, it can be detected by a trajectory termination criteria such as $$\label{eq:max_H_fluctuation_criteria}
b_n = \mathbbm{1} \left\{ \max_{0 \leq i \leq n} H({\bm{q}}_i, {\bm{p}}_i) - \min_{0 \leq i \leq n} H({\bm{q}}_i, {\bm{p}}_i)
\geq \epsilon \right\}$$ where $\mathbbm{1}$ is an indicator function. We will actually use an alternative criteria below since this leads to a simpler algorithm implementation: $$\label{eq:error_per_step_criteria}
\begin{aligned}
b_n = \mathbbm{1} \Big\{
& \left| H({\bm{q}}_i, {\bm{p}}_i) - H({\bm{q}}_{i-1}, {\bm{p}}_{i-1}) \right| \geq \epsilon \\
& \hspace{15ex} \text{ for some } \, i = 1, \ldots, n \Big\}
\end{aligned}$$ It is easy to check that the criteria and satisfy the properties and and define a valid trajectory length function $N$ of the form for Algorithm \[alg:vlt\_chmc\]. We refer to the version of VLT-CHMC based on the criteria as *rejection avoiding HMC*.
A proposal of rejection avoiding HMC recovers the usual HMC proposal with the trajectory length $N_{\rm max}$ when the fluctuation of a Hamiltonian at each step is within the error tolerance $\epsilon$. However, upon detecting the fluctuation of magnitude larger than $\epsilon$ at the step $({\bm{q}}_{i-1}, {\bm{p}}_{i-1}) \to ({\bm{q}}_i, {\bm{p}}_i)$, the algorithm proceeds to simulate the trajectory backward from $({\bm{q}}_0, {\bm{p}}_0)$ and $({\bm{q}}_0^*, {\bm{p}}_0^*) = ({\bm{q}}_i, - {\bm{p}}_i)$ to determine the sets $S$ and $S^*$ according to the rule in Step 2 of Algorithm \[alg:vlt\_chmc\].
Numerical Results {#sec:simulation}
=================
Geometrically tempered HMC with variable step size integrator
-------------------------------------------------------------
HMC is known to have a serious difficulty sampling from a multi-modal target density as the potential energy barriers among the modes prevents transition from one mode to another. To address this issue, Nishimura and Dunson [@nishimura16] propose a version of RMHMC with a mass tensor having the property $$\label{eq:gthmc_mass_tensor}
| {\bm{M}}({\bm{q}}) |^{1/2}
\propto \rho({\bm{q}})^{1 - T^{- 1}}$$ with a *temperature* parameter $T \geq 1$. It can be shown that, with such a choice of a mass tensor, RMHMC algorithm is equivalent to the usual HMC algorithm (with a constant mass tensor) applied to a tempered distribution $\tilde{\rho}(\tilde{{\bm{q}}}) \propto \rho({\bm{q}})^{1/T}$ on a manifold parametrized by $\tilde{{\bm{q}}}$. For this reason, RMHMC with the property is referred to as *geometrically tempered HMC* (GTHMC) in Ref. .
The typical velocity of the dynamics at the position ${\bm{q}}$ is given by the operator norm $\lVert {\bm{M}}({\bm{q}}) \rVert^{- 1/2}$. This quantity, and in turn the velocity of the dynamics, necessarily becomes unboundedly large in the regions where $\rho({\bm{q}})$ is small, due to the constraint . For this reason, the only practical way to approximate the dynamics underlying GTHMC algorithms is through a variable step size integrator with a step size proportional to $\lVert {\bm{M}}({\bm{q}}) \rVert^{1/2}$.
We take an example with a simple bimodal target density from Ref. . The density $\rho({\bm{q}})$ is defined as a mixture of two-dimensional Gaussians with unit-variance centered at $(4, 0)$ and $(-4, 0)$. The mass tensor is chosen as $$\label{eq:dthmc_mass_tensor}
\begin{aligned}
{\bm{M}}({\bm{q}})
&\propto \rho({\bm{q}})^{2 \gamma \left( 1 - T^{- 1} \right) } \bm{e}_1 \bm{e}_1^T \\
&\qquad + \rho({\bm{q}})^{2 \left( 1-\gamma \right) (d-1)^{- 1} \left( 1 - T^{- 1} \right) } \left( \bm{I} - \bm{e}_1 \bm{e}_1^T \right)
\end{aligned}$$ for $d^{-1} \leq \gamma \leq 1$ where $d = 2$ is the dimension of ${\bm{q}}$ and $\bm{e}_1 = (1,0)$ is a standard basis vector. The mass tensors suggested in Ref. have apparent resemblance to , but the crucial difference is that they do not satisfy and consequently offer rather limited improvement over the standard HMC.
We compare the performance of CHMC and VLT-CHMC with the explicit variable step size integrator developed in Ref. . VLT-CHMC is run with the trajectory length function . The main challenge in this example to explore the phase space along the first coordinate of ${\bm{q}}$ due to the multi-modality along this direction. Therefore the efficiency of the sampling algorithms is summarized by the effective sample sizes (ESS) along the first coordinate of ${\bm{q}}$. The ESS’s as well as the acceptance probabilities at different parameter settings of CHMC and VLT-CHMC are summarized in Table \[table:bimodal\_results\_chmc\] and \[table:bimodal\_results\_vlt\_chmc\]. As predicted by our discussion in Section \[sec:vlt\_chmc\_special\], VLT-CHMC has substantially higher acceptance probabilities and, across various parameter settings, is five times more efficient than CHMC with the optimal parameter choice. The time step size ${{\Delta s}}= .75$ for the variable step size integrator was used for all the simulations and was chosen to control the error in the Hamiltonian within a reasonable level along the trajectories. ESS’s were computed using the initial monotone sequence estimator of Geyer. [@geyer92]
Number of steps 5 10 15 20 25 30 35
----------------- ------ ------ ------ ------ ------ ------ ------
Acceptance rate 0.48 0.38 0.37 0.36 0.34 0.33 0.33
ESS 75.7 180 145 83.1 103 123 101
: ESS of CHMC along the first coordinate per $10^5$ force evaluations at the various numbers of numerical integration steps. The number of steps coincides with that of force evaluations.[]{data-label="table:bimodal_results_chmc"}
$t$ 0.50 0.75 1.00 1.25 1.50 1.75 2.00
----------------- ------ ------ ------ ------ ------ ------ ------
Number of steps 13 17 21 24 27 30 33
Acceptance rate 0.81 0.78 0.76 0.75 0.73 0.72 0.71
ESS 899 966 924 992 925 921 805
: ESS of VLT-CHMC along the first coordinate per $10^5$ force evaluations. The integration time $t$ determines the trajectory lengths through the termination criteria in . []{data-label="table:bimodal_results_vlt_chmc"}
Rejection avoiding HMC
----------------------
To illustrate the benefit of the rejection avoiding algorithm described in Section \[sec:rahmc\], we consider the problem of sampling from a probability density function $\rho(x,y) \propto \exp( - U(x,y))$ as plotted in Figure \[fig:club\_shape\_density\]. The density $\rho(x,y)$ is constructed as a (continuous) Gaussian mixture $$\rho(x,y) \propto
\int_1^{10} \frac{1}{\sigma_\mu} \exp \left(- \frac{(x-\mu)^2}{2 \, \sigma_\mu^2} - y^2 \right) {\rm d} \mu$$ where $\sigma_\mu = 0.1 + (\mu / 10)^2$. The density has a property that, along the $x$-axis, the partial derivative $\partial_y U(x,y)$ varies substantially and so does the stable step size for the leap-frog integrator typically employed in HMC. For example, the leap-frog integrator with the step size ${{\Delta t}}\geq 0.4$ approximates the Newton’s equations of motion quite accurately in the region $x > 4$, while the step size of ${{\Delta t}}\approx 0.2$ is required for a numerically stable approximation in the region $x < 2$. In in practice, such a knowledge is obviously not available to us and the appropriate step size must be determined empirically from preliminary runs of HMC. A common strategy is to pick a target acceptance rate for the HMC proposals, typically in the range $0.65 \sim 0.8$, and tune the step size accordingly. [@beskos13; @neal10; @stan15] This approach would suggest a step size well above the stability in this example, however. Figure \[fig:club\_shape\_acceptance\_rate\] shows that the acceptance rate of HMC to be quite high even for the step size ${{\Delta t}}= 0.4$. The acceptance rate can be high despite some unstable trajectories because the region where the approximation become unstable contains relatively small, though not negligible, probability. On the other hand, the performance of HMC is severely undermined by the choice of a too large step size as can be seen in Figure \[fig:club\_shape\_ess\]. The ESS’s for $10^6$ force evaluations, estimated from ten independent simulations, are shown so that the computational cost is fixed across the experiments. The error tolerance in Hamiltonian, as in , for rejection avoiding HMC is set to $\epsilon = 3$. When ${{\Delta t}}= 0.2$, less than 1% of trajectories experience the error in Hamiltonian above the tolerance, so there is no practical difference between HMC with and without rejection avoidance. However, without rejection avoidance, the ESS is reduced by the factor as large as five when increasing the step size from ${{\Delta t}}= 0.2$ to ${{\Delta t}}= 0.3$. The performance degradation is less severe for rejection avoiding HMC as the algorithm concentrates the computational efforts on the stable portions of approximated trajectories.
In summary, choosing an optimal step size for HMC is difficult in practice as the choice must be made without the detailed knowledge of a target density. A step size can appear to approximate the dynamics accurately but be above the stability limit in some regions. Rejection avoiding HMC can alleviate the effect of a suboptimal step size choice and provides far more ESS’s than the standard HMC in such situations.
![A plot of (unnormalized) probability density function $\rho(x,y) \propto \exp( - U(x,y))$ used to illustrate the benefit of rejection avoiding HMC.[]{data-label="fig:club_shape_density"}](club_shape_dist.pdf){width="1\linewidth"}
![Acceptance rate of HMC proposals at various settings of step size and integration time when sampling from the density shown in Figure \[fig:club\_shape\_density\].[]{data-label="fig:club_shape_acceptance_rate"}](club_shape_simulation_accept_rate.pdf){width=".9\linewidth"}
![ESS per $10^6$ force evaluations at various settings of step size and integration time. The ESS’s are for the mean estimation along the $x$-axis. []{data-label="fig:club_shape_ess"}](RAHMC_club_shape_simulation_summary.pdf){width=".9\linewidth"}
Acknowledgments
===============
We would like to thank Jiangfeng Lu for his feedback on a preliminary draft of the manuscript.
Derivation of limiting acceptance probability {#app:acceptance_prob_derivation}
=============================================
In this section we analyse the acceptance probability of CHMC and VLT-CHMC algorithms in the special case of RMHMC with variable step size integrators as described in Section \[sec:vlt\_chmc\_special\]. We derive explicit formulas as well as useful bounds on the acceptance probabilities in the limit ${{\Delta s}}\to 0$.
Acceptance probability of CHMC {#app:accept_prob_of_chmc}
------------------------------
When approximating a time-rescaled Hamiltonian dynamics with a reversible map ${\bm{F}}_{{\Delta s}}$ as in , the acceptance probability of the CHMC proposal from $({\bm{q}},{\bm{p}})$ is calculated by the formula $$1 \wedge \frac{\rho({\bm{R}}\circ {\bm{F}}_{{\Delta s}}^n({\bm{q}},{\bm{p}})) \left| {{\bm{D}}{\bm{F}}_{{\Delta s}}^n({\bm{q}},{\bm{p}})} \right| }{\rho({\bm{q}},{\bm{p}})}$$ In the limit ${{\Delta s}}\to 0$ and $n {{\Delta s}}\to {s'}$, the above quantity converges to $$1 \wedge \frac{\rho({\bm{R}}\circ {\bm{\Phi}}_{s'}({\bm{q}},{\bm{p}})) \left| {\bm{D}}{\bm{\Phi}}_{s'}({\bm{q}},{\bm{p}}) \right|}{\rho({\bm{q}},{\bm{p}})}$$ where ${\bm{\Phi}}_{s}$ is the solution operator of the dynamics i.e. $$\frac{\text{d} {\bm{\Phi}}_s}{\text{d} s}
= (g \circ {\bm{\Phi}}_s) \ \mathbf{f} \circ {\bm{\Phi}}_s$$ where $\mathbf{f} = (\nabla_{{\bm{p}}} H, - \nabla_{{\bm{q}}} H)$. We have $\rho \circ {\bm{\Phi}}_{s'} = \rho$ since Hamiltonian dynamics conserves the energy and so does the time-rescaled dynamics. We also have $\rho \circ {\bm{R}}= \rho$, so that $\rho({\bm{R}}\circ {\bm{\Phi}}_{s'}({\bm{q}},{\bm{p}})) = \rho({\bm{q}},{\bm{p}})$. To establish the limiting acceptance probability , therefore, it remains to show that $\left| {\bm{D}}{\bm{\Phi}}_{s'}({\bm{q}},{\bm{p}}) \right| = g({\bm{\Phi}}_{s'}({\bm{q}},{\bm{p}})) / g({\bm{q}},{\bm{p}})$. The Jacobian ${\bm{D}}{\bm{\Phi}}_s$ satisfies a matrix-valued differential equation $\frac{\partial}{\partial s} {\bm{D}}{\bm{\Phi}}_s = {\bm{D}}\mathbf{f} \circ {\bm{\Phi}}_s \, {\bm{D}}{\bm{\Phi}}_s$ and therefore Liouville’s formula tells us that $$\begin{aligned}
\left| {\bm{D}}{\bm{\Phi}}_{s'} \right|
&= \exp \left( \int_0^{s'} {\rm tr} \left( {\bm{D}}\mathbf{f} \circ {\bm{\Phi}}_s \right) {\rm d} s \right)
\end{aligned}$$ A straightforward calculation shows that ${\rm tr} \left( {\bm{D}}\mathbf{f} \circ {\bm{\Phi}}_s \right) = \frac{\partial}{\partial s} \log g \circ {\bm{\Phi}}_s$, from which the identity $\left| {\bm{D}}{\bm{\Phi}}_{s'} \right| = g \circ {\bm{\Phi}}_{s'} / g$ follows.
Acceptance probability of VLT-CHMC {#app:accept_prob_of_vlt_chmc}
----------------------------------
In the derivation below, we will follow the notations of Section \[sec:vlt\_chmc\_algorithm\_in\_special\_case\]. Namely, we set $({\bm{q}}_0^*, {\bm{p}}_0^*) = {\bm{R}}\circ {\bm{F}}^N({\bm{q}}_0,{\bm{p}}_0)$, $({\bm{q}}_i, {\bm{p}}_i) = {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0, {\bm{p}}_0)$, and $({\bm{q}}_i^*, {\bm{p}}_i^*) = {\bm{F}}_{{\Delta s}}^i({\bm{q}}_0^*, {\bm{p}}_0^*)$. The trajectory length function $N = N(t)$ is defined as in and the sets $S$ and $S^*$ as in Algorithm \[alg:vlt\_chmc\]. Note that $({\bm{q}}_0,{\bm{p}}_0)$ is fixed, but other quantities depend on ${{\Delta s}}$, including but not limited to $({\bm{q}}_i, {\bm{p}}_i)$’s, $N({\bm{q}}_0, {\bm{p}}_0)$, and $S$. We do not denote the dependence explicitly but it is implied.
We will show that the acceptance probability of the transition from $S$ to $S^*$ converges to $$\label{eq:accept_prob_limit_vltchmc}
1 \wedge \frac{g({\bm{\Phi}}_t({\bm{q}}_0, {\bm{p}}_0)) |S^*|}{g({\bm{q}}_0,{\bm{p}}_0) |S|}$$ as ${{\Delta s}}\to 0$ while $t$ fixed. Moreover, if $g({\bm{\Phi}}_t({\bm{q}}_0, {\bm{p}}_0)) < g({\bm{q}}_0, {\bm{p}}_0)$, then in the limit ${{\Delta s}}\to 0$ we have $|S| = 1$ and $$\label{eq:set_size_vlt_chmc}
\frac{g({\bm{q}}_0,{\bm{p}}_0)}{g({\bm{\Phi}}_t({\bm{q}}_0, {\bm{p}}_0))} - 1
\leq \left| S^* \right|
\leq \frac{g({\bm{q}}_0,{\bm{p}}_0)}{g({\bm{\Phi}}_t({\bm{q}}_0, {\bm{p}}_0))} + 1$$ The claimed lower bound on the acceptance probability follows immediately from and .
It is not difficult to show that ${\rm diam}(S) \to 0$ and ${\rm diam}(S^*) \to 0$ as ${{\Delta s}}\to 0$. This means that the elements of $S$ (and of $S^*$) collapse to a single state as ${{\Delta s}}\to 0$. More precisely, for all $- \ell \leq i \leq r$ and $- r^* \leq j \leq \ell^*$, $$\label{eq:set_size_collapsing}
\begin{aligned}
{\bm{F}}_{{\Delta s}}^i({\bm{q}}_0,{\bm{p}}_0)
&\to ({\bm{q}}_0,{\bm{p}}_0) \\
{\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0,{\bm{p}}_0)
&\to {\bm{R}}\circ {\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)
\end{aligned}$$ where $N_0 = N({\bm{q}}_0, {\bm{p}}_0)$ and $r, \ell, r^*, \ell^*$ are defined as in and . It follows that $$\label{subeq:denom_num_limit}
\begin{aligned}
&\rho \left({\bm{F}}_{{\Delta s}}^i({\bm{q}}_0,{\bm{p}}_0) \right) {\left| {\bm{D}}{\bm{F}}_{{\Delta s}}^{i}({\bm{q}}_{0},{\bm{p}}_{0}) \right|}
\to \rho({\bm{q}}_0,{\bm{p}}_0) \\
&\rho \left({\bm{R}}\circ {\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0,{\bm{p}}_0) \right) \left| {\bm{D}}{\bm{F}}_{{\Delta s}}^{N_0+j}({\bm{q}}_0,{\bm{p}}_0) \right| \\
&\hspace{5em} \to \rho({\bm{R}}\circ {\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) \left| {\bm{D}}{\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) \right|
\end{aligned}$$ By the same argument as in Section \[app:accept\_prob\_of\_chmc\], we can show that $$\begin{aligned}
\rho({\bm{R}}\circ {\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0))
&\left| {\bm{D}}{\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0)) \right| \\
&= \rho({\bm{q}}_0, {\bm{p}}_0) \frac{g({\bm{\Phi}}_t({\bm{q}}_0,{\bm{p}}_0))}{g({\bm{q}}_0,{\bm{p}}_0)}
\end{aligned}$$ establishing the claimed formula .
We now turn to the proof of the inequality . The intuition behind the inequality and the proof below is that the size of the set ${\left| S^* \right|}$ is roughly equal to the number of intervals of length ${{\Delta s}}\cdot g({\bm{q}}_0^*,{\bm{p}}_0^*)$ that can be fit inside the interval $\left( t, \, t + {{\Delta s}}\cdot g({\bm{q}}_0,{\bm{p}}_0) \right)$. Denote $N_0^* = N({\bm{q}}_0^*, {\bm{p}}_0^*)$. By the definition of $N_0^*$, $r^*$, and $\ell^*$, we must have $$\begin{aligned}
&\sum_{i=-\ell^*+1}^{N_0^*-1}\frac{{{\Delta s}}}{2} \left(g({\bm{q}}_{i-1}^*,{\bm{p}}_{i-1}^*) + g({\bm{q}}_i^*,{\bm{p}}_i^*)\right) \\
&\hspace{5ex} < t
< \sum_{i=r^*+1}^{N_0^*} \frac{{{\Delta s}}}{2} \left(g({\bm{q}}_{i-1}^*,{\bm{p}}_{i-1}^*) + g({\bm{q}}_i^*,{\bm{p}}_i^*)\right)
\end{aligned}$$ which implies that $$\label{eq:ineq1_for_valet_accept_prob}
\begin{aligned}
&\sum_{i=-\ell^*+1}^{r^*} \frac{1}{2} \left(g({\bm{q}}_{i-1}^*,{\bm{p}}_{i-1}^*) + g({\bm{q}}_i^*,{\bm{p}}_i^*)\right) \\
&\hspace{8ex} < \frac{1}{2} \left(g({\bm{q}}_{N_0^*-1}^*,{\bm{p}}_{N_0^*-1}^*) + g({\bm{q}}_{N_0^*}^*,{\bm{p}}_{N_0^*}^*) \right)
\end{aligned}$$ Also by the definition $N_0^*$, $r^*$, and $\ell^*$, we must have $$\begin{aligned}
&\sum_{i=r^*+2}^{N_0^*}\frac{{{\Delta s}}}{2} \left(g({\bm{q}}_{i-1}^*,{\bm{p}}_{i-1}^*) + g({\bm{q}}_i^*,{\bm{p}}_i^*)\right) \\
&\hspace{6ex} < t
< \sum_{i=-\ell^*}^{N_0^*-1} \frac{{{\Delta s}}}{2} \left(g({\bm{q}}_{i-1}^*,{\bm{p}}_{i-1}^*) + g({\bm{q}}_i^*,{\bm{p}}_i^*)\right)
\end{aligned}$$ which implies that $$\label{eq:ineq2_for_valet_accept_prob}
\begin{aligned}
&\frac{1}{2} \left(g({\bm{q}}_{N_0^*-1}^*,{\bm{p}}_{N_0^*-1}^*) + g({\bm{q}}_{N_0^*}^*,{\bm{p}}_{N_0^*}^*) \right) \\
&\hspace{8ex} < \sum_{i=-\ell^*}^{r*+1} \frac12 \left(g({\bm{q}}_{i-1}^*,{\bm{p}}_{i-1}^*) + g({\bm{q}}_i^*,{\bm{p}}_i^*)\right)
\end{aligned}$$ Since ${\rm diam}(S) \to 0$ and ${\rm diam}(S^*) \to 0$ as ${{\Delta s}}\to 0$, the inequalities and converge to $$\label{subeq:setsize_bounds}
g({\bm{q}}_0^*,{\bm{p}}_0^*) (|S^*|-1)
\leq g({\bm{q}}_0,{\bm{p}}_0)
\leq g({\bm{q}}_0^*,{\bm{p}}_0^*) (|S^*| + 1)$$ The desired inequality is obtained by rearranging the terms in the above inequality.
Finally, we turn to the proof of the fact that ${\left| S \right|} \to 1$ as ${{\Delta s}}\to 0$ when $g({\bm{q}}_0^*,{\bm{p}}_0^*) < g({\bm{q}}_0,{\bm{p}}_0)$. To this end, we only need to note that all the arguments in the proof of remain valid if we switch the role of $({\bm{q}}_i^*,{\bm{p}}_i^*)$, $r^*$, $\ell^*$ and $N_0^*$ with $({\bm{q}}_i,{\bm{p}}_i)$, $r$, $\ell$ and $N_0$. This means that the inequality still holds if we switch the role of $S^*$ with $S$ and of $({\bm{q}}_0^*,{\bm{p}}_0^*)$ with $({\bm{q}}_0,{\bm{p}}_0)$, yielding the inequality $$g({\bm{q}}_0,{\bm{p}}_0) ({\left| S \right|}-1)
\leq g({\bm{q}}_0^*,{\bm{p}}_0^*)
\leq g({\bm{q}}_0,{\bm{p}}_0) ({\left| S \right|} + 1)$$ In particular, we have ${\left| S \right|} \leq \frac{g({\bm{q}}_0^*,{\bm{p}}_0^*)}{g({\bm{q}}_0,{\bm{p}}_0)} + 1$ and hence ${\left| S \right|} = 1$.
Justification of VLT-CHMC algorithm {#app:vlt_chmc_justification}
===================================
As claimed in Section , Algorithm \[alg:vlt\_chmc\] remains a valid algorithm when we replace the reversible map ${\bm{F}}_{{\Delta s}}$ with any reversible map and the trajectory length function $N$ with any function of the form . In Section \[sec:detailed\_balance\_for\_vlt\_chmc\], the detailed balance condition of VLT-CHMC was derived using the notations of Algorithm \[alg:vlt\_chmc\]. However, it is easy to see that the same analysis carries through when we replace the reversible map ${\bm{F}}_{{\Delta s}}$ of Section \[sec:vlt\_chmc\_special\] with any reversible map as long as the set $S$ and $S^*$ satisfies . In this section, we establish the last piece in our proof of the general VLT-CHMC algorithm; the property holds whenever $N$ satisfies the short-return and order-preserving condition .
We consider a generic reversible map ${\bm{F}}$ with an associated involution ${\bm{R}}$ on a general phase space ${\bm{z}}$ as well as a generic trajectory length function $N$ satisfying the short-return and order-preserving condition. However, all the notations and definitions directly parallel those in our presentation of the special case of VLT-CHMC in Section \[sec:vlt\_chmc\_algorithm\_in\_special\_case\]. Fix ${\bm{z}}_0$ and denote ${\bm{z}}_i = {\bm{F}}^i({\bm{z}}_0)$, ${\bm{z}}_0^* = {\bm{R}}\circ {\bm{F}}^N({\bm{z}}_0)$, and ${\bm{z}}_i^* = {\bm{F}}^i({\bm{z}}_0^*)$. A trajectory function $N$ determines the sets via the formula $S = \left\{{\bm{z}}_{- \ell}, \thinspace \ldots, {\bm{z}}_r \right\}$ and $S^* = \left\{{\bm{z}}^*_{- \ell^*}, \thinspace \ldots, {\bm{z}}^*_{r^*} \right\}$ where $\ell, r, \ell^*, r^* \geq 0$ are defined as $$\label{eq:nstep_forward_backward_general}
\begin{aligned}
\ell
&= \max \left\{ i \geq 0: {\bm{F}}^N({\bm{z}}_{- i}) = {\bm{F}}^N({\bm{z}}_0) \right\} \\
r
&= \max \left\{ i \geq 0: {\bm{F}}^N({\bm{z}}_{i}) = {\bm{F}}^N({\bm{z}}_0) \right\} \\
\ell^*
&= \max \left\{ i \geq 0: {\bm{F}}^N({\bm{z}}^*_{- i}) = {\bm{F}}^N({\bm{z}}^*_0) \right\} \\
r^*
&= \max \left\{ i \geq 0: {\bm{F}}^N({\bm{z}}^*_{i}) = {\bm{F}}^N({\bm{z}}^*_0) \right\}
\end{aligned}$$
To build the intuition behind the proof, we define a partial ordering $\preceq$ on the phase space as follows: $${\bm{z}}\preceq \tilde{{\bm{z}}}
\quad \text{ if } \ {\bm{F}}^i({\bm{z}}) = \tilde{{\bm{z}}}
\text{ for } i \geq 0$$ Note that ${\bm{z}}\preceq \tilde{{\bm{z}}}$ if and only if ${\bm{R}}(\tilde{{\bm{z}}}) \preceq {\bm{R}}({\bm{z}})$, due to the reversibility of ${\bm{F}}$. With this notation, the short-return condition can be expressed as $$\label{eq:short_return_ordering}
{\bm{z}}\preceq {\bm{R}}\circ {\bm{F}}^N({\bm{z}}^*)
\quad \text{ for } \ {\bm{z}}^* = {\bm{R}}\circ {\bm{F}}^N({\bm{z}})$$ The condition can be interpreted intuitively as follows; according to the trajectory termination criteria imposed by $N$, the reverse trajectory ${\bm{z}}^*_0, {\bm{z}}^*_1, \ldots$ must terminate at ${\bm{z}}_0$ or at ${\bm{z}}_i$ for $i > 0$ before coming all the way back to ${\bm{z}}_0$. The order-preserving condition simply amounts to $${\bm{F}}^N({\bm{z}}) \preceq {\bm{F}}^N(\tilde{{\bm{z}}})
\quad \text{ if } \ {\bm{z}}\preceq \tilde{{\bm{z}}}$$
We now show how the order-preserving and short-return condition implies . By the order-preserving condition, we know that $${\bm{F}}^N ({\bm{z}}_{- \ell})
\preceq {\bm{F}}^N ({\bm{z}}_i)
\preceq {\bm{F}}^N ({\bm{z}}_{r})$$ for all $-\ell \leq i \leq r$. On the other hand, we have ${\bm{F}}^N ({\bm{z}}_{- \ell}) = {\bm{F}}^N ({\bm{z}}_{r}) = {\bm{R}}({\bm{z}}_0^*)$ by the definition of $\ell$ and $r$, so it follows that ${\bm{R}}\circ {\bm{F}}^N(\left\{{\bm{z}}_{- \ell}, \thinspace \ldots, {\bm{z}}_r \right\}) = \{{\bm{z}}_0^*\}$.
We now turn to demonstration of ${\bm{R}}\circ {\bm{F}}^N(S^*) = \{{\bm{z}}_r\}$. To this end, it suffices to show ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_0^*) = {\bm{z}}_r$ as the definition of $\ell^*$ and $r^*$ combined with the order-preserving condition implies ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_i^*) = {\bm{R}}\circ {\bm{F}}^N({\bm{z}}_0^*)$ for all $-\ell^* \leq i \leq r^*$. Since ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_r) = {\bm{z}}_0^*$, the short-return condition tells us ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_0^*) = {\bm{z}}_{r+k}$ for some $k \geq 0$. To show that $k = 0$, first observe that an application of the short-return condition to the state ${\bm{z}}_{r+k}$ implies ${\bm{z}}_0^* \preceq {\bm{R}}\circ {\bm{F}}^N({\bm{z}}_{r+k})$. On the other hand, the order-preserving condition implies ${\bm{F}}^N({\bm{z}}_r) \preceq {\bm{F}}^N({\bm{z}}_{r+k})$ and hence ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_{r+k}) \preceq {\bm{R}}\circ {\bm{F}}^N({\bm{z}}_r) = {\bm{z}}_0^*$. The preceding inequalities together show that ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_{r+k}) = {\bm{z}}_0^*$. Since $r$ was defined as the largest integer $i$ such that ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_{i}) = {\bm{z}}_0^*$, it follows that $k = 0$ and ${\bm{R}}\circ {\bm{F}}^N({\bm{z}}_0^*) = {\bm{z}}_r$.
The remaining relations in as well as the fact $r^* = 0$ can be proved similarly with repeated applications of the short-return and order-preserving properties.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We experimentally demonstrated the optical responses of the switching currents in two types of Josephson tunnel junctions: Al/AlOx/Al and Nb/AlOx/Nb. The radiation-induced switching current shifts were measured at ultra-low bath temperature ($T$ $\approx$ $16$ mK). It is observed that the Al-junction has a more sensitive optical response than the Nb-junction, which is as expected since Al electrode has a smaller superconducting gap energy. The minimum detectable radiation powers with the present Al-junction and Nb-junction are $8$ pW (corresponding to $8$ $\times 10^{5}$ incoming photons in one measurement cycle) and $2$ nW respectively. In addition, we found that the radiation-induced thermal effects are dominant in the observed optical responses. Several methods are proposed to further improve the optical responsivity, so that the josephson junction based devices could be applicable in photon detections.'
author:
- Yiwen Wang
- Pinjia Zhou
- Lianfu Wei
- Beihong Zhang
- Qiang Wei
- Jiquan Zhai
- Weiwei Xu
- Chunhai Cao
bibliography:
- 'aipsamp.bib'
nocite: '[@*]'
title: Optical responses of the switching currents in Al and Nb Josephson junctions
---
Superconducting photon detectors at near-infrared wavelengths, with photon-number resolving power, have shown great promises in quantum optics and quantum information applications. The superconducting detectors mainly include superconducting nanowire detectors (SNSPDs) [@G; @H; @A; @AA], transition-edge sensors (TESs) [@Miller; @Miller2], superconducting tunnel junctions (STJs) [@SF; @JD] and microwave kinetic inductance detectors (MKIDs) [@pkday1; @gao1; @gao2], etc.. Actually, photon detections can also be achieved through other ways, such as by measuring the changes in the critical current of a Josephson tunnel junction due to radiation. Physically, a photon with sufficient energy $h\nu$ ($>$ $2\Delta$) can directly break $\eta h\nu/2\Delta$ Cooper pairs, where $\Delta$ is the superconducting gap energy and $\eta$ the convert efficiency. Therefore, when a photon is incident on one superconducting electrode of the Josephson junction, excess quasiparticles will be excited and the Cooper pair density on the irradiated electrode will decrease. This will lead to an abrupt reduction in the critical current $I_{c}$ since [@feynman] $I_{c}\propto \sqrt{\rho_{1}\rho_{2}}$, where $\rho_{i}$ is the Cooper pair density on the $i$-th electrode. On the other hand, phonons in the substrate around the radiation center may be excited and thus cause a temperature increase nearby the junction area. This thermal effect can also reduce the critical current $I_{c}$ based on the relation [@AB] $$I_{c}R_{n}=[\pi\Delta(T)/2e]\tanh[\Delta(T)/2k_{B}T],$$ with $R_{n}$ being the normal state resistance and $T$ the bath temperature. Since the gap energy $\Delta(T)$ decreases with $T$, the critical current $I_{c}$ also decreases with $T$. Therefore, both pure pair-breaking effects and thermal effects can lead to a reduction in the critical current. This provides a feasible way to detect the incident photons via measuring the radiation-induced changes in the critical current of a Josephon junction.
Note that the well-known $ac$ Josephson effect was utilized to detect the microwave and far-infrared radiation several years ago [@shapiro]. Later, the superconducting gap voltage shifts due to visible and infrared radiation were measured in Nb/AlOx/Nb junctions [@Osterman1] and junction arrays [@Osterman2] at temperatures around $4.2$ K. Specifically, the junctions immersed in superfluid helium were observed to have lower optical responsivity compared to those in vacuum. This is because the heating effect is suppressed in liquid helium and thus the optical responses of the devices are entirely due to the pair-breaking mechanism. Other experiments [@wire; @Mont1; @Mont2; @Mont3] with Nb junctions had also verified these responses due to pure pair-breaking and thermal effects.
However, the previous experiments were all done with Nb junctions. In our experiments, we studied the optical responses of both Al- and Nb junctions. We found that the Al-junction has a more sensitive optical response than Nb-junction. This is a reasonable observation since aluminum has a smaller gap energy. Thus, a certain radiation energy can break more Cooper pairs on Al electrode. Besides, in all of the previous experiments the Josephson junctions were biased at constant currents and the gap voltage shifts were measured as the optical responses. Alternatively, we swept the bias current through the junction and measured the switching current responses to a continuous radiation at $1550$ nm. This detection approach is relatively simple and has not been reported before, as far as we know. Moreover, the previous experiments were all done at temperatures around $1$ K $\sim$ $4.2$ K while our system works in an ultra-low temperature regime, i.e., the bath temperature $T$ $\approx$ $16$ mK. Thermal noise in the circuit is minimized at such low temperatures and thus devices are expected to have more sensitive optical responses.
For our measurements, the Al/AlOx/Al junction was fabricated by electron beam double-angle evaporation, and the Nb/AlOx/Nb junction was fabricated by magnetron sputtering and ion etching. For both Al and Nb devices, the junction areas are about $6$ $\mu m^{2}$ and the top electrodes exposed for illumination are about $100$ nm thick. The chips are cut to approximately $2$ mm $\times$ $2$ mm, with Si-substrates of $0.5$ mm thick. Both junctions are slightly damped [@Stewart] and show hysteretic IV curves with small retrapping currents and sharp onsets of voltage at the maximum bias currents (i.e., the switching currents).
[f1.eps]{}
The schematics of our measurement setup are shown in Fig. 1. The measured junction is placed in a superconducting aluminum sample cell, mounted at the mixing chamber in a dilution refrigerator with base temperature around $10$ mK. Four-probe technique is used to measure the current-voltage characteristics of the devices. The waveform generator can output a voltage signal, which is applied to a resistor to generate a bias current through the junction. The voltage response is amplified by a battery-powered pre-amplifier and then fed into a timer. All electrical leads, connecting the sample cell to room temperature electronics, are filtered by low-pass RC filters and copper powder microwave filters. To radiate the junction, a single-mode optical fiber is set up from the room temperature environment down to the sample cell. A laser source is connected to the top end of the fiber and generates a steady radiation of wavelength $1550$ nm. The bottom end of the fiber is carefully aligned and fixed, so that the laser beam can focus on the top electrode of the junction. The fiber end is estimated to be about $200$ $\mu$m vertically away from the chip surface and the irradiated area is about $80$ $\mu$m in diameter. Therefore, the junction area is completely covered by light.
Due to the presence of thermal fluctuations and quantum tunneling, the junction switches from the zero-voltage state to the finite voltage state at a bias current $I_{s}$ smaller than its critical current $I_{c}$. Since this switching is a deterministic random process, the switching current $I_{s}$ shows a Lorentzian distribution [@Fulton; @Barone], which can be mainly characterized by the width $\sigma_{s}$ and mean value $\langle I_{s}\rangle$. In our experiment the switching current distribution $P(I_{s})$ is measured by using the time-of-flight method [@yiwen]. For each switching event, the bias current is ramped linearly from a value below zero up to a value higher than the critical current $I_{c}$. When the junction switches from the zero-voltage state to the finite-voltage state, the timer will be triggered to record the switching time and the corresponding switching current $I_{s}$ can be calculated from the ramping rate. The bias current is then reduced to below zero, resetting the junction to the zero-voltage state. The repetition frequency is $71.3$ Hz and the measurement cycle is repeated $2\times10^{3}$ times to obtain an ensemble of $I_{s}$, from which the distribution of switching current $P(I_{s})$ can be obtained.
[f2.eps]{}
Fig. 2 plots the measured switching current distribution, i.e., the switching probability $P(I_{s})$ as a function of the switching current $I_{s}$. Fig. 2(a) shows the switching current measurements on Nb/AlOx/Nb at temperature $16$ mK. The blue curve is the distribution in the absence of radiation, from which one can calculate the mean switching current $\langle I_{s0}(Nb)\rangle $ = $124.78$ $\mu$A and the distribution width (standard deviation) $\sigma_{s0}(Nb)$ = $102.73$ nA. The red curve corresponds to a $10$ nW radiation on the junction electrode. In this case the mean switching current shifts down to $\langle I_{s}(Nb)\rangle $ = $124.66$ $\mu$A and the distribution width $\sigma_{s}(Nb)$ = $98.00$ nA. Fig. 2(b) shows the same measurements on Al/AlOx/Al under the same experimental conditions. Without radiation, the mean switching current is $\langle I_{s0}(Al)\rangle $ = $80.44$ nA and the distribution width $\sigma_{s0}(Al)$ = $2.53$ nA. In the presence of $10$ nW radiation, the mean switching current is $\langle I_{s1}(Al)\rangle $ = $72.65$ nA and the distribution width $\sigma_{s1}(Al)$ = $2.47$ nA.
There are two ways to define the photon responsivity of the device. One is the ratio of the response to noise, i.e., $R_{a}$ $=$ $\Delta
I_{s}/\sigma_{s0}$ $=$ $(\langle I_{s0} \rangle -\langle I_{s1}
\rangle)/\sigma_{s0}$. By this way, we have $R_{a}$ $=$ $1.17$ for the Nb-junction and $R_{a}$ $=$ $3.08$ for the Al-junction, showing that the Al device has a higher photon responsivity. The second way to define photon responsivity is the relative shift of switching current, i.e., $R_{b}$ $=$ $\Delta I_{s}/I_{s0}$ = $(\langle
I_{s0}\rangle - \langle I_{s1}\rangle)/\langle I_{s0}\rangle$. In this way, we obtain $R_{b}$ = $9.6$ $\times 10^{-4}$ for the Nb-junction and $R_{b}$ $=$ $9.7$ $\times 10^{-2}$ for the Al-junction, showing again that the Al device has a more sensitive response. We take the second definition of responsivity in the following discussions.
[f3.eps]{}
We now investigate the optical responses of the switching currents under different radiation powers. To this aim we varied the light intensity and measured the corresponding average switching current at the base temperature $T$ = $16$ mK. Fig. 3 shows the relative switching current shift $\Delta I_{s}/I_{s0}$ (i.e., the responsivity $R_{b}$) as a function of the radiation power. The black squares and red circles correspond to irradiations on Al- and Nb-junction respectively. It is shown that the logarithmic switching current shift increases linearly with the logarithmic radiation power in the applied power range. By fitting the line slope, one can find that $R_{b}$ is approximately proportional to $P^{0.6}$ for the Al-junction while proportional to $P^{1.2}$ for the Nb-junction. The minimum radiation power that the Al device can detect is about $8$ pW (corresponding to $8$ $\times 10^{5}$ incoming photons per measurement cycle), which is much smaller than the minimum power of $2$ nW that Nb device can detect. The device response to low radiation power is limited by the average switching current fluctuations, which are mainly due to the inevitable low-frequency noises in the electronics.
The switching current shift increases with the radiation power, which is qualitatively similar to its bath temperature dependence. The inset of Fig. 3 shows $\Delta I_{s}/I_{s0}$ as a function of the bath temperature in the range of $30$ mK to $400$ mK, where the measured switching current shift increases with temperature. This suggests that the thermal effects are dominant in the observed radiation power dependence of $R_{b}$ at $T =$ $16$ mK. Experimentally, most of the photons are incident on the substrate rather than the superconducting electrode. Thus, the chip will be mainly heated and achieve an effective temperature greater that the bath temperature, since we are continuously pumping energy into the system. To verify that the thermal effects dominate the radiation power dependence of switching current, we moved the fiber to radiate directly on a small area of the bare substrate, which is about $0.7$ mm away from the junction area. We then performed the same switching current measurements at $16$ mK and obtained the radiation power dependence of $R_{b}$, shown in Fig. 3 (black circles). It is shown that the photon responsivity is apparently weaker, when radiating on the bare substrate of a certain distance away from the junction than that when focusing on the junction area. This is a reasonable result, which can be attributed to a nonuniform temperature distribution around the irradiated area. The effective temperature at the junction area is lower when the light spot is moved $0.7$ mm away. The $R_{b}$ exhibits the same radiation power law dependence (the same slope) for both cases of radiation on the junction and the substrate, indicating that the thermal effect is the main factor in shifting the switching current.
[f4.eps]{}
Furthermore, we measured the variations of the switching current distribution due to the changes in bath temperature and radiation power independently. Fig. 4 plots the distribution width $\sigma_{s}$ as a function of the average switching current $\langle
I_{s}\rangle $ for the Nb device. Here, the black squares correspond to the data at different bath temperatures and without radiation, while the red circles correspond to different radiation powers and at the lowest bath temperature. The inset shows the same plots for the Al device. It is seen that for both Nb and Al devices, $\sigma_{s}$ approximately follows the same function of $\langle
I_{s}\rangle$ by varying the bath temperatures or radiation powers: the distribution width has a plateau in higher switching current regime and then decreases with decreasing switching current monotonically [@yuhaifeng]. In another word, one can obtain a certain switching current distribution $P(I_{s})$ by radiating the junction with a certain power at a fixed bath temperature, and the same $P(I_{s})$ (i.e., the same $\sigma_{s}$ and $\langle
I_{s}\rangle$) can also be obtained with an un-irradiated junction by setting the bath temperature at a certain value. This suggests again that for the present devices, the optical responses of the switching current are mainly due to thermal effects and therefore the present junctions can be used as a desirable bolometer.
Principally, the weak light detection scheme in time domain is straightforward. One can bias the junction at a current slightly smaller than its switching current in the absence of radiation. If a light pulse with sufficient energy is applied, the switching current of the junction is reduced to below the bias current and then the junction will switch to a finite voltage state. Otherwise the junction will stay in zero voltage state. In this way one can judge if there are incoming photons or not.
Although the photon responsivity of the junction device demonstrated here is obviously lower than that of other superconducting detectors (such as the TESs and MKIDs), its performance in the weak light detection could be further improved by several methods. The first one is to enhance the coupling between the superconducting electrode of the junction and the incident photons. To this aim, one can use the lensed fibers to focus the light on the top electrode so that a maximum energy from the incident photons could be absorbed directly by the Cooper pairs on the electrode. Besides, the top metal electrode can be fabricated as thin as possible so that a certain radiation power can lead to a more reduction in the Cooper pair density. The second method to raise reponsivity is to reduce the thermal conductance between the chip and the sample holder to maximize the energy absorption by the whole chip. For instance, one can fabricate the junction on the substrate with low thermal conductivity (e.g., amorphous glass) or etch the back of the substrate wafer, to reduce the path of heat conduction from the chip to the sample block. By this way unnecessary radiation energy loss can be avoided effectively. The third method is to select materials with lower gap energies as the superconducting electrodes. Finally, for our experiments the fluctuations in the mean switching currents are limited by the low-frequency noises in the circuits, not by intrinsic noises of the junctions. Therefore, the device performance can be potentially improved by further reducing the noise in the measurement electronics.
Besides the optical responsivity, there are two more challenges in our studied detection system. Firstly, the detection mechanism (i.e., by measuring the switching currents of Josephson junctions) can not be made very fast, since it takes time to sweep up the current and then sweep down to reset the junction to zero voltage state. The present measurement can be done at the rate up to several KHz, which is still a lower rate. By improving the bandwidth of the measurement electronics, the measurement rate can be faster but may not easy to get up to MHz. In addition, a faster measurement (i.e., a faster current ramping rate) will broaden the distribution width of the switching currents, which can decrease the detection sensitivity. Secondly, thermal activation is greatly suppressed at ultra-low bath temperatures, but the distribution of the switching currents still has a finite width $\sigma_{q}$ due to quantum tunneling. For our tested Nb-junction sample and experimental parameters, one can calculate [@cal] $\sigma_{q}$ = $101.87$ nA, which is very close to the observed distribution width $\sigma_{s0}(Nb)$ = $102.73$ nA at $T =$ $16$ mK, indicating the quantum tunneling is dominant at ultra-low bath temperatures. This finite distribution width may be translated to high dark counts for photon counting (as shown in Fig. 2, the blue and red curves have overlaps). However, this problem disappears when the junction device is only used as an optical power meter. One can statistically average the switching currents to distinguish the incident light powers. In the case of weak low-frequency circuit noise, the fluctuations in the average switching currents could be very small and thus the junction device can be utilized as a very sensitive radiation power meter.
In summary, we experimentally investigated the optical responses of the switching currents for the Al/AlOx/Al and the Nb/AlOx/Nb Josephson junctions at ultra-low temperatures ($T$ $\approx$ $16$ mK). The radiation power dependence of the relative switching current shifts were measured for both junctions. It was found that the Al-junction has a more sensitive optical response than the Nb-junction. The minimum radiation powers that the Al and Nb devices can respond to are about $8$ pW and $2$ nW respectively. Moreover, the Al-junction has been irradiated directly and indirectly through the substrate. It was observed that, the relative switching current shifts for both cases follows the same radiation power law dependence, indicating that the thermal effects are dominant in the optical responses. Hopefully, the junction devices demonstrated here can be applied to implement photon detections in the future, once the photon responsivity can be further improved.
This work was supported in part by the National Natural Science Foundation (Grant Nos. 61301031, 61371036, 11174373, 11204249), the Fundamental Research Funds for the Central Universities (Grant No. 2682014CX087) and the National Fundamental Research Program of China (Grant No. 2010CB923104). We thank Profs. Peiheng Wu, Xuedong Hu and Yang Yu for kind supports and valuable discussions.
[1]{} G. N. Gol’tsman, O. Okunev, G. Chulkova, A. Lipatov, A. Semenov, K. Smirnov, B. Voronov, A. Dzardanov, C. Williams and R. Sobolewski, App. Phys. Lett. **79**, 705 (2001). H. Kesue, S. W. Nm, Q. Zhng, R. H. Hdfield, T. Honjo, K. Tmki and Y. Ymmoto, Nature Photonics **1**, (343) 2007. Aleksander D., Francesco M., David B., et al., Nature Photonics **2**, (302) 2008. F. Marsili, V. B. Verma, J. A. Stern, S. Harrington, A. E. Lita, T. Gerrits, I. Vayshenker, B. Baek, M. D. Shaw, R. P. Mirin and S. W. Nam, Nature Photonics **7**, (210) 2013. A. J. Miller, S. W. Nam, J. M. Martinis and A. V. Sergienko, App. Phys. Lett. **83**, 791 (2003). L. Adriana E., A. J. Miller, and S. W. Nam, Optics Express **16(5)**, 3032 (2008). S. Friedrich, M. H. Carpenter, O. B. Drury, W. K. Warburton, J.Harris, J. Hall and R. Cantor, J. Low Temp. Phys. **167**, (741) 2012. J. D. Teufel, Ph.D. Thesis, Yale University, 2008. P. K. Day, H. G. LeDuc, B. A. Mazin, A. Vayonakis, and J. Zmuidzinas, Nature(London) **425**, 817 (2003). J. Gao, M. R. Visser, M. O. Sandberg, F. C. S. da Silva, S. W. Nam, D. P. Pappas, D. S. Wisbey, E. C. Langman, S. R. Meeker, B. A. Mazin, H. G. Leduc, J. Zmuidzinas, and K. D. Irwin, App. Phys. Lett. **101**, 142602 (2012). J. Gao, Ph.D. thesis, Caltech, 2008. R. P. Feynman, B. L. Robert, and L. S. Matthew, The Feynman Lectures on Physics, Vol. **3**, Basic Books, 2011. V. Ambegaokar and A. Baratoff, Phys. Rev. Lett. **10**, 486 (1963). C. C. Grimes, P. L. Richards, and Sidney Shapiro, J. App. Phys. **39(8)**, 3905 (1968). D. P. Osterman, M. Radparvar, and S. M. Faris, IEEE Trans. Magn. **25(2)**, 1319 (1989). D. P. Osterman, P. Marr, H. Dang, C-T. Yao, and M. Radparvar, IEEE Trans. Magn. **27(2)**, 2681 (1991). M. S. Wire, L. O. Heflinger, B. J. Dalrymple, M. Leung, T. Pham, L. R. Eaton, and A. H. Silver, IEEE Trans. App. Superconduct. **3(1)**, 2107 (1993). E. Monticone, V. Lacquaniti, R. Steni, M. Rajteri, M. L. Rastello, L. Parlato, and G. Ammendola, IEEE Trans. App. Superconduct. **9(2)**, 3866 (1999). E. Monticone, M. Rajteri, R. Steni, M. L. Rastello, V. Lacquaniti, G. P. Pepe, L. Parlato, and G. Ammendola, Int. J. Mod. Phys. B. **13(09n10)**, 1283 (1998). M. Rajteri, E. Monticone, G. B. Picotto, R. Steni, M. Rastello, and V. Lacquaniti, in Proceedings ICEC 17, D. Dew-Hughes, R. G. Scurlock and J. H. P. Watson, Eds. Bristol and Phyladelphia: IOP Publishing, 711 (1998). W. C. Stewart, App. Phys. Lett. **12**, 277 (1968). T. A. Fulton and L. N. Dunkleberger, Phys. Rev. B **9**, 4760 (1974). A. Barone, R. Cristiano, and P. Silvestrini, J. App. Phys. **58**, 3822 (1985). G. Z. Sun, Y. W. Wang, J. Y. Cao, J. Chen, Z. M. Ji, L. Kang, W. W. Xu, Y. Yang, S. Y. Han, and P. H. Wu, Phys. Rev. B 77, 104531 (2008) H. F. Yu, X. B. Zhu, Z. H. Peng, Y. Tian, D. J. Cui, G. H. Chen, D. N. Zheng, X. N. Jing, L. Lu, and S. P. Zhao, Phys. Rev. Lett. **107**, 067004 (2011). The theoretical distribution of switching currents due to quantum tunnelling is given by the equation: $P(I_s)=\frac{\Gamma_{q}(I_s)}{dI/dt}\exp\left[-\frac{1}{dI/dt}\int_0^{I_s}\Gamma_{q}(I)dI\right]$, where $dI/dt$ $=$ $1.63\times 10^{-2}A/s$ is the bias current ramp rate and $\Gamma_{q}(I)$ is the well-known macroscopic quantum tunnelling (MQT) rate at bias current $I$. To calculate $\Gamma_{q}$, one needs to know the critical current $I_{c}$ and shunt capacitance $C$ of the Nb-junction. In our experiment, we obtain $I_{c}$ $=$ $126.66$ $\mu$A by fitting to the average switching current and $C$ $=$ $0.26$ pF from independent microwave resonance experiment. The distribution width $\sigma_{q}$ = $101.87$ nA can then be extracted once the distribution $P(I_s)$ is obtained.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The present paper is devoted to the boundedness of fractional integral operators in Morrey spaces defined on quasimetric measure spaces. In particular, Sobolev, trace and weighted inequalities with power weights for potential operators are established. In the case when measure satisfies the doubling condition the derived conditions are simultaneously necessary and sufficient for appropriate inequalities.'
author:
- 'Eridani$^\diamond$, Vakhtang Kokilashvili,$^\dagger$'
- |
Alexander Meskhi$^\ddagger$\
\
title: Morrey spaces and fractional integral operators
---
**
$^\diamond$Department of Mathematics,\
Airlangga University, Campus C, Mulyorejo,\
Surabaya 60115, Indonesia.\
\
$^\dagger$$^{,\ddagger}$A. Razmadze Mathematical Institute,\
M. Aleksidze St. 0193, Tbilisi 380093, Georgia.\
$^\diamond$$^{,\ddagger}$[*School of Mathematical Sciences,\
Government College University,\
68-B New Muslim Town, Lahore, Pakistan.*]{}\
+0.5cm
*Keywords*: Fractional integral operators, (weighted) Morrey spaces, trace inequality, two-weight inequality, doubling conditions, growth condition.
+0.5cm
[**2000 Mathematics Subject Classification**]{}: 42B35, 47B38. +0.5cm
Introduction
============
The main purpose of this paper is to establish the boundedness of fractional integral operators in (weighted) Morrey spaces defined on quasimetric measure spaces. We derive Sobolev, trace and two-weight inequalities for fractional integrals. In particular, we generalize: a) D. Adams [@Adm] trace inequality; b) the theorem by E. M. Stein and G. Weiss [@ST1] regarding the two-weight inequality for the Riesz potentials; c) Sobolev-type inequality. We emphasize that in the most cases the derived conditions are necessary and sufficient for appropriate inequalities.
In the paper [@GGK] (see also [@GGKK], Ch. 2) integral-type sufficient condition guaranteeing the two-weight weak-type inequality for integral operator with positive kernel defined on nonhomogeneous spaces was established. In the same paper (see also [@GGKK], Ch. 2) the authors solved the two-weight problem for kernel operators on spaces of homogeneous type.
In [@KM0] (see also [@EKM], Ch.6) a complete description of non-doubling measure $\mu$ guaranteeing the boundedness of fractional integral operator ${I_\alpha}$ (see the next section for the definition) from $L^p(\mu, X)$ to $L^q(\mu, X),\quad
1<p<q<\infty,$ was given. We notice that this result for potentials was derived in [@Ko] for potentials on Euclidean spaces. In [@KM0], theorems of Sobolev and Adams type for fractional integrals defined on quasimetric measure spaces were established. For the boundedness of fractional integrals on metric measure spaces we refer also to [@GCG]. Some two-weight norm inequalities for fractional operators on ${{\mathbf R}^n}$ with non-doubling measure were studied in [@GCM]. Further, in the paper [@KM1] necessary and sufficient conditions on measure $\mu$ governing the inequality of Stein-Weiss type on nonhomogeneous spaces were established.
The boundedness of the Riesz potential in Morrey spaces defined on Euclidean spaces was studied in [@Pe] and [@Ad1]. The same problem for fractional integrals on ${\mathbb{R}}^n$ with non-doubling measure was investigated in [@ST].
Finally we mention that necessary and sufficient conditions for the boundedness of maximal operators and Riesz potentials in the local Morrey-type spaces were derived in [@BG], [@BGG].
The main results of this paper were presented in [@ErKoMe].
It should be emphasized that the results of this work are new even for Euclidean spaces.
Constants (often different constants in the same series of inequalities) will generally be denoted by $c$ or $C$.
Preliminaries
=============
Throughout the paper we assume that $X:=(X, \rho, \mu)$ is a topological space, endowed with a complete measure $\mu$ such that the space of compactly supported continuous functions is dense in $L^1(X,\mu)$ and there exists a function (quasimetric) $\rho
:X\times X \longrightarrow [0, \infty)$ satisfying the conditions:
\(1) $\rho(x,y)>0$ for all $x\neq y,$ and $\rho(x,x)=0$ for all $x\in X$;
\(2) there exists a constant $a_0\geq 1$, such that $\rho(x,y)\leq a_0\rho(y,x)$ for all $x,\,y\in X;$
\(3) there exists a constant $a_1\geq 1$, such that $\rho(x,y)\leq a_1(\rho(x,z)+\rho(z,y))$ for all $x,\,y,\,z\in X.$
We assume that the balls $B(a,r):=\{x\in
X:\rho(x,a)<r\}$ are measurable, for $a \in X,\,r>0,$ and $0\leq\mu (B(a,r))<\infty.$ For every neighborhood $V$ of $x\in
X,$ there exists $r>0,$ such that $B(x,r)\subset V.$ We also assume that $\mu(X)=\infty,\,\mu\{a\}=0,$ and $B(a,r_2)\setminus
B(a,r_1)\neq \emptyset,$ for all $a\in X,\,\,0<r_1<r_2<\infty.$
The triple $(X, \rho, \mu)$ will be called quasimetric measure space.
Let $0<\alpha<1$. We consider the fractional integral operators ${I_\alpha},$ and $K_{\alpha}$ given by $${I_\alpha}f(x):=\int_X
f(y){\rho(x,y)}^{\alpha-1}\,d\mu(y),$$ $$K_{\alpha}f(x):=\int_X f(y){(\mu B(x,\rho(x,y)))}^{\alpha-1}\,d\mu(y),$$ for suitable $f$ on $X$.
Suppose that $\nu$ is another measure on $X$, $\lambda\geq 0$ and $1\leq p<\infty$. We deal with the Morrey space $L^{p,\lambda}(X, \nu,\mu)$, which is the set of all functions $f\in L^p_{\rm loc}(X,\nu)$ such that $${\|f\|}_{L^{p,\lambda}(X,\nu,\mu)}:=\sup_{B}{\left( \frac{1}{\mu (B)^{\lambda}} \int_{B}{|f(y)|}^p \,d\nu(y)\right)
}^{1/p}<\infty,$$ where the supremum is taken over all balls $B$.
If $\nu=\mu$, then we have the classical Morrey space $L^{p,\lambda}(X,\mu)$ with measure $\mu$. When $\nu=\mu$ and $\lambda =0$, then $L^{p,\lambda}(X,\nu,\mu)
=L^p(X,\mu)$ is the Lebesgue space with measure $\mu$.
Further, suppose that $\beta \in \mathbf{R}.$ We are also interested in weighted Morrey space $M_\beta^{p,\lambda}(X,\mu)$ which is the set of all $\mu$-measurable functions $f$ such that $${\|f\|}_{M_\beta^{p,\lambda}(X,\mu)}:=\sup\limits_{a\in X; r>0}{\left( \frac{1}{r^{\lambda}}\int_{B(a,r)}{|f(y)|}^p\rho(a,y)^\beta
\,d\mu(y)\right) }^{1/p}<\infty.$$
If $\beta=0$, then we denote $M^{p,\lambda}_{\beta}(X,\mu):= M^{p,\lambda}(X,\mu)$.
We say that a measure $\mu$ satisfies the growth condition ($\mu
\in (GC)$), if there exists $C_0>0$ such that $\mu(B(a,r))\leq C_0
r$; further, $\mu$ satisfies the doubling condition ($\mu \in$ (DC)) if $\mu(B(a,2r))\leq C_1\,\mu(B(a,r))$ for some $C_1>1.$ If $\mu \in (DC)$, then $(X, \rho, \mu)$ is called a space of homogeneous type (SHT). A quasimetric measure space $(X, \rho,
\mu)$, where the doubling condition might be failed, is also called a non-homogeneous space.
The measure $\mu$ on $X$ satisfies the reverse doubling condition ($\mu \in (RDC)$) if there are constants $\eta_1$ and $\eta_2$ with $\eta_1>1$ and $\eta_2>1$ such that $$\mu B(x, \eta_1 r) \geq \eta_2 \mu B(x, r). \eqno{(1)}$$
+0.1cm
It is known (see e.g. [@StTo], p. 11) that if $\mu \in \; (DC)$, then $\mu \in \; (RDC)$.
The next statements is from [@KM0] (see also [@EKM Theorem 6.1.1, Corollary 6.1.1] and [@Ko] in the case of Euclidean spaces).
[**Theorem A.**]{} [*Let $(X, \rho, \mu)$ be a quasimetric measure space. Suppose that $1<p<q<\infty$ and $0<\alpha<1$. Then ${I_\alpha}$ is bounded from $L^p(X)$ to $L^q(X)$ if and only if there exists a positive constant $C$ such that $$\mu(B(a,r))\le C r^{s},\quad
s=\frac{pq(1-\alpha)}{pq+p-q}, \eqno{(2)}$$ for all $a\in X$ and $r>0$.*]{}
+0.1cm
[**Corollary B.**]{} [*Let Let $(X, \rho, \mu)$ be a quasimetric measure space, $1<p<1/\alpha$ and $1/q=1/p-\alpha.$ Then ${I_\alpha}$ is bounded from $L^p(X)$ to $L^q(X)$ if and only if $\mu \in$ (GC).* ]{}
The latter statement by different proof was also derived in [@GCG] for metric spaces.
We to prove some of our statements we need the following Hardy-type transform:
$$H_a f(x):=\int_{\rho(a,y)\le \rho(a,x)}f(y)\,d\mu(y),$$ where $a$ is a fixed point of $X$ and $f\in L_{\rm loc}(X,\mu)$.
[**Theorem C.**]{} [*Suppose that $(X, \rho, \mu)$ be a quasimetric measure space, $1<p\le q<\infty$ and $V$ and $W$ are non-negative functions defined on $X\times X$. Let $\nu$ be another measures on $X$. If there exists a positive constant $C$ such that for every $a\in X$ and $t>0$, $$\left(\int_{\rho(a,y)\geq t}V(a,y)\,d\nu(y)\right)^{1/q}
\left(\int_{\rho(a,y)\le t}W(a,y)^{1-p'}\,d\mu(y)\right)^{1/p'}
\le C<\infty,$$ then there exists a positive constant $c$ such that for all $f\geq 0$ and $a\in X$ the inequality $$\left(\int_{B(a,r)} (H_af(x))^qV(a,x)\,d\nu(x)\right)^{1/q}\le
c\,\left(\int_{B(a,r)}(f(x))^pW(a,x)\,d\mu(x)\right)^{1/p}$$ holds.* ]{}
This statement was proved in [@EKM Section 1.1] for Lebesgue spaces.
[*Proof of Theorem C.*]{} Let $f\geq 0$. We define $S(s):=\int_{\rho(a,y)<s} f(y)\,d\mu(y),$ for $s\in [0,r].$ Suppose $S(r)<\infty,$ then $2^m<S(r)\le 2^{m+1},$ for some $m\in
\mathbb{Z}.$ Let $$s_j:=\sup \{t: S(t)\le 2^j\},\,\,j\le m,\quad {\rm and}\quad
s_{m+1}:=r.$$ Then it is easy to see that (see also [@EKM pp.5-8] for details) $(s_j)^{m+1}_{j=-\infty}$ is a non-decreasing sequence, $S(s_j)\le
2^j,\,S(t)\geq 2^j$ for $t>s_j,$ and $$2^j\le \int_{s_j\le \rho(a,y)\le s_{j+1}} f(y)\,d\mu(y).$$ If $\beta:=\lim\limits_{j\rightarrow -\infty}s_j,$ then $$\rho(a,x)<r \Leftrightarrow \rho(a,x)\in [0,\beta]\cup
\bigcup_{j=-\infty}^m(s_j,s_{j+1}].$$ If $S(r)=\infty,$ then we may put $m=\infty.$ Since $$0\le \int_{\rho(a,y)<\beta}f(y)\,d\mu(y)\le S(s_j)\le 2^j,$$ for every $j,$ therefore $\int_{\rho(a,y)<\beta}f(y)\,d\mu(y)=0.$ From these observations, we have
$$\int_{\rho(a,x)<r}(H_af(x))^q V(a,x)\,d\nu(x) \le
\sum_{j=-\infty}^m \int_{s_j\le \rho(a,x)\le s_{j+1}}(H_af(x))^q
V(a,x)\,d\nu(x)$$
$$\le \sum_{j=-\infty}^m \int_{s_j\le \rho(a,x)\le s_{j+1}}V(a,x)
\left(\int_{\rho(a,y)\le s_{j+1}}(f(y))\,d\mu(y)\right)^qd\nu(x).$$
Notice that
$$\int_{\rho(a,y)\le s_{j+1}} f d\mu \leq S(s_{j+2}) \le 2^{j+2}\le C\,\int_{s_{j-1}\le \rho(a,y)\le
s_j}fd\mu.$$
Using Hölder’s inequality, we find that
$$\int_{\rho(a,x)<r}(H_af(x))^qV(a,x)\,d\mu(x)$$
$$\le \sum_{j=-\infty}^m \int_{s_j\le \rho(a,x)\le s_{j+1}} V(a,x)
\left(\int_{\rho(a,y)\le s_{j+1}}(f(y))\,d\mu(y)\right)^qd\nu(x)$$
$$\le C\,\sum_{j=-\infty}^m \int_{s_j\le \rho(a,x)\le s_{j+1}}
V(a,x) \left(\int_{s_{j-1}\leq \rho(a,y)\le
s_j}(f(y))\,d\mu(y)\right)^q d\nu(x)$$
$$\le C\,\sum_{j=-\infty}^m \int_{s_j\le \rho(a,x)\le
s_{j+1}}V(a,x)\,d\nu(x)\left(\int_{s_{j-1}\leq \rho(a,y)\le
s_j}(f(y))^pW(a,y)\,d\mu(y)\right)^{q/p}$$
$$\times \left(\int_{s_{j-1}\leq \rho(a,y)\le
s_j}W(a,y)^{1-p'}\,d\mu(y)\right)^{q/p'}$$
$$\le C\,\sum_{j=-\infty}^m \int_{s_j\le
\rho(a,y)}V(a,y)\,d\nu(y)\left(\int_{\rho(a,y)\le
s_j}W(a,y)^{1-p'}\,d\mu(y)\right)^{q/p'}$$
$$\times \left(\int_{s_{j-1}\leq \rho(a,y)\le
s_j}(f(y))^p W(a,y)\,d\mu(y)\right)^{q/p}$$
$$\le C\, \sum_{j=-\infty}^m \left(\int_{s_{j-1}\leq \rho(a,y)\le
s_j}(f(y))^pW(a,y)\,d\mu(y)\right)^{q/p}$$
$$\le C\,\left(\int_{\rho(a,y)\le
r}(f(y))^pW(a,y)\,d\mu(y)\right)^{q/p}.$$
This completes the proof of the theorem. $\;\;\;\; \Box$
For our purposes we also need the following lemma (see [@KM2] for the case of ${{\mathbf R}^n}$).
[**Lemma D.**]{}
*Suppose that $(X, \rho, \mu)$ be an $SHT$. Let $0<\lambda<1\leq p<\infty$. Then there exists a positive constant $C$ such that for all balls $B_0$,*
$${\|\chi_{B_0}\|}_{L^{p,\lambda}(X,\mu)}\le C {\mu(B_0)}^{(1-\lambda)/p}.$$
[*Proof.*]{} Let $B_0:=B(x_0,r_0)$ and $B:=B(a,r).$ We have
$${\|\chi_{B_0}\|}_{L^{p,\lambda}(X, \mu)} = \sup\limits_{B}{\left(
\frac{\mu(B_0\cap B)}{{\mu(B)}^{\lambda}}\right) }^{1/p}.$$
Suppose that $B_0\cap B\neq \emptyset$. Let us assume that $r\leq r_0.$ Then (see [@StTo], Lemma 1, or [@GGKK], p.9) $ B\subset B(x_0, br_0)$, where $b= a_1(1+a_0)$. By the doubling condition it follows that $$\frac{\mu(B\cap B_0)}{\mu(B)^{\lambda}}
\leq \frac{\mu(B)}{\mu(B)^\lambda} =\mu(B)^{1-\lambda} \leq
\mu(B(x_0, br_0))^{1-\lambda}$$ $$\leq C\,\mu(B_0)^{1-\lambda}.$$
Let now $r_0<r$. Then $\mu B_0\leq c \mu B$, where the constant $c$ depends only on $a_1$ and $a_0$. Then
$$\frac{\mu(B \cap B_0)}{\mu(B)^{\lambda}}
\leq c \frac{\mu(B_0)}{\mu(B_0)^\lambda} = c \mu
(B_0)^{1-\lambda}.$$
$\Box$
+0.1cm
The next lemma may be well-known but we prove it for the completeness. +0.1cm
[**Lemma E.**]{}
*Let $(X,\rho, \mu)$ be a non-homogeneous space with the growth condition. Suppose that $\sigma>-1$. Then there exists a positive constant $c$ such that for all $a\in X$ and $r>0$, the inequality*
$$I(a, r, \sigma):= \int_{B(a,r)} \rho(a,x)^{\sigma} d\mu \leq c r^{\sigma+1}$$ holds.
+0.1cm
[**Proof.**]{} Let $\sigma \geq 0$. Then the result is obvious because of the growth condition for $\mu$. Further, assume that $-1<\sigma<0$. We have
$$I(a, r, \sigma) = \int_0^{\infty} \mu \{ x\in B(a, r): \rho(a,x)^{\sigma} >\lambda \} d\lambda$$
$$= \int_{0}^{\infty} \mu(B(a,r)\cap B(a, \lambda^{1/\sigma})) d\lambda = \int_0^{r^{\sigma}} + \int_{r^{\sigma}}^{\infty} :=
I^{(1)}(a, r, \sigma) + I^{(2)}(a, r, \sigma).$$
By the growth condition for $\mu$ we have
$$I^{(1)}(a, r, \sigma)\leq r^{\sigma} \mu(B(a,r))\leq c r^{\sigma+1},$$ while for $I^{(2)}(a, r, \sigma)$ we find that $$I^{(2)}(a, r, \sigma) \leq c \int_{r^{\sigma}}^{\infty} \lambda^{1/\sigma} d\lambda = c r^{\sigma+1}$$ because $1/\sigma< -1$. $\Box$
+0.1cm
The Following statement is the trace inequality for the operator $K_{\alpha}$ (see [@Adm] for the case of Euclidean spaces and, e.g., [@GGKK] or [@EKM], Th. 6.2.1, for an SHT). +0.1cm
[**Theorem F.**]{} [*Let $(X,\rho, \mu)$ be an SHT. Suppose that $1<p<q<\infty$ and $0<\alpha<1/p$. Assume that $\nu$ is another measure on $X$. Then $K_{\alpha}$ is bounded from $L^p(X,\mu)$ to $L^q(X,\nu)$ if and only if $$\nu B \leq c (\mu B)^{q(1/p-\alpha)}$$ for all balls $B$ in $X$.*]{}
+0.1cm
+0.1cm
Main results
============
In this section we formulate the main results of the paper. We begin with the case of an SHT.
[**Theorem 3.1.**]{} [*Let $(X, \rho, \mu)$ be an SHT and let $1<p<q<\infty$. Suppose that $0<\alpha<1/p$, $0<\lambda_1<1-\alpha p$ and $\lambda_2/q= \lambda_1 /p$. Then $K_{\alpha}$ is bounded from $L^{p,\lambda_1}(X,\mu)$ to $L^{q,\lambda_2}(X,\nu,\mu)$ if and only if there is a positive constant $c$ such that $$\nu(B)\leq c \mu(B)^{q(1/p-\alpha)},\eqno{(3)}$$ for all balls $B$.*]{}
+0.1cm
The next statement is a consequence of Theorem 3.1. +0.1cm
[**Theorem 3.2.**]{} [*Let $(X, \rho, \mu)$ be an SHT and let $1<p<q<\infty$. Suppose that $0<\alpha <1/p$, $0<\lambda_1<1-\alpha p$ and $\lambda_2/q = \lambda_1/p$. Then for the boundedness of $K_\alpha$ from $L^{p,\lambda_1}(X, \mu)$ to $L^{q,\lambda_2}(X, \mu)$ it is necessary and sufficient that $q= p/(1-\alpha p)$.* ]{}
For non-homogeneous spaces we have the following statements: +0.1cm
[**Theorem 3.3.**]{} [*Let $(X,\rho,\mu)$ be a non-homogeneous space with the growth condition. Suppose that $1< p \le q<\infty$, $1/p-1/q \le \alpha<1$ and $\alpha\neq 1/p$. Suppose also that $p
\alpha-1<\beta<p-1$, $0<\lambda_1<\beta-\alpha p +1$ and $\lambda_1q=\lambda_2p$. Then ${I_\alpha}$ is bounded from $M_\beta^{p,\lambda_1}(X,\mu)$ to $M_\gamma^{q,\lambda_2}(X,\mu)$, where $\gamma=q(1/p+\beta/p-\alpha)-1.$*]{}
+0.1cm
[**Theorem 3.4.**]{} [*Suppose that $(X,\rho,\mu)$ is a quasimetric measure space and $\mu$ satisfies condition $(2)$. Let $1<p<q<\infty$. Assume that $0<\alpha<1$, $0<\lambda_1< p/q$ and $s\lambda_1/p=\lambda_2/q$. Then the operator $I_{\alpha}$ is bounded from $M^{p,\lambda_1 s}(X,\mu)$ to $M^{q,\lambda_2}(X,
\mu)$.*]{}
+0.1cm
+0.2cm
Proof of the Main Results
=========================
In this section we give the proofs of the main results.
[*Proof of Theorem*]{} 3.1. [*Necessity.*]{} Suppose $K_\alpha$ is bounded from $L^{p,\lambda_1}(\mu)$ to $L^{q,\lambda_2}(X, \nu,
\mu).$ Fix $B_0:= B(x_0,r_0)$. For $x, y\in B_0$, we have that $$B(x,\rho(x,y))\subseteq B(x, a_1(a_0+1)r_0)\subseteq B(x_0, a_1(1+a_1(a_0+1)) r_0).$$ Hence using the doubling condition for $\mu$, it is easy to see that
$$\mu(B_0)^{\alpha} \leq c K_\alpha \chi_{B_0}(x),\quad x\in B_0.$$
Consequently, using the condition $\lambda_2/q= \lambda_1/p$, the boundedness of $K_{\alpha}$ from $L^{p,\lambda}(X,\mu)$ to $L^{q,
\lambda_2}(X,\nu, \mu )$ and Lemma D we find that
$$\mu(B_0)^{\alpha-\lambda_1 /p} \nu(B_0)^{1/q} \leq c \|K_{\alpha}\chi_{B_0}\|_{L^{q,\lambda_2}(X,\nu,\mu)}$$
$$\leq c \|\chi_{B_0}\|_{L^{p,\lambda_1}(X,\mu)} \leq c \mu(B_0)^{(1-\lambda_1)/p}.$$
Since $c$ does not depend on $B_0$ we have condition $(3)$.
Let $B:=B(a,r)$, $ \tilde{B}:=B(a,2a_1r)$ and $f\geq 0$. Write $f \in L^{p,\lambda_1}(\mu)$ as $f=f_1+f_2:=f\chi_{\tilde
B}+f\chi_{{\tilde B}^{\rm C}},$ where $\chi_B$ is a characteristic function of $B$. Then we have $$S:= \int_{B} (K_{\alpha}f(x))^q d\nu(x) \leq c \bigg(
\int_{B} (K_{\alpha}f_1(x))^q d\nu(x) + \int_{B}
(K_{\alpha}f_2(x))^q d\nu(x)\bigg) := c(S_1 + S_2).$$
Applying Theorem F and the fact $\mu \in (DC)$ we find that $$S_1 \leq \int_X (K_{\alpha} f_1)^{q}(x) d\nu(x) \leq c
\bigg(\int_{B(a, 2a_1 r)} (f(x))^p d\mu(x)\bigg)^{q/p}.$$
Now observe that if $\rho(a,x)<r$ and $\rho(a,y)>2a_1r$, then $\rho(a,y)>2a_1\rho(a,x)$. Consequently, using the facts $\mu\in
(RDC)$ (see (1)), $0<\lambda_1<1-\alpha p$ and condition (3) we have
$$S_2 \leq c \int_{B(a,r)} \bigg( \int_{\rho(a,y)>r}
\frac{f(y)}{\mu B(a, \rho(a,y)))^{1-\alpha}} d\mu(y) \bigg)^{q}
d\nu(x)$$
$$=\nu (B) \bigg[ \sum_{k=0}^{\infty} \int_{B(a, \eta_1^{k+1} r)\setminus B(a,\eta_1^k r)}
\frac{f(y)}{\mu B(a, \rho(a,y)))^{1-\alpha}}d\mu(y) \bigg]^{q}$$
$$\leq c \nu (B) \bigg[ \sum_{k=0}^{\infty} \bigg(\int_{B(a, \eta_1^{k+1}r )} (f(y))^p
d\mu(y)\bigg)^{1/p}$$
$$\times \bigg( \int_{B(a, \eta_1^{k+1}
r)\setminus B(a,\eta_1^kr)} \mu B(a, \rho(a,y))^{(\alpha-1)p'}
d\mu(y) \bigg)^{1/p'} \bigg]^{q}$$
$$\leq c \| f \|^q_{L^{p,\lambda_1}(X,\mu)} \nu (B) \bigg( \sum_{k=0}^{\infty}
\mu B(a, \eta_1^{k+1} r))^{\lambda_1/p+ \alpha-1+1/p'}\bigg)^q$$
$$\leq c \| f \|^q_{L^{p,\lambda_1}(X,\mu)} \nu (B) \mu (B)^{(\lambda_1/p+\alpha-1/p)q} \bigg(\sum_{k=0}^{\infty} \eta_2^{k
(\lambda_1/p+\alpha-1/p)}\bigg)^{q}$$
$$\leq c \| f\|^q_{L^{p,\lambda_1}(X,\mu)} \mu(B)^{q \lambda_1/p} = c \|f\|^{q}_{L^{p,\lambda_1}(X,\mu)}\mu(B)^{\lambda_2},$$ where the positive constant $c$ does not depend on $B$. Now the result follows immediately. $\Box$
+0.1cm
[*Proof of Theorem*]{} 3.2. [*Sufficency.*]{} Assuming $\alpha= 1/p-1/q$ and $\mu=\nu$ in Theorem 3.1 we have that $K_{\alpha}$ is bounded from $L^{p,\lambda_1}(X,\mu)$ to $L^{q, \lambda_2}(X, \mu)$.
[*Necessity.*]{} Suppose that $K_{\alpha}$ is bounded from $L^{p, \lambda_1}(X, \mu)$ to $L^{q, \lambda_2} (X, \mu)$. Then by Theorem 3.1 we have $$\mu(B)^{1/q-1/p+\alpha} \leq c.$$
The conditions $\mu(X)= \infty$ and $\mu\{ x\}= 0$ for all $x\in X$ implies that $\alpha= 1/p-1/q$. $\Box$ +0.1cm
[*Proof of Theorem*]{} 3.3. Let $f\geq 0$. For $x,\,a\in X,$ let us introduce the following notation:
$$E_1(x)\, :=\,
\biggl\{y:\frac{\rho(a,y)}{\rho(a,x)}<\frac{1}{2a_1}\biggr\};\;\;
E_2(x)\, :=\, \biggl\{y:\frac{1}{2a_1}\leq
\frac{\rho(a,y)}{\rho(a,x)}\leq 2a_1\biggr\};$$ $$E_3(x)\, :=\,
\biggl\{y:2a_1<\frac{\rho(a,y)}{\rho(a,x)}\biggr\}.$$
For $i=1, 2, 3$, $r>0$ and $a\in X,$ we denote $$S_i:=\int_{\rho(a,x)<r}\rho(a,x)^\gamma\left(\int_{E_i(x)}
f(y)\rho(x,y)^{\alpha-1}\,d\mu(y)\right)^q d\mu(x).$$
If $y\in E_1(x),$ then $\rho(a,x)<2a_1 a_0 \rho(x,y).$ Hence, it is easy to see that $$S_1\le C\,\int_B\rho(a,x)^{\gamma+
q(\alpha-1)}\left(\int_{\rho(a,y)<\rho(a,x)}f(y)\,d\mu(y)\right)^qd\mu(x).$$
Taking into account the condition $\gamma<(1-\alpha)q-1$ we have
$$\int_{ \rho(a,x)>t } \rho(a,x)^{\gamma+ q(\alpha-1)}
d\mu(x) = \sum_{n=0}^{\infty} \int_{B(a, 2^{k+1}t)\setminus
B(a,2^{k}t)} (\rho(a,x))^{\gamma+(\alpha-1)q}d\mu(x)$$
$$\leq c \sum_{n=0}^{\infty} (2^k t)^{\gamma +q(\alpha-1)+1} = c t^{\gamma +q(\alpha-1)+1},$$ while the condition $\beta<p-1$ implies $$\int_{\rho(a,x)<t} \rho(a,x)^{\beta( 1-p')+ 1} d\mu(x) \leq c
t^{\beta(1-p')+1}.$$ Hence
$$\sup_{a\in X, t>0} \bigg(\int_{\rho(a,x)>t} \rho(a,x)^{\gamma+ q(\alpha-1)}
d\mu(x)\bigg)^{1/q} \bigg(\int_{B(a,t)} \rho(a,y)^{\beta(1-p')}
d\mu(y)\bigg)^{1/p'} <\infty.$$
Now using Theorem C we have
$$S_1 \le c \, \bigg( \int_B \rho(a,x)^{\beta} (f(y)) \, d\mu(y) \bigg)^{q/p}
\leq c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)} r^{ \lambda_1
q/p}= c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{\lambda_2}.$$
Further, observe that if $\rho(a,y)>2 a_1 \rho(a,x)$, then $ \rho
(a,y)\leq a_1 \rho(a,x) +a_1 \rho(a,y) \leq \rho(a,y)/2 +a_1
\rho(x,y). $ Hence $ \rho(a,y)/(2a_1)\leq \rho(x,y)$. Consequently, using the growth condition for $\mu$ , the fact $\lambda_1 <\beta-\alpha p+1$ and Lemma E we find that
$$S_3 \leq c \int_{B(a,r)} \rho(a,x)^{\gamma} \bigg( \int_{\rho(a,y)> \rho(a,x)}
\frac{f(y)}{\rho(a,y)^{1-\alpha}} d\mu(y) \bigg)^{q} d\mu(x)$$
$$\leq c \int_{B(a,r)} \rho(a,x)^{\gamma} \bigg( \sum_{k=0}^{\infty} \int_{B (a, 2^{k+1} \rho(a,x))\setminus
B(a, 2^k \rho(a,x))} \frac{f(y)}{\rho(a,y)^{1-\alpha}} d\mu(y)
\bigg)^{q} d\mu(x)$$
$$\leq c \int_{B(a,r)} \rho(a,x)^{\gamma} \bigg[ \sum_{k=0}^{\infty} \bigg(
\int_{B(a, 2^{k+1}\rho(a,x))} f^p(y) \rho(a,y)^{\beta}
d\mu(y)\bigg)^{1/p}$$
$$\times \bigg(\int_{B (a, 2^{k+1}
\rho(a,x))\setminus B(a, 2^k \rho(a,x))} \rho(a,y)^{\beta(1-p')+
(\alpha-1)p'} d\mu(y) \bigg)^{1/p'}\bigg]^q d\mu(x)$$
$$\leq c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
\int_{B(a,r)} \rho(a,x)^{\gamma}$$
$$\times \bigg(
\sum_{k=0}^{\infty} (2^k \rho(a,x))^{\lambda_1/p+\alpha-1-\beta/p}
(\mu B(a, 2^{k+1} \rho(a,x)))^{1/p'}\bigg)^{q} d\mu (x)$$
$$\leq c \| f \|^q_{M^{p,\lambda_1}_{\beta}(X,\mu)}
\int_{B(a,r)} \rho(a,x)^{\gamma} \bigg( \sum_{k=0}^{\infty} (2^k
\rho(a,x))^{\lambda_1/p+\alpha-1/p-\beta/p} \bigg)^{q} d\mu (x)$$
$$\leq c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
\int_{B(a,r)} \rho(a,x)^{(\lambda_1/p+ \alpha-1/p-\beta/p)q
+\gamma} d\mu(x)$$
$$= c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
\int_{B(a,r)} \rho(a,x)^{\lambda_1 q/p -1} d\mu(x) \leq c \| f
\|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)} r^{ \lambda_1 q/p}$$ $$= c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{ \lambda_2}.$$
So, we conclude that $$S_3 \leq c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{ \lambda_2}.$$
To estimate $S_2$ we consider two cases. First assume that $\alpha<1/p$. Let
$$E_{k,r}:= \{ x: 2^k r \leq \rho(a,x) < 2^{k+1} r \};$$
$$F_{k,r}:= \{ x: 2^{k-1} r/ a_1 \leq \rho(a,x)< a_1 2^{k+2} r \}.$$
Assume that $p^{*}=p/(1-\alpha p)$. By Hölder’s inequality, Corollary B and the assumption $\gamma= q(1/p+\beta/p -\alpha)-1$ we have
$$S_2 = \sum_{k=-\infty}^{-1} \int_{E_{k,r}}\rho(a,x)^{\gamma}\bigg( \int_{E_2(x)} f(y)
\rho(x,y)^{\alpha-1} d\mu(y) \bigg)^q d\mu(x)$$
$$\leq \sum_{k=-\infty}^{-1}
\bigg(\int_{E_{k,r}}\rho(a,x)^{\gamma}\bigg( \int_{E_2(x)} f(y)
\rho(x,y)^{\alpha-1} d\mu(y) \bigg)^{p^*} d\mu(x)\bigg)^{q/p'}$$ $$\times \bigg( \int_{E_{k,r}} \rho(a,x)^{\gamma p^*/(p^* -q)}
d\mu(x)\bigg)^{(p^*-q)/p^*}$$
$$\leq c \sum_{k=-\infty}^{-1} 2^{k(\gamma+(p^*-q)/p^*)}
\bigg(\int_{X} I_{\alpha} ( f\chi_{F_{k,r}})(x))^{p^*} d\mu(x)
\bigg)^{q/p^*}$$
$$\leq c \sum_{k=-\infty}^{-1}
2^{k(\gamma+(p^*-q)/p^*)} \bigg( \int_{F_{k,r}} (f(x))^{p} d\mu(x)
\bigg)^{q/p} \leq c \bigg(\int_{B(a, 2a_1 r)} \rho(a,x)^{\beta} (f(x))^{p}
d\mu(x) \bigg)^{q/p}$$
$$\leq c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{ \lambda_1 q/p} = c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{ \lambda_2 }.$$
Let us now consider the case $1/p<\alpha<1$.
First notice that (see [@KM1])
$$\int_{E_2(x)}(\rho(x,y)^{(\alpha-1)p'} d\mu(y) \leq c
\rho(a,x)^{1+(\alpha-1)p'},$$ where the positive constant $c$ does not depend on $a$ and $x$.
This estimate and Hölder’s inequality yield
$$S_2 \leq c \sum_{k=-\infty}^{-1}
\bigg(\int_{E_{k,r}} \rho(a,x)^{\gamma+[(\alpha-1)p'+1)]q/p'}
\bigg( \int_{E_2(x)} (f(y))^p d\mu(y) \bigg)^{q/p}
d\mu(x)\bigg)^{q/p'}$$
$$\leq c \sum_{k=-\infty}^{-1}
\bigg(
\int_{E_{k,r}}\rho(a,x)^{\gamma+[(\alpha-1)p'+1)]q/p'}d\mu(x)\bigg)
\bigg(\int_{F_{k,r}} (f(y))^{p} d\mu(y) \bigg)^{q/p}$$
$$\leq c \sum_{k=-\infty}^{-1} (2^k r)^{\gamma+[(\alpha-1)p'+1)]q/p'+1}
\bigg(\int_{F_{k,r}} (f(y))^{p} d\mu(y) \bigg)^{q/p}$$
$$= c \sum_{k=-\infty}^{-1} 2^{k \beta q/ p}
\bigg(\int_{F_{k,r}} (f(y))^{p} d\mu(y) \bigg)^{q/p} \leq c \bigg(
\int_{B(a, 2a_1 r)} (f(y))^{p} \rho(a,y)^{\beta} d\mu(y) \bigg)^{q/p}$$
$$\leq c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{ \lambda_1 q/p} = c \| f \|^q_{ M^{p,\lambda_1}_{\beta}(X,\mu)}
r^{ \lambda_2 }.$$ Now the result follows immediately. $\Box$
+0.1cm
[*Proof of Theorem*]{} 3.4. Let $f\geq 0$. Suppose that $a\in X$ and $r>0$. Suppose also that $f_1= f\chi_{B(a, 2a_1 r)}$ and $f_2=f-f_1$. Then $I_{\alpha}f = I_{\alpha}f_1 + I_{\alpha}f_2$. Consequently, $$\int_{B(a,r)}(I_{\alpha}f(x))^q d\mu(x) \leq 2^{q-1} \bigg( \int_{B(a,r)}(I_{\alpha}f_1(x))^q d\mu(x)$$ $$+ \int_{B(a,r)}(I_{\alpha}f_2(x))^q d\mu(x) \bigg) := 2^{q-1}( S^{(1)}_{a,r}+ S^{(2)}_{a,r}).$$ Due to Theorem A and the condition $s\lambda_1/p= \lambda_2/q$ we have
$$S^{(1)}_{a,r} \leq c \bigg(\int_{B(a, 2a_1 r)} (f(x))^p d\mu(x)\bigg)^{q/p}$$ $$= c \bigg( \frac{1}{(2a_1 r)^{\lambda_1 s}} \int_{B(a, 2a_1 r)} (f(x))^p dx \bigg)^{q/p} r^{\lambda_1 s q/p} \leq c \| f\|^q_{M^{p,\lambda_1s}(X,\mu)} r^{\lambda_2}.$$
Now observe that if $x\in B(a, r)$ and $y\in X \setminus B(a, 2a_1
r)$, then $ \frac{\rho(a, y)}{2a_1} \leq \rho(x,y). $ Hence Hölder’s inequality, condition (2) and the condition $0<\lambda_1<p/q$ yield
$$I_{\alpha}f_2(x) = \int_{X\setminus B(a,2a_1r)} f(y)/\rho(x,y)^{1-\alpha} d\mu(y)$$
$$= \sum_{k=0}^{\infty} \bigg( \int_{B(a, 2^{k+2}a_1 r)\setminus B(a, 2^{k+1}a_1 r)} (f(y))^p
d\mu(y)\bigg)^{1/p}$$
$$\times \bigg( \int_{B(a, 2^{k+2}a_1 r)\setminus B(a, 2^{k+1}a_1
r)} \rho(a,y)^{(\alpha-1)p'}d\mu(y)\bigg)^{1/p'}$$
$$\leq c \sum_{k=0}^{\infty} \bigg( \frac{1}{(2^{k+1}a_1 r)^{\lambda_1 s}} \int_{B(a, 2^{k+1}a_1 r)} (f(y))^p d\mu(y) \bigg)^{1/p}(2^{k}a_1r)^{\lambda_1 s/p +\alpha -1 + s/p'}$$
$$\leq c \| f\|_{M^{p,\lambda_1 s}(X,\mu)}
r^{\lambda_1s /p+\alpha-1+s/p'}.$$
Consequently, by the assumptions $s\lambda_1/p =\lambda_2/q$ and $s=\frac{pq(1-\alpha)}{pq+p-q}$ we conclude that
$$S^{(2)}_{a,r}\leq c \| f\|^q_{M^{p,\lambda_1 s}(X,\mu)}
r^{(\lambda_1 s/p+\alpha-1+s/p')q +s}= c \|
f\|^q_{M^{p,\lambda_1 s}(X,\mu)} r^{\lambda_2}.$$
Summarazing the estimates derived above we finally have the desired result. $\Box$
+0.5cm
+0.5cm
[**Acknowledgement**]{}
+0.4cm
The second and third authors were partially supported by the INTAS Grant No. 05-1000008-8157 and the Georgian National Science Foundation Grant No. GNSF/ST07/3-169. +1cm
[20]{} +0.2cm
D. R. Adams, A trace inequality for generalized potentials, [*Studia Math.*]{} [**48**]{}(1973), 99-105.
D. R. Adams, A note on Reasz potentials. [*Duke Math. J.*]{} [**42**]{} (1975), No.4, 765-778.
V. Burenkov and H. V. Guliyev, Necessary and sufficient conditions for boundedness of the maximal operator in local Morrey-type spaces. [*Studia Math.*]{} [**163**]{} (2004), no. 2, 157–176.
I. V. Burenkov, H. V. Guliyev and V. S. Guliyev, Necessary and sufficient conditions for boundedness of the Riesz potential in the local Morrey-type spaces. (Russian) [*Doklady Akademii Nauk,*]{} [**412**]{}(2007), No.5, 585–589. Inglish Transl. [*Doklady Mathematics*]{} [**75**]{}(2007), No. 1, 103-–107.
D. Edmunds, V. Kokilashvili, A. Meskhi, [*Bounded and Compact Integral Operators*]{}, Mathematics and Its Applications, [**543**]{}, Kluwer Academics Publishers, Dordrecht, The Netherlands, 2002.
Eridani, V. Kokilashvili and A. Meskhi, Morrey spaces and fractional integral operators, [*Preprint No*]{}.65, [*School of Mathematical Sciences, GC University, Lahore,*]{} 2007.
J. García-Cuerva, A. E. Gatto, Boundedness properties of fractional integral operators associated to non-doubling measures, [*Studia Math.*]{} [**162**]{} (2004)(3), 245-261.
J. García-Cuerva, J. M. Martell, Two-weight norm inequalities for maximal operators and fractional integrals on non-homogeneous spaces, [*Indiana Univ. Math. J.*]{} [**50**]{} (2001)(3), 1241-1280.
I. Genebashvili, A. Gogatishvili and V. Kokilashvili, Solution of Two-weight problems for integral transforms with positive kernels. [*Georgian Math. J.* ]{} [**3**]{}(1996), No.1, 319-342.
I. Genebashvili, A. Gogatishvili, V. Kokilashvili and M. Krbec, Weight theory for integral transforms on Spaces of Homogeneous Type. [*pitman Monographs and Surveys in Pure and Applied Mathematics*]{} [**92**]{}. [*Harlow, Longman,* ]{} 1998.
G. H. Hardy, J. E. Littlewood, Some properties of fractional integrals I, [*Math. Zeit.*]{} [**27**]{} (1928), 565-606.
V. Kokilashvili, Weighted estimates for classical integral operators. [*In: Nonlinear Analysis, Function Spaces and Applications*]{} IV ([*Roudnice nad Labem*]{} 1990). [*Teubner-Texte Math.*]{} [**119**]{} [*Leipzig, Teubner,*]{} 1990, 86-103.
V. Kokilashvili, A. Meskhi, Fractional integrals on measure Spaces, [*Frac. Calc. Appl. Anal.*]{} [**4**]{} (2001), No. 4, 1-24.
V. Kokilashvili, A. Meskhi, On some Weighted Inequalities for Fractional Integrals on Non-homogeneous Spaces, [*Zeitschrift für Analysis und ihre Anwendungen*]{} [**24**]{} (2005), No. 4, 871-885.
V. Kokilashvili, A. Meskhi, S. Samko, On the inversion and characterization of the Riesz potentials in the weighted Lebesgue spaces, [*Memoirs Diff. Equations Math. Phys.*]{} [**29**]{} (2003), 31-45.
Y. Komori, T. Mizuhara, Notes on commutators and Morrey spaces, [*Hokkaido Math. J.*]{} [**32**]{} (2003), 345-353.
B. Muckenhoupt and R. L. Wheeden, Weighted norm inequalities for fractional integrals. [*Trans. Amer. Math. Soc.*]{} [**192**]{}(1974), 261-276.
J. Peetre, On the theory of ${\mathcal{L}}^{p,\lambda}$ spaces. [*J. Functional Analysis*]{} [**4**]{}(1969), 71–87.
S. Samko, A. A. Kilbas, O. I. Marichev, [*Fractional Integrals and Derivatives. Theory and Applications*]{}, Gordon and Breach, London, 1993.
Y. Sawano, H. Tanaka, Morrey spaces for nondoubling measures, [*Acta Math. Sinica*]{} [**21**]{}(2005), No. 6, 1535-1544.
E. T. Sawyer and R. L. Wheeden, Weighted inequalities for fractional integrals on Euclidean and homogeneous type spaces, [*Amer. J. Math.*]{} [**114**]{}(1992), 813-875.
E. M. Stein, G. Weiss, Fractional integrals on [*n*]{}-dimensional Euclidean spaces, [*J. Math. Mech.*]{} [**7**]{} (1958), (4), 503-514.
J. O. Strömberg and A. Torchinsky, Weighted Hardy spaces, [*Lecture Notes in Math.*]{} [**1381**]{}, [*Springer Verlag, Berlin*]{}, 1989.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Several numerical approximation strategies for the expectation-propagation algorithm are studied in the context of large-scale learning: the Laplace method, a faster variant of it, Gaussian quadrature, and a deterministic version of variational sampling ([*i.e.*]{}, combining quadrature with variational approximation). Experiments in training linear binary classifiers show that the expectation-propagation algorithm converges best using variational sampling, while it also converges well using Laplace-style methods with smooth factors but tends to be unstable with non-differentiable ones. Gaussian quadrature yields unstable behavior or convergence to a sub-optimal solution in most experiments.'
author:
- 'Alexis Roche[^1]'
title: |
On numerical approximation schemes\
for expectation propagation
---
Introduction {#sec:introduction}
============
Many machine learning tasks involve minimizing an objective that has the form of a sum: $$\label{eq:additive_objective}
u(\btheta) = \sum_{k=1}^n u_k(\btheta),$$ where $\btheta\in\mathbb{R}^d$ is an unknown vector of interest, and $u_k(\btheta)$, for $k=1,\ldots,n$, are cost functions most often associated with distinct data points, or subsets of points called “mini-batches”. Problems of type (\[eq:additive\_objective\]) include maximum likelihood parameter estimation under independent measurements, empirical risk minimization [@Vapnik-00], or composite likelihood maximization [@Varin-11].
While generic optimization algorithms may scale poorly with data size, it is possible to devise fast optimization procedures by taking advantage of the additive nature of (\[eq:additive\_objective\]). For instance, stochastic gradient descent (SGD) methods [@Bottou-10] have a natural implementation using a single cost gradient $\nabla u_k(\btheta)$ at each iteration rather than the total gradient $\nabla u(\btheta)$. This makes the computational cost of an iteration proportional to the associated mini-batch size rather than to the full data size, hence providing a massive gain in efficiency over deterministic gradient methods in large-scale learning situations.
SGD, however, assumes continuously differentiable cost functions. There are problems of interest where this assumption does not hold, such as support vector machine training [@Ertekin-11] or intensity-based image registration [@ijist:00]. In such context, a potential alternative to SGD is the expectation-propagation (EP) algorithm [@Minka-01; @Minka-05], which is also scalable to big data while relying on larger-scale function approximations. EP is fundamentally an algorithm to approximate a factorial probability distribution, providing, in particular, an approximation to the mode. Hence, EP can be used to approximately minimize an objective of the form (\[eq:additive\_objective\]) by converting it into a Boltzmann distribution: $$\label{eq:boltzmann_dist}
p(\btheta)
\equiv \frac{1}{Z(\beta)} e^{-\beta u(\btheta)}
\propto \prod_{k=1}^n f_k(\btheta),
\qquad {\rm with} \quad
f_k(\btheta) \equiv e^{-\beta u_k(\btheta)}
,$$ for some sufficiently large constant $\beta>0$. As in simulated annealing [@Kirkpatrick-84], $\beta$ could be progressively increased for exact minimization, however it will be assumed fixed in this paper.
EP exploits the decomposition of $p(\btheta)$ as a product of factors by alternatively fitting each factor $f_k(\btheta)$ with an unnormalized Gaussian $g_k(\btheta)$ via local moment matching. To approximate factor $f_k$, EP first forms the “cavity distribution”, $$\label{eq:cavity}
c_k(\btheta) = \prod_{\ell\not=k} g_\ell(\btheta),$$ which is the current approximation to the product of all factors except the one under consideration. The factor approximation $g_k$ is then computed so that $c_k g_k$ has the same moments of order 0, 1 and 2 as $c_k f_k$. We will restrict ourselves to fully factorized Gaussian approximations, in which case cross second-order moments are ignored [@Minka-05], leading to an EP variant referred to as fully factorial EP (FF-EP) in the sequel.
Since the moments essentially depend on the factor shape in the neighborhood of the cavity, the successive factor approximations are performed at adaptive, non-infinitesimal scales induced by the changing cavity distributions. Importantly, and similarly to SGD, the complexity of an EP iteration is proportional to the mini-batch size since a single factor is visited at a time. EP variants such as Averaged EP [@Dehaene-16] and Stochastic EP [@Li-15] can further save memory load by constraining all factor approximations to be identical, but rest upon the same moment matching scheme.
In comparison with SGD, the problem of computing the gradient of $u_k$ is replaced with that of computing the vector-valued integral of $c_k f_k$, which is generally well defined but may lack a closed-form expression. Unless this can be worked around by applying a functional transformation to the factor, as in the Power EP algorithm [@Minka-04b; @Minka-05], numerical approximations are needed. Both the Laplace method [@Smola-03; @Yu-06] and Gaussian quadrature [@Zoeter-05; @Yu-06] are common choices. I review these methods and introduce another method based on variational sampling [@ijasp:13], which combines Gaussian quadrature with variational approximation, and is found empirically to yield better convergence of FF-EP with non-continuously differentiable factors.
Numerical approximation schemes for FF-EP {#sec:approx_schemes}
=========================================
The basic EP building block consists of fitting an unnormalized Gaussian distribution to a single factor. Let us assume that a given factor $f_k(\btheta)$ is selected, and drop the index $k$ for clarity. The associated cavity distribution $c(\btheta)$ is computed from the other factor current approximations according to (\[eq:cavity\]), and is a product of one-dimensional Gaussians by construction; let $\bmu$ and $\S = {\rm diag}(\sigma^2_1,\ldots,\sigma^2_d)$ denote its mean and variance matrix, respectively.
In FF-EP [@Minka-05], the $(2d+1)$ moments of $c(\btheta)f(\btheta)$ up to the order 2 need to be computed, $$\label{eq:moments}
m_i = \int c(\btheta) f(\btheta) \phi_i(\btheta)d\btheta,
\qquad i=0,\ldots, 2d,$$ for the monomials of degree $0, 1, 2$: $$\label{eq:monomials}
\left\{
\begin{array}{ll}
\displaystyle \phi_0(\btheta) = 1, & \\
\displaystyle \phi_i(\btheta) = \theta_i, & i=1,\ldots,d \\
\displaystyle \phi_i(\btheta) = \theta^2_{i-d}, & i=d+1,\ldots,2d.
\end{array}
\right
.$$
The factor approximation $g(\btheta)$ is then readily given by $g(\btheta)=q(\btheta)/c(\btheta)$, where $q(\btheta)$ is the unique unnormalized factorial Gaussian distribution with same moments as (\[eq:moments\]). The problem we address here is to approximate $g(\btheta)$ when the moments are intractable.
Laplace-style approximations {#sec:laplace}
----------------------------
A strategy used in [@Smola-03; @Yu-06] is to approximate $g(\btheta)$ using the Laplace method applied to $c(\btheta)f(\btheta)$, [*i.e.*]{}, first find the value $\btheta_\star$ that maximizes $c(\btheta)f(\btheta)$, then approximate $\log f(\btheta)=-\beta u(\btheta)$ by its second-order Taylor expansion at $\btheta_\star$: $$\label{eq:simple_laplace_approx}
\log g(\btheta) = -\beta
\left[
u(\btheta_\star)
+ \nabla u(\btheta_\star)^\top (\btheta-\btheta_\star)
+ \frac{1}{2} (\btheta-\btheta_\star)^\top
{\rm diag}(\nabla \nabla^\top u(\btheta_\star))
(\btheta-\btheta_\star)
\right]
,$$ assuming that $u(\btheta)$ is twice continuously differentiable. For a fully factorized approximation, only the diagonal elements of the Hessian of $u(\btheta)$ need to be computed, hence the complexity of (\[eq:simple\_laplace\_approx\]) is linear in $d$.
However, the maximization step may be time consuming and we shall also consider a simpler variant, hereafter referred to as “quick Laplace”, whereby the Taylor expansion is performed at the cavity center $\bmu$ rather than at the maximizer $\btheta_\star$ of $c(\btheta)f(\btheta)$. Note that this makes the factor approximation independent from the cavity variance $\S$.
Gaussian quadrature {#sec:gauss_quad}
-------------------
As pointed out above, the cost derivatives may be discontinuous, or change abruptly in the cavity neighborhood, in which case they are not informative about the factor at the relevant scale of analysis. In such situations, the Laplace method can lead to poor factor approximations. An alternative is to use Gaussian quadrature to approximate the moments (\[eq:moments\]). For instance, [@Zoeter-05; @Yu-06] use a precision-3 rule: $$\label{eq:quadrature_rule}
\hat{m}_i =
\sum_{j=0}^{2d} w_j f(\btheta_j) \phi_i(\btheta_j)
\approx \int c(\btheta) f(\btheta) \phi_i(\btheta)d\btheta
,$$ which involves a weighted average of $(2d+1)$ points: $$\left\{
\begin{array}{lll}
\displaystyle \btheta_0 = \bmu, & w_0 = 1 - \frac{d}{\gamma^2} & \\
\displaystyle \btheta_j = \bmu + \gamma \sigma_j \e_j, & w_j = \frac{1}{2\gamma^2}, & j=1,\ldots,d \\
\displaystyle \btheta_j = \bmu - \gamma \sigma_{j-d} \e_{j-d}, & w_j = \frac{1}{2\gamma^2}, & j=d+1,\ldots,2d,
\end{array}
\right
.$$ where $\e_j$, for $j=1,\ldots,d$, denote the canonical vectors of $\mathbb{R}^d$, and $\gamma$ is a free parameter. As shown in [@Julier-00], this quadrature rule is exact for polynomials up to degree 3, implying that it can recover the moments up to order 2 of a constant factor (since a factor which is both linear in $\btheta$ and positive-valued on $\mathbb{R}^d$ has to be constant).
The choice $\gamma=\sqrt{3}$ minimizes the error on the fourth-order moment of a standard Gaussian [@Julier-00], and the central weight $w_0$ is then negative for $d>3$. However, [@Yu-06] argue that non-negative weights should be used to guarantee non-negative second-order moments and adopt $\gamma = \sqrt{d}$. This leads to $w_0=0$, hence discarding the central sample point. I instead set $\gamma=\sqrt{d+0.5}$ to induce uniform weights and avoid zero weights. The quadrature rule (\[eq:quadrature\_rule\]) then becomes formally similar to a Monte Carlo integral estimate using random points drawn independently from the cavity distribution $c(\btheta)$, despite that the sampling points $\theta_j$, for $j=0,\ldots, 2d+1$, are chosen in a deterministic manner.
Variational quadrature {#sec:variational_quad}
----------------------
A computational advantage of the precision-3 quadrature rule (\[eq:quadrature\_rule\]) is to be linear in $d$, requiring only $(2d+1)$ evaluations of $f(\btheta)$, but this may come at the price of fairly inaccurate moment estimates. The variational sampling method, which I originally proposed outside the context of EP [@ijasp:13], has the potential to improve over these estimates as it is exact for fully factorized Gaussian factors, just like Laplace-style approximations, yet relying on the same $(2d+1)$ evaluations as the precision-3 rule and not on differential calculus.
The key insight stems from recasting the moment-matching problem into that of minimizing the generalized Kullback-Leibler (KL) divergence [@Minka-01; @Minka-05]: $$\label{eq:kl_div}
D( c f \| c g)
=
\int
c(\btheta)
\left[
f(\btheta) \log \frac{f(\btheta)}{g(\btheta)}
- f(\btheta)+g(\btheta)
\right] d\btheta$$ over the set of unnormalized factorized Gaussian distributions, [*i.e.*]{}, functions of the form $g(\btheta)=\exp[\balpha^\top \bphi(\btheta)]$ parameterized by $\balpha \in \mathbb{R}^{2d+1}$, where $\bphi(\btheta)$ is the vector-valued function with coordinate applications given by the monomials (\[eq:monomials\]). Dropping the terms independent from $g$ in (\[eq:kl\_div\]), we see that minimizing the KL divergence with respect to $\balpha$ is equivalent to minimizing the function: $$\begin{aligned}
L(\balpha) & = &
\int c(\btheta)
\left[
-f(\btheta) \log g(\btheta)
+ g(\btheta)
\right] d\btheta
\nonumber\\
& = &
- \balpha^\top \int c(\btheta) f(\btheta) \bphi(\btheta) d\btheta
+ \int c(\btheta) e^{\balpha^\top \bphi(\btheta)} d\btheta
\label{eq:loc_cross_ent}
.\end{aligned}$$
Therefore, instead of attempting at a direct moment evaluation (\[eq:moments\]), we may consider minimizing an empirical approximation to (\[eq:loc\_cross\_ent\]) based, for instance, on the precision-3 rule: $$\begin{aligned}
\tilde{L}(\balpha) & = &
\sum_{j=0}^{2d}
w_j
\left[
-f(\btheta_j) \log g(\btheta_j)
+ g(\btheta_j)
\right]
\nonumber\\
& = &
- \balpha^\top \sum_{j=0}^{2d}
w_j f(\btheta_j) \bphi(\btheta_j)
+ \sum_{j=0}^{2d} w_j e^{\balpha^\top \bphi(\btheta_j)}
\label{eq:surrogate_kl}
.\end{aligned}$$
It turns out that (\[eq:surrogate\_kl\]) is a strictly convex function of $\balpha$, see [@ijasp:13] for a proof (more generally, strict convexity requires that the number of sample points $\btheta_j$ be at least as large as the number of real-valued moments, a condition satisfied in the present case since both are $2d+1$). Therefore, $\tilde{L}(\balpha)$ has a unique minimizer $\balpha_\star$, defining a factorized Gaussian distribution $g_\star(\btheta)$ which is hopefully close to the actual moment-matching distribution and KL minimizer.
While $\balpha_\star$ lacks a closed-form expression, it can be tracked efficiently using Newton’s method owing to the strict convexity of $\tilde{L}(\balpha)$. The gradient and Hessian expressions are then needed: $$\begin{aligned}
\nabla \tilde{L}(\balpha) & = & \sum_{j=0}^{2d} w_j \left[e^{\balpha^\top \bphi(\btheta_j)}-f(\btheta_j) \right]\bphi(\btheta_j),
\label{eq:surrogate_gradient} \\
\nabla\nabla^\top \tilde{L}(\balpha) & = & \sum_{j=0}^{2d} w_j e^{\balpha^\top \bphi(\btheta_j)} \bphi_j(\btheta_j) \bphi_j(\btheta_j)^\top
. \nonumber \end{aligned}$$
It is important to realize that, in general, the optimal fit $g_\star(\btheta)$ does [*not*]{} match the moments approximated by a direct application of Gaussian quadrature (\[eq:quadrature\_rule\]). In particular, if the factor $f(\btheta)$ is a factorized Gaussian, then $g_\star = f$ and the moments are thus recovered [*exactly*]{} unlike using Gaussian quadrature. This can be seen from the gradient formula (\[eq:surrogate\_gradient\]): if there exists $\balpha_0$ such that $f(\btheta)=\exp[\balpha_0^\top \bphi(\btheta)]$, then clearly $\nabla \tilde{L}(\balpha_0)=0$, meaning that $\balpha_\star=\balpha_0$. In some sense, variational quadrature bridges the gap between Laplace’s method and Gaussian quadrature: it is error-free for factors belonging to the EP approximating exponential family (like Laplace in the Gaussian case), and can handle non-continuously differentiable factors provided that the integrals (\[eq:moments\]) exist (like Gaussian quadrature).
My implementation of Newton’s method uses the Cholesky decomposition to invert Hessian matrices, hence the complexity of variational quadrature is in $O(d^3)$. The computational overhead with respect to Gaussian quadrature therefore increases with $d$, although it is partially compensated for by the fact that variational quadrature directly outputs an approximating factor $g_\star(\btheta)$, while Gaussian quadrature requires dividing the moment-matching distribution by the cavity distribution, which becomes more costly as $d$ increases. Experimental comparisons of timing are provided next.
Experiments {#sec:experiments}
===========
The different variants of FF-EP described in Section \[sec:approx\_schemes\] were evaluated on real data from the UCI repository (<https://archive.ics.uci.edu/ml/datasets.html>) to train linear binary classifiers according to different strategies: logistic regression, hinge regression, and a non-convex approximation to the minimization of the mean classification error.
Implementation
--------------
The FF-EP algorithm and the numerical approximation schemes described in Section \[sec:approx\_schemes\], respectively referred to as Laplace method (LA), quick Laplace method (QLA), Gaussian quadrature (GQ) and variational quadrature (VQ) in this section, were implemented in scientific Python ([www.scipy.org](www.scipy.org)). Newton’s method was used for both the optimization step in LA (see Section \[sec:laplace\]) and for VQ (see Section \[sec:variational\_quad\]) with a tolerance of $10^{-5}$ on relative parameter variations.
Each UCI dataset was formatted as a list of examples of the form $(y_k, \x_k)$, for $k=1,2,\ldots,N$, where $y_k \in \{-1,1\}$ was a binary label and $\x_k\in \mathbb{R}^d$ a feature vector. A constant value (baseline) was appended to all feature vectors. In a pre-processing step, all features were normalized to unit Euclidean norm after subtracting the mean across examples, except for the baseline.
In all experiments, the FF-EP algorithm incorporated a fixed Gaussian factor with zero mean and scalar variance matrix $25 {\bf I}_d$ modeling a prior distribution on the linear classification parameter vector $\btheta\in \mathbb{R}^d$. Factors were defined according to the chosen training cost (see below) by summing up the costs over mini-batches and exponentiating the result as in (\[eq:boltzmann\_dist\]) using a fixed “inverse temperature” parameter $\beta=1$. The corresponding factor approximations in FF-EP were initialized as identically equal to one ([*i.e.*]{}, with infinite variances).
Classification cost functions
-----------------------------
Different cost functions were considered for linear classification training:
#### Logistic loss.
This choice leads to a smooth function of $\btheta$, $$u_k(\btheta) = \log (1 + e^{-y_k \btheta^\top \x_k} ).$$
#### Hinge loss.
This is the loss classically used to train support vector machines [@Ertekin-11], $$u_k(\btheta) = \max(0, 1 -y_k \btheta^\top \x_k ).$$
#### Quasi 0–1 loss.
In some applications, the desired training objective is to minimize the mean classification error using the 0–1 loss. However, because the 0–1 loss is discontinuous and turns out to make the proposed FF-EP variants unstable, it is approximated by a continuous non-convex function: $$u_k(\btheta) = \ell_\epsilon(y_k \btheta^\top \x_k),
\qquad
{\rm with}
\quad
\ell_\epsilon(a) =
\left\{
\begin{array}{lll}
1 - \epsilon a & {\rm if} & a < 0 \\
1 - a/\epsilon & {\rm if} & 0 \leq a < \epsilon \\
0 & {\rm if} & a > \epsilon
\end{array}
\right.
,$$ for some constant $\epsilon>0$. While the case $\epsilon=1$ corresponds to the hinge loss, $\ell_\epsilon(a)$ provides a tighter approximation to the 0–1 loss for smaller $\epsilon$, see Figure \[fig:loss\_funcs\]. In these experiments, I set $\epsilon=0.1$. Note that the quasi 0–1 loss differs from the “ramp loss” [@Ertekin-11].
Note that both the hinge loss and the quasi 0–1 loss derivatives are singular, respectively on the hyperplane $\{\btheta^\top \x_k = y_k\}$ and on the hyperplanes $\{\btheta^\top \x_k = 0\}$ and $\{\btheta^\top \x_k = \epsilon y_k\}$. Derivatives at discontinuities were conventionally defined by extending $\ell'_\epsilon(a)$ at $a=0$ and $a=\epsilon$ by the half sum of its left and right limits.
![Loss functions considered for binary linear classification training.[]{data-label="fig:loss_funcs"}](loss_funcs.pdf){width=".45\textwidth"}
Small datasets
--------------
I considered three UCI classification datasets containing less than $N=1000$ examples (Haberman’s survival, Ionosphere, Wisconsin diagnostic breast cancer). For the different training objectives, the question is whether the numerically approximated FF-EP algorithm converges (knowing that convergence is not even warranted in the case of explicit updates [@Minka-05]) and, if so, how fast and how close to the global cost minimizer.
Mini-batches of size $s=10$ were used in these experiments, which enabled to loop several times over the whole set of mini-batches at reasonable computational cost. Using relatively small mini-batches, FF-EP searches for a consensus between many “weak learners” while giving several opportunities to each learner to revise its knowledge during the process.
The curves in Figure \[fig:small\_datasets\] show the variations of the total cost (computed offline) across mini-batch updates, for the different datasets, cost functions, and FF-EP approximation schemes. For comparison, the regression parameters $\btheta$ minimizing the respective costs were also computed offline using Newton’s method for the logistic loss, and Powell’s method (initialized from the logistic solution) for both the hinge and quasi 0–1 losses.
Five loops proved to be more than enough to achieve convergence in most cases using FF-EP with VQ, which was clearly the most stable among the tested approximation methods in all scenarios, and converged to close-to-optimal parameters. The slight residual fluctuations observed for the ionosphere dataset could be due to the combined effect of non-convexity and a relatively small number $N$ of examples compared to the number $d$ of features in this dataset. Both Laplace-style approximations (LA and QLA) converged well for logistic and hinge regression, although the hinge loss is not continuously differentiable, but were both unstable in quasi 0–1 regression. GQ showed oscillatory behavior or local convergence in all cases, except to some extent for logistic regression in small dimension (Haberman dataset, $d=4$).
While QLA performed comparably with LA, it was much faster, as shown by the timings reported in Table \[tab:small\_datasets\]. This is due to the function evaluations involved in the optimization step of LA, and suggests that such optimization may not be needed in EP. VQ entailed some computational overhead (depending on the parameter dimension $d$) compared to both QLA and GQ, but was also significantly faster than LA.
-- -- -- -- ----------- -------- ------- ------ -------
Model LA QLA GQ VQ
Logistic 4.78 0.59 0.34 1.05
Hinge 5.61 0.59 0.32 1.03
Quasi 0–1 11.86 0.85 0.41 1.68
Logistic 8.11 0.63 0.44 2.22
Hinge 7.67 0.67 0.40 5.56
Quasi 0–1 17.15 0.91 0.50 6.38
Logistic 9.86 0.63 0.40 1.59
Hinge 7.86 0.65 0.38 1.82
Quasi 0–1 12.82 0.91 0.49 1.81
Logistic 95.33 9.85 4.26 8.26
Hinge 117.39 10.37 3.91 9.64
Quasi 0–1 323.56 15.62 5.78 10.36
Logistic 191.53 12.20 8.14 33.97
Hinge 230.76 12.49 7.20 37.98
Quasi 0–1 628.13 17.87 9.10 33.93
-- -- -- -- ----------- -------- ------- ------ -------
: Dataset characteristics and timing on a standard single processor of different FF-EP versions for binary classification experiments. []{data-label="tab:small_datasets"}
![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](haberman_logistic.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](haberman_hinge.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](haberman_quasi_0-1.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](ionosphere_logistic.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](ionosphere_hinge.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](ionosphere_quasi_0-1.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](wdbc_logistic.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](wdbc_hinge.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for small datasets (from top to bottom: Haberman, ionosphere, breast cancer), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:small_datasets"}](wdbc_quasi_0-1.pdf "fig:"){width=".32\textwidth"}
Large datasets
--------------
Both the Bank Marketing and Census income UCI datasets were also considered for the empirical evaluation of numerical FF-EP approximations. For such training datasets containing several thousands of examples, we can afford larger mini-batches while keeping a sufficiently large number of mini-batches to take advantage of information redundancy in the data. This is motivated by the fact that larger mini-batches are presumably stronger learners.
In both datasets, we chose mini-batches of size $s=100$. Since there are about 100 times more examples than in the “small datasets”, both the size and number of mini-batches were roughly 10 times larger than in the previous classification experiments. A single loop over mini-batches was performed. In this case, memory load can be substantially reduced as there is no need to keep the approximate factors in memory; since they are initialized as uniform distributions, the cavity distribution can be updated on-the-fly as each mini-batch is visited only once.
Evolution curves of the same type as for the small dataset experiments are shown in Figure \[fig:large\_datasets\]. In these experiments, the offline minimizations were not computed due to prohibitive computational cost. Both Laplace-style methods proved more stable and more similar to VQ in quasi 0–1 regression, while GQ had a strong tendency to underestimate factor variances and quickly get stuck on a sub-optimal solution.
The better behavior of Laplace-based FF-EP can be explained by the larger number of factors (mini-batches) in these experiments, which helped stabilizing FF-EP updates, along with the fact that factors based on larger mini-batches tend to have more regular shapes. In the cases of logistic and hinge regression, convergence using either LA, QLA, or VQ occurred much before all mini-batches were visited. For the quasi 0–1 loss, QLA achieved a slightly lower cost than VQ, but did not converge as smoothly as VQ.
Table \[tab:small\_datasets\] shows that VQ ran slightly faster than QLA for the bank marketing data ($d=43$), and about 2–3 times slower for the census income data ($d=101$), consistently with theory since the complexities of VQ and QLA are respectively in $O(d^3)$ and $O(d)$. LA was, again, very slow without showing better performance than QLA and VQ.
![Evolution of total classification cost over FF-EP updates for large datasets (top: Bank marketing, bottom: Census income), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:large_datasets"}](bank_marketing_logistic.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for large datasets (top: Bank marketing, bottom: Census income), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:large_datasets"}](bank_marketing_hinge.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for large datasets (top: Bank marketing, bottom: Census income), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:large_datasets"}](bank_marketing_quasi_0-1.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for large datasets (top: Bank marketing, bottom: Census income), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:large_datasets"}](census_income_logistic.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for large datasets (top: Bank marketing, bottom: Census income), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:large_datasets"}](census_income_hinge.pdf "fig:"){width=".32\textwidth"} ![Evolution of total classification cost over FF-EP updates for large datasets (top: Bank marketing, bottom: Census income), different loss functions (from left to right: logistic, hinge and quasi 0–1), and different approximation schemes.[]{data-label="fig:large_datasets"}](census_income_quasi_0-1.pdf "fig:"){width=".32\textwidth"}
Discussion
==========
We discussed and tested several methods to approximate EP updates when analytically intractable: Laplace-style approximations, standard Gaussian quadrature and deterministic variational sampling, or variational quadrature. All methods are widely applicable as they only require that the factors can be evaluated. Laplace-style methods further require evaluating the first- and second-order derivatives: using analytic expressions is generally preferred, but derivatives can also be approximated using finite differences, yielding the same complexity as the precision-3 quadrature rule considered here.
Experimental results confirm that the Laplace-based approximation works better for EP than low-order Gaussian quadrature, as previously reported by others [@Yu-06], yet it is also significantly slower. Much computation time can be saved by avoiding the optimization step in the Laplace method, which does not seem to hamper performance in practice. However, both Laplace-based FF-EP algorithms were found to be sometimes unstable with non-smooth distributions. While the Laplace method is local in essence, both Gaussian quadrature and variational quadrature perform factor approximations at a coarser scale adaptively determined by the EP algorithm via the cavity distribution. This makes them, in principle, more robust to sudden changes in factor slope or curvature, which may confuse Laplace-style methods while being irrelevant to the overall function approximation problem.
The precision-3 Gaussian quadrature rule proved, however, too inaccurate to enable good convergence of the FF-EP algorithm in practice. While this problem could potentially be overcome using higher order quadrature rules or Monte Carlo methods at the price of increased computation time, the refinement provided by variational quadrature without sampling additional points led to a dramatic improvement. The intuitive reason is that variational quadrature is designed to be exact for Gaussian factors, and is thus generally more accurate than Gaussian quadrature, which is only exact for constant factors (see Section \[sec:gauss\_quad\]).
One drawback of variational quadrature is to scale unfavorably with parameter dimension. Using Newton’s method in the fitting step, the complexity is in $O(d^3)$, while both quick Laplace and Gaussian quadrature are linear in $d$. A possibility to reduce the complexity to linear would be to use an approximate minimization strategy. However, in moderate parameter dimension ($d<100$), the tested implementation proved experimentally faster than the Laplace method, which seems to be the current standard for numerical EP [@Smola-03; @Yu-06], and showed an order of computation time comparable with both quick Laplace and Gaussian quadrature, two methods chosen for computational speed.
The practical recommendation suggested by this work is to use the quick Laplace method in applications where the factors vary smoothly in $\btheta$ and speed is a requirement; otherwise, variational quadrature seems to be preferable. Also note that increasing the size of mini-batches is a way to reduce the computational overhead of variational quadrature with respect to quick Laplace.
[^1]: alexis.roche@{centraliens.net,gmail.com}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
8.75in 6.2in -0.8in -0.45in =cmr10 scaled0
[**ENHANCEMENT OF THE $\Lambda_b$ DECAY RATE**]{} [^1]
*Jonathan L. Rosner*
*Div. TH, CERN*
*1211 CH Geneva 23, Switzerland*
and
*Enrico Fermi Institute and Department of Physics*
*University of Chicago, Chicago, IL 60637 [^2]*
**ABSTRACT**
> The enhancement $\Delta \Gamma (\Lambda_b)$ of the $\Lambda_b$ decay rate due to four-fermion processes of weak scattering and Pauli interference is calculated within the quark model. An estimate of the relative $bu$ wave function at zero separation, $|\Psi(0)|^2_{bu}$, is obtained in terms of the $\Sigma_b^* - \Sigma_b$ hyperfine splitting, the $B^* - B$ hyperfine splitting, and the $B$ meson decay constant $f_B$. For $M(\Sigma_b^*) - M(\Sigma_b) = 56
> \pm 16$ MeV, $M(B^*) - M(B) = 46$ MeV, and $f_B = 190 \pm 40$ MeV, we find $\Delta \Gamma(\Lambda_b) = (0.025 \pm 0.013)$ ps$^{-1}$, to be compared with the observed enhancement $\Gamma(\Lambda_b) - \Gamma(B^0) = 0.20 \pm 0.05$ ps$^{-1}$. Even such a meager enhancement entails a value of $|\Psi(0)|^2_{bu}$ considerably larger than the corresponding value of $|\Psi(0)|^2_{cd}$ in the $\Lambda_c$ baryon.
The differences among lifetimes of particles containing heavy quarks are expected to become smaller as the heavy quark mass increases and free-quark estimates become more reliable. Thus, although charmed particles have lifetimes ranging from less than 0.1 ps for the $\Omega_c$ [@Oclife] to greater than 1 ps for the $D^+$ [@PDG], mesons and baryons containing $b$ quarks are expected to have lifetimes differing no more than a few percent [@Bigia; @Bigib; @NS]. For example, it is expected that the process, $b u \to
c d$ in the $\lb$ (“weak scattering”), when considered in conjunction with the partially offsetting process $b d \to c \bar u d d$ (“Pauli interference”) should lead to a small enhancement in the $\lb$ decay rate, so that $\tlb = (0.9~{\rm to}~0.95) \tb$. Some caution has been urged with regard to the four-quark matrix element in these estimates [@MSpc]. In the present Letter we present a new evaluation of this matrix element which confirms the expected smallness of the enhancement. We perform this evaluation using a hyperfine splitting sensitive to the heavy quark – light quark interaction, which has recently become possible in $b$-flavored baryons as a result of a measurement of the $\sbs - \sb$ splitting by the DELPHI Collaboration [@Delphi].
The observed $\lb$ lifetime is $\tlb = 1.20 \pm 0.07$ ps, while the $B^0$ decays more slowly: $\tb = (1.58 \pm 0.05)$ ps. Here we have averaged a compilation of world data [@Sharma] (for which $\tb = 1.18 \pm 0.07$ ps) with a new value [@CDFtb] $\tb = 1.33 \pm 0.16 \pm 0.07$ ps. The ratio of these two quantities is $\tlb/\tb = 0.76 \pm 0.05$, indicating an enhancement of the $\lb$ decay rate beyond the magnitude of usual estimates.
In the present Letter we find that, in spite of a large wave function for the $b u$ pair in the initial baryon, which we denote by $\psbu$, only $(13 \pm
7)\%$ of the needed enhancement of the $\lb$ decay rate can be explained in terms of the effects of the four-fermion matrix element. (Isospin symmetry then dictates $\psbd = \psbu$.) If we assume wave functions are similar in all baryons with a single $b$ quark and two nonstrange quarks, this quantity can be related to the hyperfine splitting $M(\sbs) - M(\sb)$, for which the DELPHI Collaboration at LEP [@Delphi] has recently quoted a large value of $56 \pm
16$ MeV. We estimate the effect of gluon exchange by performing a similar calculation for $B$ mesons, relating the $B^* - B$ splitting to the $B$ meson decay constant and taking account of differing spin and hyperfine factors in the meson and baryon systems.
A relation for the enhancement of the $\lc$ decay rate due to the weak scattering process $c d \to s u$ was first pointed out in Ref. [@BLS]. At the same order in heavy quark mass, one must also take account of Pauli interference (interference between identical quarks in the final state) [@VS; @Gub]. Thus, for the $\lb$, one considers not only the process $b u \to
c d$ (involving matrix elements between $\lb$ states of $(\bar b b)(\bar u u)$ operators), but also those processes involving matrix elements of $(\bar b
b)(\bar d d)$ operators) which contribute to interference. The net result of four-quark operators in the $\lb$ is an enhancement of the decay rate by an amount (see, e.g., Refs. [@VS; @Gub; @NS])
\[eqn:dg\] () = |V\_[ud]{}|\^2 |V\_[cb]{}|\^2 m\_b\^2(1-x)\^2 \[c\_-\^2 - (1+x)c\_+(c\_- - c\_+/2)\] . Here we have neglected light-quark masses; $x \equiv m_c^2/m_b^2$, while $c_-$ and $c_+ = (c_-)^{-1/2}$ are the short-distance QCD enhancement and suppression factors for quarks in a color antitriplet and sextet, respectively: [@QCDenh]: c\_- = \^ , = , with $n_F = 5$ the number of active quark flavors between $m_b$ and $M_W$. The $c_-^2$ term in square brackets reflects the weak scattering process $bu
\to cd \to bu$, while the remaining terms arise from destructive interference between the two intermediate $d$ quarks in the process $bd \to c \bar u dd \to
bd$.
Taking the strong interaction scale in the modified-minimal-subtraction scheme for four quark flavors to be [@KMRR] $\Lambda_{\overline{\rm MS}}^{(4)} =
200$ MeV, we find $\alpha_s(m_b^2) = 0.193$ and $\alpha_s(M_W^2) = 0.114$, and hence $c_- = 1.32,~c_+ = 0.87$. An estimate of $\psbu$ is then needed. We find it by comparing hyperfine splittings in mesons and baryons, under the assumption that the strength of the one-gluon exchange term is the same for the light quark – heavy quark pair in each system.
Our result, which we shall explain in more detail presently, is \[eqn:ps\] = 2 , where the first factor relates to color, the second to spin, and the last term is the nonrelativistic estimate of the $b \bar u$ wave function in the $B$ meson [@wf]. (Here one may use the spin-averaged value of vector and pseudoscalar masses for $M_M$.) With the DELPHI value of $M(\sbs) - M(\sb)$, the $B^* - B$ splitting of 46 MeV [@PDG], and the estimate [@fm; @lat] $f_B = 190 \pm 40$ MeV, we obtain $\psbu = (2.6 \pm 1.3) \times 10^{-2}$ GeV$^3$. This is to be compared with $\psbub = M_B f_B^2/12 = (1.6 \pm 0.7)
\times 10^{-2}$ GeV$^3$ for the $B$ meson. Our assumption of equal values of $\alpha_s$ governing the hyperfine interaction of the light – heavy pair in the meson and baryon must be viewed cautiously in this light. It is possible that $\alpha_s$ in the more compact baryonic system (since the wave function appears to be larger) is smaller than in the meson, which would further enhance the estimate of $\psbu$.
Several previous estimates of the four-quark matrix element (see, e.g., Refs. [@Bigib; @BLS; @VS; @BS]) utilized the hyperfine splitting between the $\Lambda_b$ and the $\Sigma_b$ (or the corresponding charmed states). Since this splitting involves the interaction of two light quarks as well as one light and one heavy quark, it necessarily involves a statement about light quark masses, as well as about the relation between the relative wave function of two light quarks and that of a light and a heavy quark. Our approach avoids such assumptions.
In the relation (\[eqn:dg\]) we now neglect $\sin \theta_c$ (setting $V_{ud}
=1$), and choose $m_b = 5.1$ GeV, $m_c = 1.7$ GeV, and $|V_{cb}| = 0.040 \pm
0.003$. We then find () = 0.025 0.013 [ps]{}\^[-1]{} .
The decay rates of the $B^0$ and $\lb$ are $\gb = 0.63 \pm 0.02$ ps$^{-1}$ and $\glb = 0.83 \pm 0.05$ ps$^{-1}$, differing by $\dg(\lb) = 0.20 \pm 0.05$ ps$^{-1}$. The four-quark processes noted above can explain only $(13 \pm 7)\%$ of this difference, leading to an enhancement of only $(4 \pm 2)\%$ of the total $\lb$ decay rate in contrast with the needed enhancement of $(32 \pm
8)\%$.
We now give some details of the calculation and see how well it does for the $\lc$.
The hyperfine interaction in a meson $M_{i \bar j}$ composed of a quark $i$ and an antiquark $\bar j$ leads to a mass shift [@DGG] \[eqn:mhfs\] M(M\_[i |j]{}) = \_s |(0)|\^2 , where $\hat s_i$ refers to a quark spin operator. The corresponding result for a baryon $B_{ijk}$ is \[eqn:bhfs\] M(B\_[ijk]{}) = \_s \_[i > j]{} |(0)|\_[ij]{}\^2 . The relative factor of 2 arises from the different color factors for a quark and antiquark in a meson (a triplet and antitriplet making a singlet) and two quarks in a baryon (two triplets making an antitriplet). This factor of 2 is the first term in Eq. (\[eqn:ps\]).
We take the ratio of hyperfine splittings in mesons and baryons so that light quark masses and values of $\alpha_s$ cancel out. We thus assume that (a) effective light quark masses in mesons and baryons are equal (borne out at the 20% level by phenomenological fits to meson and baryon spectra [@DGG]), and (b) the values of $\alpha_s$ governing the corresponding hyperfine splittings are similar (reasonable since in both mesons and baryons one is concerned with systems of one light and one heavy quark).
We must isolate a baryon mass shift sensitive to the interaction between a light quark and a heavy one. The $\sbs - \sb$ splitting is the appropriate quantity. In both $\sb$ and $\sbs$, the light quarks are coupled up to spin 1. The splitting then depends purely on the light quark – heavy quark interaction.
The wave function between a light quark and a heavy one is assumed, as mentioned, to be identical in the $\lb$ and in the $\sb - \sbs$ system. The two light quarks are coupled up to zero spin in the $\lb$, and hence have zero net hyperfine intreraction with the heavy quark, while the hyperfine interaction between the light quarks and the heavy one averages to zero if we take the spin-weighted average of the $\sb$ and the $\sbs$.
The value of $\langle \hat{s}_Q \cdot \hat{s}_{\bar q} \rangle$ is $(1/4,-3/4)$ for a $(^3S_1,~^1S_0)$ $Q \bar q$ meson, where $Q$ and $q$ are the heavy and light quark. For a baryon $Qqq$ with $S_{qq} = 1$, one has $\langle \hat{s}_Q
\cdot \hat{s}_q \rangle = (1/4,~-1/2)$ for states with total spin (3/2, 1/2). Thus the difference in $\hat{s}_i \cdot \hat{s}_j$ for the $\sbs - \sb$ splitting (counting a factor of 2 for the two light quarks in the baryons) is 3/2 that for the $B^* - B$ splitting. The factor of 2/3 in Eq. (\[eqn:ps\]) compensates for this ratio.
The relation [@wf] $|\Psi(0)|^2 = M_M f_M^2 /12$ for the square of the wave function of a $Q \bar q$ meson probably has important corrections of order $1/m_Q$, if lattice calculations are any guide [@lat]. These are ignored in the present discussion. They are likely to be more important when we apply the present method to charmed states.
The corresponding calculation for charmed particles makes use of the following inputs.
[*1. The $D$ meson decay constant*]{} was taken [@fm] to be $f_D =
240 \pm 40$ MeV, leading (with $M_D = 1973$ MeV) to $\pscdb = (0.95 \pm 0.32)
\times 10^{-2}$ GeV$^3$.
[*2. The $D^* - D$ splitting*]{} is assumed to be 141 MeV (the average for charged and neutral states [@PDG]).
[*3. Charmed baryon masses*]{} are taken to be $M(\sc) = 2453$ MeV [@PDG] and $M(\scs) = 2530 \pm 7$ MeV [@Ammosov].
[*4. The strong fine-structure-constant*]{} at $m_c^2$ is taken to be $\alpha_s(m_c^2) = 0.289$, consistent with the QCD scale mentioned above, leading to $c_- = 1.60,~c_+ = 0.79$.
[*5. The strange quark mass*]{} is taken to have a typical constituent-quark value, $m_s = 0.5$ GeV. We continue to neglect $u$ and $d$-quark masses for simplicity.
[*6. The CKM factors*]{} in Eq. (\[eqn:dg\]) undergo the replacements $|V_{ud}|^2 |V_{cb}|^2 \to |V_{cs}|^2 |V_{ud}|^2$, which we approximate by 1 (again neglecting $\sin \theta_c$).
The resulting matrix element $\pscd = (0.69 \pm 0.24) \times 10^{-2}$ GeV$^3$ is consistent with an estimate in Ref. [@BLS] of $\pscd = 0.74 \times
10^{-2}$ GeV$^3$. In that work, the value of $\pscd$ was assumed to be the same as for two light quarks in the charmed baryon, and was estimated using the observed $\sc - \lc$ splitting. It was also necessary to assume a specific value of $\alpha_s = 0.58$ in the hyperfine interaction expression (6), which we do not do here. Our value of $c_-$ is smaller than assumed in Ref. [@BLS] and we take account of destructive interference, leading to a smaller result for $\dg(\lc)$.
The results for systems with $c$ and $b$ quarks based on our method are summarized in Table I. Several remarks can be made.
Quantity (units) Charm Beauty
-------------------------------------------- ----------------- -------------------
$f_M$ (MeV) $240 \pm 40$ $190 \pm 40$
$|\Psi(0)|^2_{Q \bar q}~(10^{-2}$ GeV$^3$) $0.95 \pm 0.32$ $1.6 \pm 0.7$
$M(^3S_1) - M(^1S_0)$ (MeV) 141 46
$M(\Sigma^*) - M(\Sigma)$ (MeV) $77 \pm 7$ $56 \pm 16$
$|\Psi(0)|^2_{Q q}~(10^{-2}$ GeV$^3$) $0.69 \pm 0.24$ $2.6 \pm 1.3$
$c_-$ 1.60 1.32
$c_+$ 0.79 0.87
$c_-^2 - (1+x)c_+(c_- - c_+^2/2)$ 1.52 0.88
$\dg(\Lambda_Q)$ (ps$^{-1}$) $0.8\pm 0.3$ $0.025 \pm 0.013$
: Comparison of predicted squares of wave functions and decay rate enhancements for $\lc$ and $\lb$.
\(a) The difference between the central values of $|\Psi(0)|^2_{Q \bar q}$ for charm and beauty reflects the likely importance of $1/m_Q$ corrections (see, e.g., Ref. [@lat]), or – in the language of the quark model – of reduced mass effects.
\(b) The $\scs - \sc$ hyperfine splitting used in this calculation is based on one claim for observation of the $\scs$ [@Ammosov], which requires confirmation.
\(c) The value of $\psbu$ is somewhat large in comparison with the others for light-heavy systems. It would be helpful to verify the large hyperfine splitting between $\sbs$ and $\sb$ claimed by the DELPHI Collaboration [@Delphi]. The ratio of hyperfine splittings for charmed and beauty mesons is approximately 3:1, as expected if these splittings scale as $1/m_Q$. In contrast, the corresponding ratio for baryons is considerably smaller, indicating a violation of $1/m_Q$ scaling.
\(d) The enhancement of the $\lc$ decay rate is quite modest. With $\glc \approx
5$ ps$^{-1}$, to be compared with $\gdn \approx 2.4$ ps$^{-1}$ and $\gdc
\approx 1$ ps$^{-1}$, one seeks an enhancement of at least $\glc - \gdn \approx
2.6$ ps$^{-1}$. If the enhancements $\dg(\Lambda_Q)$ in Table I were about a factor of 4 larger, we could accommodate both the $\lc$ and $\lb$ decay rates, but this is not consistent with our estimates of the matrix elements and their effects on decay rates. In particular, the effect of Pauli interference is to cut the naïve estimate of the enhancement due to weak scattering alone [@BLS] by roughly a factor of 2. Hybrid logarithms [@VS], not considered here, have a relatively modest effect, leading if anything to further suppression of the enhancement for $\lb$ decay [@NS].
To summarize, we have used the hyperfine splitting between $\sbs$ and $\sb$ claimed by the DELPHI Collaboration [@Delphi] to estimate the overlap of quark wave functions between the $b$ quark and the light quarks in the $\Lambda_b$, and hence to estimate the effect of four-quark operators on its decay rate. Even though the matrix element $\psbu = \psbd$ deduced from the DELPHI result is quite large on the scale of those for heavy-light systems, one can only account for $(13 \pm 7)\%$ of the difference between the $\lb$ and $B^0$ decay rates, or an enhancement of $(4 \pm 2)\%$ of the $\lb$ decay rate. A similar approach also falls short of accounting for the corresponding enhancement for the $\lc$ decay rate. If the enhanced $\lb$ decay rate is borne out by further data, we can only speculate that strong final-state interactions which cannot be anticipated on the basis of perturbative QCD must play a role even at the rather high mass of the $\lb$.
I wish to thank G. Martinelli, M. Neubert, C. T. Sachrajda, M. Shifman, and N. G. Uraltsev for fruitful discussions, and the Physics Department of the Technion for its hospitality during the initial stages of this investigation. This work was supported in part by the United States – Israel Binational Science Foundation under Research Grant Agreement 94-00253/1 and by the United States Department of Energy under Contract No. DE FG02 90ER40560.
\#1\#2\#3[Am. J. Phys. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Ann. Phys. (N.Y.) [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Acta Phys. Polonica [**\#1**]{} (\#3) \#2 ]{} \#1\#2\#3[Ann. Rev. Nucl. Part. Sci. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Commun. Math. Phys. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Comments on Nucl. Part. Phys. [**\#1**]{} (\#3) \#2]{} 93[[*Lepton and Photon Interactions: XVI International Symposium, Ithaca, NY August 1993*]{}, AIP Conference Proceedings No. 302, ed. by P. Drell and D. Rubin (AIP, New York, 1994)]{} 89[[*CP Violation,*]{} edited by C. Jarlskog (World Scientific, Singapore, 1989)]{} 94[DPF 94 Meeting, Albuquerque, NM, Aug. 2–6, 1994]{} \#1\#2\#3[Europhys. Lett. [**\#1**]{} (\#3) \#2]{} 79[[*Proceedings of the 1979 International Symposium on Lepton and Photon Interactions at High Energies,*]{} Fermilab, August 23-29, 1979, ed. by T. B. W. Kirk and H. D. I. Abarbanel (Fermi National Accelerator Laboratory, Batavia, IL, 1979]{} 87[[*Proceeding of the 1987 International Symposium on Lepton and Photon Interactions at High Energies,*]{} Hamburg, 1987, ed. by W. Bartel and R. Rückl (Nucl. Phys. B, Proc. Suppl., vol. 3) (North-Holland, Amsterdam, 1988)]{} \#1\#2\#3[ [**\#1**]{} (\#3) \#2]{} 72[[*Proceedings of the XVI International Conference on High Energy Physics*]{}, Chicago and Batavia, Illinois, Sept. 6–13, 1972, edited by J. D. Jackson, A. Roberts, and R. Donaldson (Fermilab, Batavia, IL, 1972)]{} \#1\#2\#3[Int. J. Mod. Phys. A [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[J. Math. Phys. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[J. Phys. G [**\#1**]{} (\#3) \#2]{} 87[[*Selected Topics in Electroweak Interactions*]{} (Proceedings of the Second Lake Louise Institute on New Frontiers in Particle Physics, 15–21 February, 1987), edited by J. M. Cameron (World Scientific, Singapore, 1987)]{} 85[[*Proceedings of the International Symposium on Lepton and Photon Interactions at High Energy,*]{} Kyoto, Aug. 19-24, 1985, edited by M. Konuma and K. Takahashi (Kyoto Univ., Kyoto, 1985)]{} \#1\#2\#3[Mod. Phys. Lett. A [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Nuovo Cim. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Nucl. Phys. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3\#4[Pis’ma Zh. Eksp. Teor. Fiz. [**\#1**]{} (\#3) \#2 \[JETP Lett. [**\#1**]{} (\#3) \#4\]]{} \#1\#2\#3[Phys. Lett. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Phys. Lett. B [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Phys. Rev. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Phys. Rev. A [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Phys. Rev. D [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Phys. Rev. Lett. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Phys. Rep. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Prog. Theor. Phys. [**\#1**]{} (\#3) \#2]{} \#1\#2\#3[Rev. Mod. Phys. [**\#1**]{} (\#3) \#2]{} \#1[ ……[rp ]{}[\#1]{} ]{} 90[25th International Conference on High Energy Physics, Singapore, Aug. 2-8, 1990]{} 87[[*Proceedings of the Salt Lake City Meeting*]{} (Division of Particles and Fields, American Physical Society, Salt Lake City, Utah, 1987), ed. by C. DeTar and J. S. Ball (World Scientific, Singapore, 1987)]{} 89[[*Proceedings of the XIVth International Symposium on Lepton and Photon Interactions,*]{} Stanford, California, 1989, edited by M. Riordan (World Scientific, Singapore, 1990)]{} 82[[*Proceedings of the 1982 DPF Summer Study on Elementary Particle Physics and Future Facilities*]{}, Snowmass, Colorado, edited by R. Donaldson, R. Gustafson, and F. Paige (World Scientific, Singapore, 1982)]{} 90[[*Research Directions for the Decade*]{} (Proceedings of the 1990 Summer Study on High Energy Physics, June 25 – July 13, Snowmass, Colorado), edited by E. L. Berger (World Scientific, Singapore, 1992)]{} 90[[*Testing the Standard Model*]{} (Proceedings of the 1990 Theoretical Advanced Study Institute in Elementary Particle Physics, Boulder, Colorado, 3–27 June, 1990), edited by M. Cvetič and P. Langacker (World Scientific, Singapore, 1991)]{} \#1\#2\#3\#4[Yad. Fiz. [**\#1**]{} (\#3) \#2 \[Sov. J. Nucl. Phys. [**\#1**]{} (\#3) \#4\]]{} \#1\#2\#3\#4\#5\#6[Zh. Eksp. Teor. Fiz. [**\#1**]{} (\#3) \#2 \[Sov. Phys. - JETP [**\#4**]{} (\#6) \#5\]]{} \#1\#2\#3[Zeit. Phys. C [**\#1**]{} (\#3) \#2]{}
[99]{}
Fermilab E687 , P. L. Frabetti , ; CERN WA89 , M. I. Adamovich , .
Particle Data Group, L. Montanet , .
I. I. Bigi , in , p. 132.
I. I. Bigi, University of Notre Dame report UND-HEP-95-BIG02, hep-ph/9508408, preliminary version of a paper to be submitted to Physics Reports.
M. Neubert and C. T. Sachrajda, CERN report CERN-TH/96-19, February, 1996, in preparation.
M. Shifman, presented at Symposium on Heavy Flavours, Haifa, Israel, December, 1995 (unpublished).
DELPHI , P. Abreu , DELPHI report DELPHI 95-107 PHYS 542, June, 1995, presented by M. Feindt, C. Kreuter, A. Miagkov, and O. Podobrin at EPS-HEP 95 Conference, Brussels, 27 July - 2 August, 1995 (unpublished).
V. Sharma, presented at Symposium on Heavy Flavours, Haifa, Israel, December, 1995 (unpublished), updated version of V. Sharma and F. V. Weber, in , p. 395.
CDF , F. Abe , 1996 (unpublished).
V. Barger, J. P. Leveille, and P. M. Stevenson, .
M. B. Voloshin and M. A. Shifman, ; ; M. B. Voloshin, N. G. Uraltsev, V. A. Khoze, and M. A. Shifman, .
N. Bilić, B. Guberina, and J. Trampetić, ; B. Guberina, R. Rückl, and J. Trampetić, .
G. Altarelli and L. Maiani, ; M. K. Gaillard and B. W. Lee, ; G. Altarelli, N. Cabibbo, and L. Maiani, .
W. Kwong, P. B. Mackenzie, R. Rosenfeld, and J. L. Rosner, .
E. V. Shuryak, .
J. L. Rosner, 95-36, lectures presented at the VIII J. A. Swieca Summer School, Rio de Janeiro, Brazil, February 7 – 11, 1995, Proceedings to be published by World Scientific.
UKQCD , R. M. Baxter , ; C. W. Bernard, J. N. Labrenz, and A. Soni, [**49**]{} (1994) 2536; A. Duncan , [**51**]{} (1995) 5101.
B. Blok and M. Shifman, in [*Proceedings of the Third Workshop on Physics at a Tau-Charm Factory*]{}, Marbella, Spain, June 1993, edited by J. Kirkby and R. Kirkby (Editions Frontières, Gif-sur-Yvette, France, 1994).
A. De Rújula, H. Georgi, and S. L. Glashow, .
V. A. Ammosov , .
[^1]: To be submitted to Phys. Lett. B.
[^2]: Permanent address.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
[**Affine Poisson and affine quasi-Poisson T-duality**]{}
[\
Aix Marseille Université, CNRS, Centrale Marseille\
I2M, UMR 7373\
13453 Marseille, France]{}
**Abstract**
We generalize the Poisson-Lie T-duality by making use of the structure of the affine Poisson group which is the concept introduced some time ago in Poisson geometry as a generalization of the Poisson-Lie group. We also introduce a new notion of an affine quasi-Poisson group and show that it gives rise to a still more general T-duality framework. We establish for a class of examples that this new T-duality is compatible with the renormalization group flow.
Introduction
============
The Poisson-Lie T-duality [@KS95] is the framework which permits to construct examples of dynamically equivalent non-linear $\sigma$-models living on geometrically non-equivalent backgrounds. The basic structural ingredient underlying this kind of T-duality is the so called Drinfeld double $D$, which is a Lie group equipped with a bi-invariant metric of maximally Lorentzian (or split) signature and having two half-dimensional isotropic subgroups $K$ and $\tilde K$.
The structure of $D$ induces certain Poisson brackets on each subgroup $K$ and $\tilde K$. Those brackets are called the Poisson-Lie ones and, remarkably, the Poisson-Lie bivectors which correspond to them appear explicitely in the actions of the mutually T-dual $\sigma$-models[^1]: S=dd( (E+(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_, k(,)K;\[mmm\] S=dd( (E+(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_, k(,)K.\[mmm’\] Note that the respective targets of the $\sigma$-models and are the subgroups $K$ and $\tilde K$ of the double $D$, the bilinear form $(.,.)_\D$ on the Lie algebra $\D$ is the split metric evaluated at the unit element of the the group $D$, the linear operators $E:\tilde\K\to\K$ and $\tilde E:\K\to\tilde\K$ are constant, but the operators $\Pi(k):\tilde\K\to\K$ and $\tilde \Pi(\tilde k):\K\to\tilde\K$ depend on the points of the targets. Actually, the operators $\Pi(k)$ and $\tilde \Pi(\tilde k)$ encode the Poisson-Lie brackets on $K$ and on $\tilde K$ by the formulae {f\_1,f\_2}\_[K]{}(k)=(\^l f\_1,(k)\^l f\_2)\_; {f\_1,f\_2}\_[K]{}(k)=(\^l f\_1,(k)\^l f\_2)\_.\[171\]Here $f_1,f_2$ and $\tilde f_1, \tilde f_2$ are, respectively, functions on the groups $K$ and $\tilde K$ and the precise definitions of the $\tilde K$-valued and the $\K$-valued right-invariant differential operators $\tilde\nabla$ and $\nabla$ are given in Eqs. and . It was established in [@KS95; @KS96a] that the T-duality relates the models and if $\tilde E$ is inverse of $E$.
The duality existing in the realm of the Poisson-Lie groups expresses the fact that starting from a group $K$ and the Poisson-Lie structure $\Pi(k)$ on it, one can construct the dual group $\tilde K$ and the dual Poisson-Lie structure $\tilde\Pi(\tilde k)$ and repeating the same procedure with the pair $(\tilde K,\tilde \Pi(\tilde k))$ one gets back to the pair $(K,\Pi(k))$. It is truly remarkable that this purely geometric duality of the Poisson-Lie groups gets transported via the actions and to the dynamical T-duality in string theory.
Apart from the Poisson-Lie duality, there exists another natural geometric duality in Poisson geometry, the one which flips two Poisson-Lie groups associated to the so called the [*affine Poisson group*]{} $K$ [@DS; @Lu; @K07]. We speak then about geometric [*affine Poisson duality*]{} and the purpose of the present paper is to convert it into a stringy [*affine Poisson T-duality*]{}. The affine Poisson structure $\Pi^{\na}(k)$ on a group $K$ gives rise naturally to the existence of three other Poisson manifolds : the so-called mirror affine Poisson group $(K,\Pi^{\ma}(k))$ and also two (dual) Poisson-Lie groups $\left(\tilde K_L,\tilde\Pi^L(\tilde k_L)\right)$ and $\left(\tilde K_R,\tilde\Pi^R(\tilde k_R)\right)$. We can construct four $\sigma$-model actions of the types , for every of those four Poisson structures S\_=dd((E\^+\^(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_R]{},\[mm1\]S\_=dd( (E\^+\^[ ]{}(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_L]{},\[mm2\] S\_=dd((E\^+\^R(k\_R))\^[-1]{}\_+k\_Rk\_R\^[-1]{}, \_-k\_Rk\_R\^[-1]{})\_[\_R]{}.\[mm3\] S\_=dd((E\^+\^L(k\_L))\^[-1]{}\_+k\_Lk\_L\^[-1]{}, \_-k\_Lk\_L\^[-1]{})\_[\_L]{}.\[mm4\]
The first result of this paper is the statement that for a large class of the affine Poisson groups $(K,\Pi^\na(k))$ one can choose the linear operator $E^\na$ and, in terms of it, all other operators $E^{\ma},\tilde E^\na$ and $\tilde E^{\ma}$ in such a way that all four $\sigma$-models , , and become pairwise T-dual to each other. The most interesting T-duality is that relating the models and , since, as it will turn out, this is the only one which cannot be reduced to the standard Poisson-Lie T-duality. However, if the affine Poisson structure $\Pi^\na(k)$ is equal to its mirror $\Pi^{\ma}(k)$, then we recover the standard Poisson-Lie T-duality, i.e. the $\sigma$-models , merge to give the model and the $\sigma$-models , merge to give the model .
The second result of this paper is the generalization of the pairwise T-duality of the four $\sigma$-models , , and to the case where the structure $\Pi^{\na}(k)$ is not affine Poisson but only affine quasi-Poisson. We stress that the concept of the affine quasi-Poisson group is (apparently) a new one and we introduce it in the present paper. Why do we do it? Because while proving the affine Poisson T-duality of the four $\sigma$-models , , and , we have discovered that the duality holds even if the structure $\Pi^\na(k)$ does not give rise to a bracket of fonctions satisfying the Jacobi identity. We have then worked out what kind of violation of the Jacobi identity is compatible with the T-duality and found that it is one which is well-known in the literature under the name of the quasi-Poisson geometry. Because it is true, at the same time, that structures which would be simultaneously affine and quasi-Poisson have not yet been introduced, we devote one section of the present article to the definition of the affine quasi-Poisson groups and to the study of their basic properties.
The plan of the paper is as follows: In Section 2 and 3, we review the concept of T-duality as the symplectomorphism relating the phase spaces of two mutually dual $\sigma$-models, we recall the elements of the standard Poisson-Lie T-duality and we collect some useful facts from the mathematical literature about the theory of the affine Poisson groups. In Section 4, we establish the T-duality of the four $\sigma$-models , , and when $K$ is the affine Poisson group. In Section 5, we introduce and study the concept of the affine quasi-Poisson group and, in Section 6, we establish the pairwise T-duality of the four models , , and in the case where $K$ is the affine quasi-Poisson group. In Section 7, we expose the theory of dressing cosets [@KS96b] and show how the affine Poisson and the affine quasi-Poisson T-dualities can be interpreted from this vantage point. In Section 8, we show that for a particular class of the affine Poisson groups, the affine Poisson T-duality is compatible with the renormalisation group flow. We finish by an outlook where few open problems are formulated.
$\E$-models, Poisson-Lie T-duality and outline of generalization
================================================================
By definition, the Poisson-Lie T-duality is a symplectomorphism that maps the phase space and the Hamiltonian of the non-linear $\sigma$-model onto the phase space and the Hamiltonian of the dual $\sigma$-model . We shall refer to those models as to the Poisson-Lie $\sigma$-models.
Recall that the basic structural ingredients needed to construct the dual pair of the Poisson-Lie $\sigma$-models and are a $2d$-dimensional Lie algebra $\D$ (the Drinfeld double Lie algebra), a symmetric ad-invariant non-degenerate bilinear form $(.,.)_\D$ on $\D$ with the split signature $(d,d)$, two $d$-dimensional Lie subalgebras $\K$ and $\tilde\K$ of $\D$ such that the restriction of the form $(.,.)_\D$ vanish on both of them, and, finally, a $d$-dimensional linear subspace $\E$ of $\D$ such that the restriction of $(.,.)_\D$ on $\E$ is strictly positive definite. All those data are needed in order to write down a duality invariant description of the common first order dynamics of the Poisson-Lie $\sigma$-models and in terms of the so called $\E$-models [@KS96a; @K15].
The points of the phase space of the $\E$-model are maps $l:S^1\to D$ from a circle parametrized by the variable $\sigma$ into the Drinfeld double $D$, or, in other words, they are the elements of the loop group $LD$ of the Drinfeld double. The symplectic form $\omega_{LD}$ of the $\E$-model is given by the following simple formula \_[LD]{}:=-d(l\^[-1]{}dl,\_(l\^[-1]{}dl))\_,\[342\]where the symbol $\oint$ stands for integration over the loop parameter. Finally, the Hamiltonian $Q_\E$ is given by the formula Q\_=d(\_ll\^[-1]{},\_ll\^[-1]{})\_.\[345\]Recall that we have defined $\E$ as the $d$-dimensional subspace of the $2d$-dimensional Drinfeld double Lie algebra $\D$, such that the restriction of the bilinear form $(.,.)_\D$ on $\E$ is strictly positive definite. By the abuse of notation, we denote by $\E$ also the linear operator $\E:\D\to\D$, self-adjoint with respect to the bilinear form $(.,.)_\D$, which has the subspace $\E$ as the eigenspace for the eigenvalue $+1$ and the orthogonal complement subspace $\E^\perp$ as the eigenspace for the eigenvalue $-1$.
For the purpose of this paper, it will be sufficient to consider a simpler variant of the Poisson-Lie T-duality (actually, the one originally introduced in [@KS95; @KS96a]) for which the Drinfeld double is perfect in the sense of the Footnote 1. In this case, it is particularly simple to construct the symplectomorphisms corresponding to the arrows in the following chain of dynamical systems: (T\^\*LK,,\_)(LD,\_[LD]{},Q\_)(T\^\*LK,,\_).\[Sch1’\] Here $\omega$ is the canonical symplectic form on the cotangent bundle of the loop group $T^*LK$: =d(d, k\^[-1]{}dk),kLK, L\^\*\[sf\] and the arrow pointing from $(LD,\omega_{LD},Q_\E)$ to the left is obtained by writing $l\in LD$ as the product l=kh, kLK, hLK.\[dd\] Indeed, inserting the decomposition into , we obtain easily \_[LD]{}=d(d(\_hh\^[-1]{},k\^[-1]{}dk)\_).\[sf1\] To make coincide the forms and , it is now sufficient to remark that $\partial_\sigma \tilde h\tilde h^{-1}$ is the element of $L\tilde\K$ and can be therefore identified with $\beta\in L\K^*$ via the non-degenerate bilinear form $(.,.)_\D$.
Inserting the decomposition into , we obtain also the Hamiltonian $h_\E$ featuring in the scheme : h\_(k,) =d(\_kk\^[-1]{}+ kk\^[-1]{},(\_kk\^[-1]{}+ kk\^[-1]{}))\_,\[hpm\]where we view $\beta$ as the element of $L\tilde\K$.
The first order action of the dynamical system $(T^*LK,\omega,\H_\E)$ is now given by the data and : S\_=dd(,k\^[-1]{}\_k)\_- dh\_(k,). \[foa\]The dependence of $S_\E$ on $\beta$ is quadratic, it is therefore easy to eliminate $\beta$ which gives the second order action of the Poisson-Lie $\sigma$-model : S\_(k)=dd ((E+(k))\^[-1]{}\_+kk\^[-1]{}, \_-kk\^[-1]{})\_.\[ea\]Note that $E:\tilde\K\to\K$ is the linear operator such that its graph $\{\tilde x+E\tilde x,\tilde x\in\tilde\K\}$ coincides with the subspace $\E$ in the double $\D$ and we recall that $\Pi(k):\tilde\K\to\K$ encodes the Poisson-Lie bracket of two functions $f_1,f_2$ on the group $K$ in the sense of the formula: {f\_1,f\_2}\_[K]{}(k)=(\^l f\_1,(k)\^l f\_2)\_.\[spls’\]Here $\nabla^{l}$ is $\tilde\K$-valued differential operator acting on the functions on $K$ as (\^l f, x)\_(k):= (\^l\_[x]{}f)(k)\_[s=0]{}, x.\[431\]
The Poisson-Lie $\sigma$-model dual to the model can be obtained from and in the same way starting from the dual decomposition of the Drinfeld double: l=kh, kLK,hLK,\[ddd\] The result is the Poisson-Lie $\sigma$-model S\_(k)=dd((E+(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_,\[ead\]Here $\tilde E:\K\to\tilde\K$ is the linear operator such that its graph $\{x +\tilde Ex, x\in\K\}$ coincides with the subspace $\E$ in the double $\D$, which in fact means that $\tilde E=E^{-1}$. Of course, $\tilde\Pi(\tilde k):\K\to\tilde\K$ encodes the Poisson-Lie structure on the group $\tilde K$ in the sense of the formula {f\_1,f\_2}\_[K]{}(k)=(\^l f\_1,(k)\^l f\_2)\_,\[dspl\]where $\tilde\nabla^{l}$ is $\K$-valued differential operator acting on the functions on $\tilde K$ as (\^l f,x)\_(k):= (\^l\_[x]{}f)(k)= \_[s=0]{}, x.\[39\]
We now draw another scheme, which replaces that and encapsulates the generalization of the Poisson-Lie T-duality introduced in the present paper:
$$\begin{aligned}
(T^*LK,\omega,\H_{\E_L})\longleftrightarrow (LD_L,&\omega_{LD_L},Q_{\E_L})\longleftrightarrow (T^*L\tilde K_L,\tilde\omega_L,\tilde\H_{\E_L}) \nonumber
\\ \updownarrow \hskip 2.7cm &\updownarrow\hskip 9pc\updownarrow\label{Sch2} \\
(T^*LK,\omega,\H_{\E_R})\longleftrightarrow (LD_R,&\omega_{LD_R},Q_{\E_R})\longleftrightarrow (T^*L\tilde K_R,\tilde\omega_R,\tilde\H_{\E_R}) .\nonumber\end{aligned}$$
The horizontal and the vertical bidirectional arrows in the scheme stand for the Hamiltonian-preserving symplectomorphisms between the dynamical systems represented by the triples $(P,\omega,\H)$, where $P$ is the phase space of the dynamical system, $\omega$ is its symplectic form and $\H$ is its Hamiltonian. In particular, the notation $(T^*LK,\omega,\H_{\E_R})$ means that the cotangent bundle $T^*LK$ of the loop group $LK$ is the phase space of a non-linear $\sigma$-model living on the target manifold $K$, the symplectic form $\omega$ is the canonical one on the cotangent bundle and the Hamiltonian is $\H_{\E_R}$. Of course, the triplet $(T^*LK,\omega,\H_{\E_R})$ is extracted from the action of the $\sigma$-model by the standard manipulations, which start from defining the canonical momenta and end by writing up the first-order Hamiltonian description of the dynamics.
Let us now identify all dynamical systems located at the corners of the scheme . The lower-left corner $(T^*LK,\omega,\H_{\E_R})$ represent the $\sigma$-model , the upper-left corner $(T^*LK,\omega,\H_{\E_L})$ is the $\sigma$-model , the lower-right corner $(T^*L\tilde K_R,\tilde\omega_R,\tilde\H_{\E_R})$ corresponds to the action and the upper-right corner $(T^*L\tilde K_L,\tilde\omega_L,\tilde\H_{\E_L})$ to the action . Actually, the target of the $\sigma$-model corresponding to the given corner can be always read off from the first entry of the triple, e.g. $T^*LK$ means that the phase space of the $\sigma$-model is the cotangent bundle of the loop group $LK$, therefore the target of the corresponding $\sigma$-model is the group $K$.
The dynamical systems $(LD_L,\omega_{LD_L},Q_{\E_L})$ and $(LD_R,\omega_{LD_R},Q_{\E_R})$ located in the middle column of the scheme are the $\E$-models. In general, they live on non-isomorphic Drinfeld doubles $D_L\ne D_R$ but the scheme represents a non-trivial generalization of the Poisson-Lie T-duality even if $D_L$ is isomorphic to $D_R$ with only the subspaces $\E_L$ and $\E_R$ being different.
The scheme represents first of all the superposition of two standard Poisson-Lie dualities, one on the upper line of the scheme and the other downstairs. Of course, if we want to apply the scheme to the context of the $\sigma$-models ,, and , we have to prove that the $\sigma$-model is Poisson-Lie T-dual to the $\sigma$-model via some intermediate $\E$-model $(LD_R,\omega_{LD_R},Q_{\E_R})$, while is Poisson-Lie T-dual to via some $(LD_L,\omega_{LD_L},Q_{\E_L})$. Indeed, after getting familiar with the properties of the affine (quasi-)Poisson groups in Sections 3 and 5, we shall be able to prove those facts in Sections 4 and 6.
Now we turn our attention to the up-down arrows in the scheme which indicate that there exists the Hamiltonian-preserving symplectomorphism relating the $\E$-models $(LD_L,\omega_{LD_L},Q_{\E_L})$ and $(LD_R,\omega_{LD_R},Q_{\E_R})$. The existence of the non-trivial symplectomorphisms $LD_L\to LD_R$ pulling back the Hamiltonian $Q_{\E_R}$ onto the Hamiltonian $Q_{\E_L}$ is by no means an obvious thing, we shall nevertheless show in the present paper that such symplectomorphisms do exist for many choices of the Drinfeld doubles $\D_L,\D_R$. This fact leads to a substantial enlargement of the non-Abelian T-duality group as defined in [@LO] since the vertical arrow symplectomorphisms in the scheme makes possible to “travel” between the Poisson-Lie T-dualities based on the different subspaces $\E_L\subset \D_L$ and $\E_R\subset \D_R$. In particular, it follows from this that all four $\sigma$-models ,, and are pairwise T-dual to each other.
Our procedure to prove the pairwise T-duality of the models ,, and will be therefore as follows: given an affine (quasi-)Poisson group $K$, we find the linear operators $E^\na$, $E^{\ma}$, $\tilde E^\na$ and $\tilde E^{\ma}$ such that three things hold: 1) The $\sigma$-model is Poisson-Lie T-dual to the $\sigma$-model ; 2) the model is Poisson-Lie T-dual to ; 3) there exists the symplectic automorphism of the symplectic manifold $(T^*LK,\omega)$ which pull backs the Hamiltonian $\H_{\E_R}$ to the Hamiltonian $\H_{\E_L}$. Note in this respect, that from the existence of all horizontal symplectomorphisms and of the left vertical symplectomorphism in the scheme , the existence of all remaining vertical symplectomorphisms is automatically guaranteed. In particular, the right vertical symplectomorphisms exists and establishes the affine (quasi-)Poisson T-duality relating the $\sigma$-models and .
Affine Poisson groups
=====================
The theory of the affine Poisson groups was developed in [@DS; @Lu; @K07] and the reader is either invited to consult those references for finding proofs of some of the statements which we review in the present section or they may extract them as the special cases of the full-fledged proofs given in Section 5 in the more general case of the affine quasi-Poisson groups. We mention also that some parts of the theory (like e.g. the Poisson-Lie cohomology) are not included in this short review since they do not play an apparent role in the T-duality story.
Let $K$ be a Lie group equipped with a Poisson bracket $\{f_1,f_2\}^\na$ defined for any pair of smooth functions $f_1,f_2$ on $K$ : {f\_1,f\_2}\^df\_1df\_2, \^.The section $\pi^\na$ of the bundle $\Lambda^2TK$ is called the Poisson bivector corresponding to the Poisson bracket $\{.,.\}^\na$. The fact that the Poisson bracket $\{.,.\}^\na$ verifies the Jacobi identity is equivalent to the fact that the Schouten bracket of the bivector $\pi^\na$ with itself vanishes \_K=0.\[sch\] Recall that the Schouten bracket $[.,.]_\K$ of bivectors is defined by the linear extension of the bracket \_K:=\[u\_1,v\_1\]\_Ku\_2v\_2+u\_1\_Kv\_2- v\_1\_Ku\_2-v\_1u\_1\_K,\[247\]where $[u_1,v_1]_K$ stands for the standard Lie bracket of vector fields.
The Poisson structure $\pi^\na$ on the group manifold $K$ is called [*affine*]{}, if the Lie derivatives of $\pi^\na$ with respect to the right-invariant and left-invariant vector fields satisfy, respectively Ł\_[\^l\_[t\_i]{}]{}\^= - \^[L]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[pma\]Ł\_[\^r\_[t\_i]{}]{}\^= \^[R]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}.\[pmb\]Here $t_i$ is some basis of the Lie algebra $\K$ of the group $K$, while $\ ^{L}\tilde c_i^{\ jk}$ and $\ ^{R}\tilde c_i^{\ jk}$ are, respectively, the structure constants of two Lie algebras denoted as $\tilde K_L$ and $\tilde K_R$. The quantities $\nabla^l_{t_j}$ and $\nabla^r_{t_j}$ are, respectively, the right-invariant and the left-invariant vector fields on $K$ which take value $t_i$ et the unit element $e_K$ of the group $K$. Explicitely, the vector fields $\nabla^l_{t_j}$ and $\nabla^r_{t_j}$ act on the functions on $K$ as (\^l\_[t\_i]{}f)(k)\_[s=0]{},(\^r\_[t\_i]{}f)(k)\_[s=0]{}.The crucial role in the theory of the affine Poisson groups is played by the value of the bivector field $\pi^\na$ at the unit element $e_K$ of the group $K$. We denote this value as $m$: m:=\^(e\_K)m\^[ij]{}t\_it\_j. Remark that $m$ is naturally viewed as the element of $\Lambda^2\K$ where $\K$ is the Lie algebra of $K$. In the special case when $m$ vanishes, the affine Poisson group $(K,\pi^\na)$ is called the [*Poisson-Lie group*]{} and the finite-dimensional Lie algebras $\tilde\K_L$ and $\tilde\K_R^{opp}$ are isomorphic. If $m$ does not vanish, then $\pi^\na$ is not a Poisson-Lie structure on $K$ but it is true, however, that [*two different*]{} Poisson-Lie structures $\pi^L$, $\pi^R$ on the group $K$ can be naturally constructed out of the affine Poisson structure $\pi^\na$. They are given by[^2] the formulae \^L=L\_\*m-\^m\^[ij]{}\^r\_[t\_i]{}\^r\_[t\_j]{}-\^, \^R=\^-R\_\*m \^-m\^[ij]{}\^l\_[t\_i]{}\^l\_[t\_j]{},\[lr\]where $L_*m$ and $R_*m$ are, respectively, the left- and right-invariant bivector fields on $K$ obtained by the left and right transport of the element $m$ to the whole group manifold. They verify Ł\_[\^l\_[t\_i]{}]{}\^L= \^[L]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[266\]Ł\_[\^r\_[t\_i]{}]{}\^L= \^[L]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{};\[267\]Ł\_[\^l\_[t\_i]{}]{}\^R= \^[R]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[268\]Ł\_[\^r\_[t\_i]{}]{}\^R= \^[R]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}.\[269\]
Note also, that the Poisson bivector $\pi^{\ma}$ given by the formula \^:= R\_\*m+L\_\*m-\^,\[263\]defines another affine Poisson structure on $K$ called the “mirror” one, which satisfies Ł\_[\^l\_[t\_i]{}]{}\^= - \^[R]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[270\]Ł\_[\^r\_[t\_i]{}]{}\^= \^[L]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}.\[271\]The two Poisson-Lie structures associated to the mirror affine Poisson structure are the same as those associated to the original one but their roles are reversed, the left one becomes the right one and vice versa.
Drinfeld showed that each Poisson-Lie group $K$ of dimension $d$ can be embedded in a maximally isotropic way as a subgroup in a $2d$-dimensional Lie group $D$ called the Drinfeld double of $K$. The “maximally isotropic” means that there is a symmetric non-degenerate ad-invariant bilinear form $(.,.)_\D$ defined on the Lie algebra $\D$ of $D$ which identically vanishes when restricted to the Lie algebra $\K$ of the subgroup $K$. The structure of the Drinfeld double “remembers” the Poisson-Lie bivector on $K$, which can be extracted out of it from the way how certain another maximally isotropic $d$-dimensional subgroup $\tilde K$ of $D$ is embedded in $D$. The role of the groups $K$ and $\tilde K$ in the Poisson-Lie story is then interchangeable, which means that $\tilde K$ is also the Poisson-Lie group and the Poisson-Lie bivector on $\tilde K$ can be again extracted from the way how $K$ and $\tilde K$ are embedded in the Drinfeld double. In fact, any pair of maximally isotropic embeddings of two $d$-dimensional subgroups $H$ and $\tilde H$ in $D$ defines the mutually dual Poisson-Lie structures on $H$ and $\tilde H$, provided that $\D$ can be written as the direct sum of vector subspaces Lie($H)\oplus$Lie($\tilde H$).
It follows that one can associate to every affine Poisson group $(K,\pi^\na)$ two Drinfeld doubles $D_L$ and $D_R$, which are, respectively, the Drinfeld doubles of the Poisson-Lie groups $(K,\pi^L)$ and $(K,\pi^R)$. The double $D_L$ has naturally two maximally isotropic subgroups $K$ and $\tilde K_L$ , while $D_R$ has subgroups $K$ and $\tilde K_R$. The Lie groups $\tilde K_L$ and $\tilde K_R$ are called the left and the right dual groups of the affine Poisson group $K$ and their Lie algebras are isomorphic to the Lie algebras $\tilde\K_L$ and $\tilde\K_R$ with the structure constants $\ ^{L}\tilde c_i^{\ jk}$ and $\ ^{R}\tilde c_i^{\ jk}$.
For completeness, we detail how the left and right Poisson-Lie structures $\pi^L$ and $\pi^R$ as well as the mirror affine Poisson structure $\pi^{\ma}$ associated to the affine Poisson structure $\pi^\na$ are extracted from the structure of the Drinfeld doubles of the affine Poisson group $K$ [@K07]. We start from the perspective of the double $D_L$, we pick a basis $t_i, i=1,...,d$ of the Lie algebra $\K\subset\D_L$ and the dual basis $T^i_L, i=1,...,d$ of $\tilde\K_L\subset\D_L$ in such a way that (t\_i,T\^j\_L)\_[\_L]{}=\_i\^j.\[20\]We can then write \^[L]{}c\_i\^[ jk]{}=(\[T\_L\^j,T\_L\^k\],t\_i)\_[\_L]{}. The affine Poisson bracket $\{.,.\}^{\na}_K$ corresponding to the affine Poisson bivector $\pi^{\na}$ is then given by the formula[^3] {f\_1,f\_2}\_K\^(k)=m\^[ij]{}\^r\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2+(T\^i\_L,Ad\_kT\^j\_L)\_[\_L]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2;\[pba\] and the Poisson-Lie bracket $\{.,.\}^{L}_K$ corresponding to the Poisson-Lie bivector $\pi^{L}$ is given by the formula {f\_1,f\_2}\_K\^L(k)=-(T\^i\_L,Ad\_kT\^j\_L)\_[\_L]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2.\[plr\]Here $f_1,f_2$ are smooth functions on $K$ and the Einstein summation convention applies.
From the point of view of the double $D_R$, we pick a basis $t_i, i=1,...,d$ of the Lie algebra $\K\subset\D_R$ and the dual basis $T^i_R, i=1,...,d$ of $\tilde\K_R\subset\D_R$ in such a way that (t\_i,T\^j\_R)\_[\_R]{}=\_i\^j.\[308\]We can then write \^[R]{}c\_i\^[ jk]{}=(\[T\_R\^j,T\_R\^k\],t\_i)\_[\_R]{}.The mirror affine Poisson bracket $\{.,.\}^{\ma}_K$ corresponding to the mirror affine Poisson bivector $\pi^{\ma}$ is then given by the formula {f\_1,f\_2}\_K\^(k)=m\^[ij]{}\^r\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2+(T\^i\_R,Ad\_kT\^j\_R)\_[\_R]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2\[pbam\] and the Poisson bracket $\{.,.\}^{R}_K$ corresponding to the Poisson-Lie bivector $\pi^{R}$ is given by {f\_1,f\_2}\_K\^R(k)=-(T\^i\_R,Ad\_kT\^j\_R)\_[\_R]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2.\[314\]
We finish this section by proving the following useful relation {f\_1S,f\_2S}\_K\^={f\_1,f\_2}\_K\^S,\[331\] where $S:K\to K$ is defined as $S(k):=k^{-1}$. First of all, we show that for the Poisson-Lie structure $\pi^L$ given by the formula it holds {f\_1S,f\_2S}\_K\^L=-{f\_1,f\_2}\_K\^LS.\[334\] This follows from the repeated use of the following obvious identity \^l\_[t\_i]{}(fS)=-(\^r\_[t\_i]{}f)S,\[336\]indeed, we have $$\{f_1\circ S,f_2\circ S\}^L_K(k)=-(T^i_L,Ad_kT^j_L)_{\D_L}\nabla^l_{t_i}(f_1\circ S)\nabla^r_{t_j}(f_2\circ S)=$$= -(Ad\_[k\^[-1]{}]{}T\^i\_L,T\^j\_L)\_[\_L]{}((\^r\_[t\_i]{}f\_1)S)((\^l\_[t\_j]{}f\_2)S)= -{f\_1,f\_2}\_K\^L(k\^[-1]{}).Recall that we have $\pi^{\ma}=\pi^L+R_*m$ and $\pi^\na=-\pi^L+L_*m$, therefore for proving , we have to show that m\^[jk]{} \^l\_[t\_j]{}(f\_1S)\^l\_[t\_k]{}(f\_2S)=m\^[jk]{} (\^r\_[t\_j]{}f\_1)S (\^r\_[t\_k]{}f\_1)S.But this follows from Eq..
Affine Poisson T-duality
========================
The strategy of our presentation in the present section is as follows: we first assume that the T-duality between the models and indeed holds and then we infer from this assumption that all $\sigma$-models ,, and are pairwise T-dual to each other and fit into the scheme . In subsection 4.2, we formulate a sufficient condition to be fulfilled by the data $K,\Pi^{\na},E^{\na},E^{\ma}$ in order that the assumed duality between and really takes place, and, in Subsection 4.3, we construct a large class of examples of the affine Poisson groups and of the operators $E^{\na}$ and $E^{\ma}$ for which this sufficient condition is satisfied.
Pairwise T-duality of four $\sigma$-models
------------------------------------------
Let $\tilde K_L$ and $\tilde K_R$ be the dual groups of a given affine Poisson group $(K,\pi^\na)$ and let $D_L$ and $D_R$ be their respective Drinfeld doubles. Consider the $\sigma$-models and living on $K$: S\_(k)=d((E\^+\^(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_R]{},\[af1\]S\_(k)=d((E\^+\^[ ]{}(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_L]{}.\[af2\]where $E^{\na}:\tilde\K_R\to\K$ and $E^{\ma}:\tilde\K_L\to\K$ are some invertible linear operators, and $\Pi^{\na}(k):\tilde\K_R\to\K$ and $\Pi^{ \ma}(k):\tilde\K_L\to\K$ are, respectively, the affine Poisson structure and the mirror affine Poisson structure on $K$ in the sense of the formulae {f\_1,f\_2}\^\_K=( \^R\^l f\_1,\^(k) \^R\^l f\_2)\_[\_R]{},\[mor1\]{f\_1,f\_2}\^\_K=( \^L\^l f\_1,\^( k) \^L\^l f\_2)\_[\_L]{}.\[mor2\]Here $\ \!^R\nabla^l$ and $\ \!^L\nabla^l$ are, respectively, $\tilde\K_R$- and $\tilde\K_L$-valued differential operators acting on the functions on $K$ as ( \^[L]{}\^l f, x)\_[\_L]{}:= \^l\_[x]{} f,( \^[R]{}\^l f, x)\_[\_R]{}:= \^l\_[x]{} f, x.\[sab\]
[**Assertion 1**]{}: [*If the $\sigma$-model is dual to the $\sigma$-model for some choice of the linear operators*]{} $E^\na$ [*and*]{} $E^{\ma}$, [*then there exist operators*]{} $\tilde E^{\na}$ [*and*]{} $\tilde E^{\ma}$ [*such that all $\sigma$-models ,, and are pairwise T-dual to each other.*]{}
Let us prove the [**Assertion 1**]{}. We set M\^R=\^(e\_K), M\^L=\^(e\_K)\[mlr’\] and rewrite the pair of the actions , equivalently as S\_(k)=d((E\^+M\^R+\^R(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_R]{},\[eadl\]S\_(k)=d((E\^+M\^L+\^L(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_L]{},\[eadr\]where $\Pi^L(k):\tilde\K_L\to\K$ and $\Pi^R(k):\tilde\K_R\to\K$ stand, respectively, for the left and right Poisson-Lie structures on $K$, associated to the affine Poisson structure in the sense of the formulae {f\_1,f\_2}\^L\_K=( \^L\^l f\_1,\^L(k) \^L\^lf\_2)\_[\_L]{},\[mor11\]{f\_1,f\_2}\^[R]{}\_K=( \^R\^l f\_1,\^R( k) \^R\^l f\_2)\_[\_R]{}.\[mor22\]The crux of the proof of the [**Assertion 1**]{} is now obvious, because the model rewritten as is of the type , hence it is Poisson-Lie T-dualizable. Similarly, the model rewritten as is also of the type , hence it is also Poisson-Lie T-dualizable. The standard Poisson-Lie T-duality reviewed in Section 2 then gives, respectively, the dual models of and of : S\_(k\_R)=d((E\^+\^R(k\_R))\^[-1]{}\_+k\_Rk\_R\^[-1]{}, \_-k\_Rk\_R\^[-1]{})\_[\_R]{},\[ear”\]S\_(k\_L)=d((E\^+\^L(k\_L))\^[-1]{}\_+k\_Lk\_L\^[-1]{}, \_-k\_Lk\_L\^[-1]{})\_[\_L]{}.\[eal”\] Here $\tilde k_L,\tilde k_R$ are, respectively, $\tilde K_L$,$\tilde K_R$-valued $\sigma$-model fields, $\tilde\Pi^L(\tilde k_L)$, $\tilde\Pi^R(\tilde k_R)$ are the Poisson-Lie structures on $\tilde K_L,\tilde K_R$ dual to the Poisson-Lie structures $\Pi^L(k)$, $\Pi^R(k)$ on $K$, and $\tilde E^\na$, $\tilde E^{\ma}$ are the inverse operators to $E^\na+M^R$ and to $E^{\ma}+M^L$. The models and coincide with the models and , therefore we conclude that all $\sigma$-models ,, and are pairwise T-dual to each other. We note, moreover, that the subspaces $\E_R$ and $\E_L$ featuring in the scheme are, respectively $\E_L=\{\tilde x +(E^{\ma}+M^L)\tilde x,\tilde x\in\tilde\K_L\}$ and $\E_R=\{\tilde x +(E^{\na}+M^R)\tilde x,\tilde x\in\tilde\K_R\}$.
Sufficient condition for the affine Poisson T-duality
-----------------------------------------------------
So far we have established, that all $\sigma$-models ,, and are pairwise T-dual to each other if we succeed to associate a symplectomorphism to the left vertical arrow in the scheme , or, in other words, if we prove that the models and are T-dual to each other. The principal result of the present subsection is the [**Assertion 2**]{} which states the sufficient condition on the data $K,\Pi^\na,E^\na,E^{\ma}$ guaranteeing the existence of the seeken left vertical symplectomorphism. In order to formulate this condition, we need to define two bilinear forms $(.,.)_\na$ and $(.,.)_{\ma}$ on the Lie algebra $\K$: (x,y)\_:=(x,(E\^)\^[-1]{}y)\_[\_R]{},(x,y)\_:=(x,(E\^)\^[-1]{}y)\_[\_L]{},x,y.\[lem\]
[**Assertion 2**]{}: [*If the bilinear form*]{} $(.,.)_\na$ [*is symmetric, non-degenerate and ad$_\K$-invariant and, moreover, if it coincides with the bilinear form*]{} $(.,.)_{\ma}$ [*then the $\sigma$-model is T-dual to the $\sigma$-model .* ]{}
In order to prove [**Assertion 2**]{}, we first remark that it holds E\^ \^R\^l=E\^ \^L\^l= \^\^l,\[plk\] where the operators $\ \!^R\nabla^l$ and $\ \!^L\nabla^l$ where defined in and we define a $\K$-valued differential operator $\ \!^\na{\bm\nabla}^l$ acting on the functions on $K$ as ( \^\^l f, x)\_:= \^l\_[x]{} f, x.\[bfs\] To see e.g. that the first of the relations indeed holds, we rewrite the left-hand-side of the second of Eqs. as ( \^\^l f, x)\_:= \^l\_[x]{} f = ( \^[R]{}\^l f, x)\_[\_R]{}=([E\^]{}\^[-1]{}E\^ \^[R]{}\^l f, x)\_[\_R]{}= (E\^ \^[R]{}\^l f, x)\_ and we finish up the argument by invoking the non-degeneracy of the bilinear forms $(.,.)_{\D_R}$ and $(.,.)_\K$.
Our next goal is to prove the following relation \^(k)= Ad\_k\^(k\^[-1]{})([E\^]{})\^[-1]{}Ad\_[k\^[-1]{}]{}E\^,\[cru\] where the operators $\Pi^{\ma}(k)$ and $\Pi^{\na}(k)$ where defined respectively in Eqs. and . To prove , we employ the identity . By using the relations , we can rewrite the affine Poisson brackets and as {f\_1,f\_2}\^\_K=([E\^]{}\^[-1]{}E\^ \^R\^l f\_1,\^(k)[E\^]{}\^[-1]{}E\^ \^R\^l f\_2)\_[\_R]{}= ( \^\^l f\_1,\^(k)[E\^]{}\^[-1]{} \^\^l f\_2)\_, \[blb1\] and, similarly, {f\_1,f\_2}\^\_K= ( \^\^l f\_1,\^(k)[E\^]{}\^[-1]{} \^\^l f\_2)\_.\[blb2\] We obtain the desired identity from the relations , , and also from the following identity Ad\_k( \^\^rf(k))= \^\^lf(k). Let us rewrite the $\sigma$-model actions and as follows $$S_\na(k)= \jp\int d\tau \oint \biggl({E^\na}^{-1}\Bigl(1+\Pi^\na(k){E^\na}^{-1}\Bigr)^{-1}\partial_+k k^{-1}, \partial_- k k^{-1}\biggr)_{\D_R} =$$=d((1+\^(k)[E\^]{}\^[-1]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_\[e1\] $$S_{\ma}(k)= \jp\int d\tau \oint \biggl({E^{\ma}}^{-1}\Bigl(1+\Pi^{\ma}(k){E^{\ma}}^{-1}\Bigr)^{-1}\partial_+k k^{-1}, \partial_- k k^{-1}\biggr)_{\D_L} =$$=d((1+\^(k)[E\^]{}\^[-1]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_.\[e2\] Using the crucial identity as well as the hypothesis that the ad$_\K$-invariant bilinear forms $(.,.)_\na$ and $(.,.)_{\ma}$ coincide, we find finally S\_(k)=S\_(k\^[-1]{}).\[ama\] The equation may look surprising because it proves the T-duality between the $\sigma$-models and in the simplest possible way; indeed, the model becomes the model by the simple field redefinition $k\to k^{-1}$ and the corresponding T-duality symplectomorphism is therefore just the so called “point canonical transformation”. In the standard T-duality story, the point canonical transformation is normally not considered to be an interesting one since it establishes the dynamical equivalence of two $\sigma$-models just by field redefinitions, so what is the point in discovering that the left vertical symplectomorphism in the scheme relating the models and is just the point canonical transformation? In fact, the crux of the affine Poisson T-duality is the [*nontriviality of the right vertical symplectomorphism*]{} relating the $\sigma$-models and . This right vertical symplectomorphism is not the point canonical transformation since it is the composition of the upper horizontal, left vertical and lower horizontal symplectomorphisms appearing in the scheme and the horizontal symplectomorphisms, being the Poisson-Lie T-dualities, are not the point canonical transformations.
Summarizing: if the sufficient condition formulated in [**Assertion 2**]{} holds then there is the nontrivial T-duality transformation relating the $\sigma$-models living on the dual groups $\tilde K_L$ and $\tilde K_R$ of the given affine Poisson group $K$: S\_(k\_R)=d((E\^+\^R(k\_R))\^[-1]{}\_+k\_Rk\_R\^[-1]{}, \_-k\_Rk\_R\^[-1]{})\_[\_R]{}.\[eal’\] S\_(k\_L)=d((E\^+\^L(k\_L))\^[-1]{}\_+k\_Lk\_L\^[-1]{}, \_-k\_Lk\_L\^[-1]{})\_[\_L]{}.\[ear’\]
This is the principal statement of the affine Poisson T-duality story.
Examples of the affine Poisson T-duality
----------------------------------------
If an affine Poisson group $(K,\pi^\na)$ together with some operators $E^\na,E^{\ma}$ satisfy the sufficient condition of [**Assertion 2**]{} then the Lie algebra $\K$ is quadratic, which means that there is a symmetric ad$_\K$-invariant non-degenerate bilinear form $(.,.)_\K$ on $\K$. Moreover, we restrict our attention to such affine Poisson structures $\pi^\na$ on the quadratic[^4] group $K$ for which the associated Drinfeld doubles $D_L$ and $D_R$ are isomorphic as the Lie groups, and both Lie algebras $\D_L$ and $\D_R$ have the structure of the Lie algebra $\D_\e$ defined as follows: the elements of $\D_\e$ are pairs $(x_1,x_2)$, $x_1,x_2\in\K$ and the Lie algebra commutator is given by the expression \_=(\[x\_1,y\_1\]+,\[x\_1,y\_2\]+\[x\_2,y\_1\]).\[str\]The symmetric non-degenerate ad-invariant bilinear form $(.,.)_{\D_\e}$ on $\D_\e$ does not depend on the real negative parameter $\e$ and it is given by the formula ((x\_1,x\_2),(y\_1,y\_2))\_[\_]{}:= (x\_2,y\_1)\_+(x\_1,y\_2)\_.\[bil\]The Lie algebra $\K$ is embedded into $\D_\e$ in the maximally isotropic way as $(\K,0)\subset \D_\e$.
As we already know, the affine Poisson structure $\pi^\na$ induces two maximally isotropic Lie subalgebras $\tilde\K_L$, $\tilde\K_R$ of the respective doubles $\D_L$ and $\D_R$. Since we consider the case $\D_L=\D_R=\D_\e$, the both Lie algebras $\tilde\K_L$ and $\tilde\K_R$ must be maximally isotropic subalgebras of $\D_\e$. Recall that $\D_\e$ as the vector space (but not necessarily as the Lie algebra) can be written in terms of the direct sums \_=\_L=\_R.\[dso\]This fact as well as the structure of the Lie bracket on the double $\D_\e$ entail the existence of two linear operators $R^r, R^l:\K\to\K$ such that the Lie subalgebras $\tilde\K_L,\tilde K_R\subset \D_\e$ can be written as the following graphs \_L ={(-R\^lx,x),x},\_R ={(-R\^rx,x),x}.\[max\] Moreover, the fact that the restrictions of the bilinear form $(.,.)_\D$ on the both $\tilde\K_L$ and $\tilde K_R$ must vanish implies the skew-symmetry of the both operators $R^r,R^l$: (R\^rx,y)\_=-(x,R\^ry)\_,(R\^lx,y)\_=-(x,R\^ly)\_, x,y. Finally, the fact that $\tilde\K_L$ and $\tilde K_R$ are the Lie subalgebras of $\D_\e$ sets the following constraints on the operators $R^r$ and $R^l$: =R\^r(\[R\^rx,y\]+\[x,R\^ry\])-, x,y, \[YBr\] =R\^l(\[R\^lx,y\]+\[x,R\^ly\])-, x,y, \[YBl\] The constraints and are known under the name of the Yang-Baxter equations.
Actually, the knowledge of two skew-symmetric solutions $R^r$ and $R^l$ of the Yang-Baxter equation is all what is needed for reconstructing the affine Poisson structure $\pi^\na$ on the quadratic Lie group $K$. The affine Poisson bracket $\{.,.\}^\na$, its mirror $\{.,.\}^{\ma}$ and the associated left and write Poisson-Lie brackets $\{.,.\}^L$, $\{.,.\}^R$ can be easily extracted from Eqs. , and and they are given by the formulae {f\_1,f\_2}\^\_K:=(\^r f\_1,R\^r\^r f\_2)\_+(\^l f\_1,R\^l\^l f\_2)\_.\[aff\]{f\_1,f\_2}\^\_K:=(\^r f\_1,R\^l\^r f\_2)\_+(\^l f\_1,R\^r\^l f\_2)\_,\[afg\]{f\_1,f\_2}\^L\_K:=(\^r f\_1,R\^l\^r f\_2)\_-(\^l f\_1,R\^l\^l f\_2)\_,\[l\]{f\_1,f\_2}\^[R]{}\_K:=(\^r f\_1,R\^r\^r f\_2)\_-(\^l f\_1,R\^r\^l f\_2)\_.\[r\]Here we have defined the $\K$-valued differential operators ${\bm\nabla}^r, {\bm\nabla}^l$ acting on the functions on $K$ as (\^r f, x)\_:= \^r\_[x]{} f (\^l f, x)\_:= \^l\_[x]{} f , x.\[bft’\]Working with the affine Poisson structure given by Eq. , can we find the operators $E^\na:\tilde\K_R\to\K$ and $E^{\ma}:\tilde\K_L\to\K$ which would fulfil the sufficient condition of [**Assertion 2**]{}? The answer to this question is affirmative; in the two cases, it is given by: E\^(-R\^rx,x):=a(x,0), E\^(-R\^lx,x):=a(x,0), a<0.\[ho\]Using the definition , it is then easy to verify for every $x,y\in \K$ that it holds (x,y)\_((x,0),(E\^)\^[-1]{}(y,0))\_=((x,0),(-R\^ry,y))\_=(x,y)\_,(x,y)\_((x,0),(E\^)\^[-1]{}(y,0))\_=((x,0),(-R\^ly,y))\_=(x,y)\_, hence the sufficient condition for the affine Poisson T-duality is indeed satisfied.
It is instructive to cast the $\sigma$-model actions and in terms of the Yang-Baxter operators $R^l,R^r$. For that, we combine the formulae , with , to find \^(k)([E\^]{})\^[-1]{}=R\^l+R\^r\_[k\^[-1]{}]{}, \^(k)([E\^]{})\^[-1]{}=R\^r+R\^l\_[k\^[-1]{}]{},\[ger\]where R\^l\_[k\^[-1]{}]{}:=Ad\_k R\^l Ad\_[k\^[-1]{}]{}, R\^r\_[k\^[-1]{}]{}:=Ad\_k R\^r Ad\_[k\^[-1]{}]{}. Using then the formulae , and , we infer S\_(k)=d((a+ R\^l+R\^r\_[k\^[-1]{}]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_,\[kad\]S\_(k)=d((a+R\^r+R\^l\_[k\^[-1]{}]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_,\[kbd\]There is no universal way to rewrite the actions of the dual $\sigma$-models $\tilde S_\na(\tilde k_R)$ and $\tilde S_{\ma}(\tilde k_L)$ in terms of the operators $R^r,R^l$ since the very structure of the groups $\tilde K_L,\tilde K_R$ depends implicitely on $R^r,R^l$. On the other hand, the operators $R^r$, $R^l$ are very useful if we wish to describe explicitely the subspaces $\E_L$, $\E_R$ underlying, respectively, the Poisson-Lie T-dualities relating the models with and with . We find with the help of the formulae , and that M\^R(E\^)\^[-1]{}= M\^L(E\^)\^[-1]{}=R\^l+R\^r,\[mlr\]hence \_R={x\_R +(E\^+M\^R)x\_R,x\_R\_R}={(E\^)\^[-1]{}x +(E\^+M\^R)(E\^)\^[-1]{}x , x}={(ay+R\^ly,y),y},\[ell\] \_L={x\_L +(E\^+M\^L)x\_L,x\_L\_L}={(E\^)\^[-1]{}x +(E\^+M\^L)(E\^)\^[-1]{}x , x}={(ay+R\^ry,y),x}.\[err\]Note that as far as $R^r\neq R^l$, the subspaces $\E_L$ and $\E_R$ do not coincide and the scheme hence represents the generalization of the Poisson-Lie T-duality (recall that the Poisson-Lie T-duality is based on the unique subspace). In the special case when $R^l=R^r$, the dual groups $\tilde K^L$ and $\tilde K^R$ coincide as well as do the subspaces $\E_L$ and $\E_R$. This means that the $\sigma$-models and coincide too and the upper line of the affine Poisson T-duality scheme merges in this case with the lower line to become the Poisson-Lie T-duality scheme . In conclusion, if $R^r=R^l$, the affine Poisson T-duality becomes the standard Poisson-Lie T-duality.
The affine Poisson T-duality and the Drinfeld twist operators
-------------------------------------------------------------
In this subsection, we choose for $\K$ a compact simple Lie algebra equipped with its Killing-Cartan form $(.,.)_\K$ and for its Drinfeld double we take $\D_\e$ for $\e=0$ (cf. Eq.). We now pick a Cartan subalgebra $\gH\in\K$ and we consider the subspace $\gH^\perp\subset \K$ which is perpendicular to $\gH$ with respect to the Killing-Cartan form $(.,.)_\K$. We shall call any skew-symmetric operator $R:\K\to\K$ the Drinfeld twist operator, if $\gH^\perp\subset {\rm Ker}(R)$ and Im$(R)\subset\gH$. Any Drinfeld twist operator verifies automaticallly the Yang-Baxter condition because of the commutativity of the Cartan subalgebra. In the case when the affine Poisson structure is given by the Drinfeld twist operators $R^l,R^r$, the $\sigma$-models and dual to the models and can be written more explicitly and we present here the corresponding formulae.
We start with the description of the Lie group $D_0$ integrating the Lie algebra $\D_0$ (the commutator of $\D_0$ is given by Eq. for $\e=0$). The element of $D_0$ are the pairs $(k,\kappa)$, where $k\in K$ and $\kappa\in\K$, the group multiplication in $D_0$ is given by (k\_1,\_1)(k\_2,\_2)=(k\_1k\_2,\_1+Ad\_[k\_1]{}\_2)\[gl\]and the inverse element by (k,)\^[-1]{}=(k\^[-1]{},-Ad\_[k\^[-1]{}]{}).\[ind\]In what follows, we shall moreover need explicit expressions for the Maurer-Cartan forms on $D_0$ as well as those for the adjoint action of the group $D_0$ on the Lie algebra $\D_0$. We have, respectively, for the left- and right-invariant forms (k,)\^[-1]{}d(k,)=(k\^[-1]{}dk,Ad\_[k\^[-1]{}]{}( d)),d(k,)(k,)\^[-1]{}=(dkk\^[-1]{},d+\[,dkk\^[-1]{}\])\[dif\] and, for the adjoint action Ad\_[(k,)]{}(x\_1,x\_2)= (Ad\_kx\_1,Ad\_kx\_2+\[,Ad\_kx\_1\]).\[adj\]The elements of the subgroup $K$ of $D_0$ have the form $(k,0)$ and the elements of the subgroups $\tilde K_L$ and $\tilde K_R$ have, respectively, the form K\_L={ (e\^[-R\^l]{},),},K\_R={ (e\^[-R\^r]{},),}.\[dgs\]The reader may verify by direct computations, that the definitions yield the Poisson brackets , , and via Eqs. , , and .
In order to write down explicitly the actions and, we first represent them in the following form S\_=dd((1+(E\^+M\^R)\^R(k\_R))\^[-1]{}(E\^+M\^R)\_+k\_Rk\_R\^[-1]{}, \_-k\_Rk\_R\^[-1]{})\_[\_R]{}.\[ealb\] S\_=dd((1+(E\^+M\^L)\^L(k\_L))\^[-1]{}(E\^+M\^L)\_+k\_Lk\_L\^[-1]{}, \_-k\_Lk\_L\^[-1]{})\_[\_L]{}.\[earb\] We find from the formulae and \_k\_L k\_L\^[-1]{}=(-R\^l \_, \_-\[,R\^l \_\]),\_k\_R k\_R\^[-1]{}=(-R\^r \_, \_-\[,R\^r \_\])\[605\]and from the formulae and , we infer (E\^+M\^R)\_k\_Rk\_R\^[-1]{}=((a+R\^l+R\^r)(\_-\[,R\^r \_\]),0);(E\^+M\^L)\_k\_Lk\_L\^[-1]{}=((a+R\^l+R\^r)(\_-\[,R\^l \_\]),0).Thus, taking into account the formula , we can rewrite Eqs. and in terms of the bilinear form $(.,.)_\K$. S\_=dd((1+(E\^+M\^R)\^R(k\_R))\^[-1]{}(a+R\^l+R\^r)(\_+-\[,R\^r \_+\]), \_--\[,R\^r \_-\])\_.\[ealb’\] S\_=dd((1+(E\^+M\^L)\^L(k\_L))\^[-1]{}(a+R\^l+R\^r)(\_+-\[,R\^l \_+\]), \_--\[,R\^l \_-\])\_.\[earb’\] It remains to determine the operators $(E^\na+M^R)\tilde\Pi^R(\tilde k_R)$ and $(E^{\ma}+M^L)\tilde\Pi^L(\tilde k_L)$, which are both the endomorphisms of the vector space $\K$. We deduce from the formulae and , that {f\_1,f\_2}\_[K\_L]{}(k\_L)=(\^l f\_1,T\^i\_L)\_[\_0]{}(t\_i,\^L(k\_L)t\_j)\_[\_0]{}(T\^L\_j,\^l f\_2)\_[\_0]{}=(t\_i,\^L(k\_L)t\_j)\_[\_0]{} \^l\_[T\^i\_L]{} f\_1 \^l\_[T\^j\_L]{} f\_2;\[10l\]{f\_1,f\_2}\_[K\_R]{}(k\_R)=(\^r f\_1,T\^i\_R)\_[\_0]{}(t\_i,\^R(k\_R)t\_j)\_[\_0]{}(T\^R\_j,\^l f\_2)\_[\_0]{}= (t\_i,\^R(k\_R)t\_j)\_[\_0]{} \^l\_[T\^i\_R]{} f\_1 \^l\_[T\^j\_R]{} f\_2,\[10r\]where $t_i\in\K$ is the orthonormal basis on the compact simple Lie algebra $\K$ verifying (t\_i,t\_j)\_=-\_[ij]{},\[nr\] and the basis $T^i_L\in\tilde\K_L$ and $T^i_R\in\tilde\K_R$ were introduced in and in . On the other hand, from Eqs., we infer {f\_1,f\_2}\_[K\_L]{}(k\_L)=-(t\_i,Ad\_[k\_L]{}t\_m)\_[\_0]{}(Ad\_[k\_L]{}T\_L\^m,t\_j)\_[\_0]{}\^l\_[T\^i\_L]{} f\_1\^l\_[T\^j\_L]{} f\_2;\[11l\]{f\_1,f\_2}\_[K\_R]{}(k\_R)=-(t\_i,Ad\_[k\_R]{}t\_m)\_[\_0]{}(Ad\_[k\_R]{}T\_R\^m,t\_j)\_[\_0]{}\^l\_[T\^i\_R]{} f\_1\^l\_[T\^j\_R]{} f\_2;\[11r\]Combining Eqs., with Eqs., , we find (t\_i,\^L(k\_L)t\_j)\_[\_0]{} =-(t\_i,Ad\_[k\_L]{}t\_m)\_[\_0]{}(Ad\_[k\_L]{}T\_L\^m,t\_j)\_[\_0]{},\[12l\](t\_i,\^R(k\_R)t\_j)\_[\_0]{}=-(t\_i,Ad\_[k\_R]{}t\_m)\_[\_0]{}(Ad\_[k\_R]{}T\_R\^m,t\_j)\_[\_0]{}.\[12r\]We can rewrite the formulae , without resorting to the choices of the basis by writing \^L(k\_L)=-\_LAd\_[k\_L]{}\_L Ad\_[k\^[-1]{}\_L]{}\_L;\[13l\]\^R(k\_R)=-\_RAd\_[k\_R]{}\_R Ad\_[k\^[-1]{}\_R]{}\_R.\[13r\]Here all the operators $\J_L,\J_R,\tilde \J_L,\tilde \J_R:\D_0\to\D_0$ are projectors; $\J_L$ projects on $\K$ with the kernel $\tilde\K_L$, $\J_R$ projects on $\K$ with the kernel $\tilde\K_R$, $\tilde \J_L$ projects on $\tilde\K_L$ with the kernel $\K$ and $\tilde \J_R$ projects on $\tilde\K_R$ with the kernel $\K$. With the choice of the basis $t_i,T_L^i,T_R^i$ as before we have, in particular \_L A=(A,T\^i\_L)\_[\_0]{} t\_i,\_R A=(A,T\^i\_R)\_[\_0]{} t\_i, \_L A= (A,t\_i)\_[\_0]{} T\^i\_L,\_R A= (A,t\_i)\_[\_0]{} T\^i\_R,A\_0.Thus, for the subalgebras \_L={(-R\^l,),},\_R={ (-R\^r,),}\[dga\]we find \_L(x,y)=(x+R\^ly,0), \_R(x,y)=(x+R\^ry,0), \_L (x,y)= (-R\^ly,y), \_R (x,y)= (-R\^ry,y), (x,y)\_0.\[pro\]Using the formulae , , , , and , we find \^L()x=-(-R\^l(\[,x\]-\[,R\^l\[,x\]\]),\[,x\]-\[,R\^l\[,x\]\] ),x;\^R()x=-(-R\^r(\[,x\]-\[,R\^r\[,x\]\]),\[,x\]-\[,R\^r\[,x\]\] ),x.Furthermore, from the formulae and we infer (E\^+M\^R)\^R()x=-(a+R\^l+R\^r)(\[,x\]-\[,R\^r\[,x\]\]); (E\^+M\^L)\^L()x= -(a+R\^l+R\^r)(\[,x\]-\[,R\^l\[,x\]\]).We can now finally rewrite the formulae and in the final form S\_=d(((a+R\^l+R\^r)\^[-1]{}-O\^r() [ad]{}\_)\^[-1]{}O\^r()\_+, O\^r()\_-)\_;\[efia\] S\_=d(((a+R\^l+R\^r)\^[-1]{}-O\^l() [ad]{}\_)\^[-1]{}O\^l()\_+, O\^l()\_-)\_,\[efib\] where the operators $O^{l,r}(\kappa):\K\to\K$ are defined as O\^l():= 1 - [ad]{}\_R\^l, O\^r():= 1 - [ad]{}\_R\^r. One of the important results of the present article is the statement that the $\sigma$-models and are dual to each other and the T-duality which relies them is not the Poisson-Lie one but the more general affine Poisson T-duality. Of course, if we want that this new duality be really interesting, we have to show that the $\sigma$-models and cannot be rendered equivalent by field redefinitions, or, equivalently, by a point canonical transformation. To show that, we start the argument by noting that when $R^r=R^l=R$ (which corresponds to the special case when affine Poisson T-duality becomes the Poisson-Lie T-duality) then the $\sigma$-models and obviously coincide. The whole question is what happens if we get out slightly of this special case and consider a deformation $R^l=R+\delta R$ and $R^r=R-\delta R$. Do the deformed models and remain the same up to field redefinitions? If yes, this would mean that there exists a vector field $\B$ on the target which would verify the following condition Ł\_\^[+]{}=\^+-\^-,\[652\] where $\T^\pm$ is defined as the following section of the tensor product of the tangent bundle of the target with itself: \^:=(((a+2R)\^[-1]{}-(1 - [ad]{}\_(RR))\_)) \^[RR]{}\^l, \^[RR]{}\^l)\_.\[655\] Here the $\K$-valued vector fields $\ ^{R\pm\delta R}\tilde\nabla^l$ are dual to the right-invariant Maurer-Cartan forms $\left(1 - {\rm ad}_\kappa\circ (R\pm\delta R)\right)d\kappa$. Explicitely, we have in the orthonormal basis $t_i$ (cf. Eq.) \^[RR]{}\^l t\_i \^[RR]{}\^l\_[t\_i]{},\[656\] where \^[RR]{}\^l\_[t\_i]{}=(\_[pi]{}+(\[(RR)t\_i,\],t\_p)\_) \_[\_p]{}\[658\] and the coordinates $\kappa_p$ on the target space are defined by the decomposition =\_p t\_p. Note that the geometric quantities $\T^\pm$ are dual with respect to those appearing in the Lagrangian and we have chosen them in order to get rid of the uncomfortable inverse.
Let us rewrite the formulae , and more invariantly. For that, we define a $\K$-valued partial derivative $\d_\kappa$ as \_=t\_p\_[\_p]{}. Then \^[RR]{}\^l=\_+(RR)[ad]{}\_\_\[666\] and \^:=( ((a+2R)\^[-1]{}(1+(RR)\_)-\_),(1+(RR)\_))\_. \[670\] We find also up to first order $$\T^+-\T^-=$$=( ((a+2R)\^[-1]{}(1+ R \_)-\_), 2(R)\_)\_+ ( ((a+2R)\^[-1]{}2(R)\_) ),(1+ R \_))\_. \[671\] We shall now look for the vector field $\B$ satisfying the condition in the form =(B(),)\_,where $B(\kappa)$ is a $\K$-valued function on the target. By counting the powers of the variable $\kappa$ on both sides of Eq., we find that $B(\kappa)$ must be at most linear in $\kappa$ and it must be also homogeneous because for $\kappa=0$ the quantities $\T^+$ and $\T^-$ coincide. This means B()=B, where $B:\K\to\K$ is a linear operator. If we now work out the condition and set $\kappa=0$, we obtain ((a+2R)\^[-1]{}B\^\*\_,\_)\_+ ((a+2R)\^[-1]{}\_,B\^\*\_)\_=0,\[679\] where $B^*$ is adjoint to $B$ with respect to the bilinear form $(.,.)_\K$. From Eq. we then deduce (a+2R)B+B\^\*(a+2R)=0,\[683\] and, by taking the adjoint of this relation, also (a-2R)B+B\^\*(a-2R)=0.\[684\] By adding as well as substracting Eqs. and , we infer that $B$ is anti-Hermitian and it commutes with $R$: B\^\*=-B,=0.\[688\] We use the relation for writing the part of the condition linear in $\kappa$ as follows ((\_)\_B\_+(\_)\_B \_,\_)\_= ((R)\_\_+\_(R)\_,\_)\_,\[690\] where (\_)\_B\_B -B \_+\_[(B)]{} .We now rewrite the condition as ([2a+4R]{})(a(\_)\_B -4(R)\_) +(a(\_)\_B-4\_(R)) ([2a+4R]{})=0\[695\] and the conjugated one as ([2a-4R]{})(-a(\_)\_B -4(R)\_) +(-a(\_)\_B-4\_(R)) ([2a-4R]{})=0,\[698\] Adding and subtracting and , we find -(R)\_-\_(R)+R(\_)\_B+(\_)\_B R=0a\^2(\_)\_B - 4 R(R)\_-4\_(R)R=0.Inserting the second relation into the first one we obtain -a\^2(R)\_-a\^2\_(R) +4R(R)\_R+4R\^2(R)\_+4R \_(R)R+4\_(R)R\^2=0.This relation must be true for all $a$ which means that $\delta R$ must anticommute with the adjoint action of every element of $\K$ which is impossible. This implies that the vector field $\B$ does not exist and the actions $\tilde S_\na$ and $\tilde S_{\ma}$ given by Eqs. and cannot be made equivalent by field redefinitions.
Affine quasi-Poisson groups
===========================
Recall from Ref. [@AK] that the quasi-Poisson $\K$-space is a manifold $(M,\pi)$ on which acts a Lie quasi-bialgebra $\K$ in a way compatible with the quasi-Poisson bivector $\pi$. Recall that the Lie quasi-bialgebra $\K$ is an ordinary Lie algebra $(\K,[.,.])$ supplied with additional anti-symmetric bracket $[.,.]^*$ on the dual space $\K^*$ and equipped also with a completely antisymmetric trilinear form $\chi:\Lambda^3\K^*\to\br$. The structures $[.,.]$, $[.,.]^*$ and $\chi$ must be compatible in the sense that the direct sum of the vector spaces $\D^q:=\K\oplus\K^*$ has to be ordinary Lie algebra equipped with the following commutator \_[\^q]{}=(\[x,y\]+x\^\*-y\^\*+(,,.))(\[,\]\^\*-+).\[qdc\]Here e.g. the expression $x\circ[\beta,.]^*$ has to be interpreted as the element of $\K$ which acts on the elements of $\K^*$ as x\^\*,:=x,\[,\]\^\*,\^\*.We notice, that if the trilinear form $\chi$ vanishes then the Lie quasi-bialgebra $\K$ is just the standard Lie bialgebra, that is the bracket $[.,.]^*$ gives the Lie commutator on the dual space $\K^*$ and the commutator is the one of the standard Drinfeld double $\D$ of $\K$. If the quantity $\chi$ does not vanish, the anti-symmetric bracket $[.,.]^*$ may be but need not be a Lie commutator on $\K^*$, nevertheless the bracket on $\D^q:=\K\oplus\K^*$ is always an honest Lie commutator. We shall refer to $\D^q$ as to the quasi-Drinfeld double of the Lie quasi-bialgebra.
We shall see concrete examples of the Lie quasi-bialgebras in the next section, here we continue for the moment the general exposure. An action $\rho$ of the Lie quasi-bialgebra $\K$ on the manifold $(M,\pi)$ is called quasi-Poisson, if it holds Ł\_[(x)]{}=-(f(x));\[qp1\] \_M=().\[qp2\] Here $\L_{\rho(x)}$ stands for the Lie derivative, $\tilde f:\K\to\K\wedge\K$ is the map dual to the bracket $[.,.]^*:\K^*\wedge\K^*\to\K^*$ and $[.,.]_M$ is the Schouten bracket on the manifold $M$. We recall that, for decomposable bivectors, the Schouten bracket is defined as \_M:=\[u\_1,v\_1\]\_Mu\_2v\_2+u\_1\_Mv\_2- v\_1\_Mu\_2-v\_1u\_1\[bsc\]and the general bracket is obtained by linearity. Of course, $[u_1,v_1]_M$ stands for the Lie bracket of vector fields.
It may be illuminating to write the conditions , for the quasi-Poisson action in some basis $t_i$ of the Lie quasi-bialgebra $\K$. We have Ł\_[(t\_i)]{}=-f\_i\^[ jk]{}(t\_j)(t\_k);\[qp3\] \_M=\^[ijk]{}(t\_i)(t\_j)(t\_k).\[qp4\]
Let us now define the affine quasi-Poisson group $K$ as the Lie group equipped with a bivector $\pi^\qa$, on which the standard left and right group multiplications induce the quasi-Poisson actions (possibly with respect to two non-isomorphic Lie quasi-bialgebras) in the sense of the conditions and . The left action $\rho_L$ and the right one $\rho_R$ are obviously described by the invariant vector fields \_L(t\_i)=\^l\_[t\_i]{},\_R(t\_i)=-\^r\_[t\_i]{},\[ga7\]so that we require in full analogy with and Ł\_[\^l\_[t\_i]{}]{}\^= - \^[L]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[pmc\]Ł\_[\^r\_[t\_i]{}]{}\^= \^[R]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}.\[pmd\]Here $t_i$ is some basis of the Lie algebra $\K$ of the group $K$, $\ ^{L}\tilde c_i^{\ jk}$ and $\ ^{R}\tilde c_i^{\ jk}$ are, respectively, the structure constants of the brackets $[.,.]^*_L$ and $[.,.]^*_R$. The condition in this context becomes \_K= -\_R\^[ijk]{}\^r\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_l]{}=\_L\^[ijk]{}\^l\_[t\_i]{}\^l\_[t\_j]{}\^l\_[t\_k]{}.\[qp6\]
It is clear that our definition of the affine quasi-Poisson group implies that the Lie quasi-bialgebra $\K_L:=(K,[.,.],[.,.]_L^*,\chi_L)$ acting from the left is not necessary isomorphic to the Lie quasi-bialgebra $\K_R:=(K,[.,.],[.,.]_R^*,\chi_R)$ acting for the right, however, the fact that the both left and right actions are simultaneously quasi-Poisson with respect to the same bivector $\pi^\qa$ implies some restrictions on the possibles structures of the Lie quasi-bialgebras $\K_L$ and $\K_R$. In particular, it must hold \_L=-\_R,and, moreover, $\chi$ must be ad$_\K$-invariant. This is not all, however. It turns out also that the dual brackets $[.,.]_L^*$ and $[.,.]_R^*$ must be both Lie commutators, i.e. the Jacobi identity must hold for them. To see this, we calculate the Lie derivatives of Eq. with respect to the left-invariant and right-invariant vector fields $\nabla^r_{t_i}$ and $\nabla^l_{t_i}$. Because of the invariance of the trilinear form $\chi$, the Lie derivatives of the right-hand-side vanish, while for the Lie derivatives of the left-hand-side, we obtain successively from Eqs. , and $$\L_{\nabla^l_{t_i}}[\pi^\qa,\pi^\qa]_K=2[\L_{\nabla^l_{t_i}}\pi^\qa,
\pi^\qa]= -\ ^{L}\tilde c_i^{\ jk}[\nabla^l_{t_j}\wedge \nabla^l_{t_k},\pi^\qa]= -2\ ^{L}\tilde c_i^{\ jk}\nabla^l_{t_j}\wedge [\nabla^l_{t_k}, \pi^\qa]_K=$$ \^[L]{}c\_i\^[ jk]{} \^[L]{}c\_k\^[ mn]{}\^l\_[t\_j]{}\^l\_[t\_m]{}\^l\_[t\_n]{}=( \^[L]{}c\_i\^[ jk]{} \^[L]{}c\_k\^[ mn]{}+ \^[L]{}c\_i\^[ mk]{} \^[L]{}c\_k\^[ nj]{}+ \^[L]{}c\_i\^[ nk]{} \^[L]{}c\_k\^[ jm]{})\^l\_[t\_j]{}\^l\_[t\_m]{}\^l\_[t\_n]{}=0.\[ji\] Thus we infer \^[L]{}c\_i\^[ jk]{} \^[L]{}c\_k\^[ mn]{}+ \^[L]{}c\_i\^[ mk]{} \^[L]{}c\_k\^[ nj]{}+ \^[L]{}c\_i\^[ nk]{} \^[L]{}c\_k\^[ jm]{}=0,which is nothing but the Jacobi identity for the structure constants $\ ^{L}\tilde c_i^{\ jk}
$. The Jacobi identity for the structure constants $\ ^{R}\tilde c_i^{\ jk}$ can be established similarly.
As in the case of the theory of the affine Poisson groups treated in Section 3, also in the affine quasi-Poisson case the crucial role is played by the value of the bivector field $\pi^\qa$ at the unit element $e_K$ of the group $K$. We again denote this value as $m$: m:=\^(e\_K). Our next goal is to show, that there are two Poisson-Lie structures $\pi^L$ and $\pi^R$ naturally associated with the affine quasi-Poisson structure $\pi^\qa$. They are given by the formulae \^L=L\_\*m-\^,\^R=\^-R\_\*m.\[717\] Let us prove e.g. that $\pi^L$ is the Poisson-Lie structure. For that, we have to prove the following three identities: \_K=0;\[1d\]Ł\_[\^l\_[t\_i]{}]{}\^L= \^[L]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[2d\]Ł\_[\^r\_[t\_i]{}]{}\^L= \^[L]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}.\[3d\]The condition which is the easiest to prove is Eq., because $\nabla^r_{t_i}$ commutes with $\nabla^l_{t_j}$, hence Ł\_[\^l\_[t\_i]{}]{}\^L=Ł\_[\^l\_[t\_i]{}]{}(m\^[jk]{} \^r\_[t\_j]{}\^r\_[t\_k]{}-\^)=-Ł\_[\^l\_[t\_i]{}]{}\^= \^[L]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{},\[aa\]where in the last equality we have used Eq..
Now we prove Eq. . We first note that the bivector $\pi_L$ vanishes at the group origin $e_K$, therefore the Lie derivatives $\L_{\nabla^l_{t_i}}\pi^L$ and $\L_{\nabla^r_{t_i}}\pi^L$ coincide at $e_K$. This means that in order to prove Eq., it is sufficient to prove that $\L_{\nabla^r_{t_i}}\pi^L$ is a left-invariant bivector, because such bivectors are completely determined by their values at the group origin. The proof of the left invariance of $\L_{\nabla^r_{t_i}}\pi^L$ is simple, indeed, we find from Eq. Ł\_[\^r\_[t\_i]{}]{}\^L=Ł\_[\^r\_[t\_i]{}]{}(m\^[jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}-\^)=(- \^[R]{}c\_i\^[ lk]{} +f\_[ij]{}\^[ l]{}m\^[jk]{}-f\_[ij]{}\^[ k]{}m\^[jl]{} )\^r\_[t\_l]{}\^r\_[t\_k]{}.\[bb\]As a bonus, we have obtained the following identity relying the left and the right dual structure constants \^[L]{}c\_i\^[ lk]{}=- \^[R]{}c\_i\^[ lk]{} +f\_[ij]{}\^[ l]{}m\^[jk]{}-f\_[ij]{}\^[ k]{}m\^[jl]{},where $f_{ij}^{\ \ k}$ are the structure constants of the Lie algebra $\K$.
In ordre to prove , we need first to establish the following relation (- \^[L]{}c\_j\^[ kn]{}+ \^[R]{}c\_j\^[ kn]{})m\^[ij]{}+ (- \^[L]{}c\_j\^[ ni]{}+ \^[R]{}c\_j\^[ ni]{})m\^[kj]{}+ (- \^[L]{}c\_j\^[ ik]{}+ \^[R]{}c\_j\^[ ik]{})m\^[nj]{}=2\^[kni]{}.\[fre\]To do it, let us calculate the value of the Schouten bracket $[\pi^\qa,\pi^\qa]_K$ et the unit element $e_K$. For that, we represent the bivector $\pi^\qa$ in terms of the left and the right trivialisation of the bundle $\Lambda^2TK$ as \^ \^r\^[ij]{}\^r\_[t\_i]{}\^r\_[t\_j]{} \^l\^[kn]{}\^l\_[t\_k]{}\^l\_[t\_n]{}.Because $\nabla^r_{t_i}$ commutes with $\nabla^l_{t_j}$, we find \_K= \^r\^[ij]{}(\^r\_[t\_j]{} \^l\^[kn]{})\^r\_[t\_i]{}\^l\_[t\_k]{}\^l\_[t\_n]{}+ \^l\^[ij]{}(\^l\_[t\_j]{} \^r\^[kn]{})\^l\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_n]{},\[bl4\]which, with the help of Eqs. and , gives at the group unit $e_K$ \_K\_[e\_K]{}=(- \^[L]{}c\_j\^[ kn]{}+ \^[R]{}c\_j\^[ kn]{})m\^[ij]{}t\_it\_k t\_n.From this and from Eq., we infer finally the desired identity .
We now have from Eqs. and \_K=\[\^r\_[t\_i]{},\^\]\_K+\[\^r\_[t\_i]{},\^L \]\_K=( \^[R]{}c\_i\^[ kn]{}+ \^[L]{}c\_i\^[ kn]{})\^r\_[t\_k]{}\^r\_[t\_n]{}, hence we find \_K=\_K=m\^[ij]{}( \^[R]{}c\_j\^[ kn]{}+ \^[L]{}c\_j\^[ kn]{}) \^r\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_n]{}.\[cc\]Then we have from \_K=\_K=m\^[ij]{} \^[R]{}c\_j\^[ kn]{}\^r\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_n]{}.\[ff\]By using Eqs., , and , we find finally \_K=\[L\_\*m-\^,L\_\*m-\^\]\_K=(\^[ikn]{}- m\^[ij]{}(- \^[L]{}c\_j\^[ kn]{}+ \^[R]{}c\_j\^[ kn]{}) ) \^r\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_n]{}=0. \[dos\]The bivector $\pi^L\equiv L_*m-\pi^\qa$ thus defines the Poisson-Lie structure on the group $K$ and it can be shown similarly that the bivector $\pi^R\equiv\pi^\qa-R_*m$ also defines the Poisson-Lie structure.
Given the affine quasi-Poisson structure $\pi^\qa$, does it exist the mirror affine quasi-Poisson structure $\pi^\aq$ similarly as it is the case in the ordinary affine Poisson case? We now show that the answer to this question is affirmative. We define this mirror affine quasi-Poisson structure by the formula \^:= R\_\*m+L\_\*m-\^,\[760\]in full analogy with Eq.. Of course, we must show that $\pi^{\aq}$ indeed verifies the affine quasi-Poisson conditions , and with the role of the left and the right inversed. In other words, we have to prove that Ł\_[\^l\_[t\_i]{}]{}\^= - \^[R]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{};\[764\]Ł\_[\^r\_[t\_i]{}]{}\^= \^[L]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{};\[765\]\_K=- \^[ijk]{}\^l\_[t\_i]{}\^l\_[t\_j]{}\^l\_[t\_k]{}=-\^[ijk]{}\^r\_[t\_i]{}\^r\_[t\_j]{}\^r\_[t\_k]{}.\[762\] We start by proving Ł\_[\^r\_[t\_i]{}]{}\^=Ł\_[\^r\_[t\_i]{}]{}(\^L+R\_\*m)=Ł\_[\^r\_[t\_i]{}]{}\^L= \^[L]{}c\_i\^[ jk]{}\^r\_[t\_j]{}\^r\_[t\_k]{}, where we have used Eq. as well as the fact that the right-invariant vector fields commute with the left-invariant ones. We prove similarly Ł\_[\^l\_[t\_i]{}]{}\^=Ł\_[\^l\_[t\_i]{}]{}(L\_\*m-\^R)=-Ł\_[\^l\_[t\_i]{}]{}\^R=- \^[R]{}c\_i\^[ jk]{}\^l\_[t\_j]{}\^l\_[t\_k]{}. Finally, we establish the identity by using Eqs. , and $$[\pi^\aq,\pi^\aq]_K=\left[L_*m-\pi^R,L_*m-\pi^R\right]_K=[\pi^R,\pi^R]_K+
[L_*m,L_*m]_K-2[L_*m,\pi^R]=$$ = \[L\_\*m,L\_\*m\]\_K-2\[L\_\*m,\^\]\_K= - m\^[ij]{}(- \^[L]{}c\_j\^[ kn]{}+ \^[R]{}c\_j\^[ kn]{}) \^r\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_n]{}=- \^[ikn]{} \^r\_[t\_i]{}\^r\_[t\_k]{}\^r\_[t\_n]{}.
It thus turns out that we can associate to every affine quasi-Poisson group $(K,\pi^\qa)$ two Drinfeld doubles[^5] $D_L$ and $D_R$, which are, respectively, the Drinfeld doubles of the Poisson-Lie groups $(K,\pi^L)$ and $(K,\pi^R)$. The double $D_L$ has naturally two maximally isotropic subgroups $K$ and $\tilde K_L$ , while $D_R$ has subgroups $K$ and $\tilde K_R$. We call the Lie groups $\tilde K_L$ and $\tilde K_R$ the left and the right dual groups of the affine quasi-Poisson group $K$; their Lie algebras are isomorphic to the Lie algebras $\tilde\K_L$ and $\tilde\K_R$ with the structure constants $\ ^{L}\tilde c_i^{\ jk}$ and $\ ^{R}\tilde c_i^{\ jk}$.
For completeness, we detail how the left and right Poisson-Lie structures $\pi^L$ and $\pi^R$ as well as the mirror affine quasi-Poisson structure $\pi^{\aq}$ associated to the affine Poisson structure $\pi^\qa$ are extracted from the structure of the Drinfeld doubles $D_L$ and $D_R$ of the affine quasi-Poisson group $K$. We start from the perspective of the double $D_L$, we pick a basis $t_i, i=1,...,d$ of the Lie algebra $\K\subset\D_L$ and the dual basis $T^i_L, i=1,...,d$ of $\tilde\K_L\subset\D_L$ in such a way that (t\_i,T\^j\_L)\_[\_L]{}=\_i\^j\[834\]We can then write \^[L]{}c\_i\^[ jk]{}=(\[T\_L\^j,T\_L\^k\],t\_i)\_[\_L]{}. The affine quasi-Poisson bracket $\{.,.\}^{\qa}_K$ corresponding to the affine quasi-Poisson bivector $\pi^{\qa}$ is then given by the formula {f\_1,f\_2}\_K\^(k)=m\^[ij]{}\^r\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2+(T\^i\_L,Ad\_kT\^j\_L)\_[\_L]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2;\[pba’\] and the Poisson-Lie bracket $\{.,.\}^{L}_K$ corresponding to the Poisson-Lie bivector $\pi^{L}$ is given by the formulae {f\_1,f\_2}\_K\^L(k)=-(T\^i\_L,Ad\_kT\^j\_L)\_[\_L]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2.\[plr’\]Here $f_1,f_2$ are smooth functions on $K$ and the Einstein summation convention applies.
From the point of view of the double $D_R$, we pick a basis $t_i, i=1,...,d$ of the Lie algebra $\K\subset\D_R$ and the dual basis $T^i_R, i=1,...,d$ of $\tilde\K_R\subset\D_R$ in such a way that (t\_i,T\^j\_R)\_[\_R]{}=\_i\^j.\[845\]We can then write \^[R]{}c\_i\^[ jk]{}=(\[T\_R\^j,T\_R\^k\],t\_i)\_[\_R]{}.The mirror affine quasi-Poisson bracket $\{.,.\}^{\aq}_K$ corresponding to the mirror affine Poisson bivector $\pi^{\aq}$ is then given by the formula {f\_1,f\_2}\_K\^(k)=m\^[ij]{}\^r\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2+(T\^i\_R,Ad\_kT\^j\_R)\_[\_R]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2\[pbam’\] and the Poisson bracket $\{.,.\}^{R}_K$ corresponding to the Poisson-Lie bivector $\pi^{R}$ is given by {f\_1,f\_2}\_K\^R(k)=-(T\^i\_R,Ad\_kT\^j\_R)\_[\_R]{}\^l\_[t\_i]{}f\_1\^r\_[t\_j]{}f\_2.\[314’\]
[**Remark:**]{} [The affine quasi-Poisson formulae and look identical than the affine Poisson formulae and , however, they are different in the sense that the matrix $m^{ij}$ has to be chosen in such a way that the brackets $\{.,.\}^\aq$ and $\{.,.\}^\na$ be quasi-Poisson in the first case and Poisson in the second.]{}
We finish this section by stating that the following useful relation holds {f\_1S,f\_2S}\_K\^={f\_1,f\_2}\_K\^S,\[876\] where $S:K\to K$ is defined as $S(k):=k^{-1}$. The proof of this fact is similar to the one presented at the end of Section 3 in the affine Poisson case .
Affine quasi-Poisson T-duality
==============================
General framework
-----------------
This section constitutes the generalization of Section 4 which was devoted to the affine Poisson T-duality.
Let $(K,\pi^\qa)$ be the affine quasi-Poisson group, $D_L$ and $D_R$ its Drinfeld doubles and $\tilde K_L\subset D_L$ and $\tilde K_R\subset D_R$ the dual groups of $K$ in the sense of Section 5. Consider then two $\sigma$-models living on $K$ S\_(k)=d((E\^+\^(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_R]{};\[817\]S\_(k)=d((E\^+\^(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_L]{},\[818\] one $\sigma$-model on $\tilde K_R$ S\_(k\_R)=d((E\^+\^R(k\_R))\^[-1]{}\_+k\_Rk\_R\^[-1]{}, \_-k\_Rk\_R\^[-1]{})\_[\_R]{},\[820\]and, finally, one on $\tilde K_L$ S\_(k\_L)=d((E\^+\^L(k\_L))\^[-1]{}\_+k\_Lk\_L\^[-1]{}, \_-k\_Lk\_L\^[-1]{})\_[\_L]{}.\[822\] Here the fields-valued operators $\Pi^\qa(k):\tilde\K_R\to\K$, $\Pi^\aq (k):\tilde\K_L\to\K$, $\tilde\Pi^R(\tilde k_R):\K\to\tilde\K_R$ and $\tilde\Pi^L(\tilde k_L):\K\to\tilde\K_L$ characterize the affine quasi-Poisson structures $\pi^\qa$, $\pi^\aq$ on $K$ and the Poisson-Lie structures $\tilde\pi^L$,$\tilde \pi^R$ on $\tilde K_L$ and $\tilde K_R$ dual to the Poisson-Lie structures $\pi^L$,$\pi^R$ on $K$ in the sense of the relations: \^,df\_1df\_2(k)={f\_1,f\_2}\^\_K(k)=( \^R\^l f\_1,\^(k) \^R\^l f\_2)\_[\_R]{},\[821\]\^,df\_1df\_2(k)={f\_1,f\_2}\^\_K(k)=( \^L\^l f\_1,\^( k) \^L\^l f\_2)\_[\_L]{},\[822b\] \^R,df\_1df\_2(k\_R) ={f\_1,f\_2}\_[K\_R]{}(k\_R)=( \^R\^l f\_1,\^R(k\_R) \^R\^l f\_2)\_[\_R]{}.\[827\] \^L,df\_1df\_2(k\_L) ={f\_1,f\_2}\_[K\_L]{}(k\_L)=( \^L\^l f\_1,\^L(k\_L) \^L\^l f\_2)\_[\_L]{},\[826\]
Recall in this respect that $\ \!^L\nabla^r$ and $\ \!^R\nabla^r$ are, respectively, $\tilde\K_L$ and $\tilde\K_R$-valued differential operators acting on functions on $K$ as ( \^[L]{}\^l f, x)\_[D\_L]{}:= \^l\_[x]{} f, ( \^[R]{}\^l f, x)\_[D\_R]{}:= \^l\_[x]{} f, x,\[824\]and $\ \!^{L}\tilde\nabla^{r}$, $\ \!^{R}\tilde\nabla^{r}$ are $\K$-valued differential operators acting respectively on the functions on $\tilde K_L$ and $\tilde K_R$ as (\^L\^l f,x\_L)\_[\_L]{}(k\_L):= (\^l\_[x\_L]{}f)(k\_L)= \_[s=0]{}, x\_L\_L;\[831\](\^R\^l f,x\_R)\_[\_R]{}(k\_R):= (\^l\_[x\_R]{}f)(k\_R)= \_[s=0]{}, x\_R\_R.\[832\]
As in the affine Poisson context studied in Section 4, also in the case of the affine quasi-Poisson T-duality the idea is to find suitable invertible linear operators $E^\qa$, $E^\aq$, $\tilde E^\qa$ and $\tilde E^{\aq}$ in such a way that the four $\sigma$-models , , and are pairwise T-dual to each other.
In fact, the liberty of choice exists only for the operator $E^\qa:\tilde\K_R\to\K$ because the operators $E^\aq:\tilde\K_L\to\K$, $\tilde E^\qa:\K\to\tilde\K_R$ and $\tilde E^{\aq}:\K\to\tilde\K_L$ turn out to be determined from it. For the case of $\tilde E^\qa$ and $\tilde E^{\aq}$, this can be seen by rewriting the $\sigma$-model actions and equivalently as S\_(k)=d((E\^+M\^R+\^R(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_R]{},\[841\]S\_(k)=d((E\^+M\^L+\^L(k))\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_[\_L]{},\[842\]where M\^R=\^(e\_K), M\^L=\^(e\_K)\[843\] and $\Pi^L(k):\tilde\K_L\to\K$ and $\Pi^R(k):\tilde\K_R\to\K$ stand, respectively, for the left and right Poisson-Lie structures on $K$ (associated to the affine quasi-Poisson structure) in the sense of the formulae \^L,df\_1df\_2(k) ={f\_1,f\_2}\^L\_K(k)=( \^L\^l f\_1,\^L(k) \^L\^lf\_2)\_[\_L]{},\[844\] \^R,df\_1df\_2(k) ={f\_1,f\_2}\^R\_K(k)=( \^R\^l f\_1,\^R(k) \^R\^lf\_2)\_[\_R]{}.\[846\]We then observe that the $\sigma$-model is related by the standard Poisson-Lie T-duality to the model if the operator $\tilde E^\qa$ is inverse to $E^\qa+M_R$; and is Poisson-Lie T-dual to if $\tilde E^\aq$ is inverse to $E^{\aq}+M_L$.
We thus remark that if the $\sigma$-model is T-dual to then all four $\sigma$-models , , and are pairwise T-dual to each other. From the point of view of the scheme , on the up-right vertex of the scheme there is the $\sigma$-model , on the up-left there is , on the down-left is the model and, finally, on the down-right there is the $\sigma$-model . The up-right vertex is linked with the up-left vertex by the standard Poisson-Lie T-duality based on the subspace $\E_L=\{\tilde x +(E^\aq+M^L)\tilde x,\tilde x\in\tilde\K_L\}$ and the down-right vertex is linked with the down-left vertex by the standard Poisson-Lie T-duality based on the subspace $\E_R=\{\tilde x +(E^{\qa}+M^R)\tilde x,\tilde x\in\tilde\K_R\}$, which means that the existence of all horizontal arrow symplectomorphisms is established.
Sufficient condition for the affine quasi-Poisson T-duality
-----------------------------------------------------------
For what choice of the data $K,\Pi^\na,E^\qa,E^{\aq}$ the $\sigma$-model is T-dual to ? In order to formulate the sufficient condition for this T-duality we need to define two bilinear forms $(.,.)_\qa$, $(.,.)_{\aq}$ on the Lie algebra $\K$: (x,y)\_:=(x,(E\^)\^[-1]{}y)\_[\_R]{},(x,y)\_:=(x,(E\^)\^[-1]{}y)\_[\_L]{},x,y.\[867\] We then have the following assertion
[**Assertion 3**]{}: [*If the bilinear form*]{} $(.,.)_\qa$ [*is symmetric, non-degenerate and ad$_\K$-invariant and, moreover, if it coincides with the bilinear form*]{} $(.,.)_{\aq}$ [*then the $\sigma$-model is T-dual to the $\sigma$-model .* ]{}
The proof of [**Assertion 3**]{} is in every aspect similar to the one of Assertion 2 in Section 5, we shall be therefore brief. We first remark that it holds E\^ \^R\^l=E\^ \^L\^l= \^\^l,\[879\] where the operators $\ \!^L\nabla^l$ and $\ \!^R\nabla^l$ where defined in and we define the $\K$-valued differential operator $\ \!^\qa{\bm\nabla}^l$ acting on the functions on $K$ as ( \^\^l f, x)\_:= \^l\_[x]{} f, x.\[882\] To see e.g. that the first of the relations indeed holds, we rewrite the left-hand-side of the second of Eqs. as ( \^\^l f, x)\_:= \^l\_[x]{} f = ( \^[R]{}\^l f, x)\_[\_R]{}=([E\^]{}\^[-1]{}E\^ \^[R]{}\^l f, x)\_[\_R]{}= (E\^ \^[R]{}\^l f, x)\_ and we finish the argument by invoking the non-degeneracy of the bilinear forms $(.,.)_{\D_R}$ and $(.,.)_\qa$.
We shall need also the following relation \^(k)= Ad\_k\^(k\^[-1]{})([E\^]{})\^[-1]{}Ad\_[k\^[-1]{}]{}E\^,\[890\] which can be obtained from the identity much in the same way as the relation was obtained from the identity in Section 5.2.
We can finally rewrite the $\sigma$-model actions and as follows $$S_\qa(k)= \jp\int d\tau \oint \biggl({E^\qa}^{-1}\Bigl(1+\Pi^\qa(k){E^\qa}^{-1}\Bigr)^{-1}\partial_+k k^{-1}, \partial_- k k^{-1}\biggr)_{\D_R} =$$=d((1+\^(k)[E\^]{}\^[-1]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_\[894\] $$S_{\aq}(k)= \jp\int d\tau \oint \biggl({E^{\aq}}^{-1}\Bigl(1+\Pi^{\aq}(k){E^{\aq}}^{-1}\Bigr)^{-1}\partial_+k k^{-1}, \partial_- k k^{-1}\biggr)_{\D_L} =$$=d((1+\^(k)[E\^]{}\^[-1]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_.\[895\] Using the crucial identity as well as the hypothesis that the ad$_\K$-invariant bilinear forms $(.,.)_\qa$ and $(.,.)_{\aq}$ coincide, we find finally S\_(k)=S\_(k\^[-1]{}).\[898\]We conclude that if the sufficient condition of [**Assertion 3**]{} holds, then the T-duality between the $\sigma$-models and is realized simply by the field redefinition $k\to k^{-1}$ and the corresponding T-duality symplectomorphism is therefore just the point canonical transformation. On the other hand, the affine quasi-Poisson T-duality between the models and is nontrivial because it is the composition of two horizontal and one (left) vertical symplectomorphism in the scheme and only the left vertical one is point-like.
Examples of the affine quasi-Poisson T-duality
----------------------------------------------
If the affine quasi-Poisson group $(K,\pi^\qa)$ and the operators $E^\qa,E^{\aq}$ satisfy the sufficient condition of [**Assertion 3**]{} then the Lie algebra $\K$ is quadratic, which means that there is the symmetric ad$_\K$-invariant non-degenerate bilinear form $(.,.)_\K$ on $\K$. Let us therefore restrict our effort to construct affine quasi-Poisson structures to the case of the groups $K$ which have quadratic Lie algebras $\K$. We first define two Lie quasi-bialgebra structures on $\K$ and then construct the affine quasi-Poisson structure $\pi^\qa$ compatible with them in the sense of the conditions , and .
Let $R^r:\K\to\K$ and $R^l:\K\to\K$ be two Yang-Baxter operators, which means that they are both skew-symmetric with respect to the bilinear form $(.,.)_\K$ and they verify the Yang-Baxter identities with different values of the parameter $\e$: =R\^r(\[R\^rx,y\]+\[x,R\^ry\])-\_r\[x,y\], x,y; \[913\] =R\^l(\[R\^lx,y\]+\[x,R\^ly\])-\_l\[x,y\], x,y. \[914\] Consider also an element $\chi\in\Lambda^3\K$ defined by the relation (,xyz)\_(’,x)\_(”,y)\_(”’,z)\_:=(\_r-\_l)(\[x,y\],z)\_,x,y,x,\[916\] where we have used the Sweedler notation $\chi=\chi'\otimes\chi''\otimes\chi'''$. The definition of the affine quasi-Poisson structure requires the element $\chi$ to be $\K$-invariant, which is indeed the case because the ad$_\K$-invariance of $\K$ implies that it holds for every $u,x,y,z\in\K$ (\[ad\_ux,y\],z)\_+(\[x,ad\_uy\],z)\_+(\[x,y\],ad\_uz)\_= 0. Let us now identify the spaces $\K^*$ and $\K$ via the non-degenerate form $(.,.)_\K$, that is, we have the identification map $J:\K^*\to\K$ defined as (J(),x)\_=,x.We define the left Lie quasi-bialgebra structure on $\K$ as the quadruple $\K_L:=(\K,[.,.],[.,.]_L^*,\chi)$ where the $3$-form $\chi\in\Lambda^3\K$ is given by Eq. and the dual commutator $[.,.]^*_L$ on $\K^*$ is given by \^\*\_L:=-J\^[-1]{}(\[R\^lJ(),\]+\[J(),R\^lJ()\]).\[928\]Similarly, the right Lie quasi-bialgebra structure on $\K$ is the quadruple $\K_R:=(\K,[.,.],[.,.]_R^*,-\chi)$ where the dual commutator $[.,.]^*_R$ on $\K^*$ is given by \^\*\_R:=-J\^[-1]{}(\[R\^rJ(),\]+\[J(),R\^rJ()\]).\[932\]Before defining the affine quasi-Poisson structure $\pi^\qa$ compatible with the actions of the two Lie quasi-bialgebras just introduced, we make a little digression and try to understand the structure of the respective quasi-Drinfeld doubles $D^q_L$ and $D^q_R$ of the Lie quasi-bialgebras $\K_L$ and $\K_R$. We start with $\D^q_L$ and remark that the direct application of the general formula gives the following result $$[x\oplus\alpha, y\oplus \beta]_{\D^q_L}=\left([x,y]-[x,R^lJ(\beta)]+R^l[x,J(\beta)]+[y,R^lJ(\alpha)]-R^l[y,J(\alpha)] +(\e_r-\e_l)[J(\alpha),J(\beta)]\right)\oplus$$(-J\^[-1]{}\[J(),J()\]\^\*-J\^[-1]{}\[J(),x\]+J\^[-1]{}\[J(),y\]).\[934\]This formula looks quite cumbersome but there is a simple isometric isomorphism $\ups_L$ from $\D^q_L$ into the Lie algebra $\D_{\e_r}$ with the commutator defined by Eq. \_[\_r]{}=(\[x\_1,y\_1\]+\_r\[x\_2,y\_2\],\[x\_1,y\_2\]+\[x\_2,y\_1\])\_r.\[938\]The map $\ups_L$ is given by \_L(x)=(x-R\^lJ(),J())\_rand the Yang-Baxter property of the operator $R^l$ is the only thing needed for proving that the map $\ups_L$ is indeed the isomorphism of the Lie algebras $\D^q_L$ and $\D_{\e_r}$. Similarly, the map $\ups_R$ given by \_R(x)=(x-R\^rJ(),J())\_lrealizes the isometric isomorphism of the Lie algebras $\D^q_R$ and $\D_{\e_l}$.
We notice that the structures of the quasi-doubles $\D^q_L$ and $\D^q_R$ of the Lie quasi-bialgebras $\K_L$ and $\K_R$ do not depend on the particular form of the operators $R^l$ and $R^r$ but only on the parameters $\e_r$ and $\e_l$. On the other hand, the ad-invariant non-degenerate symmetric bilinear forms $(.,.)_{\D_{\e_r}}$ and $(.,.)_{\D_{\e_l}}$ do not depend even on the real parameters $\e_r,\e_l$ and they are given by the formulae ((x\_1,x\_2)\_r,(y\_1,y\_2)\_r)\_[\_[\_r]{}]{}:= (x\_2,y\_1)\_+(x\_1,y\_2)\_.\[918\]((x\_1,x\_2)\_l,(y\_1,y\_2)\_l)\_[\_[\_l]{}]{}:= (x\_2,y\_1)\_+(x\_1,y\_2)\_.\[919\]We also note that the Lie algebra $\K$ is embedded into $\D_{\e_r}$ and into $\D_{\e_l}$ in the same way, i.e. $(\K,0)\subset \D_{\e_r(\e_l)}$.
The affine quasi-Poisson bracket $\{.,.\}^\qa$ compatible with the left and right actions of the Lie quasi-bialgebras $\K_L$ and $\K_R$ is then given by {f\_1,f\_2}\^\_K:=(\^r f\_1,R\^r\^r f\_2)\_+(\^l f\_1,R\^l\^l f\_2)\_,\[955\] which the $\K$-valued differential operators ${\bm\nabla}^r, {\bm\nabla}^l$ acting on the functions on $K$ were defined in Eq.. The formula looks identical as in the affine Poisson case but now the operators $R^l,R^r$ verify the Yang-Baxter equations with different values of the parameters $\e_l\neq\e_r$. In consequence, the bracket is not Poisson but just quasi-Poisson. Let us indeed verify that the bracket is affine quasi-Poisson according to the defining relations , and .
In order to verify the conditions and , we choose the orthonormal basis $t_i$ on $\K$ and write the bivector $\pi^\qa$ corresponding to the bracket as \^=(t\_i,R\^rt\_j)\_[t\_i]{}\^r\^r\_[t\_j]{}+(t\_i,R\^lt\_j)\_[t\_i]{}\^l\^l\_[t\_j]{}. Since the left-invariant vector fields on $K$ commute with the right-invariant ones, we find, respectively, the desired results Ł\_[\^l\_[t\_k]{}]{}\^= -(t\_i,R\^lt\_j)\_(\_[\[t\_k,t\_i\]]{}\^l\^l\_[t\_j]{}+\_[t\_i]{}\^l\^l\_[\[t\_k,t\_j\]]{})= (\[t\_i,R\^lt\_j\]+\[R\^lt\_i,t\_j\],t\_k)\_\_[t\_i]{}\^l\_[t\_j]{}\^l.\[966\] Ł\_[\^r\_[t\_k]{}]{}\^=(t\_i,R\^rt\_j)(\_[\[t\_k,t\_i\]]{}\^r\^r\_[t\_j]{}+\_[t\_i]{}\^r\^r\_[\[t\_k,t\_j\]]{})=-(\[t\_i,R\^rt\_j\]+\[R\^rt\_i,t\_j\],t\_k)\_\_[t\_i]{}\^r\_[t\_j]{}\^r.\[967\] because the dual structure constants $\ ^{L}\tilde c_i^{\ jk}$ and $\ ^{R}\tilde c_i^{\ jk}$ featuring in and are given by the dual commutators and of the elements of the basis $t_i$.
It remains to calculate the Schouten bracket $$\jp[\pi^\qa,\pi^\qa]_K=$$=(t\_i,R\^rt\_j)\_(t\_k,R\^rt\_m)\_\_K +(t\_i,R\^lt\_j)\_(t\_k,R\^lt\_m)\_\_K.\[973\] We use the formula and the Yang-Baxter identities and to obtain the desired result $$\jp[\pi^\qa,\pi^\qa]_K=-\frac{1}{6}\left([R^lt_i,R^lt_j]-R^l\left([R^lt_i,t_j]+[t_i,R^lt_j]\right),t_k\right)_\K \nabla_{t_i}^l\wedge \nabla_{t_j}^l\wedge\nabla_{t_k}^l +$$+ (\[R\^rt\_i,R\^rt\_j\]-R\^r(\[R\^rt\_i,t\_j\]+\[t\_i,R\^rt\_j\]),t\_k)\_\_[t\_i]{}\^r\_[t\_j]{}\^r\_[t\_k]{}\^r=(\_r-\_l)(\[t\_i,t\_j\],t\_k)\_\_[t\_i]{}\^r\_[t\_j]{}\^r\_[t\_k]{}\^r.\[975\] The mirror affine quasi-Poisson bracket $\{.,.\}^{\aq}$ and the associated left and write Poisson-Lie brackets $\{.,.\}^L$, $\{.,.\}^R$ can be easily extracted from Eqs. and and they are given by the formulae {f\_1,f\_2}\^\_K:=(\^r f\_1,R\^l\^r f\_2)\_+(\^l f\_1,R\^r\^l f\_2)\_,\[978\]{f\_1,f\_2}\^[L]{}\_K:=(\^r f\_1,R\^l\^r f\_2)\_-(\^l f\_1,R\^l\^l f\_2)\_,\[979\]{f\_1,f\_2}\^[R]{}\_K:=(\^r f\_1,R\^r\^r f\_2)\_-(\^l f\_1,R\^r\^l f\_2)\_.\[980\]The fact that the brackets $\{.,.\}^L$ and $\{.,.\}^R$ are Poisson-Lie can be also verified by a direct calculation which amounts to replacing appropriately $R^r$,$R^l$ in Eqs. , , and . For example, for $\pi^L$ we replace $R^r$ by $R^l$ and $R^l$ by $-R^l$ and the result is Ł\_[\^l\_[t\_k]{}]{}\^L =- (\[t\_i,R\^lt\_j\]+\[R\^lt\_i,t\_j\],t\_k)\_\_[t\_i]{}\^l\_[t\_j]{}\^l;\[981\] Ł\_[\^r\_[t\_k]{}]{}\^L =-(\[t\_i,R\^lt\_j\]+\[R\^lt\_i,t\_j\],t\_k)\_\_[t\_i]{}\^r\_[t\_j]{}\^r;\[982\] \_K =(\_l-\_l)(\[t\_i,t\_j\],t\_k)\_\_[t\_i]{}\^r\_[t\_j]{}\^r\_[t\_k]{}\^r=0.\[983\] Therefore $\pi^L$ and, similarly, $\pi^R$ are indeed the Poisson-Lie structures.
What are the Drinfeld doubles $D_L$ and $D_R$ for the Poisson-Lie groups $(K,\pi^L)$ and $(K,\pi^R)$? Well, these are the doubles $D_{\e_l}$ and $D_{\e_r}$, respectively, where the corresponding maximally isotropic Lie subalgebras $\tilde \K_L$ and $\tilde \K_R$ are given as the graphs of the operators $R^l$ and $R^r$: \_L ={(-R\^lx,x)\_l, x},\_R ={(-R\^rx,x)\_r, x}.\[990\] It can be checked that the explicit formulae and applied to the subspaces reproduce the formulae and as they should.
Can the affine quasi-Poisson structure serve as the basis for a viable example of the affine quasi-Poisson T-duality? The answer to this question is affirmative because it is easy to verify that in this case the sufficient condition of [**Assertion 3**]{} holds. Indeed, the following choice of the operators $E^\qa:\tilde\K_R\to\K$ and $E^{\aq}:\tilde\K_L\to\K$ does the job E\^(-R\^rx,x)\_r:=a(x,0)\_r, E\^(-R\^lx,x)\_l:=a(x,0)\_l, a<0.\[994\]Using the definitions and , it is then easy to verify for every $x,y\in \K$ that there hold the equalities of the following inner products (x,y)\_((x,0)\_r,(E\^)\^[-1]{}(y,0)\_r)\_[D\_[\_r]{}]{} =((x,0)\_r,(-R\^ry,y)\_r)\_[\_[\_r]{}]{}=(x,y)\_,(x,y)\_((x,0)\_l,(E\^)\^[-1]{}(y,0)\_l)\_[\_[\_l]{}]{}=((x,0)\_l,(-R\^ly,y)\_l)\_[\_[\_l]{}]{}=(x,y)\_, hence the sufficient condition for the affine quasi-Poisson T-duality is indeed satisfied.
It is instructive to cast the $\sigma$-model actions and in terms of the Yang-Baxter operators $R^l,R^r$. Similarly as in Section 5.3, we find \^(k)([E\^]{})\^[-1]{}=R\^l+R\^r\_[k\^[-1]{}]{}, \^(k)([E\^]{})\^[-1]{}=R\^r+R\^l\_[k\^[-1]{}]{},\[ger’\]where R\^l\_[k\^[-1]{}]{}:=Ad\_k R\^l Ad\_[k\^[-1]{}]{}, R\^r\_[k\^[-1]{}]{}:=Ad\_[k]{} R\^r Ad\_[k\^[-1]{}]{}. Finally, we infer S\_(k)=d((a+ R\^r+R\^l\_[k\^[-1]{}]{})\^[-1]{}\_+k k\^[-1]{}, \_- k k\^[-1]{})\_;\[008\]S\_(k)=d((a+ R\^l+R\^r\_[k\^[-1]{}]{})\^[-1]{} \_+k k\^[-1]{}, \_- k k\^[-1]{})\_.\[009\]The affine quasi-Poisson formulae and seem to coincide with the affine Poisson ones and , but the difference resides in the fact that in the affine quasi-Poisson case the operators $R^l$ and $R^r$ verify the Yang-Baxter conditions and with $\e_l\neq\e_r$.
Of course, there is no universal way to rewrite the actions of the dual $\sigma$-models $\tilde S_{\qa}(\tilde k_R)$ and $\tilde S_{\aq}(\tilde k_L)$ in terms of the Yang-Baxter operators $R^r,R^l$ since the very structure of the groups $\tilde K_L,\tilde K_R$ depends implicitely on $R^r,R^l$. The formulae and represent the maximum that can be achieved in this direction, on the other hand, the operators $R^r$ and $R^l$ are useful if we wish to describe explicitely the subspaces $\E_L\subset \D_{\e_l}$ and $\E_R\subset\D_{\e_r}$ underlying, respectively, the Poisson-Lie T-dualities relating the model with the model and the model with the model . We find with the help of the formulae and that M\^R(E\^)\^[-1]{}= M\^L(E\^)\^[-1]{}=R\^r+R\^l,\[014\]hence \_R= {(ay+R\^ly,y)\_r,y},\_L= {(ay+R\^ry,y)\_l,x}.\[017\]
We conclude this section by listing several distinguished choices of the Yang-Baxter operators $R^l,R^r$ leading to the affine quasi-Poisson T-duality. We first consider the case $R^l=0$ and $R^r$ arbitrary in which case the action is that of the Yang-Baxter $\sigma$-model introduced by the present author in [@K02; @K09]. This case is distinguished by the fact that the Drinfeld doubles $\D_L$ and $\D_R$ are not isomorphic.
If one of the operators $R^l, R^r$ is a multiple of the other then the corresponding action describes another particular $\sigma$-model introduced by the present author in [@K09; @K14] in connection with the question of integrability of nonlinear $\sigma$-models. Finally, we may consider an affine quasi-Poisson generalisation of the affine Poisson T-duality based on the Drinfeld twist operators (cf. Section 4.4). In this case we have R\^l=\_l R\^c+R\_t\^l,R\^r=\_r R\^c+R\_t\^r, \_l,\_r,where $R_t^l$, $R_t^r$ are the Drinfeld twist operators introduced at the beginning of Section 4.4 and $R^c:\K\to\K$ is the so-called canonical Yang-Baxter operator defined as (cf. Ref. [@K09]): R\^cT\^=0,R\^cB\^=C\^,R\^cC\^=-B\^.\[022\]Here $T^\mu$ is a basis of the Cartan subalgebra $\gH$ of the real simple compact Lie algebra $\K$ and the basis of $\gH^\perp\subset \K$ is chosen as $ B^{\al},C^{\al}$, $\al>0$ where B\^=(E\^+E\^[-]{}), C\^=(E\^-E\^[-]{})and $E^{\pm\al}$ are the step generators of $\K^\bc$.
It is not difficult to check that the skew-symmetric operator $R^c$ verifies the Yang-Baxter identity =R\^c(\[R\^cx,y\]+\[x,R\^cy\])+\[x,y\], x,y\[II\] and, using this fact, to verify that $R^l=\alpha_l R^c+R_t^l$ and $R^r=\alpha_r R^c+R_t^r$ are also the Yang-Baxter operators with $\e_l=-\alpha_l^2$ and $\e_r=-\alpha_r^2$, respectively.
Dressing cosets
===============
There exists a generalization of the Poisson-Lie T-duality introduced as the Poisson-Lie counterpart of the standard non-Abelian T-duality for the cases where the non-Abelian isometry group does not act freely on the target space [@KS96b]. This so-called “dressing cosets” construction gives rise seemingly to the same dual pair of the $\sigma$-models and as the standard Poisson-Lie T-duality but the linear operators $E,\tilde E$ featuring in the Lagrangians are such that the both models and develop a gauge symmetry reducing the common dimension of their targets. In particular, if the target groups $K$ and $\tilde K$ are $d$-dimensional and the dimension of the gauge group $F$ is $p$, then after the gauge fixing the targets of the $\sigma$-models and become effectively $(d-p)$-dimensional.
As an example of the dressing cosets construction, consider the dual pair of the $\sigma$-models and for the case when $K$ is a simple compact group and $\tilde K$ is the Lie algebra $\K$ with the (Abelian) group structure given by the vector space addition. The Drinfeld double is the group $D_0$ with the multiplication law given by Eq. , the Poisson-Lie bivector $\Pi(k)$ on $K$ then trivally vanishes and the Poisson-Lie bivector on $\tilde K$ is given by the adjoint action of the Lie algebra. The actions of the $\sigma$-models and in this particular case thus become S=dd( E\_+k k\^[-1]{}, \_- k k\^[-1]{})\_, k(,)K;\[1131\] S=dd( (E- [ad]{}\_)\^[-1]{}\_+, \_- )\_, (,).\[1132\] For the linear operator $\tilde E$ we pick the orthogonal projector on the subspace $\gH^\perp\subset \K$ perpendicular to the Cartan subalgebra $\gH$. Then it is not difficult to see that both $\sigma$-models and develop the gauge symmetry with respect to the action of the Cartan torus $\mathbb T\subset K$ (the Lie algebra of the Cartan torus is the Cartan subalgebra $\gH$). In particular, an element $f(\tau,\sigma)$ of the gauge group acts as kfk, \_f.For example, for the group $K=SU(2)$ the models and have both two-dimensional targets and correspond to the dual pairs obtained by different methods already in the early days of the non-Abelian T-duality [@OQ92; @GR93; @AABL93; @Hew96]. Other examples of the dressing cosets have been studied in [@S98; @S99; @KP99; @CM06; @HT; @SST15; @BTW; @HS; @SV].
The dressing cosets method may not look like a substantial generalisation of the Poisson-Lie T-duality, because it boils down just to the study of some special linear operators $E$ for which the dual $\sigma$-models and develop the gauge symmetry. It is, however, the very purpose of the present section to show that the dressing cosets generalization of the Poisson-Lie T-duality is not as mild one as it may seem, since it covers all examples of the affine (quasi-)Poisson T-duality which we have constructed in the previous sections!
A remark is in order at this point of the exposition. When we say that the affine (quasi-)Poisson generalization of the Poisson T-duality is already included in the dressing cosets construction [@KS96b], we do not mean that we have exposed in the previous sections the story which had been already known. Actually, the existence of the pairwise T-duality of the four $\sigma$-models , , , associated to the affine (quasi-)Poisson geometry constitutes the genuinely new result of the present article. Of course, we could have also presented this result as a new nontrivial application of the old dressing cosets method but this would not respect the logic of the things. Indeed, in reality, we have discovered the affine (quasi-)Poisson T-duality by reasoning making no reference to the dressing cosets and only thanks to the insights obtained in this way we were able to find out [*à posteriori*]{} that there is beyond all that an appropriate variant of the dressing cosets construction. It seems difficult to imagine to move in the opposite direction. Indeed, it looks counterintuitive to suspect that by taking particular singular values of the operators $E,\tilde E$ in the actions and the quadruple of the four pairwise T-dual $\sigma$-models , , , could emerge, and it neither seems evident, how the dressing cosets construction could explain that two $\E$-models living on different Drinfeld doubles can be dynamically equivalent. Nevertheless, it is precisely what eventually happens...
To move on forward, we have first to review the first order formalism of the dressing cosets. As in the case of the standard Poisson-Lie T-duality, the construction is underlied by a particular Drinfeld double which we now denote as $\mathbb D$. Three subgroups of the double $\dd$ enter into the game: two of them are the half-dimensional isotropic subgroups $K$ and $\tilde K$ as before but there is also a subgroup $F$, which is isotropic too (i.e. the restriction of the bilinear form $(.,.)_{{\rm Lie}(\dd)}$ on the Lie algebra $\F$ of the group $F$ vanishes). The group $F$ may have whatever dimension smaller than the half of the dimension of the double and plays the role of the gauge group.
The phase space $L\dd_F$ of the dressing coset $\E$-model is now the space of the elements $l(\sigma)$ of the loop group $L\dd$, for which it holds (\_ll\^[-1]{},)\_[[Lie]{}()]{}=0.The (pre)symplectic form of the dressing coset $\E$-model is just the restriction to $L\dd_F$ of the symplectic form \_[LD]{}=-d(l\^[-1]{}dl,\_(l\^[-1]{}dl))\_[[Lie]{}()]{}\[1158\]and the Hamiltonian of the dressing coset $\E$-model looks the same as in the case of the standard Poisson-Lie T-duality Q\_=d(\_ll\^[-1]{},\_ll\^[-1]{})\_[[Lie]{}()]{}.\[1160\]However, now the linear operator $\E$ has different properties. Denoting by $\F^\perp$ all elements in the Lie algebra Lie$(\dd)$ orthogonal to $\F$, we require that the operator $\E:\F^\perp\to\F^\perp$ is self-adjoint, its kernel contains $\F$, the bilinear form $(.,\E.)_{{\rm Lie}(\dd)}$ on $\F^\perp$ is positive semi-definitive, the image of the operator $\E^2-{\rm Id}$ is contained in $\F$ and, finally, $\E$ must commute with the adjoint action of the Lie algebra $\F$ on the vector space $\F^\perp$.
By decomposing the elements $l(\sigma)$ of the phase space $L\dd_F$ in two ways as $l=k\tilde h$ and $l=\tilde kh$ and by eliminating the fields $\tilde h$ and $h$, we obtain, respectively, the $\sigma$-models and , where the linear operators $E$ and $\tilde E$ are obtained by decomposing the image of the operator $\E+{\rm Id}$ as (+[Id]{})={x +Ex,x}={ x +E x, x}.The reader can find the details of this procedure in the original paper [@KS96b].
In order to produce the affine (quasi-)Poisson T-duality scheme out from the formalism of the dressing cosets, we choose the Drinfeld double ${\mathbb D}$ of the following form D=D\_[\_l]{}D\_[\_r]{}where $D_{\e_l}$ and $D_{\e_r}$ are two Drinfeld doubles of the group $K$ defined by Eq.. The invariant split bilinear form $(.,.)_{{\rm Lie}(\dd)}$ is given by ((u\_l,u\_r),(v\_l,v\_r))\_[[Lie]{}()]{}:=(u\_l,v\_l)\_[\_[\_l]{}]{}+(u\_r,v\_r)\_[\_[\_r]{}]{}, u\_l,v\_l, u\_r,v\_rand the correct choice of the gauge Lie subalgebra $\F\subset$ Lie $(\mathbb D)$ is ={(x,x),x}.\[1177\]Detailing the elements $u_l\in{\D_{\e_l}}$ and $u_r\in{\D_{\e_r}}$ as u\_l=(x\_l,y\_l), u\_r=(x\_r,y\_r), x\_l,y\_l,x\_r,y\_r,we first rewrite as ={((x,0),(x,0)), x}and then we find the subspace $\F^\perp$: \^={((x\_l,y),(x\_r,-y)), x\_l,x\_r,y}.It remains to choose the operator $\E:\F^\perp\to\F^\perp$. In turns out, that the following one has all needed properties and does the job ((x\_l,y),(x\_r,-y)):=((y,(x\_l-x\_r)),(-y,(x\_r-x\_l))).How do we obtain the four $\sigma$-models , , and from the first order formulae and , knowing that the standard dressing cosets construction gives rise just to the two $\sigma$-models and ? Well, the key trick is to use here the construction known as the Poisson-Lie T-plurality [@KS97; @KS97b; @vU02] which amounts to the statement that there are as many dynamically equivalent $\sigma$-models with different geometries extracted from the first order formalism and , as is the number of maximally isotropic subgroups of the Drinfeld double $\dd$ not related by internal automorphisms. If there are just two maximally isotropic subgroups $K$ and $\tilde K$, then there are just two mutually dual $\sigma$-models and . However, our double $\dd=D_{\e_l}\times D_{\e_r}$ has more than two maximally isotropic subgroups; we shall actually need three of them: $\tilde K_L\times \tilde K_R$, $\tilde K_L\times K$ and $K\times \tilde K_R$, where $\tilde K_L\subset D_{\e_l}$ and $\tilde K_R\subset D_{\e_r}$ are the maximally isotropic subgroups of the respective doubles $D_{\e_l}$ and $D_{\e_r}$ (their Lie algebras are described by Eq. ).
Following the general dressing cosets construction [@KS96b], the triple of the pairwise T-dual $\sigma$-models have for their respective targets the double cosets $F\backslash \dd/(\tilde K_L\times \tilde K_R)$, $F\backslash \dd/(\tilde K_L\times K)$ and $F\backslash \dd/(K\times \tilde K_R)$ or, equivalently, $F\backslash (K\times K)$, $F\backslash (K\times \tilde K_R)$ and $F\backslash (\tilde K_L\times K)$. Furthermore, the gauge group $F$ is isomorphic to $K$, therefore the targets $F\backslash (K\times \tilde K_R)$ and $F\backslash (\tilde K_L\times K)$ become simply $\tilde K_R$ and $\tilde K_L$ and a straightforward computation shows that the $\sigma$-model living on those targets are nothing but the $\sigma$-models and . Concerning the remaining target $F\backslash (K\times K)$, we can fix the gauge in two different ways: either we obtain the target $\{e_K\}\times K$ or $K\times \{e_K\}$, where $e_K$ stands for the unit element of the group $K$. Those two gauge choices turn out to result at the $\sigma$-models and living on the target $K$. All in all, the affine (quasi-)Poisson T-duality of the four $\sigma$-models , , and fits into the dressing cosets construction.
Renormalisation group flow and the Drinfeld twist operators
===========================================================
It was discovered in [@VKS; @SfS] that the ultraviolet corrections to the actions of the Poisson-Lie $\sigma$-models and can be absorbed by appropriate redefinitions of the linear operators $E$ and $\tilde E$. Moreover, those redefinitions respect the T-duality between the $\sigma$-models. Said in other words, the renormalisation group flow of the operators $E$ and $\tilde E$ calculated separately from the models and respects the duality condition requiring that $E$ is inverse to $\tilde E$. Actually, the RG flow of the pair of the mutually dual $\sigma$-models and can be calculated also from the duality invariant data $(LD,\omega_{LD}, Q_\E)$ introduced in Section 2. The quantity which flows in the duality invariant description is the subspace $\E$, and this flow was described by the following elegant formula derived in [@SST]: =k(\_[AC]{}\_[BF]{}-\_[AC]{}\_[BF]{})(\^[KD]{}\^[HE]{}-\^[KD]{}\^[HE]{})f\_[KH]{}\^[ C]{}f\_[DE]{}\^[ F]{}. \[Sf\]Here $s$ is the flow parameter, $k$ a constant, the capital Latin indices refer to the choice of a basis $T_A$ in the Lie algebra $\D$: \_[AB]{}:=(T\_A,T\_B)\_, \_[AB]{}:=(T\_A,T\_B)\_,=f\_[AB]{}\^[C]{}T\_Cand they are respectively lowered and raised with the help of the tensor $\eta_{AB}$ and its inverse. Recall also, that $\E$ is the self-adjoint linear operator $\E:\D\to\D$ which has the subspace $\E$ as the eigenspace for the eigenvalue $+1$ and the orthogonal complement subspace $\E^\perp$ as the eigenspace for the eigenvalue $-1$.
Up to an irrelevant normalization constant, the flow formula can be cast in the basis-independent way as follows: =¶\_+\[\[¶\_+,¶\_-\]\]¶\_-+¶\_-\[\[¶\_+,¶\_-\]\]¶\_+,\[flo\]where the operators $\P_\pm$ are defined as ¶\_=(1),and the double bracket $[[.,.]]:S^2\D\times S^2\D\to S^2\D$ is defined on the symmetric product $S^2\D$ as :=\[A’,B’\].Here we use the Sweedler notation $A=A'\otimes A''$, $B=B'\otimes B''$ and we view the self-adjoint operators $\P_\pm$ as the elements of $S^2\D$ in the sense of the formula ¶\_x:=¶\_’(¶”\_,x)\_, x.
We are now going to show that the affine Poisson T-duality based on the affine Poisson bracket is compatible with the renormalisation group flow if we choose for the Yang-Baxter operators $R^r,R^l$ any pair of the Drinfeld twist operators studied in Section 5.4[^6]. In other words, we show in the present section, that the subspaces $\E_L,\E_R\in\D_0$ flow in a compatible way. We recall the context: we consider the compact simple Lie algebra $\K$ equipped with its Killing-Cartan form $(.,.)_\K$ and for its Drinfeld double we take $\D_\e$ for $\e=0$ (cf. Eq.). We pick the Cartan subalgebra $\gH\in\K$ and we consider the subspace $\gH^\perp\subset \K$ which is perpendicular to $\gH$ with respect to the Killing-Cartan form $(.,.)_\K$. We recall that any skew-symmetric operator $R:\K\to\K$ is the Drinfeld twist operator, if $\gH^\perp\subset {\rm Ker}(R)$ and Im$(R)\subset\gH$. Any Drinfeld twist operator verifies the Yang-Baxter condition for $\e=0$ because of the commutativity of the Cartan subalgebra.
Let us solve the flow equation with the initial conditions $\E_L$ and $\E_R$ introduced, respectively, in Eqs. and . We are going to argue that the equation implies the following simple flow of the subspaces $\E_L$ and $\E_R$ : \_L(s)= {(a(s)y+R\^ry,y),y},\[elt\] \_R(s)= {(a(s)y+R\^ly,y),x},\[ert\]notably, we remark that only the parameter $a$ flows and its dependence $a(s)$ on the flow parameter $s$ is the same for the case of the initial conditions $\E_L$ as well as $\E_R$.
Let us first see that the flow $\E_L(s)$ given by Eq. fulfils, for a suitable choice of the function $a(s)$, the equation with the initial condition $\E_L(0)=\E_L$. For that, pick a basis $t_i$ on the compact simple Lie algebra $\K$ such that (t\_i,t\_j)\_=-\_[ij]{}\[nr’\]and then choose the following basis $E_i(s)$ and $E_i^\perp(s)$ of the subspaces $E_{L}(s)$ and $E_L^\perp(s)$ respectively: E\_i(s)=(a(s)t\_i+R\^rt\_i,t\_i), E\_i\^(s)=(-a(s)t\_i+R\^rt\_i,t\_i).\[du2\]Note that it holds (E\_i(s),E\_j(s))\_=\_[ij]{}, (E\_i\^(s),E\_j\^(s))\_=-\_[ij]{}, (E\_i(s),E\_j\^(s))\_=0which means that the operators $\E_L(s),\P_\pm(s) :\D\to\D$ viewed as the elements of $S^2\D$ can be written as \_L(s)=E\_i(s)E\_i(s)+E\_i\^(s)E\_i\^(s),¶\_+(s)=E\_i(s)E\_i(s), ¶\_-(s)=-E\_i\^(s)E\_i\^(s)\[ii\]and the Einstein summation convention holds.
By differentiating Eqs. , we find =-E\_i\^, =-E\_i,hence =-(E\_iE\_i\^+E\_i\^E\_i)\[lhs\]Now we use Eqs. and calculate the right-hand-side of the flow equation ¶\_+\[\[¶\_+,¶\_-\]\]¶\_-+¶\_-\[\[¶\_+,¶\_-\]\]¶\_+=(E\_i,\[E\_k,E\^\_l\])\_(E\_j\^,\[E\_k,E\^\_l\])\_(E\_j\^E\_i+E\_iE\_j\^).\[bof\]We find from , and (E\_i,\[E\_k,E\^\_l\])\_= -(t\_i,\[t\_k,t\_l\])\_-(t\_i,\[t\_k,R\^rt\_l\])\_,(E\_j\^,\[E\_k,E\^\_l\])\_=-(t\_j,\[t\_k,t\_l\])\_+(t\_j,\[R\^rt\_k,t\_l\])\_.We infer from the properties of the Drinfeld twist operators the following identities (t\_j,\[R\^rt\_k,t\_l\])\_(t\_i,\[t\_k,R\^rt\_l\])\_=0,-\[t\_k,\[R\^rt\_k,t\_j\]\]=0,which makes possible to rewrite Eq. as ¶\_+\[\[¶\_+,¶\_-\]\]¶\_-+¶\_-\[\[¶\_+,¶\_-\]\]¶\_+=-(t\_i,\[t\_k,t\_l\])\_(t\_j,\[t\_k,t\_l\])\_(E\_j\^E\_i+E\_iE\_j\^).\[bof’\]We have from the normalization condition as well as from the ad-invariance of the bilinear form $(.,.)_\K$: (t\_i,\[t\_k,t\_l\])\_(t\_j,\[t\_k,t\_l\])\_=-(\[t\_i,t\_k\],\[t\_j,t\_k\])\_=(t\_i,\[t\_k,\[t\_k,t\_j\]\])\_=-c(t\_i,t\_j)\_=c\_[ij]{},where $c$ is the value of the quadratic Casimir in the adjoint representation. The last relation allows us to conclude that ¶\_+\[\[¶\_+,¶\_-\]\]¶\_-+¶\_-\[\[¶\_+,¶\_-\]\]¶\_+=- (E\_i\^E\_i+E\_iE\_i\^).\[rhs\]Comparing the evaluation of the left-hand-side of the flow equation with the evaluation of the right-hand-side , we find the complete agreement provided it holds =, \[fa\]which is the flow equation for the parameter $a$. This means that the flowing subspace $\E_L(s)$ given by Eq. is the solution of the flow equation if the function $a(s)$ fulfils the differential equation . The remarkable thing is that the exactly analogous calculation with $R^r$ replaced by $R^l$ yields the result that also $\E_R(s)$ given by Eq. is the solution of the flow equation if the function $a(s)$ fulfils the same differential equation . This fact completes the proof of the compatibility of the affine Poisson T-duality with the renormalisation group flow for the case where the Yang-Baxter operators $R^l$ and $R^r$ are the Drinfeld twist operators.
Outlook
=======
It would be interesting to find out whether there exists examples of the affine Poisson or of the affine quasi-Poisson T-duality which would be based on more general Drinfeld doubles than those defined by the commutator . A way to solve this problem would consist in careful inspection of possible affine (quasi)-Poisson structures on an arbitrary Lie group with the goal to determine whether they admit solutions of the sufficient conditions for the duality formulated in [**Assertion 2**]{} of Section 5.2 and in [**Assertion 3**]{} of Section 7.2. Another possibility would be to use the dressing coset construction for more general Drinfeld doubles.
The relation of the affine quasi-Poisson T-duality to the double field theory in the spirit of Ref. [@Ha] seems also to be an interesting issue to work out.
[99]{}
[^1]: The formulae and are valid for the so called [*perfect*]{} Drinfeld doubles for which there exists a diffeomorphism $\Upsilon: D\to K\times \tilde K$ composing to identity with the group multiplication map $m: K\times \tilde K\to D$, i.e. it must hold that $m\circ \Upsilon$ is the identity map from $D$ to $D$. In the present paper, we shall consider only the doubles $D$ verifying this property.
[^2]: Some overall signs in the definitions and are conventional. In particular, we have flipped the signs of $\pi^{\ma}$ and of $\pi^R$ with respect to Ref. [@K07] in order to have the uniform positive sign in front of all the Poisson structures appearing in the formulae ,, and .
[^3]: The formula may suggest that adding any left-invariant $m$-term to a Poisson-Lie structure produces the affine Poisson structure but this is not true. In fact, the $m$-term has to be such that the affine Poisson structure be indeed Poisson, that is, the corresponding bracket of functions has to satisfy the Jacobi identity. An explicit condition for the good $m$-term can be written down but we do not need to know it for the T-duality story and the interested reader can find it e.g. in Ref.[@K07].
[^4]: A Lie group $K$ is said quadratic if its Lie algebra $\K$ is quadratic, that is, if it exists an ad-invariant symmetric non-degenerate bilinear form on $\K$.
[^5]: It has to be stressed that the Drinfeld doubles $\D_L$ and $\D_R$ need not to be respectively isomorphic to the quasi-Drinfeld doubles $\D^q_L$ and $\D^q_R$ of the Lie quasi-bialgebras $\K_L$ and $\K_R$.
[^6]: Very recently, there appeared Ref. [@SV] in which it is claimed that the renormalization group flow is compatible with T-duality for every dressing coset. The contents of the present section can be therefore interpreted as an illustration of this general fact.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Deep neural networks (DNNs) have recently achieved great success in a multitude of classification tasks. Ensembles of DNNs have been shown to improve the performance. In this paper, we explore the recent state-of-the-art DNNs used for image classification. We modified these DNNs and applied them to the task of acoustic scene classification. We conducted a number of experiments on the TUT Acoustic Scenes 2017 dataset to empirically compare these methods. Finally, we show that the best model improves the baseline score for DCASE-2017 Task 1 by 3.1% in the test set and by 10% in the development set.'
address: |
Seernet Technologies, LLC\
{venkatesh.duppada, sushant.hiray}@seernet.io
bibliography:
- 'refs.bib'
title: Ensemble of deep neural networks for acoustic scene classification
---
Deep learning, Convolutional Neural Networks, Recurrent Neural Networks, Ensemble
Introduction {#sec:intro}
============
Acoustic scene classification(ASC) [@barchiesi2015acoustic] focuses on recognizing various environmental sounds from a sound sample. The broader category of Computational Auditory Scene Analysis aims to model the human auditory system and its mechanisms to detect, identify, separate and segregate sounds in the same way that humans do[@wang2006computational]. Amongst the existing acoustic classification tasks, ASC is a challenging task since various environmental sounds have similar background noises and span a wider range of frequencies. ASC is particularly interesting to researchers due to its wide applications in audio tagging [@cai2006flexible], audio indexing using wearable devices [@shah2012lifelogging], robot navigation systems [@chu2006scene] etc.
Image classification and detection have been remarkably popular in recent years. The popularity is attributed primarily to the availability of large annotated standard datasets [@deng2009imagenet]. Audio classification and detection have not attracted a similar level of attention. The DCASE challenge is a step towards creating a standard dataset for researchers. The 2013 iteration [@giannoulis2013detection] included challenges for scene classification and synthetic acoustic classification. The DCASE Challenge 2016 [@heittola2016dcase2016] comprised of 4 tasks: ASC, Acoustic Event Detection, Sound Event Detection in Real Life Audio and Domestic Audio Tagging. The DCASE Challenge 2017 [@DCASE2017challenge] comprises of 4 tasks: ASC, detection of rare sound events, sound event detection in real life audio, large scale weakly supervised sound event detection for smart cars.
In this work, we focus on the challenge 1 of DCASE Challenge 2017. The goal of ASC is to classify a test recording into one of the provided 15 predefined classes that characterizes the environment in which it was recorded. This challenge is in continuation of the previous year’s challenge with additional data.
Early work in this area focused primarily on using classifiers such as GMM-HMM [@chum2013ieee], tree bagger classifiers [@olivetti2013wonders], support vector machines [@geiger2013large]. Some of these classifiers cannot effectively model temporal dynamics of audio. With the recent advancements in deep learning, many new DNN architectures have been studied which are better at encoding the temporal nature. Also DNNs are better at abstracting the large feature sets, usually associated with audio clips. Quite a few teams experimented with various deep learning approaches in the previous challenge. The challenge winners[@eghbal2016cp] for DCASE 2016 extracted MFCC and i-vectors and used Deep Convolutional Neural Networks (DCNN) to improve the performance over conventional baseline model. [@valenti2016dcase] proposed a vanilla DCNN with a fully connected layer at the end, whereas [@lidy2016cqt] pre-processed the audio clips using CQT before using a DCNN. [@bae2016acoustic] experimented with a parallel combination of CNN and LSTM for the final classification.
In this work, we apply various state-of-the-art DCNNs from image classification tasks. We modified these architectures to apply them on the audio dataset for ASC. Specifically, we use the following architectures: (1). LeNet (2). SqueezeNet (3). 1-D CNN. We also tried some deeper networks such as: Highway Networks, Densely Connected CNN. However, due to the lack of availability of a large annotated dataset, these variants didn’t perform well. We also experimented with multiple types of feature extractors: extracting log-mel spectrograms by varying the hop-lengths and the sampling frequency. We compare the results from these DCNN models with the baseline DNN model. Finally, we also show that creating an ensemble of these networks gives a much better performance on the fore mentioned task, effectively improving the baseline model’s accuracy by 10% with the final macro accuracy being 84.8%.
Data {#sec:data}
====
In this work, we use **TUT Acoustic Scenes 2017** dataset [@Mesaros2016_EUSIPCO] for development. The dataset is a collection of recordings from various acoustic scenes all from distinct locations. For each recording location 3-5 minute long audio recordings are captured and are split into 10 seconds which act as unit of sample for this task. All the audio clips are recorded with 44.1 kHz sampling rate and 24 bit resolution. Readers can find more information about the data and how it is collected from here[^1]. The development dataset comprises of 52 minutes of audio for each label amounting to 13 hours of entire dataset. On the contrary, the 2016 dataset comprised of over 9.75 hours of development dataset. Apart from the increase in dataset size, the datasets also differ in their classification units. The 2017 dataset has a classification unit of 10 seconds whereas the 2016 dataset has 30 seconds. These differences need to be observed before comparing the results from the 2017 edition with those of 2016 edition.
System Architecture {#sec:system_description}
===================
![Audio Processing Pipeline[]{data-label="fig:results"}](DCASEBlock){width="\columnwidth"}
For each audio clip, (from development and evaluation set), the processing pipeline consists of the following steps: 1) Pre-process the audio with various transformations to standardize the input, 2) Extract features from the pre-processed audio, 3) Train the DCNN models using the features from step 2, 4) To get the final classification for each file, we fuse the predictions from each segment.
In the following subsections, we explore each of these steps.
Audio Preprocessing {#ssec:preprocessing}
-------------------
Before being passed to the feature extractor, a few preprocessing steps are carried out on the original audio. In this task, the audio files have two sets of channels (one for left ear and one for right ear), so we convert the audio to mono by averaging the channels. We also normalize the amplitude to lie between -1 and 1.
Feature Extraction {#ssec:features}
------------------
Spectrograms provide a visual way of representing the signal strength over time at various frequencies. Mel-Spectrograms maps the equally spaced spectrogram frequencies into bins according to human ear perception, hence using mel-spectrogram as input has shown lot of success in various tasks like speech recognition, speaker identification etc. In this task, we experiment with two variants of feature vectors:
1. The original audio was downsampled to 16 kHz. We compute log mel-spectrogram with 0.025 seconds as window length and 0.010 seconds as hop length with 64 mel frequencies. The above processing results in log mel-spectrogram with 999 frames and 64 mel frequency bins. We then splits these into non overlapping windows of 111 frames which acts as input to our neural networks. Hence, we effectively split the raw audio into 9 segments.
2. In this variant, the audio was kept in its original sampling frequency of 44.1 kHz. We compute log mel-spectrogram with 0.046 seconds as window length and 0.023 seconds as hop length with 64 mel frequencies. The above processing results in log mel-spectrogram with 431 frames and 64 mel frequency bins. We then splits these into non overlapping windows of 43 frames which acts as input to our neural networks. Hence, we effectively split the raw audio into 10 segments.
Classification
--------------
In this work, we apply a variety of DCNNs. They are described in detail in Section \[sec:cnn\]. Once the feature vectors are extracted from the audio file, it is split into smaller segments as described in sub-section \[ssec:features\]. Each model is trained using mini-batches of size 256. We used Adadelta optimizer [@zeiler2012adadelta] with an initial learning rate of 1. Dropout layers significantly improve performance. Hence, we added dropout rate of 0.5. All our models were trained for 200 epochs, and the models with best validation accuracy were chosen.
Fusion
------
Since we split the raw audio clip into multiple segments, to get the final class prediction, we average predictions across all segments of an audio clip. This implies averaging 9 segments for the feature vector of type 1, and 10 segments for the feature vector of type 2 as mentioned in subsection \[ssec:features\].
Ensemble
--------
Finally, we create an ensemble using the model variants mentioned in Section \[sec:cnn\]. We trained 10 different models with 5 distinct architectures. We sort the models via the macro-accuracy computed using [@mesaros2016metrics]. Finally, we chose the best models which had the maximum variance, and whose accuracies were better than the baseline model.
We train our deep learning models with the Keras library [@chollet2015keras] using Tensorflow [@tensorflow2015-whitepaper] backend on cloud GPU’s with 60 GB RAM hosted on Floydhub. [^2]
Convolutional Neural Networks {#sec:cnn}
=============================
Convolutional Neural Networks is yet another DNN which gained popularity in image classification tasks. DCNNs are an extension to DNN with a few changes to the architecture. The major difference is the usage of convolutions. A simple DCNN comprises of a set of layers stacked together. These are namely convolutional, pooling, optional fully connected layers at the end, and finally the output layer.
- The convolutional layers consists of a kernel (also called as filter). These kernels perform non-linear operations over a limited receptive field. The kernel is tiled across the entire input space, resulting in the creation of a feature map. The primary motivation for this operation is that the function which the kernel learns are independent of the position in which they are found. Hence, we reduce the number of parameters required for the neural network. The parameters for this layer are: context of the kernel (width x height), depth of the kernel and the strides. A typical convolutional layer will consist of numerous kernels. The depth depends on the number of kernels present in the previous layer. In the case of 1D convolution, the context of the kernel is defined only by its width.
- The pooling layers are added to further reduce the dimensionality of the feature maps. The pooling layer essentially provide summaries over each context window, thus enhancing the network invariance to transitional shifts in the input patterns. Usually, pooling layers are placed after every convolution layer, however, in some DCNN variants, these are placed scantily. There are multiple types of pooling layers, max pooling, average pooling, global max pooling global average pooling, etc. The most salient feature of the pooling layer is that it doesn’t add any additional parameters in the network and in turn helps reduce the dimensionality of the data significantly. Similar to the conv layer, pooling layer also have a context and a stride. In most typical scenarios, the strides are configured such that we pool over non-overlapping windows.
This hierarchical structure consisting of alternating feature extraction layers and pooling layers allows CNNs to operate on multiple timescales. In the upcoming sub-sections we describe the different DCNN architectures we’ve experimented with. Since the annotated dataset for the competition was small in size, we’ve used DCNN architectures which require less number of parameters.
**LeNet** **SqueezeNet** **1D CNN**
---------------- --------------------- ---------------
8x3x3-BN-ReLu 64x3x3-BN-ReLu 64x5 -BN-ReLu
MP: 3x2 MP: 2x2 MP: 3
16x3x3-BN-ReLu FIRE(sq:16, ex:64) 128x5-BN-ReLu
MP: 3x2 FIRE(sq:16, ex:64) MP: 3
32x3x3-BN-ReLu MP: 2x2 256x5-BN-ReLu
Dropout: 0.5 FIRE(sq:32, ex:128) Dropout: 0.5
FC: 512 FIRE(sq:32, ex:128) FC: 512
Softmax: 15 MP: 2x2 Softmax: 15
FIRE(sq:48, ex:192)
FIRE(sq:64, ex:256)
Dropout: 0.5
15x1x1-BN-ReLu
GlobalAvgPool
: The various model architectures experimented in the paper and described in Section \[sec:cnn\]. Some abbreviations used are:- BN: Batch Normalization, MP: MaxPooling Layer, FC: Fully Connected Layer, FIRE: fire module described in \[ssec:squeeze\], sq: squeeze, ex: expand.[]{data-label="tab:model_config"}
LeNet
-----
We experimented with 3 variants of the LeNet architecture [@lecun1998gradient] by varying the filter sizes keeping everything else constant. We used the filter sizes of 3x3, 5x5, 7x7. A small building block of this architecture comprises of Conv-BN-Relu-Pool layer. We stacked 3 copies of the building blocks where the number of filters became twice of the previous block. Finally, we flatten the layer, add a fully connected layer of size 512 and use softmax for the final class label. The architecture is described in Table \[tab:model\_config\] for filter size: 3x3.
SqueezeNet {#ssec:squeeze}
----------
SqueezeNet [@SqueezeNet] is a DCNN which achieves AlexNet-level accuracy on ImageNet with 50x fewer parameters. SqueezeNet comprises of multiple stacked building blocks called Fire Modules. A Fire module is comprised of: a squeeze convolution layer (which has only 1x1 filters), feeding into an expand layer that has a mix of 1x1 and 3x3 convolution filters. The original SqueezeNet architecture comprises of 8 stacked Fire Modules where the number of filters keeps increasing as the depth increases. We modified the existing architecture by reducing the depth and stacking 6 Fire Modules instead. We experimented with various FIRE module configurations, the best configuration is described in Table \[tab:model\_config\].
1D CNN
------
While applying CNN to audio dataset, it is important to understand that the two axes of the spectrogram have different meanings (time v/s frequency), which is essentially not the case for images. Hence, technically it makes sense to convolve only in a single dimension as opposed to both the dimensions in case of images. Hence, in this architecture we experiment with convolution in only single dimension (time). This architecture is inspired from [@NIPS2013_5004] [^3] with the major variation being, in the penultimate layer that variation uses various statistics, we used the entire layer and added a single hidden layer. This was primarily because we didn’t want to overfit with large number of parameters.
Results and Analysis {#sec:results}
====================
From feature extraction side we experimented with sampling rate, window length and hop length when converting audio to log mel-spectrogram. If the window length is too big, frequency resolution with respect to time is reduced. If the window is too small, frequency patters across time are missed. This is a trade-off, one can have high a resolution in time or a high resolution in frequency but not both. So, we experimented with various window and hop lengths. The human ear can listen to frequencies between 0Hz and 5kHz clearly without any effort, so, it is a common practice to downsample the audio. Hence, we experimented with the original as well as the downsampled audio files.
On the model architecture side, we experimented with LeNet, SqueezeNet and 1D CNNs with different kernel sizes. In LeNet we used filters of size 3x3, 5x5 and 7x7 to understand how filter size affects the classification accuracy.
The model **CNN-V1** is trained on log mel-spectrograms of audio files without down-sampling with 3x3 conv filters whereas all the other models are trained on log mel-spectrograms of down-sampled (16kHz) audio files. While training we observed that the model overfits very quickly when using actual files because it contains lot of information which may not be relevant to task at hand results in poor accuracy. The models trained on down-sampled audio files are robust to overfitting and hence perform better.
The models **CNN-V2-1**, **CNN-V2-2** and **CNN-V2-3** are based on LeNet architecture, where the only difference is the size of the conv filters used. Conv filters of size 3x3, 5x5 and 7x7 are used in them respectively. The conv filter sizes affect the model performance. The filters which are either too big or too small cannot understand the time-frequency patterns. Hence, we experimented with various filter sizes and we can see that the classification accuracy deteriorates with the increase in the filter size. Due to recent success of **SqueezNet** with very less number of parameters we tried this architecture out of curiosity.
We finally tried **1D CNN** because a spectrogram is not like an image, it represents frequency in one axis and time in one axis. Even though we can use it like an image sometimes model may benefit from treating them separately.
After training all the models we create an ensemble by choose best three models and combined their individual predictions using geometric mean. Since the models are learning different because of their architectures (3x3 vs 5x5 etc) or input features (16kHz-44.1kHz log mel-spectrogram) we see a improvement in accuracy of about 3 percent compared to best individual model.
We plotted confusion matrix in figure \[fig:conf\_matrix\] to analyze the conflicting acoustic scenes. The most confused pairs are: **Home** is confused with **Library**, and **Park** is confused with **Residential area**. We can clearly see why the model is struggling to identify these pairs because these scenes are closely related to one another acoustically.
This can be reduced by increasing the audio sample length from 10 seconds so that we have some more context to identify these pairs correctly. Among the most accurately predicted acoustic scenes there are **office**, **car**, **city\_center**. These scenes have unique fingerprint to them and could be accurately classified. On the other hand, **park**, **cafe restaurant**, **residential\_area** are least accurately predicted probably due to acoustic closeness of these scenes.
**Acoustic Scene** **Base** **CNN-V1** **CNN-V2-1** **CNN-V2-2** **CNN-V2-3** **SqueezeNet** **1D CNN** **Ensemble**
-------------------- ------------- ------------- -------------- -------------- -------------- ---------------- ------------ --------------
Beach 75.3 - 40.7 87.2 - 13.0 83.3 - 8.3 83.7 - 10.2 85.3 84.6 75.0 87.5 - 13.9
Bus 75.3 - 38.9 93.9 - 35.2 84.9 - 39.8 92.9 - 49.1 86.9 85.9 84.3 90.1 - 42.6
Cafe / Restaurant 57.7 - 43.5 66.0 - 51.9 72.1 - 57.4 64.4 - 45.4 64.1 53.2 55.8 72.8 - 57.4
Car 97.1 - 64.8 94.6 - 88.0 97.8 - 96.3 94.2 - 77.8 97.4 93.6 97.8 98.4 - 85.2
City center 90.7 - 79.6 82.1 - 85.2 90.1 - 75.9 91.7 - 89.8 93.6 84.9 83.7 93.6 - 85.2
Forest path 79.5 - 85.2 85.9 - 86.1 87.5 - 88.0 93.6 - 85.2 91.3 82.4 76.0 91.7 - 87.0
Grocery store 58.7 - 49.1 79.2 - 52.8 90.1 - 58.3 90.1 - 54.6 85.6 87.2 89.4 92.9 - 57.4
Home 68.6 - 76.9 70.1 - 68.5 74.5 - 79.6 74.2 - 81.5 69.5 77.7 72.0 74.8 - 83.3
Library 57.1 - 30.6 81.7 - 25.0 81.7 - 34.3 85.9 - 38.9 78.8 73.7 77.2 87.2 - 35.2
Metro station 91.7 - 93.5 93.6 - 28.7 69.9 - 23.1 71.8 - 97.2 65.1 67.9 69.9 86.5 - 63.9
Office 99.7 - 73.1 92.6 - 72.2 99.7 - 86.1 99.0 - 94.4 99.0 97.8 98.1 98.4 - 88.9
Park 70.2 - 32.4 60.6 - 35.2 70.8 - 40.7 62.8 - 25.0 59.9 56.1 60.9 66.3 - 31.5
Residential area 64.1 - 77.8 70.2 - 82.4 70.8 - 78.7 75.6 - 80.6 67.9 72.4 69.6 73.4 - 81.5
Train 58.0 - 72.2 61.9 - 71.3 74.4 - 74.1 66.3 - 75.0 67.3 42.0 55.1 75.3 - 72.2
Tram 81.7 - 57.4 78.5 - 60.2 80.4 - 57.4 77.9 - 56.5 81.1 77.2 82.1 83.3 - 60.2
Average Accuracy 74.8 - 61.0 79.9 - 57.0 81.9 - 59.9 81.6 - 64.1 79.5 75.8 76.4 84.8 - 63.0
![Confusion Matrix[]{data-label="fig:conf_matrix"}](confusion_matrix.pdf){width="\columnwidth"}
Future Work & Conclusion
========================
Currently, the training suffers because of over fitting due to lack of enough data. In future we’ll try to avoid this using data augmentation methods as these methods have shown to boost the performance and prevent over-fitting in [@han2016acoustic] [@salamon2017deep]. In this work, we experimented with the window length and the hop length of mel-spectrogram to study its affect on acoustic scene classification. We also tried different DCNN architectures for the same and finally created an ensemble of these deep CNN models to improve upon the baseline model by 10%.
[^1]: <http://www.cs.tut.fi/sgn/arg/dcase2017/challenge/task-acoustic-scene-classification>
[^2]: https://www.floydhub.com/
[^3]: http://benanne.github.io/2014/08/05/spotify-cnns.html
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'I review some recent progress made in our understanding of galaxy evolution and the cosmic history of star formation. The [*Hubble Deep Field*]{} (HDF) imaging survey has achieved the sensitivity to capture the bulk of the extragalactic background light from discrete sources. No evidence is found in the optical number-magnitude relation down to $AB=29$ mag for a large amount of star formation at high redshifts. A census of the ultraviolet and blue “dropouts”, which requires the inclusion of the effects of intergalactic attenuation on the colors of cosmologically distant galaxies, appears to confirm this basic conclusion. The emission history of the universe at ultraviolet, optical, and near-infrared wavelengths can be modeled from the present epoch to $z\approx 4$ by tracing the evolution with cosmic time of the galaxy luminosity density, as determined from several deep spectroscopic samples and the HDF. The global spectrophotometric properties of field galaxies are well fitted by a simple stellar evolution model, defined by a time-dependent star formation rate (SFR) per unit comoving volume and a universal initial mass function which is relatively rich in massive stars. The SFR density is found to rise sharply, by about an order of magnitude, from a redshift of zero to a peak value at $z\approx 1.5$ in the range 0.12–0.17 $\sfrd$, to fall again by a factor of 2 (4) out of a redshift of 3 (4). Since only 10% of the current stellar content of galaxies is produced at $z>2.5$, a rather low cosmic metallicity is predicted at these early times, in good agreement with the observed enrichment history of the damped Lyman-$\alpha$ systems. The biggest uncertainty is represented by the poorly constrained amount of starlight that was absorbed by dust and reradiated in the IR at early epochs. A “monolithic collapse” scenario, where half of the present-day stars formed at $z>2.5$ and were shrouded by dust, can be made consistent with the global history of light, but appears to overpredict the metal mass density at high redshifts.'
author:
- Piero Madau
title: The Hubble Deep Field and the Early Evolution of Galaxies
---
\#1[to 0pt[\#1]{}]{}
Introduction
============
Much observing time has been devoted in the past few years to the problem of the detection of galaxies at high redshifts, as it was anticipated that any knowledge of their early luminosity and color evolution would set important constraints on the history of structure and star formation in the universe. While it has become now clear that blank-sky surveys for strong -emitting primeval galaxies are not particularly efficient (Djorgovski & Thompson 1993), the method of obtaining multicolor broadband observations of the emitter’s rest-frame UV and optical stellar continuum has been successfully applied to select galaxies at cosmological distances in ground-based surveys (Steidel & Hamilton 1992; Steidel 1996a; see review by Pettini in this proceedings) and in the [*Hubble Deep Field*]{} (HDF) (Madau 1996, hereafter M96; Steidel 1996b; Lowenthal 1997). Together with the tremendous progress in our understanding of faint galaxy data at $z\lta 1$ made possible by the recent completion of several comprehensive ground-based spectroscopic surveys (Lilly 1995; Ellis 1996; Cowie 1996), the identification of star-forming galaxies at $2\lta z\lta 4$ has provided new clues to some key questions of galaxy formation and evolution studies: Is there a characteristic epoch of star and metal formation in galaxies? What fraction of the luminous baryons observed today were already locked into galaxies at early epochs? Are high-$z$ galaxies obscured by dust? Do spheroids form early and rapidly? Is there a “global” IMF? Through the systematic study of galaxies at increasing cosmological lookback times it has become possible to reconstruct the history of stellar birthrate [*directly*]{}, as opposite to the “classical” method where one studies the resolved stellar populations of the Milky Way and nearby galaxies and infers their evolutionary history from fossil records – well-known examples of this more traditional approach are nuclear cosmochronology, the color-magnitude diagram of globular clusters, the cooling sequence of white dwarfs (see Renzini 1993 and references therein).
In this talk I will review the broad picture that has recently emerged from the “direct” method, focusing on the emission properties at ultraviolet, optical, and near-IR wavelengths of the galaxy population [*as a whole*]{}. I will show how the combination of HST deep imaging and ground-based spectroscopy offers now an exciting first glimpse to the history of the conversion of neutral gas into stars in field galaxies. In the following, all magnitudes will be given in the AB system, and a flat cosmology with $q_0=0.5$ and $H_0=50\,$km s$^{-1}$ Mpc$^{-1}$ will be adopted.
The Hubble Deep Field
=====================
As the best view to date of the optical sky at faint flux levels, the HDF imaging survey has rapidly become a key testing ground for models of galaxy evolution. The field, an undistinguished portion of the northen sky at high galactic latitudes, was imaged for approximately 150 orbits from 18 to 30 December 1995 with the Wide Field Planetary Camera onboard the [*Hubble Space Telescope*]{}. With its depth – reaching 5-$\sigma$ limiting AB magnitudes of roughly 27.7, 28.6, 29.0, and 28.4 in $U, B, V,$ and $I$ [^1] (Williams 1996) – and four-filter strategy in order to detect Lyman-break galaxies at various redshifts, the HDF offers the opportunity to study the galaxy population in unprecedented detail.
Galaxy Counts and the EBL
-------------------------
There are about 3,000 galaxies in the HDF, corresponding to $2\times 10^6$ deg$^{-2}$ to $V\approx 29$ mag. The galaxy counts as a function of AB isophotal magnitude are shown in Figure 1 in the F300W, F450W, F606W, and F814W bandpasses (the number corresponds to the central wavelength in nm) for all galaxies with signal-to-noise ratio $S/N>3$ within the band (Williams 1996). A compilation of existing ground-based data is also shown, together with the predictions of no-evolution models, i.e. models in which the absolute brightness, volume density, and spectra of galaxies do not change with time. The HDF counts are plotted down to the 80% completeness limit and are found to agree reasonably well with previous surveys, to within $20\%$ in the magnitude range $22<AB<26$. In all four bands, the slope $\alpha$ of the differential galaxy counts, $\log N(m)=\alpha m$, flattens at faint magnitudes, e.g., from $\alpha=0.45$ in the interval $21<B<25$ to $\alpha=0.17$ for $25<B<29$. This feature cannot be due to the reddening of distant sources as their Lyman break gets redshifted into the blue passband, since the fraction of Lyman-break galaxies at $B\sim 25$ is only of order 10% (cf. Guhathakurta 1990). Moreover, an absorption-induced loss of sources could not explain the similar flattening of the number-magnitude relation observed in the $V$ and $I$ bands. Rather, the change of slope suggests a decline in the surface density of luminous galaxies beyond $z\sim 1.5$.
The contribution of known galaxies to the extragalactic background light (EBL) – an indicator of the total optical luminosity of the universe – can be calculated directly by integrating the emitted flux times the differential galaxy number counts down to the detection threshold. The leveling off of the counts is clearly seen in Figure 2, where the function $i_\nu=10^{-0.4(m+48.6)}\times N(m)$ is plotted against apparent magnitude in all bands (Pozzetti 1997). While counts having a logarithmic slope of $\alpha\ge0.40$ continue to add to the EBL at the faintest magnitudes, it appears that the HDF survey has achieved the sensitivity to capture the bulk of the extragalactic light from discrete sources (an extrapolation of the observed counts to brighter and/or fainter magnitudes would typically increase the sky brightness by less than 20%). To $AB=29$, the sky brightness from resolved galaxies in the $I$-band is $\approx 2\times 10^{-20}\iunits$, increasing roughly as $\lambda^2$ from 2000 to 8000 Å. The flattening of the number counts has the interesting consequences that the galaxies that produce $\sim 60\%$ of the blue EBL have $B<24.5$. They are then bright enough to be identified in spectroscopic surveys, and are indeed known to have median redshift $\langle
z\rangle=0.6$ (Lilly 1995). The quite general conclusion is that there is no evidence in the number-magnitude relation down to very faint flux levels for a large amount of star formation at high redshift. Note that these considerations do not constrain the [*rate*]{} of starbirth at early epochs, only the total (integrated over cosmic time) amount of stars – hence background light – being produced, and [*neglect the effect of dust reddening*]{}.
Intergalactic Absorption
------------------------
A more direct way to track galaxy evolution at early epochs is through a census of the HDF “dropouts”. At faint magnitudes, the interpretation and detailed modeling of the observations require the self-consistent inclusion of the effect of intergalactic attenuation on galaxy colors. Absorption by intervening material has been known for quite some time to distort our view of objects at cosmological distances. It has been realized only recently, however, that the accumulated line-blanketing and Lyman-continuum absorption from the forest clouds and Lyman-limit systems along the path to high redshifts can be efficiently used to identify galaxies at $z\gta2$ (Madau 1995; Steidel & Hamilton 1992). Although other spectral features, such as the $4000\,$Å and $912\,$Å breaks which characterize the integrated spectra of stellar populations, with the latter possibly enhanced by self-absorption from interstellar gas within the galaxy itself, can and have been used as tracers of redshifts, the model predictions of their magnitude are sensitive to the unknown physical and evolutionary state of the galaxy, i.e., its star formation history, age, and distribution, and hence are subject to substantial uncertainties. By contrast, the “reddening” effect due to atomic processes in cosmological distributed QSO absorption systems is ubiquitous and can be reliably taken into account. While stochastic in nature, r.m.s. fluctuations away from the mean opacity are bound to be modest in most situations, due to the broadband nature of the adopted filter set.
The magnitude and cosmological importance of intergalactic absorption on galaxy spectra can be effectively illustrated by including its effect in the standard $k$-correction term needed to translate the galaxy magnitude at Earth into its rest-frame value, $$k(z)=-2.5\log\left[(1+z){L(\nu)\over L(\nu_{\rm obs})}\langle
e^{-\tau}\rangle \right],$$ where $L(\nu)$ is the specific power emitted by a source at redshift $z$, $\nu_{\rm obs}=\nu/(1+z)$, $\langle e^{-\tau}\rangle$ is the cosmic transmission averaged over all lines of sight, and we have assumed no intrinsic luminosity evolution. Figure 3 shows the $k$-correction in the $U$ and $B$ HDF bands as a function of redshift for synthetic spectra of galaxies which well reproduce the colors of present-day ellipticals and spirals. The effect at high redshift is huge. Due to intergalactic attenuation alone, the $k$-correction in the F300W bandpass increases by as much as 4 magnitudes between $z\approx 2$ and $z\approx 3.5$, giving origin to a “UV dropout”. In the F450W band, the increase is barely noticeable at $z\approx3$, but becomes very large above $z\approx 4$, thereby producing a “blue dropout”.
Ultraviolet and Blue Dropouts
-----------------------------
Ground-based observations have used color techniques which are sensitive to the presence of a Lyman-continuum break superposed to an otherwise flat UV spectrum to identify galaxies at $z\approx 3$ (Steidel & Hamilton 1992; Steidel 1996a). New photometric criteria for robustly selecting Lyman-break galaxies have been developed based on the HDF color system, providing what appear to be largely uncontaminated samples of star-forming galaxies at high redshifts (M96). The HDF ultraviolet passband – which is bluer than the standard ground-based $U$ filter – permits the identification of star-forming galaxies in the interval $2<z<3.5$ (Figure 4). Galaxies in this redshift range predominantly occupy the top left portion of the $\ub$ vs. $\bi$ color-color diagram because of the attenuation by the intergalactic medium and intrinsic extinction. Galaxies at lower redshift can have similar $U-B$ colors, but they are typically either old or dusty, and are therefore red in $B-I$ as well. Of order 100 (200) ultraviolet dropouts can be identified in the HDF which are brighter than $B=27$ ($B=29$), approximately 25% (20%) of the total. To date, about 25 among the brightest ones have spectroscopically confirmed redshift in the range $2.0<z<3.4$. The color-selection region is illustrated in Figure 5. The $UBI$ criteria isolate objects that have relatively blue colors in the optical, but a sharp drop into the UV. In analogous way, the blue passband allows the selection of candidate star-forming galaxies in the interval $3.5<z<4.5$. Only $\sim 20$ (60) $B$ dropouts down to $V=28$ ($V=30$) have been identified in the $\bv$ vs. $\vi$ plane (M96). The brightest one has recently been confirmed through deep Keck spectroscopy to be at $z=4.02$ (Dickinson 1997), consistent with the photometric predictions.
The UV continuum emission from a galaxy with significant ongoing star formation is entirely dominated by late-O/early-B stars on the main sequence, which have masses $\gta 10\msun$ and lifetimes $t_{MS}\lta
2\times 10^7\,$yr. After an initial transient phase where the UV flux rises rapidly and the turnoff mass drops below $10\msun$, a steady state is reached where the measured luminosity becomes proportional to the instantaneous SFR and independent of the past star formation history (see Madau, Pozzetti, & Dickinson 1997). Figure 6 depicts the present-epoch “SFR function” – which describes the number of star-forming galaxies as a function of their ongoing SFR –, together with the derived distribution of stellar birthrates for the HDF ultraviolet and blue dropouts. The comparison shows the sign of a significant density evolution – a rapid increase in $\phi_*$.
Evolution of the Galaxy Luminosity Density
==========================================
The integrated light radiated per unit volume from the entire galaxy population is an average over cosmic time of the stochastic, possibly short-lived star formation episodes of individual galaxies, and should follow a relatively simple dependence on redshift. In the UV – where it is proportional to the global star formation rate (SFR) – its evolution should provide information, e.g., on the mechanisms which may prevent the gas within virialized dark matter halos to radiatively cool and turn into stars at early times, or on the epoch when galaxies exhausted their reservoirs of cold gas. From a comparison between different wavebands it should be possible to set constraints on the average initial mass function (IMF) and dust content of galaxies.
The comoving luminosity density, $\rho_\nu(z)$, from the present epoch to $z\approx 4$ is shown in Figure 7 in five broad passbands centered around 0.15, 0.28, 0.44, 1.0, and 2.2 . The data are taken from the $K$-selected wide-field redshift survey of Gardner (1997), the $I$-selected CFRS (Lilly 1996) and $B$-selected Autofib (Ellis 1996) surveys, the photometric redshift catalog for the HDF of Connolly (1997) – which take advantage of deep infrared observations by Dickinson (1997) – and the color-selected UV and blue “dropouts” of M96 (see also Madau 1997). They have all been corrected for incompleteness by integrating over the best-fit Schechter function in each redshift bin, $$\rho_\nu(z)=\int_0^\infty L\phi(L,z)dL=\Gamma(2+\alpha)\phi_*L_*. \label{eq:ld}$$ As from the Connolly (1997) and M96 data sets it is not possible to reliably determine the faint end slope of the luminosity function, a value of $\alpha=-1.3$ has been assumed at each redshift interval for comparison with the CFRS sample (Lilly 1995). The error bars are typically less than 0.2 in the log, and reflect the uncertainties present in these corrections and, in the HDF $z>2$ sample, in the volume normalization and color-selection region. In the $K$-band, the determination by Gardner (1997) agrees to within 30% with Cowie (1996), and we have assigned an error of 0.1 in the log to the estimate of the local luminosity density at 2.2 . Despite the obvious caveats due to the likely incompleteness in the data sets, different selection criteria, and existence of systematic uncertainties in the photometric redshift technique, the spectroscopic, photometric, and Lyman-break galaxy samples appear to provide a remarkably consistent picture of the emission history of field galaxies. The UV luminosity density rises sharply, by about an order of magnitude, from a redshift of zero to a peak at $z\approx
1.5$, to fall again by a factor of 2 (4) out to a redshift of 3 (4) (M96; Lilly 1996; Connolly 1997). This points to a rapid drop in the volume-averaged SFR in the last 8–10 Gyr, and to a redshift range $1\lta z\lta
2$ in which the bulk of the stellar population was assembled. The decline in brightness at late epochs is shallower at longer wavelengths, as galaxies becomes redder with cosmic time, on the average.
Stellar Population Synthesis Modeling
=====================================
Stellar population synthesis has become a standard technique to study the spectrophotometric properties of galaxies. In the following, I will make extensive use of the latest version of Bruzual & Charlot (1993) isochrone synthesis code, optimized with an updated library of stellar spectra (Bruzual & Charlot 1997), to predict the time change of the spectral energy distribution of a stellar population. The uncertanties linked to the underlying stellar evolution prescriptions and the lack of accurate flux libraries do not typically exceed 35% (Charlot, Worthey, & Bressan 1966). I will consider three possibilities for the IMF, $\phi(m)\propto m^{-1-x}$: a Salpeter (1955) function ($x=1.35$), a Scalo (1986) function, which is flatter for low-mass stars and significantly less rich in massive stars than Salpeter, and an intermediate case with $x=1.7$. In all models the metallicity is fixed to [*solar*]{} values and the IMF is truncated at 0.1 and 125 $\msun$.
An interesting question now arises as to whether a simple stellar evolution model, defined by a time-dependent SFR per unit volume and a constant IMF, may reproduce the global UV, optical, and near-IR photometric properties of the universe as given in Figure 7. In a stellar system with arbitrary star formation rate, the luminosity density at time $t$ is given by the convolution integral $$\rho_\nu(t)=\int^t_0 L_\nu(\tau)\times {\rm SFR}(t-\tau)d\tau,
\label{eq:rho}$$ where $L_\nu(\tau)$ is the specific luminosity radiated per unit initial mass by a generation of stars with age $\tau$. In the instantaneous recycling approximation (Tinsley 1980), the total stellar mass density produced at time $t$ is $$\rho_s(t)=(1-R)\int_0^t {\rm SFR}(t)dt,$$ where $R$ is the mass fraction of a generation of stars that is returned to the interstellar medium, $R\approx 0.3, 0.15,$ and 0.2 for a Salpeter, $x=1.7$, and Scalo IMF, respectively.
In computing the time evolution of the spectrophotometric properties of a stellar population in comoving volumes large enough to be representative of the universe as a whole, our first task is to relate the observed UV emission to a mean star formation rate. This is done by assuming a universal IMF and then fitting a smooth function to the UV continuum emissivity at various redshifts. By construction, all models will therefore produce, to within the errors, the right amount of ultraviolet light. Bruzual and Charlot’s synthesis code can then be used to predict the cosmic emission history at long wavelenghts. It is fair to point out some of the limitations of this approach at the outset. (1) It focuses on the emission properties of “normal”, optically-selected field galaxies which are only moderately affected by dust – a typical spiral emits 30% of its energy in the far-infrared region (Soifer & Neugebauer 1991) –. Starlight which is completely blocked from view even in the near-IR by a large optical depth in dust is not recorded by this technique, and the associated baryonic mass and metals missed from our census. The contribution of infrared-selected dusty starbursts to the integrated stellar mass density cannot be large, however, for otherwise the current limits to the energy density of the mid- and far-infrared background would be violated (Puget 1996; Kashlinsky, Mather, & Odenwald 1996; Fall, Charlot, & Pei 1996; Guiderdoni 1997). Locally, infrared luminous galaxies are known to produce only a small fraction of the IR luminosity of the universe (Soifer & Neugebauer 1991). (2) While the method bypasses the ambiguities associated with the study of morphologically-distint samples whose physical significance remains unclear, by the same token it does not provide any [*direct*]{} information on the processes which shaped the Hubble sequence. (3) Although in all our calculations the IMF extends from 0.1 to 125 $\msun$, by modeling the rest-frame galaxy luminosity density from 0.15 to 2.2 we are actually only sensitive to stars within the mass range from $\sim 0.8$ to about 20$\msun$. This introduces non-negligible uncertainties in our estimates of the total amount of stars and metals produced. (4) No attempt has been made to include the effects of cosmic chemical evolution on the predicted galaxy colors. All the population synthesis models assume solar metallicity and will therefore generate colors at early epochs that are too red. (5) The uncertanties present in the estimates of the UV luminosity density from the identification of Lyman-break galaxies in the HDF are quite large, and the data points at $z>2$ should still be regarded as tentative. This is especially true for the faint blue dropout sample at $\langle z\rangle=4$, where only one spectroscopic confirmation has been obtained so far. On the other hand, there is no evidence for a gross mismatch at the $z\approx 2$ transition between the photometric redshift sample of Connolly (1997) and the M96 ultraviolet dropout sample.
Salpeter IMF
------------
Figure 7 shows the model predictions for the evolution of $\rho_\nu$ at rest-frame ultraviolet to near-infrared frequencies. In the absence of dust reddening, this relatively flat IMF generates spectra that are too blue to reproduce the observed mean galaxy colors. The shape of the predicted and observed $\rho_\nu(z)$ relations agrees better to within the uncertainties if some amount of dust extinction, $E(B-V)=0.06$, is included. In this case, the observed UV luminosities must be corrected upwards by a factor of 1.4 at 2800 Å and 2.1 at 1500 Å. As expected, while the ultraviolet emissivity traces remarkably well the rise, peak, and sharp drop in the instantaneous star formation rate (the smooth function shown in the inset on the upper-right corner of the figure), an increasingly large component of the longer wavelenghts light reflects the past star formation history. The peak in the luminosity density at 1.0 and 2.2 occurs then at later epochs, while the decline from $z\approx 1$ to $z=0$ is more gentle than observed at shorter wavelengths. The total stellar mass density at $z=0$ is $\rho_s(0)=3.7\times 10^8\mdden$, with a fraction close to 65% being produced at $z>1$, and only 20% at $z>2$. In the assumed cosmology, about half of the stars observed today are more than 9 Gyr old, and only 20% are younger than 5 Gyr.[^2]
x=1.7 IMF
---------
Figure 8 shows the model predictions for a $x=1.7$ IMF and negligible dust extinction. While able to reproduce quite well the $B$-band emission history and consistent within the error with the local $K$-band light, this model slightly underestimates the 1 luminosity density at $z\approx 1$. The total stellar mass density today is larger than in the previous case, $\rho_s(0)=6.2\times 10^8\mdden$.
Scalo IMF
---------
The fit to the data is found to be much poorer for a Scalo function, since this IMF generates spectra that are too red to reproduce the observed mean galaxy colors, as already noted by Lilly (1996). Because of the relatively large number of solar mass stars formed, it produces too much long-wavelength light by the present epoch (Madau 1997). The addition of dust reddening would obviously make the fit even worse.
The Brightness of the Night Sky
-------------------------------
Our modeling of the data points to a redshift range $1\lta z\lta 2$ where the bulk of the stellar mass was actually produced. The uncertainties in the determination of the luminosity density at that epoch are, however, quite large. At $z\approx 1$, the increase in the “estimated” emissivity (i.e., corrected for incompleteness by integrating over the best-fit Schechter function) over that “directly” observed in the CFRS galaxy sample is about a factor of 2 (Lilly 1996). Between $z=1$ and $z=2$, the peak in the average SFR is only constrained by the photometric redshifts of Connolly (1997) and by the HDF UV dropout sample, both of which may be subject to systematic biases.
An important check on the inferred emission history of field galaxies comes from a comparison of the EBL produced by known galaxies (see §2.1) and the predicted mean surface brightness of the night sky, $$I_\nu={1\over 4 \pi}\int_0^\infty dz {dl\over dz}\rho_{\nu'}(z)$$ where $\nu'=\nu(1+z)$ and $dl/dz$ is the cosmological line element. The results are plotted in Figure 9. The overall agreement is remarkably good, with the model spectra being only slightly bluer, by about 20–30%, than the observed EBL. The straightforward conclusion of this exercise is that the star formation histories depicted in Figures 7 and 8 appear able to account for the entire background light recorded in the galaxy counts down to the very faint magnitudes probed by the HDF.
The Colors of High-Redshift Galaxies
------------------------------------
Figure 10 shows a comparison between the HDF data and the model predictions for the evolution of galaxies in the $\ub$ vs. $\vi$ color-color plane according to the star formation histories of Figures 7 and 8. The fact that the Salpeter IMF, $E(B-V)=0.06$ model reproduces quite well the rest-frame UV colors of high-$z$ galaxies, while a dust-free $x=1.7$ IMF generates $\vi$ colors that are 0.2 mag too blue, suggests the presence of some amount of dust extinction in Lyman-break galaxies at $z\sim 3$ (cf. Meurer 1997).
The Stellar Mass Density Today
------------------------------
The best-fit models discussed in §4 generate a present-day stellar mass density in the range between 4 and 6 $\times 10^8\mdden$. Although one could in principle lower this number by adopting a top-heavy IMF, richer in massive UV-producing stars, in practice a significant amount of dust reddening – hence of “hidden” star formation – would then be required to match the observed galaxy colors. The net effect of this operation would be a baryonic mass comparable to the estimate above and a far too large infrared background (see below). It appears therefore that the observed galaxy emission history implies a stellar mass density at the present epoch in the interval $0.005\lta
\Omega_sh_{50}^2\lta 0.009$. The stellar mass-to-light ratios range from 4.5 in the $B$-band and 0.9 in $K$ for a Salpeter function, to 8.1 in $B$ and 1.5 in $K$ for a $x=1.7$ IMF. Note that these values are quite sensitive to the lower-mass cutoff of the IMF, as very-low mass stars can contribute significantly to the mass but not to the integrated light of the whole stellar population. A lower cutoff of 0.2$\msun$, instead of the 0.1$\msun$ adopted, would decrease the mass-to-light ratio by a factor of 1.3 for a Salpeter function, 1.6 for $x=1.7$, and 1.1 for a Scalo IMF.
Star Formation at High Redshift: Monolithic Collapse Versus Hierarchical Clustering Models
==========================================================================================
The significant uncertanties present in the estimates of the star formation density at $z>2$ have already been mentioned. The biggest one is probably associated with dust reddening, but, as the color-selected HDF sample includes only the most actively star-forming young objects, one could also imagine the existence of a large population of relatively old or faint galaxies still undetected at high-$z$. The issue of the amount of star formation at early epochs is a non trivial one, as the two competing models, “monolithic collapse” versus hierarchical clustering, make very different predictions in this regard. From stellar population studies we know in fact that about half of the present-day stars are contained into spheroidal systems, i.e., elliptical galaxies and spiral galaxy bulges (Schechter & Dressler 1987). In the “monolithic” scenario these formed early and rapidly, experiencing a bright starburst phase at high-$z$ (Eggen, Lynden-Bell, & Sandage 1962; Tinsley & Gunn 1976; Bower 1992). In hierarchical clustering theory instead ellipticals form continuosly by the merger of disk/bulge systems (Kauffman 1993; White & Frenk 1991), and most galaxies never experience star formation rates in excess of a few solar masses per year (Baugh 1997). The star formation histories discussed in § 4 produce only 10% of the current stellar content of galaxies at $z>2.5$, in apparent agreement with hierarchical clustering cosmologies. In fact, the tendency to form the bulk of the stars at relatively low redshifts is a generic feature not only of the $\Omega_0=1$ CDM cosmology, but also of successful low-density CDM models (cf Figure 21 of Cole 1994; Baugh 1997).
It is then of interest to ask how much larger could the volume-averaged SFR at high-$z$ be before its fossil records – in the form of long-lived, near solar-mass stars – became easily detectable as an excess of $K$-band light at late epochs. In particular, is it possible to envisage a toy model where 50% of the present-day stars formed at $z>2.5$ and were shrouded by dust? The predicted emission history from such a model is depicted in Figure 11. To minimize the long-wavelength emissivity associated with the radiated ultraviolet light, a Salpeter IMF has been adopted. Consistency with the HDF data has been obtained assuming a dust extinction which increases rapidly with redshift, $E(B-V)=0.0067(1+z)^{2.2}$. This results in a correction to the rate of star formation of a factor $\sim 5$ at $z=3$ and $\sim 15$ at $z=4$. The total stellar mass density today is $\rho_s(0)=5.0\times 10^8\mdden$ ($\Omega_sh_{50}^2=$0.007).
Overall, the fit to the data is still acceptable, showing how the blue and near-IR light at $z<1$ are relatively insensitive to significant variations in the SFR at high redshifts, and are then, because of the short timescale available at $z\gta 2$, [*relatively poor indicators of the star formation history at early epochs*]{}. Note, however, that the adopted extinction-redshift relation implies negligible reddening at $z\lta 1$. Relaxing this – likely unphysical – assumption would cause the model to significantly overproduce the $K$-band local luminosity density. We have also checked that a larger amount of hidden star formation at early epochs would generate too much blue, 1 and 2.2 light to be still consistent with the observations. An IMF which is less rich in massive stars would only exacerbate the discrepancy.
Constraints from the Mid- and Far-Infrared Background
-----------------------------------------------------
Ultimately, it should be possible to set some constraints on the total amount of star formation hidden by dust over the entire history of the universe by looking at the cosmic infrared background (CIB). From an analysis of the smoothness of the [*COBE*]{} DIRBE maps, Kashlinsky (1996) have recently set an upper limit to the CIB of 10–15 nW m$^{-2}$ sr$^{-1}$ at $\lambda=$10–100 assuming clustered sources which evolve according to typical scenarios. An analysis using data from [*COBE*]{} FIRAS by Puget (1996) has produced a tentative detection at a level of 3.4 ($\lambda/400
\micron)^{-3}$ nW m$^{-2}$ sr$^{-1}$ in the 400–1000 range. By comparison, the integrated light that is reprocessed by dust in the model depicted in Figure 7 is close to 8 nW m$^{-2}$ sr$^{-1}$. The monolithic collapse scenario of Figure 11 generates about 6.5 nW m$^{-2}$ sr$^{-1}$ instead. While both these models appear to be consistent with the data (given the large uncertainties associated with the subtraction of foreground emission and the spectral shape of the CIB), it is clear that a significantly larger amount of hidden star formation at early and/or late epochs would overproduce the IR background (Fall 1996; Guiderdoni 1997).
Metal Production
----------------
We may at this stage use our set of models to establish a cosmic timetable for the production of heavy elements ($Z\ge 6$) in relatively bright field galaxies. What we are interested in here is the universal rate of ejection of newly synthesized material. In the approximation of instantaneous recycling, the metal ejection rate per unit comoving volume can be written as $${\dot \rho_Z}=y(1-R)\times {\rm SFR}, \label{eq:rhoz}$$ where the [*net*]{}, IMF-averaged yield of returned metals is $$y={\int mp_{\rm zm}\phi(m)dm\over (1-R)\int m\phi(m)dm},$$ $p_{\rm zm}$ is the stellar yield, i.e., the mass fraction of a star of mass $m$ that is converted to metals and ejected, and the dot denotes differentiation with respect to cosmic time.
The predicted end-products of stellar evolution, particularly from massive stars, are subject to significant uncertainties. These are mainly due to the effects of initial chemical composition, mass-loss history, the mechanisms of supernova explosions, and the critical mass, $M_{\rm BH}$, above which stars collapse to black holes without ejecting heavy elements into space (Maeder 1992; Woosley & Weaver 1995). The IMF-averaged yield is also very sensitive to the choice of the IMF slope and lower-mass cutoff. Observationally, the best-fit “effective yield” (derived assuming a closed box model) is 0.025$Z_\odot$ for Galactic halo clusters, 0.3$Z_\odot$ for disk clusters, 0.4$Z_\odot$ for the solar neighborhood, and 1.8$Z_\odot$ for the Galactic bulge (Pagel 1987). The latter value may represent the universal true yield, while the lower effective yields found in the other cases may be due, e.g., to the loss of enriched material in galactic winds.
Figure 12 shows the total mass of metals ever ejected, $\rho_Z$, versus redshift, i.e., the sum of the heavy elements stored in stars and in the gas phase as given by the integral of equation (\[eq:rhoz\]) over cosmic time. The values plotted have been computed from the star formation histores depicted in Figures 7 and 11, and have been normalized to $y\rho_s(0)$, the mass density of metals at the present epoch according to each model. A characteristic feature of the two competing scenarios is the rather different average metallicity expected at high redshift. For comparison, we have also plotted the [*gas metallicity*]{}, $Z_{\rm DLA}/Z_\odot$, as deduced from observations by Pettini (1997) of the damped Lyman-$\alpha$ systems (DLAs). At early epochs, when the gas consumption into stars is still low, the metal mass density predicted from these models gives, in a closed box model, a measurement of the metallicity of the gas phase. If DLAs and star-forming field galaxies have the same level of heavy element enrichment (Pei & Fall 1995), then one would expect a rough agreement between $Z_{\rm DLA}$ and the model predictions at $z\gta 3$. This is not true at $z\lta 2$, when a significant fraction of heavy elements is locked into stars. Without reading too much into this comparison (note the large error bars associated with the data points), it does appear that the monolithic collapse model tends to overpredict the cosmic metallicity at high redshifts as sampled by the DLAs.
While the detection with NICMOS of the established stellar populations surrounding the regions of star formation observed in the HDF at $z\sim
2$ should shed some light on the questions addressed in this talk, it will ultimately take the Next Generation Space Telescope to see the visible light emitted by stars at $z=2-5$, and to effectively open much of the universe to a direct study of galaxy formation and the history of the conversion of neutral gas into stars.
I would like to thank G. Bruzual, S. Charlot, A. Connolly, M. Pettini, and my collaborators, L. Pozzetti and M. Dickinson, for many stimulating discussion on various topics related to this talk. Support for this work was provided by NASA through grant AR-06337.10-94A from the Space Telescope Science Institute, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555.
Baugh, C. M., Cole, S., Frenk, C. S., & Lacey, C. G. 1997, , submitted
Bower, R. G., Lucey, J. R., & Ellis, R. S. 1992, , 254, 589
Bruzual, A. G., & Charlot, S. 1993, , 405, 538
Bruzual, A. G., & Charlot, S. 1997, in preparation
Charlot, S., Worthey, G., & Bressan, A. 1996, , 457, 625
Cohen, J. G., Cowie, L. L., Hogg, D. W., Songaila, A., Blandford, R., Hu, E. M., & Snopbell, P. 1996, , 471, L5
Cole, S., Aragón-Salamanca, A., Frenk, C. S., Navarro, J. F., & Zepf, S. E. 1994, , 271, 781
Connolly, A. J., Szalay, A. S., Dickinson, M. E., SubbaRao, M. U., & Brunner, R. J. 1997, , in press
Cowie, L. L., Songaila, A., Hu, E. M., & Cohen, J. G. 1996, , 112, 839
Djorgovski, S., & Thompson, D. 1993, in IAU Symp. 149, The Stellar Populations in Galaxies, ed. A. Renzini & B. Barbuy (Dordrecht: Kluwer), 337
Dickinson, M. E. 1997, private communication
Dickinson, M. E., 1997, in preparation
Ellis, R. S., Colless, M., Broadhurst, T., Heyl, J., & Glazebrook, K. 1996, , 280, 235
Eggen, O. J., Lynden-Bell, D., & Sandage, A. R. 1962, , 136, 748
Fall, S. M., Charlot, S., & Pei, Y. C. 1996, , 464, L43
Gallego, J., Zamorano, J., Arag[’o]{}n-Salamanca, A., & Rego, M. 1995, , 455, L1
Gardner, J. P., Sharples, R. M., Frenk, C. S., & Carrasco, B. E. 1997, , 480, L99
Guhathakurta, P., Tyson, J. A., & Majewski, S. R. 1990, , 357, L9
Guiderdoni, B., Bouchet, F. R., Puget, J.-L., Lagache, G., & Hivon, E. 1997, Nature, in press
Kashlinsky, A., Mather, J. C., & Odenwald, S. 1996, , 473, L9
Kauffmann, G., White, S. D. M., & Guiderdoni, B. 1993, , 264, 201
Lilly, S. J., Tresse, L., Hammer, F., Crampton, D., & Le F[é]{}vre, O. 1995, , 455, 108
Lilly, S. J., Le F[é]{}vre, O., Hammer, F., & Crampton, D. 1996, , 460, L1
Lowenthal, J. D., Koo, D. C., Guzman, R., Gallego, J., Phillips, A. C., Faber, S. M., Vogt, N. P., Illingworth, G. D., & Gronwall, C. 1997, , 481, 673
Madau, P. 1995, , 441, 18
Madau, P. 1997, in Star Formation Near and Far, eds. S. S. Holt & G. L. Mundy, (AIP: New York), p. 481
Madau, P., Ferguson, H. C., Dickinson, M. E., Giavalisco, M., Steidel, C. C., & Fruchter, A. 1996, , 283, 1388 (M96)
Madau, P., Pozzetti, L., & Dickinson, M. E. 1997, , submitted Maeder, A. 1992, A&A, 264, 105
Meurer, G. R., Heckman, T. M., Lehnert, M. D., Leitherer, C., & Lowenthal, J. 1997, , in press
Pagel, B. E. J. 1987, in The Galaxy, eds. G. Gilmore & B. Carswell (Reidel: Dordrecht), p. 341
Pei, Y. C., & Fall, S. M. 1995, , 454, 69
Pettini, M., Smith, L. J., King, D. L., & Hunstead, R. W. 1997, , in press
Pozzetti, L., Madau, P., Ferguson, H. C., Zamorani, G., & Bruzual, G. A. 1997, , submitted
Puget, J.-L., Abergel, A., Bernard, J.-P., Boulanger, F., Burton, W. B., Desert, F.-X., & Hartmann, D. 1996, A&A, 308, L5
Renzini, A. 1993, in Ann. NY Acad. Sci., No. 688, Proc. Texas/Pascos Symp on Relativistic Astrophysics & Particle Cosmology, ed. C. W. Akerlof & M. A. Srednicki, 124
Salpeter, E. E. 1955, , 121, 161
Scalo, J. N. 1986, Fundam. Cosmic Phys., 11, 1
Schechter, P. L., & Dressler, A. 1987, , 94, 56
Soifer, B. T., & Neugebauer, G. 1991, , 101, 354
Steidel, C. C., Giavalisco, M., Pettini, M., Dickinson, M. E., & Adelberger, K. 1996a, , 462, L17
Steidel, C. C., Giavalisco, M., Dickinson, M. E., & Adelberger, K. 1996b, , 112, 352
Steidel, C. C., & Hamilton, D. 1992, , 104, 941
Tinsley, B. M. 1980, Fundam. Cosmic Phys., 5, 287
Tinsley, B. M., & Gunn, J. E. 1976, , 203, 52
White, S. D. M., & Frenk, C. S. 1991, , 379, 25
Williams, R. E., 1996, , 112, 1335
Woosley, S. E., & Weaver, T. A. 1995, , 101, 181
[^1]: This is roughly three magnitudes fainter than the deepest ground based images in the red bands, two magnitudes deeper in the blue, and one magnitude deeper in the ultraviolet.
[^2]: Note that, contrary to the measured number densities of objects and rates of star formation, the integrated stellar mass density does not depend on the assumed cosmological model.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'R. Liseau'
bibliography:
- 'ALMA.bib'
date: 'Received ; accepted'
title: '[**$\alpha$**]{}Centauri revisited: 2nd epoch ALMA observations'
---
[The observational study of stars in the sub-millimetre regime has only rather recently begun and was made possible mainly by the Atacama Large Millimeter/submillimeter Array (ALMA). The emission mechanisms of this radiation from normal Main-Sequence stars and its physical significance for the outer atmospheric layers is the topic of intense contemporary study.]{} [Our previous ALMA observations of the [$\alpha \, {\rm Cen}$]{}tauri binary system detected the submm emission originating in the chromospheres of these solar-type stars. Observations at another epoch are aiming at further characterising these atmospheric layers and their behaviour with time. In addition, we were aiming at clarifying the status of the recently discovered U source and its relation to the [$\alpha \, {\rm Cen}$]{} system.]{} [The comparison of data from two epochs should present the basis for more advanced theoretical modelling of the chromospheres of [$\alpha \, {\rm Cen}$]{}A and B. Proper motion data of the U source should establish its relation to the [$\alpha \, {\rm Cen}$]{} system, and U’s submm spectral energy distribution (SED) should provide information about its physical nature.]{} [In Cycle4, both stars were again detected in the same bands as in the earlier Cycle2. These early data suggested a flattening of the SED towards longer wavelengths. By analogy with the Sun, this was not expected. Eventually, it turned out to be caused by an obsolete calibration, but this has now been remedied. Each SED exhibits now a single spectral slope over the entire frequency range (90 to 675GHz). For the U source, the upper limits on its proper motion (pm) are much smaller than the pm of [$\alpha \, {\rm Cen}$]{}, which essentially excludes any physical relationship with the binary.]{} [The second epoch ALMA observations of [$\alpha \, {\rm Cen}$]{} did not confirm the flattening of the SED in the lowest frequency bands that was reported before. Rather, this was the result of an inadequate flux calibration using the minor planet Ceres. Over the entire frequency range observed with ALMA, the SEDs from Cycle4 can be fit by power laws of the form $S_{\!\nu} \propto \nu^{\,\alpha}$ with $\alpha = 1.76 \pm 0.01$ for [$\alpha \, {\rm Cen}$]{}A and $\alpha = 1.71 \pm 0.02$ for [$\alpha \, {\rm Cen}$]{}B. For the infrared/submm background object U applies $\alpha = 2.55 \pm 0.14$. If this emission from U is due to dust, its opacity exponent $\beta = \alpha - 2$ would be about 0.5, indicative of particle sizes that are larger than those of the interstellar medium ($\beta_{\rm \,ISM} \sim 2$), but comparable to those found in circumstellar discs.]{}
Introduction
============
During ALMA-Cycle2, Alpha Centauri ([$\alpha \, {\rm Cen}$]{}) has been observed in 2014 and 2015. The results of these successful observations have been communicated by @liseau2015 and @liseau2016. For the further study of the chromospheric emissions from the binary stars, a new observing campaign was initiated during Cycle4 (2016 - 2017), again exploiting all of the previously available frequency bands, i.e. bands3, 4, 6, 7, 8 and 9 in the range 90 to 675GHz (3 to 0.4mm). This generated second epoch datasets for the study of the chromospheres of the G2V ([$\alpha \, {\rm Cen}$]{}A) and K1V star ([$\alpha \, {\rm Cen}$]{}B) and of the nature of the mysterious and unidentified object U, then situated about 5[$^{\prime \prime}$]{} north of A. The discovery of U was reported in the 2016 paper.
In the next section, Sect.2, we will discuss the new observations and the data reduction, with particular emphasis on the calibrations. The results for [$\alpha \, {\rm Cen}$]{}A and B are provided in Sect.3, which are discussed in Sect.4. In Sect.5, we turn our attention toward the U source. Finally, in Sect.6, we briefly present our main conclusions.
Observations and data reduction
===============================
The observations of Cycle2 (C2, ID 2013.1.00170.S) have already been described and analysed by @liseau2015 and @liseau2016 and will not be repeated here, unless required by the context, (see, e.g., the upper half of Table\[Antennae\]).
The Cycle4 (C4, ID 2016.1.00441.S) observations using the number of antennas shown in the lower half of Table\[Antennae\] were performed during 2016 and 2017. Due to phase errors in the Band3 and 4 observations, these data had to be re-acquired, which resulted in the delayed access to the data. For all observations, the number of telescopes was larger during C4 and, therefore, a higher sensitivity, by about 25% to 40%, could be expected. The parameters of the synthesised beams are also provided in that table, where $a$ and $b$ refer to the major and minor axes, respectively, of the elliptical Gaussian beam at half power.
[ ccccr]{}\
Cycle & Obs. Date & $N$ & Synth. Beam & PA\
Band & yyyy-mm-dd & Ant. & $a \times b$ ([$^{\prime \prime}$]{})$^2$& ([$^{\circ}$]{})\
C2 & & & &\
3 & 2014-07-03 & 30 & $1.81 \times 1.22$ & 19\
7 & 2014-07-07 & 32 & $ 0.43\times 0.28$ & 47\
9 & 2014-07-18 & 31 & $0.22\times 0.16$ & 36\
6 & 2014-12-16 & 35 & $1.64\times 1.07$ & 71\
4 & 2015-01-18 & 34 & $3.16\times 1.67$ & 82\
8 & 2015-05-02 & 37 & $0.77\times 0.68$ & $-70$\
C4 & & & &\
4 & 2016-12-19 & 42 & $1.38 \times 1.01 $ & $-25$\
6 & 2016-12-30 & 44 & $0.81 \times 0.70$ & $-50$\
9 & 2017-03-22 & 42 & $0.67 \times 0.51$ & 60\
7 & 2017-03-26 & 43 & $1.28 \times 1.08$ & 41\
8 & 2017-04-07 & 43 & $1.08 \times 0.97$ & 51\
3 & 2017-05-08 & 47 & $0.73 \times 0.50$ & $-28$\
[ c c clll]{}\
& Obs. Date &\
\
& yyyy-mm-dd & Phase & Bandpass & Flux\
C2 & & & &\
Bd & & & &\
3 & 2014-07-03 & J1617-5848 & J1427-4206 & Ceres\
7 & 2014-07-07 & J1617-5848 & J1427-4206 & Titan\
9 & 2014-07-18 & J1617-5848 & J1508-4953 & Ceres\
6 & 2014-12-16 & J1408-5712 & J1427-4206 & J1427-421\
4 & 2015-01-18 & J1617-5848 & J1617-5848 & Ceres\
8 & 2015-05-02 & J1617-5848 & J1427-4206 & Titan\
C4 & & & &\
Bd & & & &\
4 & 2016-12-19 & J1424-6807 & J1617-5848 & Ganymed\
6 & 2016-12-30 & J1424-6807 & J1617-5848 & Callisto\
9 & 2017-03-22 & J1424-6807 & J1266-0547 & Titan\
7 & 2017-03-26 & J1427-4206 & J1424-6807 & Titan\
8 & 2017-04-07 & J1424-6807 & J1617-5848 & Callisto\
3 & 2017-05-08 & J1424-6807 & J1617-5848 & J1617-5848\
The selection of the ALMA configuration, i.e. the maximum extent of the baselines, should ensure that the stars, having apparent diameters less than 10mas, would remain spatially unresolved at all observing frequencies. As a result, the observation of point sources should render the reduction of the interferometric data relatively straightforward, using a source model corresponding to the synthesised elliptical Gaussian telescope beam to obtain best-fits to the observed visibilities.
ALMA Data Calibration
=====================
The visibilities were calibrated following standard procedures using the CASA package (Common Astronomy Software Application, version 5.1.1). The calibration sources and their functions during our 2014/15 and 2016/17 observing campaigns, respectively, are shown in Table\[cal\]. For complex gain calibration and bandpass, quasars were generally used. These are listed in the table under “Phase” and “Bandpass”, respectively. Flux calibration was achieved by mostly observing asteroids and moons in the solar system, but occasionally quasars were also used (see column “Flux” in Table\[cal\]).
The C2 observations were all re-reduced in the same way as those of C4. That should guarantee an overall homogeneous data set for both cycles (see, e.g., Table\[ceres\]).
[ l l ]{}\
Band 3: $S_{97.5\,{\rm GHz}}$ (mJy) & Band 4: $S_{145\,{\rm GHz}}$ (mJy)\
\
Cycle 2 &\
A: $3.37\pm 0.01$ \[S/N=281\] & $6.33\pm 0.08$ \[S/N=83\]\
B: $1.59\pm 0.02$ \[S/N=80\] & $2.58\pm 0.08$ \[S/N=34\]\
\
Cycle 4 &\
A: $3.159\pm 0.016$ \[S/N=197\] & $5.255\pm 0.037$ \[S/N=142\]\
B: $1.549\pm 0.016$ \[S/N=97\] & $2.481\pm 0.042$ \[S/N=59\]\
Flux calibration with Ceres
---------------------------
From the Cycle2 data, it came as a complete surprise that the spectral energy distribution (SED) appeared to change slope at the lowest ALMA frequencies, viz. in bands 3 and 4 [@liseau2016].
Based on the solar analogy, one would have expected this not to happen before at much lower frequencies, i.e. in the radio regime. In the paper by @liseau2016 [Table2], the data for Bands 3 (2014-07-03) and 4 (2015-01-18) have been calibrated with an obsolete model for the asteroid Ceres. These observations took place before January 2015, and a Ceres model with a constant brightness temperature with frequency (185K) was used. Thereafter, thermo-physical models have been applied [@butler2012 see also: Appendix C in the CASA User Manual, https://casa.nrao.edu/docs/UserMan/casa\_cookbook014.html].
Absolute flux calibration
-------------------------
In order to determine absolute fluxes, the absolute systematic uncertainties need to be assessed and accounted for: currently, estimates and goals place these at better than 10% in B3 (calibration from the quasar J1617-5848), and according to @butler2012 better than 5-10% in B4 (calibration from Ganymede), better than 10% in B6 (calibration from Callisto), perhaps better than 20% in B7 (calibration from Titan, no spectral lines), better than 30% in B8 (calibration from Callisto) and maybe better than 30-50% in B9 (calibration from Titan, no spectral lines).
The comparison of the absolute flux densities derived for the C4 data with those of C2 (both previous and current) is shown in Fig.\[C4\_C2\].
Results
=======
Also in Cycle4, both stars, [$\alpha \, {\rm Cen\,A}$]{} and [$\alpha \, {\rm Cen\,B}$]{}, were detected at high signal to noise. As expected, the reconstructed stellar images were those of point sources outlining the synthesized elliptical telescope beams at the different frequencies. The adopted flux calibrations resulted in the values displayed in Tables\[posA\] and \[posB\].
For [$\alpha \, {\rm Cen}$]{}A, the results are presented in Table\[posA\] which contains the data for both the re-reduced C2 and C4, ordered chronologically. The first column, designated Bd, shows the ALMA band identification, the frequency of which is given in column two.The Gregorian observing date and the Modified Julian Date follow, where MJD = JD - 2400000.5. The measured J2000.0 equatorial coordinates (ICRS) with their error estimates are listed in columns 5 to 8. Finally, the flux densities and their statistical errors are found in columns 9 and 10. The corresponding information for [$\alpha \, {\rm Cen}$]{}B is given in Table\[posB\].
Prior to 2015, Band9 observations (2014-07-18) were also flux calibrated using Ceres and also showed over-estimated flux densities compared to those taken in 2017-03-22. However, given the absolute uncertainties of up to 30-50% in Band9 (calibration from Titan), this mismatch is formally within the errors.
[ ccc ccc ccr c ]{}\
Bd& $\nu$ & Obs. Date & MJD & RA (J2000) &$\delta_{\rm RA}$ & Dec (J2000) & $\delta_{\rm Dec}$ & $S_{\nu}$ & $\delta\,{S_{\nu}}$\
& GHz & yyyy mm dd&JD - 2400000.5&14h 39m + s & s & $-(60$[$^{\circ}$]{} 49[$^{\prime}$]{} $+$[$^{\prime \prime}$]{}) & [$^{\prime \prime}$]{}& mJy & mJy\
3 & 91.487&2014 07 03& 56841.0557552&28.800&0.0004& 58.14 & 0.003 & 3.159 & 0.016\
7 & 337.487 & 2014 07 07 & 56845.1131701 & 28.802 & 0.0003 & 58.18 & 0.002 & 25.989 & 0.191\
9 & 675.002 & 2014 07 18 & 56856.0551458 & 28.777 & 0.0004 & 58.08 & 0.003 & 76.945 & 1.647\
6 & 224.992 & 2014 12 16 & 57007.4699005 & 28.716 & 0.0008 & 57.61 & 0.004 & 9.373 & 0.082\
4 & 138.987 & 2015 01 10 & 57040.5804693 & 28.663 & 0.0014 & 57.99 & 0.006 & 5.255 & 0.042\
8 & 398.987 & 2015 05 02 & 57144.1387865 & 28.455 & 0.0003 & 58.29 & 0.002 & 30.850 & 0.175\
4 & 138.987 & 2016 12 19 & 57741.4962025 & 27.654 & 0.0005 & 56.78 & 0.004 & 5.685 & 0.037\
6 & 224.992 & 2016 12 29 & 57752.4800353 & 27.636 & 0.0003 & 56.79 & 0.002 & 12.675 & 0.063\
9 & 675.002 & 2017 03 21 & 57834.3408501 & 27.511 & 0.0007 & 57.42 & 0.005 & 92.226 & 1.542\
8 & 398.987 & 2017 03 26 & 57838.2734034 & 27.499 & 0.0004 & 57.46 & 0.003 & 34.969 & 0.188\
7 & 337.487 & 2017 03 26 & 57838.2947216 & 27.498 & 0.0003 & 57.38 & 0.002 & 26.608 & 0.075\
3 & 91.487 & 2017 05 07 & 57881.1324039 & 27.374 & 0.0003 & 57.41 & 0.003 & 2.652 & 0.019\
[ ccc ccc ccr c ]{}\
Bd& $\nu$ & Obs. Date & MJD & RA (J2000) &$\delta_{\rm RA}$ & Dec (J2000) & $\delta_{\rm Dec}$ & $S_{\nu}$ & $\delta\,{S_{\nu}}$\
& GHz & yyyy mm dd &JD - 2400000.5&14h 39m + s & s & $-(60$[$^{\circ}$]{} 49[$^{\prime}$]{} $+$[$^{\prime \prime}$]{}) & [$^{\prime \prime}$]{}& mJy & mJy\
3 & 91.487 & 2014 07 03 & 56841.0557552 & 28.252 & 0.0004 & 57.14 & 0.003 & 1.549 & 0.016\
7 & 337.487 & 2014 07 07 & 56845.1131701 & 28.240 & 0.0003 & 57.28 & 0.002 & 11.322 & 0.248\
9 & 675.002 & 2014 07 18 & 56856.0551458 & 28.216 & 0.0004 & 57.16 & 0.003 & 23.150 & 0.825\
6 & 224.992 & 2014 12 16 & 57007.4699005 & 28.743 & 0.0008 & 56.41 & 0.004 & 4.171 & 0.083\
4 & 138.987 & 2015 01 10 & 57040.5804693 & 28.163 & 0.0014 & 56.44 & 0.006 & 2.481 & 0.042\
8 & 398.987 & 2015 05 02 & 57144.1387865 & 27.943 & 0.0003 & 56.80 & 0.002 & 13.404 & 0.188\
4 & 138.987 & 2016 12 19 & 57741.4962025 & 27.259 & 0.0005 & 53.85 & 0.004 & 2.844 & 0.035\
6 & 224.992 & 2016 12 29 & 57752.4800353 & 27.244 & 0.0003 & 53.83 & 0.002 & 6.089 & 0.056\
9 & 675.002 & 2017 03 21 & 57834.3408501 & 27.134 & 0.0007 & 54.26 & 0.005 & 42.870 & 1.534\
8 & 398.987 & 2017 03 26 & 57838.2734034 & 27.125 & 0.0004 & 54.30 & 0.003 & 16.050 & 0.188\
7 & 337.487 & 2017 03 26 & 57838.2947216 & 27.124 & 0.0003 & 54.24 & 0.002 & 12.352 & 0.076\
3 & 91.487 & 2017 05 07 & 57881.1324039 & 27.009 & 0.0003 & 54.16 & 0.003 & 1.354 & 0.019\
The SEDs of [$\alpha \, {\rm Cen}$]{}A and B
--------------------------------------------
At the lower frequencies, the SEDs follow essentially the same trend as that determined at higher frequencies, i.e. exhibiting $S_{\!\nu}$[$\sim$]{}$\nu^{\,\alpha}$ over nearly two orders of magnitude in flux density (cf. Fig.\[SED\]). For instance, $\alpha$ equals two for blackbody, i.e. optically thick, radiation in the Rayleigh-Jeans regime.
Linear regression [@press1986] of the C4 Band3 to 9 data results in $d \log S_{\!\nu}/d \log \nu = 1.76 \pm 0.01$ for [$\alpha \, {\rm Cen}$]{}A and $1.71 \pm 0.02$ for [$\alpha \, {\rm Cen}$]{}B, respectively (Fig.\[SED\]). The SEDs appear flatter than those of blackbodies and also of what was obtained before. However, the results of the 2016 paper seem less significant. Clearly, with the current uncertain status of the absolute intensity calibration, our monitoring can at present not be used to meaningfully assess the level of stellar variability of [$\alpha \, {\rm Cen}$]{} in the submm (see Fig.\[C4\_C2\]).
The sky motions of the stars [$\alpha \, {\rm Cen}$]{}A and B
-------------------------------------------------------------
The positional measurements of the components of the [$\alpha \, {\rm Cen}$]{} binary, together with the graphical representation of their ephemerides, are shown in Fig.\[ABU\_ephem\]. These data are also provided in Tables \[posA\] and \[posB\]. The ephemerides were calculated adopting the stellar data from @kervella2016. Included in the ephemeris computations are the contributions from the orbital motions, the annual parallaxes and the proper motions of the binary stars, leading to complex patterns as shown in Fig.\[ABU\_ephem\]. The agreement of the theoretical results with the observational data is generally satisfactory.
The nature of the U source
==========================
The ephemerides of U
--------------------
In Fig.\[ABU\_ephem\] also the corresponding data for the U source are shown, in black for the C2 and in green for the C4 observations. The source was firmly discovered in Band8 in 2015, only some 5[$^{\prime \prime}$]{} north of [$\alpha \, {\rm Cen\,A}$]{} [@liseau2016]. At the distance of the solar sibling, this would correspond to a projected orbit midway between Jupiter and Saturn in the solar system. As such, U seemed very intriguing and caught our interest.
Mainly because of detector saturation by the bright [$\alpha \, {\rm Cen}$]{}, this anonymous object had not been noticed before at any wavelength and its nature was thus undetermined, i.e. whether it was physically associated with [$\alpha \, {\rm Cen}$]{} or an object in the fore- or background. Consequently, the object was termed U, meaning unidentified. To gain insight, data from at least another epoch would be required.
In Table\[posU\], the data for the C2 and C4 observations are compiled. As previously mentioned, also during C4 was [$\alpha \, {\rm Cen}$]{} observed in all bands. However, the U source would have been outside the primary beam of Band9 and, hence, no data are available for this source in that band. The identifications are based on positional coincidences across all the bands. In Figure\[U\_pos\], the observations are shown for Bands8 and 7 in C2 (left) and C4 (right), respectively, demonstrating the reliability of this approach. On the right-hand side, the Band8 and 7 images of 2017 are particularly compelling, as these data were taken on the same day and within 30 minutes: in both data sets is the U source situated exactly in the same place in the sky (see also Table\[posU\]).
### The proper motion of U
The apparent sky motions of U with time do not allow any meaningful determination of its parallax. Its distance remains therefore unknown, but of course, it is much larger than that of [$\alpha \, {\rm Cen}$]{}. The C2 data for Bands7, 9 and 8, and the C4 data for Bands6, 8 and 7 are collected around essentially the same Right Ascension and displaced by + in Declination. Therefore, over the course of 1.9 years, any proper motion of U is limited to $\mu_{\alpha} < -$15masyr$^{-1}$ and $\mu_{\delta} <$ 100masyr$^{-1}$. These numbers essentially reflect the astrometric accuracy of the ALMA data. Anyway, these limits are very much different from the proper motion of [$\alpha \, {\rm Cen}$]{}, i.e. $\mu_{\alpha} = -$3620masyr$^{-1}$ and $\mu_{\delta} =$ 694masyr$^{-1}$ [@kervella2016], and we conclude that U is (quasi-)stationary, implying that U is not part of the nearby high-proper-motion [$\alpha \, {\rm Cen}$]{} system.
[ ccc cc ll]{}\
Date & Bd & $\nu$ & R.A. (2000) & Dec. (J2000) & $S_{\nu}$ & $\delta S_{\nu}$\
yyyy mm dd & & GHz & hh mm ss.sss & [$^{\circ}$]{} [$^{\prime}$]{} [$^{\prime \prime}$]{} & mJy & mJy\
2014 07 03 & 3 &97.5&14 39 28.458& $-60$ 49 52.83 & 0.0554 & 0.0011\
2015 01 18 & 4 & 145 & 14 39 28.539 & $-60$ 49 52.79 & 0.1565 & 0.0015\
2014 12 16 & 6 & 233 & 14 39 28.444 & $-60$ 49 53.01 & 0.592 & 0.07\
2014 07 07 & 7 & 343.5 & 14 39 28.507 & $-60$ 49 52.66 & 1.34 & 0.4\
2015 05 02 & 8 & 405 & 14 39 28.507 & $-60$ 49 52.74 & 3.2 & 0.5\
201407 18 & 9 & 679 & 14 39 28.504 & $-60$ 49 52.70 & 6.7 & 1.3\
2016 12 30 & 6 & 233 & 14 39 28.510 & $-60$ 49 52.42 & 0.559 & 0.09\
2017 03 26 & 8 & 405 & 14 39 28.505 & $-60$ 49 52.50 & 1.47 & 0.6\
2017 03 26 & 7 & 343.5 & 14 39 28.505 & $-60$ 49 52.50 & 1.86 & 0.4\
2016 12 19 & 4 & 145 & $\cdots$ & $\cdots$ & $<0.08$ &$1\,\sigma$\
2017 05 08 & 3 &$\phantom{1}97.5$&$\cdots$ & $\cdots$ & $<0.017$ &$1\,\sigma$\
The SED of U
------------
As shown in Table\[posU\], flux data could be extracted for all bands during C2, whereas in C4, the source was not detected in Bands3 and 4 and, in Band9, it was outside the field of view. The resulting SED is shown in Fig.\[U\_SED\], together with a regression fit to the logarithmic data. The slope of that line is $2.55 \pm 0.14$. If due to thermal emission by dust, an opacity exponent $\beta$ of about 0.5 would be implied, where $\kappa_{\nu} \propto \nu^{\,\,\beta}$. This is similar to the dust found in, e.g., the debris disc around [$\beta \, {\rm Pic}$]{} [@liseau2003], where the dust particles are significantly larger than those found in the diffuse interstellar medium, for which $\beta \sim 2$.
Conclusions
===========
We have obtained 2nd epoch data for [$\alpha \, {\rm Cen}$]{}tauri in ALMA-Bands 3, 4, 6, 7, 8 and 9 (92 to 675GHz). The Band3 data of ALMA Cycle4 had to be re-taken due to problems with the phase calibration, and that delayed the analysis. But the main reason for the delay was our discovery that the already published data (Cycle 2) for bands 3 and 4 were faulty: these were based on a flux calibration using Ceres that was eventually abandoned. Unfortunately, we were not made aware of that by the ALMA project, but found out in the course of the analysis of the Cycle 4 data.
Here, we provide a complete re-reduction and analysis. With the aim of assessing the level of chromospheric time variability, the comparison of the results from the earlier C2 campaign with those from C4 casts doubt on the quality of the ALMA data even for the nearest stars to the Sun.
The mysterious object U that was discovered in May 2015, then about 5[$^{\prime \prime}$]{} north of [$\alpha \, {\rm Cen\,A}$]{}, has been re-observed during Cycle4 in all bands, except in Band9, where it fell outside the primary beam. These second epoch data were examined in order to establish whether U shared the proper motion of [$\alpha \, {\rm Cen}$]{}. It does not. Hence, U is not related to the [$\alpha \, {\rm Cen}$]{} system. Nevertheless, U appears to be an interesting object in its own right.
We are grateful to Drs. Paresh Prema and Steve Bell at the HM Nautical Almanac Office, UK, for their kind help with essential data. We also wish to thank Dr. Sebastien Muller at the Nordic ALMA Regional Centre (ARC node) for his great help with the ALMA data. The Nordic ARC node is funded through Swedish Research Council grant No 2017-00648. Dr. Dirk Petry at the ALMA User Support at ESO assisted with the Ceres calibration issues. This paper makes use of the following ALMA data: ADS/JAO.ALMA\#2013.1.00170.S and \#2016.1.00441.S. ALMA is a partnership of ESO (representing its member states), NSF (USA), and NINS (Japan), together with NRC (Canada) and NSC and ASIAA (Taiwan), in cooperation with the Republic of Chile. The Joint ALMA Observatory is operated by ESO, AUI/NRAO, and NAOJ.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
It is known that a lattice is representable as a ring of sets iff the lattice is distributive. $\operatorname{\textbf{CRL}}$ is the class of bounded distributive lattices (DLs) which have representations preserving arbitrary joins and meets. $\operatorname{\textbf{jCRL}}$ is the class of DLs which have representations preserving arbitrary joins, $\operatorname{\textbf{mCRL}}$ is the class of DLs which have representations preserving arbitrary meets, and $\operatorname{\textbf{biCRL}}$ is defined to be $\operatorname{\textbf{jCRL}}\cap \operatorname{\textbf{mCRL}}$. We prove $$\operatorname{\textbf{CRL}}\subset \operatorname{\textbf{biCRL}}=\operatorname{\textbf{mCRL}}\cap\operatorname{\textbf{jCRL}}\subset \operatorname{\textbf{mCRL}}\neq\operatorname{\textbf{jCRL}}\subset \operatorname{\textbf{DL}}$$ where the marked inclusions are proper.
Let $L$ be a DL. Then $L\in\operatorname{\textbf{mCRL}}$ iff $L$ has a distinguishing set of complete, prime filters. Similarly, $L\in\operatorname{\textbf{jCRL}}$ iff $L$ has a distinguishing set of completely prime filters, and $L\in\operatorname{\textbf{CRL}}$ iff $L$ has a distinguishing set of complete, completely prime filters.
Each of the classes above is shown to be *pseudo-elementary* hence closed under ultraproducts. The class $\operatorname{\textbf{CRL}}$ is not closed under elementary equivalence, hence it is not elementary.
author:
- Robert Egrot
- Robin Hirsch
title: Completely representable lattices
---
[Introduction]{}\[S:intro\] An *atomic representation* $h$ of a Boolean algebra $ B$ is a representation $h\colon B\to\operatorname{\wp}(X)$ (some set $X$) where $h(1)=\bigcup {\{h(a)\colon a\mbox{ is an atom of } B\}}$. It is known that a representation of a Boolean algebra is a complete representation (in the sense of a complete embedding into a field of sets) if and only if it is an atomic representation and hence that the class of completely representable Boolean algebras is precisely the class of atomic Boolean algebras, and hence is elementary [@HirHod97]. This result is not obvious as the usual definition of a complete representation is thoroughly second order. The purpose of this note is to investigate the possibility of corresponding results for the class of bounded, distributive lattices. The situation is a little more complex in this case as in the absence of Boolean complementation a representation of a (distributive) lattice may be complete with respect to one of the lattice operations but not the other.
It turns out (theorem \[not elem\]) that the class $\operatorname{\textbf{CRL}}$ of completely representable bounded, distributive lattices is *not* elementary, however, building on early work in lattice theory by Birkhoff [@Bir33], and Birkhoff and Frink [@BirFri48] it is possible to characterise complete representability of a lattice in terms of the existence of certain prime filters (or dually using prime ideals). Using this characterisation an alternative proof of the identification of the completely representable Boolean algebras with the atomic ones is provided. It is also shown that $\operatorname{\textbf{CRL}}$, and the classes of (bounded, distributive) lattices that have representations respecting either or both arbitrary infima and suprema *are* pseudo-elementary, and thus closed under ultraproducts. Using the well known fact that a class is elementary if and only if it is closed under isomorphism, ultraproducts and ultraroots it follows that $\operatorname{\textbf{CRL}}$ is not closed under ultraroots. The question of whether this holds for the other classes of lattices under consideration, and thus whether they are elementary, remains open at this time and is the subject of ongoing investigation.
[Representations of bounded, distributive lattices]{}
Let $L$ be a bounded, distributive lattice. A *representation* of $L$ is an embedding $h\colon L\to\operatorname{\wp}(X)$ for some set $X$, where $\operatorname{\wp}(X)$ is considered as a ring of sets, under the operations of set union and intersection. When such a representation exists we say that $L$ is *representable*.
For simplicity we shall assume that our representations $h\colon L\to \operatorname{\wp}(X)$ are ‘irredundant’, that is, for all $x\in X$ there is some $a\in L$ with $x\in h(a)$. For irredundant representations $h\colon L\to \operatorname{\wp}(X)$ the ‘inverse image’ $h^{-1}[x]=\{a\in L\colon x\in h(a)\}$ of any point $x\in X$ is a prime filter, with closure under finite meets coming from finite meet preservation by the representation, and primality coming from finite join preservation. Upward closure can be derived from either of these preservation properties using the equivalent definitions of the order relation in a lattice. Conversely, any set $K$ of prime filters of $L$ with the property that for every pair $a\neq b \in L$ there exists $f\in K$ with either $a\in f$ and $b\not \in f$ or vice versa determines a representation $h_K\colon L \to \operatorname{\wp}(K)$ using $h_K(a)=\{f\in K\colon a\in f\}$ (note that for $f\in K$ we have $h_K^{-1}[f]=f$). For ease of exposition later we introduce a definition for sets of sets generalising the condition for filters given above.
A set $S\subseteq\operatorname{\wp}(L)$ is *distinguishing* over $L$ iff for every pair $a\neq b \in L$ there exists $s\in S$ with either $a\in s$ and $b\not \in s$ or vice versa.
Using this definition we state the results of the preceding discussion as a simple theorem:
A bounded distributive lattice $L$ is representable if and only if it has a distinguishing set of prime filters.
As a consequence of the prime ideal theorem for distributive lattices we have:
A bounded lattice is representable if and only if it is distributive.
Henceforth, all lattices under consideration are bounded and distributive. We now discuss representations preserving arbitrary meets and/or joins.
A lattice map $f\colon L_1\to L_2$ is *meet-complete* if for all $S\subseteq L_1$ where $\operatorname{\bigwedge}S$ exists in $L_1$ we have $f(\operatorname{\bigwedge}_{L_1} S)=\operatorname{\bigwedge}_{L_2} f[S]$.
A similar definition is made for *join-complete*. When a map is both *meet-complete* and *join-complete* we say it is *complete*. When a bounded, distributive lattice has a meet-complete representation we say it is *meet-completely representable*, and we make similar definitions for join-complete and complete representations. We shall call the class of all bounded, distributive lattices $\operatorname{\textbf{DL}}$, the class of all completely representable lattices $\operatorname{\textbf{CRL}}$, the classes of meet and join-completely representable lattices $\operatorname{\textbf{mCRL}}$ and $\operatorname{\textbf{jCRL}}$ respectively, and the class of lattices with both a meet-complete and a join-complete representation $\operatorname{\textbf{biCRL}}$.
\[duality\] A lattice $L$ has a meet-complete representation iff its order dual $L^{\delta}$ has a join-complete representation.
If $h\colon L\to \operatorname{\wp}(P)$ is a representation, where $P$ is some distinguishing set of prime filters of $L$, then the map $\bar{h}\colon L^{\delta}\to \operatorname{\wp}(P)$, $a\mapsto -h(a)$ is also a representation. If $h$ is meet-complete then by De Morgan $\bar{h}(\operatorname{\bigvee}_{\delta} S) = -h(\operatorname{\bigwedge}S) = - \bigcap h[S] = -\bigcap -\bar{h}[S] = \bigcup \bar{h}[S]$ (here ‘$-$’ denotes set theoretic complement).
An ideal $I$ of a lattice $L$ is complete if whenever $\operatorname{\bigvee}S$ exists in $L$ for $S\subseteq I$ then $\operatorname{\bigvee}S\in I$. Similarly a filter $F$ of $L$ is complete if whenever $\operatorname{\bigwedge}T$ exists in $L$ for $T\subseteq F$ then $\operatorname{\bigwedge}T \in F$.
A prime ideal $I$ of $L$ is completely prime if whenever $\operatorname{\bigwedge}T\in I$ for some $T\subseteq L$ then $I\cap T \not= \emptyset$. Similarly, a prime filter $F$ of $L$ is completely prime if whenever $\operatorname{\bigvee}S\in F$ for some $S\subseteq L$ then $F\cap S \neq \emptyset$.
\[ideal lemma\] If $F$ is a prime filter of $L$ and $I=L\setminus F$ is its prime ideal complement then $F$ is complete iff $I$ is completely prime, and $I$ is complete iff $F$ is completely prime.
Using $I=L\setminus F$ we can rewrite the definition of completeness of $I$ as $\operatorname{\bigvee}S \in F \implies F\cap S \not=\emptyset$. Similarly we can write completeness for $F$ as $\operatorname{\bigwedge}T \in I \implies T\cap I \not=\emptyset$.
\[main2\] Let $L$ be a bounded, distributive lattice. Then:
1. $L$ has a meet-complete representation iff $L$ has a distinguishing set of complete, prime filters,
2. $L$ has a join-complete representation iff $L$ has a distinguishing set of completely-prime filters,
3. $L$ has a complete representation iff $L$ has a distinguishing set of complete, completely-prime filters,
We prove 1), the rest follows from theorem \[duality\] and lemma \[ideal lemma\]. For the left to right implication, let $h$ be an irredundant meet-complete representation of $L$ over some domain $D$. Since $h$ is an irredundant representation, ${\{h^{-1}[d]\colon d\in D\}}$ is a distinguishing set of prime filters. Also, if $S\subseteq h^{-1}[d]$ then $d\in \bigcap h[S]$ and by completeness of $h$ if $\operatorname{\bigwedge}S$ exists then $d\in h(\operatorname{\bigwedge}S)$ so $\operatorname{\bigwedge}S\in h^{-1}[d]$, hence each of these prime filters is complete. Conversely, let $K$ be a distinguishing set of complete, prime filters. Define a map $h\colon L\to\wp(K)$ by $h(l)={\{p\in K\colon l\in p\}}$. Because $K$ is a distinguishing set of prime filters, $h$ is a representation. Furthermore, since each $p\in K$ is complete, if $S\subseteq L$ and $\operatorname{\bigwedge}S$ exists then for all $p\in K,$ $$\begin{aligned}
p\in h(\operatorname{\bigwedge}S)&\iff\operatorname{\bigwedge}S\in p\\
&\iff S\subseteq p\\
&\iff p\in \bigcap h[S]\end{aligned}$$ so $h(\operatorname{\bigwedge}S)=\bigcap h[S]$ and $h$ is a complete representation.
In the light of lemma \[ideal lemma\] it’s straightforward to prove an analogous result to theorem \[main2\] using ideals in place of filters.
We briefly turn our attention to the special case of Boolean algebras. Recall that a bounded lattice $(L, 0, 1, \wedge, \vee)$ is *complemented* iff for all $s\in L$ there is $s'\in L$ such that $s\vee s'=1$ and $s\wedge s'=0$. Since there can be at most one complement to an element, we may write $-s$ instead of $s'$.
If $L$ is complemented then its prime filters are precisely its ultrafilters, moreover the following are equivalent:
1. $U$ is a principal ultrafilter of $L$,
2. $U$ is a complete ultrafilter of $L$,
3. $U$ is a completely-prime ultrafilter of $L$.
It’s easy to see that the ultrafilters of a BA are precisely its prime filters. Clearly $1)\implies 2)$. Let $U$ be an ultrafilter. If $U$ is complete it must contain a non-zero lower bound $s$ and thus be principal (otherwise it would contain the complement of that lower bound, but $s\leq -s\Rightarrow s=0$), so $2)\implies 1)$. For any $S\subseteq L$ we write $-S$ for ${\{-s\colon s\in S\}}$. The infinite De Morgan law for Boolean algebras (see e.g. [@Sik69 section 19]) gives $-\operatorname{\bigvee}S=\operatorname{\bigwedge}-S$ so if $U$ is complete then $S\cap U=\emptyset \implies-\operatorname{\bigvee}S \in U \implies \operatorname{\bigvee}S\notin U$, so $2)\implies3)$. Similarly, if $U$ is completely-prime then $\operatorname{\bigwedge}S \notin U \implies -\operatorname{\bigwedge}S \in U \implies \operatorname{\bigvee}-S \in U \implies -s \in U $ for some $s\in S\implies S\not \subseteq U$, so $3)\implies 2)$.
We have as a corollary the following result (see [@HirHod97 corollary 6] for the equivalence of the first two parts).
\[BooCor\] For a Boolean algebra $B$ the following are equivalent:
1. $B$ is atomic,
2. $B$ is completely representable,
3. $B$ is meet-completely representable,
4. $B$ is join-completely representable.
Turning our attention back to the lattice case we now give some examples to illustrate the relationships between the classes we have defined.
\[exI\]*A distributive lattice both meet-completely representable and join-completely representable but not completely representable.* Let $L=[0,1]\subseteq \operatorname{\mathbb{R}}$. Then by taking $\{ [x, 1]\colon x\in L\}$ we obtain a distinguishing set of complete, prime filters, and by taking $\{ (x, 1]\colon x\in L\}$ we obtain a distinguishing set of completely-prime filters.\
However, if $F$ is a complete filter of $L$ then $\operatorname{\bigwedge}F\in F$ (by completeness properties of $L$ and $F$) and, since $\operatorname{\bigwedge}F=\operatorname{\bigvee}\{x\in L\colon x<\operatorname{\bigwedge}F\}$, $F$ cannot be completely-prime.
\[exII\]*A distributive lattice neither meet nor join-completely representable.* In view of corollary \[BooCor\] we can take any Boolean algebra that fails to be atomic.
\[exIII\]*A distributive lattice join-completely representable but not meet-completely representable.* Let $L$ be the lattice $(\overline{\mbox{\( \mathbb N \)}}\times\overline{\mbox{\( \mathbb N \)}}) \cup{\{0\}}$ shown in figure \[fig:w2\], where $\overline{\mbox{\( \mathbb N \)}}$ is the set of non-positive integers under the usual ordering and the element $0$ is a lower bound for the whole lattice. Then $L$ has no complete, prime filters, but all its filters are completely-prime, hence by theorem \[main2\] it has a join-complete representation but no meet-complete representation.
$$\xymatrix@dr{\bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-->}[r]\ar@{-}[d] & \\
\bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-->}[r]\ar@{-}[d] &\\
\bullet\ar@{-}[r]\ar@{-->}[d]&\bullet\ar@{-}[r]\ar@{-->}[d]&\bullet\ar@{-->}[r]\ar@{-->}[d] &\\
&&&\bullet 0
}$$
Examples \[exI\], \[exII\] and \[exIII\] (and its dual) give us the following:
$$\tag{$\dag$} \label{dagger} \operatorname{\textbf{CRL}}\subset \operatorname{\textbf{biCRL}}=\operatorname{\textbf{mCRL}}\cap\operatorname{\textbf{jCRL}}\subset \operatorname{\textbf{mCRL}}\neq\operatorname{\textbf{jCRL}}\subset \operatorname{\textbf{DL}}$$
There is a relationship between the existence of types of complete representation and the join and meet-densities of the sets of join and meet-irreducibles of $L$, which we make precise in the following proposition.
\[irred\] Let $L$ be a bounded, distributive lattice. Define $J(L)$ and $M(L)$ to be the sets of join-irreducible and meet-irreducible elements of $L$ respectively, and $J^\infty(L)$ and $M^\infty(L)$ to be the sets of completely join/ meet-irreducibles of $L$ respectively, then:
1. If the set $J(L)$ is join-dense in $L$ then $L$ has a meet-complete representation, dually if the set $M(L)$ is meet-dense in $L$ then $L$ has a join-complete representation. When $L$ is complete then if $L$ has a meet/join-complete representation the sets $J(L)/M(L)$ are join/meet-dense in $L$.
2. If either $J^\infty(L)$ is join-dense in $L$ or $M^\infty(L)$ is meet-dense in $L$ then $L$ has a complete representation. When $L$ is complete it is also true that whenever $L$ has a complete representation $J^\infty$ and $M^\infty$ are join and meet-dense in $L$ respectively.
For the first part of 1, we just take the sets of principal filters/ideals generated by the join/meet-irreducibles respectively, for the second we note that the generator of each filter/ideal must be join/meet-irreducible. For the first part of 2 we note that if we take the sets of principal filters/ideals generated by $J^\infty(L)$ and $M^\infty(L)$ respectively we obtain distinguishing sets of completely-prime filters/ideals, and for the second part the generator of each filter/ideal will be completely join/meet-irreducible.
Note that the full converses to proposition \[irred\] (i.e. when $L$ is not complete) do not hold, so e.g. in a completely representable lattice $L$, $J^{\infty}(L)$ need not be join-dense, as the following example illustrates.
\[ex2\] $L$ is the lattice with domain $(\overline{\mbox{\( \mathbb N \)}}\times\overline{\mbox{\( \mathbb N \)}})\cup{\mbox{\( \mathbb N \)}}$ as shown in figure \[fig:w3\], where $\overline{\mbox{\( \mathbb N \)}}$ is the set of non-positive integers under their usual ordering and each element of ${\mbox{\( \mathbb N \)}}$ is less than each element of $(\overline{\mbox{\( \mathbb N \)}}\times\overline{\mbox{\( \mathbb N \)}})$. For $-n\in \overline{\mbox{\( \mathbb N \)}}$, the set $[-n, 0]\times\overline{\mbox{\( \mathbb N \)}}$ is a complete, completely-prime filter (with no infimum) and similarly $\overline{\mbox{\( \mathbb N \)}}\times [-n, 0]$ is also complete, completely-prime. Hence $L$ has a distinguishing set of complete, completely-prime filters but $J^{\infty}(L)=J(L)={\mbox{\( \mathbb N \)}}$ is not join dense in $L$.
$$\begin{array}{c}
\xymatrix@dr{\bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-->}[r]\ar@{-}[d] & \\
\bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-}[r]\ar@{-}[d] & \bullet\ar@{-->}[r]\ar@{-}[d] &\\
\bullet\ar@{-}[r]\ar@{-->}[d]&\bullet\ar@{-}[r]\ar@{-->}[d]&\bullet\ar@{-->}[r]\ar@{-->}[d] &\\
&&&
}
\\
\xymatrix{ \; \\
\bullet\ar@{-->}[u]\\
\bullet\ar@{-}[u]\\
\bullet\ar@{-}[u]
}
\end{array}$$
We end this section with a note about canonical extensions.
A complete lattice $L$ is completely representable if and only if it is doubly algebraic (a complete lattice is algebraic if every element can be written as a join of compact elements, a complete lattice is doubly algebraic if both it and it’s order dual are algebraic).
It is known, see e.g. [@CrawDil73], that a lattice $L$ is doubly algebraic if and only if it is complete, completely distributive, and $J^\infty(L)$ and $M^\infty(L)$ are join/meet-dense. When $L$ is completely representable it inherits complete distributivity from its representation and, by proposition \[irred\], has the required density properties. Conversely, by the same proposition, the density properties of algebraicity and dual algebraicity are both sufficient conditions for complete representability.
The canonical extension of any bounded distributive lattice is completely representable.
The canonical extension $L^\sigma$ of a bounded distributive lattice $L$ can be defined as a doubly algebraic lattice into which $L$ embeds in a certain way (see e.g. [@GehJon04 theorem 2.5] for details).
[HSP, elementarity and pseudo-elementarity]{} \[SECpseud\] Since a subalgebra of an atomic Boolean algebra need not be atomic we know that none of the classes in is closed under subalgebras, and thus cannot be varieties, or even quasi-varieties. Similarly, given an atomic Boolean algebra $B$ we can define an equivalence relation $R$ on $B$ by $xRy\iff |\{a\in At(B)\colon a\leq x\}\triangle \{a\in At(B)\colon a\leq y\}|<|\omega|$, that is, if and only if the symmetric difference of the sets of atoms beneath each element is finite. It can easily be shown that $R$ is a congruence, and in the case where $B$ is the complete, atomic Boolean algebra on $\omega$ generators the resulting $\frac{B}{R}$ is isomorphic to the countable atomless Boolean algebra, and thus none of classes in can be closed under homomorphic images. We can say something positive about closure under direct products, which we express in the following lemma:
The classes in are all closed under taking direct products.
We do the proof for $\operatorname{\textbf{mCRL}}$, the others are similar. Suppose $\{L_i\}_I$ is a family of lattices in $\operatorname{\textbf{mCRL}}$. Let $f\neq g \in \prod_I L_i$, then we can choose $j\in I$ with $f(j)\neq g(j)$, and by the assumption of meet-complete representability there is a complete, prime filter $\gamma$ distinguishing $f(j)$ and $g(i)$. Define sets $S_i\subseteq L_i$ by $S_j=\gamma$ and $S_i=L_i$ for all $i\neq j$, then $S=\prod_I S_i$ is a complete, prime filter distinguishing $f$ and $g$.
As ‘being atomic’ is a first order property for Boolean algebras, it follows immediately from corollary \[BooCor\] that the class of completely representable Boolean algebras is elementary. The aim here is to investigate to what extent similar results hold for the classes in . Our first result is negative:
\[not elem\] **CRL** is not closed under elementary equivalence.
The lattice $L=[0,1]\subseteq \operatorname{\mathbb{R}}$ from example \[exI\] is not in **CRL**, however the lattice $L'=[0,1]\cap \operatorname{\mathbb{Q}}$ *is* in $\operatorname{\textbf{CRL}}$ as for every irrational $r$ the set $\{ a\in L'\colon a> r\}$ is a complete, completely-prime filter. $L$ and $L'$ are elementarily equivalent as $\operatorname{\mathbb{R}}$ and $\operatorname{\mathbb{Q}}$ are.
We can, however, show that all the classes in are at least *pseudo-elementary*. In particular we shall demonstrate that $\operatorname{\textbf{mCRL}}$ is precisely the first order reduct of the class of models of a theory in two-sorted FOL, and thus is pseudo-elementary (the proof can be readily adapted for the other classes). We proceed as follows.
\[PEclass\]*(Pseudo-elementary class)* Given a first order signature $\operatorname{\mathscr{L}}$, a class $\operatorname{\mathscr{C}}$ of $\operatorname{\mathscr{L}}$ structures is pseudo-elementary if there are
1. a two-sorted language $\operatorname{\mathscr{L}}^+$, with disjoint sorts $\operatorname{\mathbb{A}}$ and $\operatorname{\mathbb{S}}$, containing $\operatorname{\mathbb{A}}$-sorted copies of all symbols of $\operatorname{\mathscr{L}}$, and
2. an $\operatorname{\mathscr{L}}^+$ theory $U$
with $\operatorname{\mathscr{C}}=\{M^{\operatorname{\mathbb{A}}}\upharpoonright_{\operatorname{\mathscr{L}}}\colon M\models U \}$, where $M^{\operatorname{\mathbb{A}}}$ is a structure in the sublanguage of $\operatorname{\mathscr{L}}^+$ containing only $\operatorname{\mathbb{A}}$-sorted symbols whose domain contains only $\operatorname{\mathbb{A}}$-sorted elements, $M^{\operatorname{\mathbb{A}}}\upharpoonright_{\operatorname{\mathscr{L}}}$ is the $\operatorname{\mathscr{L}}$ reduct of $M^{\operatorname{\mathbb{A}}}$ obtained easily by identifying the symbols of $\operatorname{\mathscr{L}}$ with their $\operatorname{\mathbb{A}}$-sorted counterparts in $\operatorname{\mathscr{L}}^+$, and $\{M^{\operatorname{\mathbb{A}}}\upharpoonright_{\operatorname{\mathscr{L}}}\colon M\models U \}$ being thus the class of $\operatorname{\mathbb{A}}$-sort $\operatorname{\mathscr{L}}$ reducts of models of $U$.
See [@HirHod02 section 9] for more information on this definition, and for proof of its equivalence with single-sorted definitions of pseudo-elementarity.
Now, let $\operatorname{\mathscr{L}}=\{+,\cdot,0,1\}$ be the language of bounded, distributive lattices in FOL. Define the two-sorted language $\operatorname{\mathscr{L}}^+=\operatorname{\mathscr{L}}\cup\{\in\}$, where $\in$ is a binary predicate whose first argument takes variables of the $\operatorname{\mathbb{A}}$ sort and whose second takes variables of the $\operatorname{\mathbb{S}}$ sort. Let the original functions of $\operatorname{\mathscr{L}}$ be wholly $\operatorname{\mathbb{A}}$-sorted in $\operatorname{\mathscr{L}}^+$ (the $\operatorname{\mathbb{A}}$ sort is meant to represent lattice elements and the $\operatorname{\mathbb{S}}$ sort sets of these elements). Define binary $\operatorname{\mathbb{A}}$-sorted predicates $\leq$ and $\geq$, and binary $\operatorname{\mathbb{S}}$-sorted predicate $\subseteq$ in $\operatorname{\mathscr{L}}^+$ in the obvious way. For simplicity we will write $\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}$ for $\in(\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}},\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})$, and similar for $\leq$, $\subseteq$ etc.\
Define additional predicates $P$, $I$ and $C$ as follows:
- $P(\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})$ if and only if each of the following properties hold:
1. $\forall \operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\Big(\big((\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\wedge (\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\geq \operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}})\big)\rightarrow (\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\Big)$
2. $\forall\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\Big(\big((\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \wedge (\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\big)\rightarrow (\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\cdot \operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\Big)$
3. $\forall\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\Big((\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}+\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\rightarrow \big((\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\vee (\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\big)\Big) $
$P$ is meant to capture the property of being a prime filter.\
- $I(\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}},\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})$ if and only if $$\forall \operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\Big((\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \rightarrow (\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\leq\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}})\Big)\wedge \forall\operatorname{\mathit{z}^{\operatorname{\mathbb{A}}}}\Big(\big((\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \rightarrow (\operatorname{\mathit{z}^{\operatorname{\mathbb{A}}}}\leq\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}})\big)\rightarrow (\operatorname{\mathit{z}^{\operatorname{\mathbb{A}}}}\leq\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}) \Big).$$ $I$ corresponds to the notion of an element being the infimum of a set.
- $C(\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})$ if and only if $\forall \operatorname{\mathit{t}^{\operatorname{\mathbb{S}}}}\forall\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\Big( \big( (\operatorname{\mathit{t}^{\operatorname{\mathbb{S}}}}\subseteq \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\wedge I(\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}},\operatorname{\mathit{t}^{\operatorname{\mathbb{S}}}})\big)\rightarrow (\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \Big)$, so $C$ specifies a limited form of completeness.
Now, let $T$ be the $\operatorname{\mathscr{L}}$ theory of bounded, distributive lattices. Define $T^+$ as the natural translation of $T$ into the language $\operatorname{\mathscr{L}}^+$ plus the following additional axioms:
1. \[ax:1\]$\forall \operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\Big(\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\neq\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\rightarrow \exists\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}\Big( \big(P(\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\wedge C(\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \big) \wedge \big(\big((\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \wedge (\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\notin \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\big)\vee \big((\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \wedge (\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\notin \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\big)\big) \Big)\Big)$
2. \[ax:2\]$\forall \operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\exists \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}\forall \operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}\Big( (\operatorname{\mathit{y}^{\operatorname{\mathbb{A}}}}>\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}})\leftrightarrow (y\in \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}) \Big)$
3. \[ax:3\] $\forall \operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}}\operatorname{\mathit{t}^{\operatorname{\mathbb{S}}}}\exists \operatorname{\mathit{u}^{\operatorname{\mathbb{S}}}}\forall\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\Big( \big((\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{s}^{\operatorname{\mathbb{S}}}})\wedge (\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in\operatorname{\mathit{t}^{\operatorname{\mathbb{S}}}})\big) \leftrightarrow (\operatorname{\mathit{x}^{\operatorname{\mathbb{A}}}}\in \operatorname{\mathit{u}^{\operatorname{\mathbb{S}}}}) \Big)$
The first of these axioms forces the $\operatorname{\mathbb{S}}$ sort into providing a distinguishing set of ‘complete’ (with respect to $\operatorname{\mathbb{S}}$) prime filters, and the second and third force the existence of sufficiently many elements of $\operatorname{\mathbb{S}}$ that this notion of completeness is equivalent to actual completeness, as the lemma below demonstrates.
\[PE\] The class $\{M^{\operatorname{\mathbb{A}}}\upharpoonright_{\operatorname{\mathscr{L}}}\colon M\models T^+ \}$ of $\operatorname{\mathscr{L}}$-reducts of models of $T^+$ is precisely the class of meet-completely representable bounded, distributive lattices.
Clearly if $L$ is in $\operatorname{\textbf{mCRL}}$ its elements satisfy $T$, and $(L, \operatorname{\wp}(L), \in)$ satisfy $T^+$, where $\in$ is ordinary set membership. Conversely, if $A=M^{\operatorname{\mathbb{A}}}\upharpoonright_{\operatorname{\mathscr{L}}}$ for some model $M$ of $T^+$ then by axiom \[ax:1\] of $T^+$ the (interpretation of) the $\in$ predicate naturally defines a distinguishing set $K$ of prime filters of $A$. We claim that each prime filter in $K$ is complete. For the claim, let $p\in K$ and $s\subseteq p$ with $x=$inf$(s)$. We must show that $x\in p$. If $x\in s$ then this is immediate, so we suppose not: $x\not\in s$. We consider the following cases:
1. $x=$ inf$\{y\colon y>x\}$: then $s\subseteq \{y\colon y>x\}\cap p \subseteq \{y\colon y>x\}$ so $x=$ inf$(s)\geq $ inf$(\{y\colon y>x\}\cap p)\geq x$ and thus inf$(\{y\colon y>x\}\cap p)=x$, but clearly $\{y\colon y>x\}\cap p \subseteq p$ and by axioms \[ax:2\] and \[ax:3\] of $T^+$ also corresponds to an element of the $\operatorname{\mathbb{S}}$ sort. Therefore, by definition of the predicate $C$ we have $x\in p$, as required.
2. $x\neq $ inf$\{y\colon y>x\}$: Let $z$ be a lower bound for $\{y\colon y>x\}$, suppose $z\not \leq x$. Then $x\vee z$ is a lower bound for $\{y\colon y>x\}$ and is contained in $\{y\colon y>x\}$. In light of this assume wlog that inf$\{y\colon y>x\}= z>x$. Then, as $x=$ inf$(s)$, we have $s\subseteq \{y\colon y>x\}$ and thus $s$ has $z$ as a lower bound, but this a contradiction as $x<z$, so this case cannot arise.
We deduce that $x\in p$, so $p$ is complete, as claimed. Since $T^+$ demands $A$ be a bounded, distributive lattice we have $A\in\operatorname{\textbf{mCRL}}$, by theorem \[main2\](1).
We have now proved the following:
\[pseudelem\] $\operatorname{\textbf{mCRL}}$ is pseudo-elementary.
Lemma \[PE\] shows the condition of definition \[PEclass\] hold for $\operatorname{\textbf{mCRL}}$.
It is not difficult to see how analogous results can also be proved for $\operatorname{\textbf{jCRL}}$, $\operatorname{\textbf{biCRL}}$ and $\operatorname{\textbf{CRL}}$ using a similar method.
[Ultraproducts and ultraroots]{} \[SECur\] We begin this section by stating two well known facts from model theory:
\[she cor\] A class $\mathscr{C}$ of similar structures is elementary iff it is closed under isomorphism, ultraproducts and ultraroots.
$\operatorname{\mathscr{C}}$ is pseudo-elementary $\implies$ $\operatorname{\mathscr{C}}$ is closed under ultraproducts.
For proof of the first see e.g. [@Hodg93 corollary 9.5.10] or [@ChaKei90 theorem 4.1.12 and corollary 4.3.13]. The proof of the second is simple, see e.g. [@ChaKei90 exercise 4.1.17].
In view of the facts above and the material in the preceding section, since $\operatorname{\textbf{CRL}}$ is pseudo-elementary, and closed under isomorphism, but is not elementary, it cannot be closed under ultraroots. $\operatorname{\textbf{mCRL}}$, $\operatorname{\textbf{jCRL}}$ and $\operatorname{\textbf{biCRL}}$ will be elementary if and only if they are closed under ultraroots. Note that $\operatorname{\textbf{mCRL}}$ is elementary iff $\operatorname{\textbf{jCRL}}$ is elementary (by duality), and therefore $\operatorname{\textbf{mCRL}}$ is elementary $\implies\operatorname{\textbf{biCRL}}$ is elementary (as $\operatorname{\textbf{biCRL}}=\operatorname{\textbf{mCRL}}\cap\operatorname{\textbf{jCRL}}$). It is not known which, if any, of $\operatorname{\textbf{biCRL}}$, $\operatorname{\textbf{mCRL}}$ and $\operatorname{\textbf{jCRL}}$ are closed under ultraroots but it is possible to state some conditions on a lattice $L$ which must necessarily hold if $L\not\in X$ but an ultrapower of $L$ belongs to $X$ (where $X=\operatorname{\textbf{biCRL}}, \operatorname{\textbf{mCRL}}$ or $\operatorname{\textbf{jCRL}}$).
First of all in order for the ultraproduct $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ to be meet-completely representable $L$ must be $\vee(\operatorname{\bigwedge})$-distributive, i.e. for $a\in L,\; S\subseteq L$ if both sides of the equation below are defined then they are equal $$a\vee\operatorname{\bigwedge}S =\operatorname{\bigwedge}_{s\in S}(a\vee s)$$ as we shall see in the next proposition. Note that the converse to this is false as, for example, every Boolean algebra is $\vee(\operatorname{\bigwedge})$-distributive (see e.g. [@Rom08 theorem 5.13] for a proof) but not necessarily atomic, so not necessarily meet-completely representable by corollary \[BooCor\]. We will use the following notation and lemma:
- For $a \in L$ define $\bar{a}\in \prod_I L$ by $\bar{a}(i)=a$ for all $i\in I$.
- Fix some ultrafilter $U$ over $I$. For $x\in\prod_IL$ we write $[x]$ for ${\{y\in\prod_IL\colon \colon {\{i\colon x(i)=y(i)\}}\in U\}}$.
- For $S\subseteq L$ define $S^*=\{[x]\in \operatorname{ \prod_{\mathit{U} } \mathit{L}}\colon \{i\in I\colon x(i)\in S\}\in U\}$.
- For $T\subseteq\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ define $T_*=\{a\in L\colon [\operatorname{\bar{\mathit{a}}}]\in T\}$.
\[inf exist\] Let $S\subseteq L$ and suppose $\operatorname{\bigwedge}S$ exists in $L$. Then $\operatorname{\bigwedge}(S^*)$ exists in $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ and equals $[\overline{\operatorname{\bigwedge}S}]$.
This can be proved by defining an additional predicate ‘$S$’ in the language of lattices meant to correspond to ‘being an element of the set $S$’, the result then following easily from Łoś’ theorem. An alternative algebraic proof is as follows: Clearly $[\overline{\operatorname{\bigwedge}S}]$ is a lower bound for $S^*$. Suppose $[z]$ is another such lower bound and $[z]\not \leq [\overline{\operatorname{\bigwedge}S}]$. Then $\{i\in I\colon z(i)\not \leq \operatorname{\bigwedge}S\}\in U$, so $\{i\in I\colon \exists s_i\in S$ with $z(i)\not \leq s_i\}\in U$, $=u$ say (as $\operatorname{\bigwedge}S$ is the greatest lower bound of $S$). Define $x$ by $x(i)=s_i$ for $i\in u$ and $x(i)=\operatorname{\bigwedge}S$ otherwise. Then $[x]\in S^*$ but $[z]\not \leq [x]$, but this contradicts the assertion that $[z]$ is a lower bound.
\[dist root\] The class of $\vee(\operatorname{\bigwedge})$-distributive bounded lattices is closed under ultraroots.
By lemma \[inf exist\] if there is some $A\cup\{b\}\subseteq L$ with $b\vee\operatorname{\bigwedge}A\neq\operatorname{\bigwedge}(b\vee A)$ then $\operatorname{\bigwedge}A^* \vee [\operatorname{\bar{\mathit{b}}}]=[\bar {\operatorname{\bigwedge}A}]\vee [\operatorname{\bar{\mathit{b}}}]=[(\overline{\operatorname{\bigwedge}A) \vee} b]\neq [\overline{\operatorname{\bigwedge}(A\vee b})]=\operatorname{\bigwedge}(A^*\vee [\operatorname{\bar{\mathit{b}}}])$, so if $L$ is not $\vee(\operatorname{\bigwedge})$-distributive then neither is $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$.
\[ur1\] If $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ has a meet-complete representation then $L$ is $\vee(\operatorname{\bigwedge})$-distributive.
This follows from corollary \[dist root\] and the fact that when $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ is in $\operatorname{\textbf{mCRL}}$ it inherits $\vee(\operatorname{\bigwedge})$-distributivity from its representation.
By duality a similar result holds for $\operatorname{\textbf{jCRL}}$, and hence for $\operatorname{\textbf{biCRL}}$. In order for $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ to be in $\operatorname{\textbf{mCRL}}$ but $L$ not to be it turns out $L$ must satisfy an infinite density property, which we make precise in the next proposition.
\[ur2\] If $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ has a meet-complete representation but $L$ does not then there is a pair $x<y$ such that for every pair $a<b\in [x,y]$ there is some $c$ with $a<c<b$.
If $L$ is not in $\operatorname{\textbf{mCRL}}$ then there is a pair $x,y,\in L$ that cannot be distinguished by a complete, prime filter. Wlog assume $x<y$. Since $\operatorname{ \prod_{\mathit{U} } \mathit{L}}$ is in $\operatorname{\textbf{mCRL}}$, for each pair $a<b\in [x,y]$ there is a complete, prime filter $\gamma$ distinguishing $[\operatorname{\bar{\mathit{a}}}]$ and $[\operatorname{\bar{\mathit{b}}}]$. It’s easy to show that $\gamma_*$ is a prime filter of $L$ with $b\in \gamma_*$ and $a\notin \gamma_*$ (and thus $y\in \gamma_*$ and $x\notin \gamma_*$). Let $a<b$ and $(a,b)=\emptyset$ and suppose $S\subseteq \gamma_*$. Then for each $[z]\in S^*$ we have must have $[z]\vee[\operatorname{\bar{\mathit{a}}}]=[\operatorname{\bar{\mathit{b}}}]$, and thus by primality $S^*\subseteq \gamma$. So by lemma \[inf exist\] $\operatorname{\bigwedge}S\in \gamma_*$, and so $\gamma_*$ is complete, which is a contradiction as we assumed $x$ and $y$ could not be distinguished by a complete, prime filter.
Again by duality the same result holds for join-complete representations. Note that if we could find a counter example $(L, \prod_U L)$ where $\operatorname{ \prod_{\mathit{U} } \mathit{L}}\in\operatorname{\textbf{mCRL}}, \;$ $L\not\in\operatorname{\textbf{mCRL}}$, we could restrict to the sublattice bounded by $x$ and $y$, so we lose nothing by assuming that $x$ and $y$ are the lower and upper bounds respectively, and that the whole lattice therefore has this density property.
We have seen that the class of completely representable Boolean algebras is atomic (indeed finitely axiomatisable) and that the class $\operatorname{\textbf{CRL}}$ of completely representable lattices is not.
None of the classes $\operatorname{\textbf{jCRL}}, \operatorname{\textbf{mCRL}}, \operatorname{\textbf{biCRL}}$ is elementary.
We intend to prove this in a subsequent article.
[99]{} Birkhoff, G.: On the combination of subalgebras. Proc. Camb. Math. Soc. **29**, 441–464 (1933)
Birkhoff, G., Frink Jr., O.: Representations of lattices by sets. Trans. Am. Math. Soc. **64**, 299–316 (1948)
Chang, C., Keisler, H.: Model Theory, 3rd edn. North Holland, Amsterdam (1990)
Crawley, P., Dilworth, R.: Algebraic theory of lattices. Prentice Hall, Englewood Cliffs N.J (1973)
Gehrke, M., Jónsson, B.: Bounded distributive lattice expansions. Math. Scand. **94**, 13–45 (2004)
Hirsch, R., Hodkinson, I.: Complete representations in algebraic logic. J. Symb. Logic **62**, 816–847 (1997)
Hirsch, R., Hodkinson, I.: Relation Algebras by Games. North-Holland, Amsterdam (2002)
Hodges, W.: Model Theory. Cambridge University Press, Cambridge (1993)
Roman, S.: Lattices and Ordered Sets. Springer, New York (2008)
Sikorski, R.: Boolean Algebras, 3rd edition. Springer-Verlag, Berlin (1969)
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This paper explores energy-, momentum-, density-, and positivity-preserving spatio-temporal discretizations for the nonlinear Landau collision operator. We discuss two approaches, namely direct Galerkin formulations and discretizations of the underlying infinite-dimensional metriplectic structure of the collision integral. The spatial discretizations are chosen to reproduce the time-continuous conservation laws that correspond to Casimir invariants and to guarantee the positivity of the distribution function. Both the direct and the metriplectic discretization are demonstrated to have exact H-theorems and unique, physically exact equilibrium states. Most importantly, the two approaches are shown to coincide, given the chosen Galerkin method. A temporal discretization, preserving all of the mentioned properties, is achieved with so-called discrete gradients. Hence the proposed algorithm successfully translates all properties of the infinite-dimensional time-continuous Landau collision operator to time- and space-discrete sparse-matrix equations suitable for numerical simulation.'
author:
- Eero Hirvijoki
- 'Joshua W. Burby'
- Michael Kraus
bibliography:
- 'bibfile.bib'
title: 'Energy-, momentum-, density-, and positivity-preserving spatio-temporal discretizations for the nonlinear Landau collision operator with exact H-theorems'
---
Introduction {#sec:introduction}
============
This paper is devoted to investigating the discretization of the collisional relaxation problem encountered in plasmas. There, a distribution function $f(\bm{v},t):\mathbb{R}^3\times\mathbb{R}_{\geq 0}\mapsto \mathbb{R}_{\geq 0}$ is assumed to evolve according to the equation $$\begin{aligned}
\label{eq:landau}
\frac{\partial f}{\partial t}=\frac{\partial}{\partial \bm{v}}\cdot\int \limits_{\mathbb{R}^3}\mathbb{Q}(\bm{v}-\bm{v}')\cdot\left( f(\bm{v}')\frac{\partial f}{\partial\bm{v}}-f(\bm{v})\frac{\partial f}{\partial \bm{v}'}\right)d\bm{v}',\end{aligned}$$ corresponding to the dynamics driven by the nonlinear Landau collision operator [@Landau:1936]. The dyad $\mathbb{Q}(\bm{\xi})=(\mathbb{I}-\bm{\hat{\xi}}\bm{\hat{\xi}})/|\bm{\xi}|$ in the above expression is an inversely scaled projection matrix with an eigenvector $\bm{\xi}$ corresponding to zero eigenvalue, and $\bm{\hat{\xi}}=\bm{\xi}/|\bm{\xi}|$. For practical purposes, the discussion in this paper will be limited to the single-species collisional relaxation problem, and we will consider only the velocity-space evolution, using normalized units to avoid unnecessary clutter. Nothing, however, prevents generalizing our results to multiple species, if need be.
The above collisional relaxation problem can be formulated in a weak sense, given an arbitrary time-independent test function $u(\bm{v})$. The weak formulation is $$\begin{aligned}
\label{eq:weakform}
\frac{d}{dt}M(u, f)=C_f(u , \ln f),\end{aligned}$$ where the symmetric, bilinear forms $M$ and $C_f$ are defined according to $$\begin{aligned}
M(g,h)&=\int \limits_{\mathbb{R}^3} g h\,d\bm{v},\\
C_f(g,h)&=-\frac{1}{2}\iint \limits_{\mathbb{R}^3\times\mathbb{R}^3}\left(\frac{\partial g}{\partial \bm{v}}-\frac{\partial g}{\partial \bm{v}'}\right)\cdot f(\bm{v})\mathbb{Q}(\bm{v}-\bm{v}')f(\bm{v}')\cdot\left(\frac{\partial h}{\partial \bm{v}}-\frac{\partial h}{\partial \bm{v}'}\right) d\bm{v}d\bm{v}'.\end{aligned}$$ Provided that the distribution function $f$ is nonnegative, the form $C_f(u,w)$ is negative semidefinite, with a left-right null-space $\phi=\{|\bm{v}|^2,\bm{v},1\}$.
Using the weak formulation, it is straightforward to show that the functions $\phi(\bm{v})=\{|\bm{v}|^2,\bm{v},1\}$ generate invariant forms $M(\phi,f)$ with respect to the dynamics and correspond to the conservation laws of energy, momentum and density. These follow directly from the null-space of $C_f(u,w)$. Consequently, the condition $C_f(u,\ln f)=0$, with respect to arbitrary $u$, requires that $\ln f$ is a linear combination of the functions $\phi=\{|\bm{v}|^2,\bm{v},1\}$, corresponding to a Maxwellian equilibrium state. Furthermore, because $M(1,f)$ is an invariant and $C_f(u,u)\leq 0$, one finds that $\partial_t M(-\ln f,f)\ge 0$, which completes the H-theorem (monotonic entropy production that vanishes only for the equilibrium state). The final property of the Landau collision integral is that it preserves the positivity of the distribution function: assuming $f$ to be at least twice differentiable and non-negative, then, at a point $\bm{v}^{\star}$ where $f(\bm{v}^{\star})=0$, $\partial_{\bm{v}}f(\bm{v}^{\star})=0$, and $\partial^2_{\bm{v}\bm{v}}f(\bm{v}^{\star})$ is positive semi-definite, the evolution equation provides $$\begin{aligned}
\frac{\partial f(\bm{v}^{\star})}{\partial t}=\int\limits_{\mathbb{R}^3}\mathbb{Q}(\bm{v}^{\star}-\bm{v}')f(\bm{v}')d\bm{v}':\frac{\partial^2f(\bm{v}^{\star})}{\partial\bm{v}\partial\bm{v}}\geq 0.\end{aligned}$$
With this paper, we target spatio-temporal discretization methods that preserve the above mentioned properties exactly, to machine precision. We shall take two different routes which, in the end, are shown to coincide. In Sec. \[sec:direct-Galerkin\], we propose a Galerkin discretization of the weak form that “accidentally” succeeds in achieving all the desired properties in a spatially discrete but time-continuous system. In Sec. \[sec:metriplectic-structure\], we provide a mathematical explanation for this accident, demonstrating that the proposed spatial Galerkin discretization in fact defines a finite-dimensional metriplectic structure (for an introduction to metriplectic dynamics, see [@Morrison:1986vw; @Grmela:1984dn]). In Sec. \[sec:metriplectic-discretization\], we demonstrate how the very same spatial discretization is obtained directly from the underlying infinite-dimensional metriplectic structure of the Landau collision operator. Finally, in Sec. \[sec:temporal\], we propose a temporal integration method based on the concept of discrete gradients [@Quispel:1996; @McLachlan:1999; @CohenHairer:2011] which is shown to preserve all the desired properties to machine precision. The paper is concluded in Sec. \[sec:discussion\].
Direct Galerkin discretization {#sec:direct-Galerkin}
==============================
In the past and present, direct Galerkin methods for the Landau collision operator have received and are receiving significant attention from applied mathematicians [@Buet_LeThanh:hal-00092543; @Yoon:2014:POP; @Taitano:2015JCP; @Hager:2016:JCP]. Based on our previous work [@Hirvijoki:2017ei], we have learned that straightforward Galerkin discretizations of the form $$\begin{aligned}
\label{eq:old-discretization}
f_h(\bm{v},t)=\sum_if^i(t)\psi_i(\bm{v}),\end{aligned}$$ manage to preserve the conservation laws exactly, as long as the basis $\{\psi_i\}_{i\in I}$ is capable of representing quadratic functions exactly within the domain of support for the chosen basis. The conservation laws are thus somewhat trivial to achieve with polynomial second-order finite-element methods. Unfortunately, the preservation of positivity and the existence of an H-theorem are much trickier. Discretizations of type in general cannot guarantee the strict non-negativeness of $f_h$. This quickly turns into realizability issues in simulations and is against the basic principles of physics. After some reflection, we have found a simple, yet elegant solution to the positivity-preservation problem that happens to be consistent with conservation laws and the H-theorem. The recipe is simple – the root idea can be tracked down to the logarithm present in . This section is devoted to describing the recipe in detail.
We begin by choosing an *ab initio* positive discretization $$\begin{aligned}
\label{eq:new-discretization}
f_h=\exp(g_h), && g_h=\sum_{i\in I} g^i(t)\psi_i(\bm{v}),\end{aligned}$$ with $\{\psi_i\}_{i\in I}$ a second order Galerkin basis with compact support, and $\{g^i\}_{i\in I}$ the degrees of freedom for $g_h$. Choosing a test function $u=\psi_i$, direct substitution of to the weak formulation provides us with a nonlinear matrix equation $$\begin{aligned}
\label{eq:Galerkin-matrix}
\sum_{j\in I} M(\psi_i, f_h \psi_j) \frac{dg^j}{dt} = \sum_{j\in I} C_{f_h}(\psi_i , \psi_j)g^j \quad \forall\, i\in I,\end{aligned}$$ which is a linearly-implicit expression for the equations of motion of the degrees of freedom. Here, the integrals within the forms $M$ and $C_f$ are naturally limited to the domain of support for the basis. Also, note that the square matrices $M(\psi_i, f_h \psi_j)$ and $C_{f_h}(\psi_i , \psi_j)$ depend on the degrees of freedom via $f_h$, and that while $M(\psi_i, f_h \psi_j)$ is sparse, $C_{f_h}(\psi_i , \psi_j)$ is not. Equation will be our work horse throughout the rest of the paper.
Next we show that the discretization proposed above preserves the conservation laws exactly. We start by noting that the total energy, total momentum, and density can be written as $$\begin{aligned}
\label{eq:energy-momentum-density}
E=\sum_{i\in I}e^iM(\psi_i, f_h), && \bm{P}=\sum_{i\in I}\bm{v}^i M(\psi_i, f_h), && N=\sum_{i\in I}1^i M(\psi_i, f_h),\end{aligned}$$ where the coefficients $\{e^i\}_{i\in I}$, $\{\bm{v}^i\}_{i\in I}$, and $\{1^i\}_{i\in I}$ correspond to the degrees of freedom, or more precisely, to the expansion coefficients with respect to the chosen Galerkin basis for the functions $(|\bm{v}|^2,\bm{v},1)$, i.e., $$\begin{aligned}
|\bm{v}|^2=\sum_{i\in I} e^i\psi_i(\bm{v}) &&\bm{v}&=\sum_{i\in I} \bm{v}^i\psi_i(\bm{v}), && 1=\sum_{i\in I} 1^i\psi_i(\bm{v}).\end{aligned}$$ This follows from the request that the basis $\{\psi_i\}_{i\in I}$ exactly reproduces quadratic functions. Also note that these coefficients are unique for any polynomial Galerkin basis. Hence the time derivatives of energy, momentum, and density vanish identically $$\begin{aligned}
&\frac{dE}{dt}=\sum_{i,j\in I}e^i M(\psi_i, f_h \psi_j) \frac{dg^j}{dt}=\sum_{j\in I} C_{f_h}\left(\sum_{i\in I} e^i\psi_i , \psi_j\right)g^j=0,\\
&\frac{d\bm{P}}{dt}=\sum_{i,j\in I}\bm{v}^i M(\psi_i, f_h \psi_j) \frac{dg^j}{dt}=\sum_{j\in I} C_{f_h}\left(\sum_{i\in I} \bm{v}^i\psi_i , \psi_j\right)g^j=0,\\
&\frac{dN}{dt}=\sum_{i,j\in I}1^i M(\psi_i, f_h \psi_j) \frac{dg^j}{dt}=\sum_{j\in I} C_{f_h}\left(\sum_{i\in I} 1^i\psi_i , \psi_j\right)g^j=0,\end{aligned}$$ where we have used the equations of motion for the degrees of freedom , the bilinearity and the null-space of the form $C_f(u,w)$, and the requested property that the basis $\{\psi_i\}_{i\in I}$ reproduces quadratic functions exactly.
To prove the existence and uniqueness of an equilibrium state, as well as the H-theorem, we first note that $\{e^i\}_{i\in I}$, $\{\bm{v}^i\}_{i\in I}$, and $\{1^i\}_{i\in I}$ are the only eigenvectors of the matrix $C_{f_h}(\psi_i , \psi_j)$, that correspond to zero eigenvalues. This follows from the fact that, for the chosen second-order polynomial Galerkin basis, there exists only one unique set of coefficients, namely $\{e^i\}_{i\in I}$, $\{\bm{v}^i\}_{i\in I}$, and $\{1^i\}_{i\in I}$, in terms of which the null space $\phi=\{|\bm{v}|^2,\bm{v},1\}$ of the operator $C_f(u,w)$ can be expressed. Hence the equilibrium state $g^i_{\text{eq}}$ must be a linear combination $$\begin{aligned}
g^i_{\text{eq}}=a\, e^i + \bm{b}\cdot\bm{v}^i+c\, 1^i,\end{aligned}$$ which, within the support of our finite-element basis, corresponds to the numerical distribution function $$\begin{aligned}
f_{h,\text{eq}}=\exp\left(a|\bm{v}|^2+\bm{b}\cdot\bm{v}+c\right).$$ Note that this expression is to be evaluated only within the supporting domain for the basis $\{\psi_i\}_{i\in I}$. Outside, it has no meaning. Furthermore, because the energy, momentum, and density are conserved, the coefficients $a$, $\bm{b}$, and $c$ are uniquely determined in terms of the moments of a given initial state. To conclude the H-theorem, we note that the entropy $S=-\int f_h\ln f_h d\bm{v}$ can be written as $$\begin{aligned}
\label{eq:finite-dimensional-entropy}
S=-\sum_{i\in I}M(f_h,\psi_i)g^i.\end{aligned}$$ It’s time derivative then becomes $$\begin{aligned}
\frac{dS}{dt}&=-\sum_{i\in I}M(f_h,\psi_i)\frac{dg^i}{dt}-\sum_{i,j\in I}\frac{dg^j}{dt}M(\psi_jf_h,\psi_i)g^i\nonumber\\
&=-\frac{dN}{dt}-\sum_{i,j\in I}C_{f_h}(\psi_j,\psi_i)g^ig^j\nonumber\\
&\geq 0.\end{aligned}$$ The last line follows from the density conservation and the fact that the form $C_f(u,w)$ is negative semidefinite, with the only nontrivial zero solution being a linear combination of the operator’s null-space, corresponding to the equilibrium state. This concludes our proof for the direct Galerkin discretization.
Finally, we make some additional remarks. First, any generic map $$\begin{aligned}
f_h=\Psi(g_h), && g_h=\sum_ig^i(t)\psi_i(\bm{v}),\end{aligned}$$ with $\Psi:\mathbb{R}\mapsto\mathbb{R}_{\geq 0}$ and combined with a polynomial basis of second order, would succeed in reproducing the conservation laws and the non-negativity constraint. Second, any Galerkin basis that uniquely contains the functions $\phi(\bm{v})=\{|\bm{v}|^2,\bm{v},1\}$ within the support of the basis would suffice to provide the desired properties. For numerical purposes, it is convenient to use bases with compact support, though, as is implicit. The exception would be a basis for which $M(\psi_i,f_h\psi_i)$ would become diagonal. Lastly, if the dyad $\mathbb{Q}$ were replaced with something more complicated, as in the case of a relativistic collision operator, all of the above could be generalized as long as the chosen Galerkin basis would exactly and uniquely reproduce the null-space of the corresponding $C_f(u,w)$, within the domain of support for the basis.
Metriplectic structure of the direct Galerkin discretization {#sec:metriplectic-structure}
============================================================
At first, it seems as if the properties of our discretization scheme were a pure coincidence. There exists, however, a deeper level to the collision operator and its structure preserving discretizations. Specifically, the nonlinear Landau collision operator can be cast into an infinite-dimensional metriplectic system where the conservation laws correspond to so-called Casimir invariants [@Morrison:1986vw]. Consequently, a careful discretization of the structure provides a finite-dimensional metriplectic system. As it happens, the discretization described previously “accidentally” defines a finite-dimensional metriplectic structure, and that the very same structure can be derived systematically, by discretizing the infinite-dimensional metriplectic structure. In this section, we demonstrate the correspondence of the direct Galerkin discretization to a finite-dimensional metriplectic system.
Let us start by multiplying equation with the inverse of the matrix $M(\psi_i, f_h \psi_j)$, which leads to $$\begin{aligned}
\frac{dg^k}{dt} = \sum_{i,j\in I}M^{-1}(\psi_k, f_h \psi_i)\,C_{f_h}(\psi_i , \psi_j)\,g^j, \quad \forall\, k\in I.\end{aligned}$$ Next we use the finite-dimensional entropy , compute it’s derivative with respect to $g^{\ell}$, and invert for the vector $$\begin{aligned}
\label{eq:entropy-derivative}
g^j+1^j=-\sum_{\ell\in I} M^{-1}(\psi_{j},f_h\psi_{\ell})\frac{\partial S}{\partial g^{\ell}}, \quad \forall\, j\in I.\end{aligned}$$ In the next step, we use the fact that the vector $1^j$ is an eigenvector of the matrix $C_{f_h}(\psi_i,\psi_j)$ with a zero eigenvalue. This provides us the equations of motion in the form $$\begin{aligned}
\label{eq:metriplectic-equation-of-motion}
\frac{dg^k}{dt} = -\sum_{\ell\in I}G_{k\ell}(g)\frac{\partial S}{\partial g^{\ell}}, \quad \forall\, k\in I,\end{aligned}$$ where we have collected the individual matrices together and defined $$\begin{aligned}
\label{eq:metric-tensor}
G_{k\ell}(g)=\sum_{i,j\in I} M^{-1}(\psi_k, f_h \psi_i)\,C_{f_h}(\psi_i , \psi_j)\, M^{-1}(\psi_{j},f_h\psi_{\ell}).\end{aligned}$$ To reveal the metriplectic structure in it’s full glory, we consider time derivatives of generic functions $U(g)$ that depend only on the degrees-of-freedom $g=\{g^k\}_{i\in I}$, so that $$\begin{aligned}
\frac{d U(g)}{dt}=\sum_{k\in I}\frac{\partial U}{\partial g^k}\frac{d g^k}{dt}.\end{aligned}$$ With the help of this identity, we may cast into $$\begin{aligned}
\label{eq:metriplectic-structure}
\frac{d U}{dt} = -\sum_{k,\ell\in I}\frac{\partial U}{\partial g^k} G_{k\ell}(g)\frac{\partial S}{\partial g^{\ell}}, \quad \forall\, k\in I,\end{aligned}$$
The metriplectic structure of then follows from the facts that (i) the matrix $G_{k\ell}(g)$ is symmetric and negative semidefinite, (ii) it has the finite-dimensional energy, momentum, and density (defined in ) as Casimir invariants due to the conditions $$\begin{aligned}
\sum_{i\in I}\frac{\partial E(g)}{\partial g^i}\,G_{ij}(g)=0, &&
\sum_{i\in I}\frac{\partial \bm{P}(g)}{\partial g^i}\,G_{ij}(g)=0, &&
\sum_{i\in I}\frac{\partial N(g)}{\partial g^i}\,G_{ij}(g)=0, \quad \forall\, j\in I,\end{aligned}$$ and that (iii) it has a unique equilibrium state $$\begin{aligned}
g^i_{\text{eq}}=a\,e^i+\bm{b}\cdot\bm{v}^i+c\,1^i,\end{aligned}$$ with the coefficients $a$, $\bm{b}$, and $c$ defined from the initial state, and that the equilibrium state satisfies the condition $$\begin{aligned}
\sum_{i\in I}\frac{\partial S(g_{\text{eq}})}{\partial g^i}\,G_{ij}(g_{\text{eq}})=0, \quad \forall\, j\in I,\end{aligned}$$ To complete the cycle, we next demonstrate how , and consequently , can be obtained directly from discretizing the underlying infinite-dimensional metriplectic structure of the Landau collision operator.
Discretization of the infinite-dimensional metriplectic structure {#sec:metriplectic-discretization}
=================================================================
The infinite-dimensional metriplectic structure of the nonlinear Landau collision operator has been known for quite some time [@Morrison:1986vw]. Here we explain its connection to the weak formulation and demonstrate how discretization of it provides the same equations of motion as the direct Galerkin discretization introduced in Sec. \[sec:direct-Galerkin\], when the same discretization is assumed for the distribution function as there.
In terms of a negative semidefinite, symmetric bracket defined according to $$\begin{aligned}
({\mathcal{A}},{\mathcal{B}})[f]=C_f\left(\frac{\delta{\mathcal{A}}}{\delta f},\frac{\delta{\mathcal{B}}}{\delta f}\right),\end{aligned}$$ and an entropy functional defined as $$\begin{aligned}
{\mathcal{S}}[f]=M(-\ln f,f),\end{aligned}$$ the dynamics of $f$ that reproduce can be recovered from the functional differential equation $$\begin{aligned}
\label{eq:infinite-dimensional-motion}
\frac{d{\mathcal{U}}[f]}{dt}=({\mathcal{U}},-{\mathcal{S}}),\end{aligned}$$ by choosing a functional ${\mathcal{U}}[f]=M(u,f)$, and requiring the resulting equation to hold for all time-independent $u(\bm{v})$. This is left as an exercise for the reader to verify. Furthermore, entropy ${\mathcal{S}}[f]$, energy ${\mathcal{E}}[f]=M(|\bm{v}|^2,f)$, momentum $\bm{{\mathcal{P}}}[f]=M(\bm{v},f)$, and density ${\mathcal{N}}[f]=M(1,f)$ trivially satisfy $$\begin{aligned}
\frac{dS}{dt}\geq 0, && ({\mathcal{E}},{\mathcal{A}})=0, && ({\mathcal{P}},{\mathcal{A}})=0, && ({\mathcal{N}},{\mathcal{A}})=0,\end{aligned}$$ for arbitrary functionals ${\mathcal{A}}$, and the equilibrium state corresponding to $d{\mathcal{S}}/dt=0$ is achieved if and only if the functional derivative of the entropy is a linear combination of the functional derivatives of the Casimirs ${\mathcal{E}}$, ${\mathcal{P}}$, and ${\mathcal{N}}$, corresponding to a Maxwellian.
Before we proceed with the discretization, it is useful to investigate what happens if we use a map $f=\Psi(g)$. Since $$\begin{aligned}
\label{eq:modified-functional-derivative}
\frac{\delta {\mathcal{A}}[f]}{\delta f}=\frac{1}{\Psi'(g)}\frac{\delta {\mathcal{A}}[\Psi(g)]}{\delta g},\end{aligned}$$ A metric bracket with respect to the function $g$ is obtained after the substitution $$\begin{aligned}
({\mathcal{A}},{\mathcal{B}})[g]=C_{\Psi(g)}\left(\frac{1}{\Psi'(g)}\frac{\delta {\mathcal{A}}}{\delta g},\frac{1}{\Psi'(g)}\frac{\delta {\mathcal{B}}}{\delta g}\right).\end{aligned}$$ Our goal is to restrict $g$ to live within some finite-dimensional function space, basically choosing $$\begin{aligned}
g_h(\bm{v},t)=\sum_{i\in I} g^i(t)\psi_i(\bm{v}),\end{aligned}$$ and still preserve the Casimirs and the correct equilibrium state. This indicates that the expression should be chosen so that it becomes exact for the Casimirs and the entropy when evaluated with respect to $g_h$, so that the null-space of $C_f(u,w)$ can be exploited. For the Casimirs, this issue has been discussed in detail in [@Kraus-Hirvijoki-2017; @Hirvijoki-Kraus-Burby:2018arXiv]. To obtain the exactness also for the entropy functional corresponding to Maxwell-Boltzmann statistics, it turns out that the choice $\Psi(x)=\exp(x)$ is essential. With these guidelines, following [@Kraus-Hirvijoki-2017; @Hirvijoki-Kraus-Burby:2018arXiv] then leads to the conclusion that $$\begin{aligned}
\frac{\delta {\mathcal{A}}[f_h]}{\delta f}=\frac{1}{\exp(g_h)}\frac{\delta {\mathcal{A}}[\exp(g_h)]}{\delta g}=\sum_{i,j\in I}\frac{\partial {\mathcal{A}}_h(g)}{\partial g^i}M^{-1}(\psi_i,f_h\psi_j)\psi_j,\end{aligned}$$ where ${\mathcal{A}}_h(g)={\mathcal{A}}[f_h]={\mathcal{A}}[\exp(g_h)]$. The finite-dimensional bracket with respect to the degrees of freedom $\{g^i\}_{i\in I}$ is then obtained by substituting the discrete functional derivative to the infinite-dimensional bracket providing $$\begin{aligned}
({\mathcal{A}}_h,{\mathcal{B}}_h)_h(g)=({\mathcal{A}},{\mathcal{B}})[f_h]=\sum_{k,\ell\in I}\frac{\partial {\mathcal{A}}_h(g)}{\partial g^k} G_{k\ell}(g)\frac{\partial {\mathcal{B}}_h(g)}{\partial g^{\ell}},\end{aligned}$$ where the matrix $G_{k\ell}(g)$ is the one defined in . Finally, using , we obtain $$\begin{aligned}
\frac{d{\mathcal{U}}_h(g)}{dt}=({\mathcal{U}}_h,-{\mathcal{S}}_h)_h(g)=-\sum_{k,\ell\in I}\frac{\partial {\mathcal{U}}_h}{\partial g^k} G_{k\ell}(g)\frac{\partial {\mathcal{S}}_h}{\partial g^{\ell}},\end{aligned}$$ which is the same result as given in (simply replace ${\mathcal{U}}_h(g)$ with $U(g)$ etc.). All the properties are hence proven the same way as before. This concludes our discretization of the infinite-dimensional metriplectic structure.
Temporal discretization {#sec:temporal}
=======================
Thus far, we have managed to convert the infinite-dimensional Landau collision operator to a finite-dimensional, time-continuous ordinary differential equation which has been shown to respect all of the properties present in the infinite-dimensional system. As a next step, we propose a novel integration method, that translates all of these desired properties to discrete time, providing a fully conservative and thermodynamically consistent set of equations that can be implemented on a computer. While recently there has been an ambitious attempt towards integration methods for metriplectic systems analoguous of symplectic integrators for Hamiltonian systems [@Ottinger:2018], we take a different route here.
To begin, we introduce the so-called discrete gradient methods [@Quispel:1996; @McLachlan:1999; @CohenHairer:2011], that are often times used to construct integrators for Hamiltonian systems while numerically preserving first integrals, e.g., the Hamiltonian, to machine precision. Given an ordinary differential equation of the form $$\begin{aligned}
\frac{dg^k}{dt} = -\sum_{\ell\in I}G_{k\ell}(g)\frac{\partial S}{\partial g^{\ell}}, \quad \forall\, k\in I,\end{aligned}$$ and denoting time instances with subscripts according to $g(\delta t)=g_1$ and $g(0)=g_0$, discrete gradient methods temporally approximate the above ODE system according to $$\begin{aligned}
\label{eq:temporal-discretization}
\frac{g^k_1-g^k_0}{\delta t}=-\sum_{\ell\in I}\overline{G}_{k\ell}[g_0,g_1]\overline{\frac{\partial S}{\partial g^{\ell}}}[g_0,g_1], \quad \forall\, k\in I.\end{aligned}$$ The operator $\overline{\partial A/\partial g^{\ell}}[g_0,g_1]$ is referred to as the discrete gradient, and it is required to satisfy the properties $$\begin{aligned}
\label{eq:discrete-gradient-definition}
\sum_{\ell\in I}\overline{\frac{\partial A}{\partial g^{\ell}}}[g_0,g_1]\, (g_1^{\ell}-g_0^{\ell})=A(g_1)-A(g_0), &&
\overline{\frac{\partial A}{\partial g^{\ell}}}[g,g]=\frac{\partial A}{\partial g^{\ell}}(g).\end{aligned}$$ Many such operators are known in the literature [@HartenLaxLeer:1983; @Gonzalez:1996]. Furthermore, requiring $\overline{G}_{k\ell}(g,g)=G_{k\ell}(g)$ guarantees that the limit $\delta t\rightarrow 0$ collapses Eq. to the correct time-continuous ordinary differential equation.
Proceeding, and using the definition of the discrete gradient as well as the time-discrete evolution equation , we note that the temporally-discrete evolution of any function $U(g)$ now satisfies $$\begin{aligned}
U(g_1)-U(g_0)=-\delta t \sum_{k,\ell\in I}\overline{\frac{\partial U}{\partial g^{k}}}[g_0,g_1]\overline{G}_{k\ell}[g_0,g_1]\overline{\frac{\partial S}{\partial g^{\ell}}}[g_0,g_1].\end{aligned}$$ Hence, as long as the matrix operator $\overline{G}_{k\ell}[g_0,g_1]$ is negative semidefinite, entropy production will be guaranteed, according to $$\begin{aligned}
S(g_1)-S(g_0)=-\delta t\sum_{k,\ell\in I}\overline{\frac{\partial S}{\partial g^{k}}}[g_0,g_1]\overline{G}_{k\ell}[g_0,g_1]\overline{\frac{\partial S}{\partial g^{\ell}}}[g_0,g_1]\geq 0.\end{aligned}$$
Next we note the important result that will hint us on how to define the operator $\overline{G}_{k\ell}[g_0,g_1]$. For all Casimirs $C=\{E,\bm{P},N\}$, i.e., energy, momentum, and density, the derivative with respect to the degrees of freedom can be written in a convenient form, namely $$\begin{aligned}
\frac{\partial C}{\partial g^k}=\sum_{k\in I}c^i M(\psi_i,f_h\psi_k),\end{aligned}$$ with $c^i=\{e^i,\bm{v}^i,1^i\}$. A discrete gradient of the Casimirs is thus defined according to $$\begin{aligned}
\overline{\frac{\partial C}{\partial g^i}}[g_0,g_1]=\sum_{k\in I}c^k \overline{M}_{ki}[g_0,g_1],\end{aligned}$$ where $\overline{M}_{ki}[g_0,g_1]$ is required to satisfy the condition $\overline{M}_{ki}[g,g]=M(\psi_i,f_h\psi_k)$. The specific form of the matrix $\overline{M}_{ij}[g_0,g_1]$ depends on the chosen discrete gradient. We will provide a particularly convenient, explicit form soon. With these remarks, we see that the temporally-discrete evolution of the Casimirs satisfies $$\begin{aligned}
C(g_1)-C(g_0)=-\delta t\sum_{i\in I}\sum_{k,\ell\in I}c^{i} \overline{M}_{i k}[g_0,g_1]\overline{G}_{k\ell}[g_0,g_1]\overline{\frac{\partial S}{\partial g^{\ell}}}[g_0,g_1].\end{aligned}$$ If we are to achieve the discrete-time Casimir invariance $C(g_1)-C(g_0)=0$ for all possible state vectors $(g_1,g_0)$, we must choose $$\begin{aligned}
\label{eq:approximate-G}
\overline{G}_{k\ell}[g_0,g_1]=\sum_{i,j\in I} \overline{M}^{-1}_{ki}[g_0,g_1]\,C_{f_{h,1/2}}(\psi_i , \psi_j)\, \overline{M}^{-1}_{j\ell}[g_0,g_1],\end{aligned}$$ where $\overline{M}^{-1}_{ij}[g_0,g_1]$ is the inverse of the matrix $\overline{M}_{ij}[g_0,g_1]$. This is also a valid choice: since we required $\overline{M}_{ij}[g,g]=M(\psi_i,f_h\psi_j)$, we trivially have $\overline{M}^{-1}_{ij}[g,g]=M^{-1}(\psi_i,f_h\psi_j)$, and hence $\overline{G}_{k\ell}(g,g)=G_{k\ell}(g)$. The choice then provides the desired result $$\begin{aligned}
C(g_1)-C(g_0)=-\delta t\sum_{i\in I}\sum_{j,\ell\in I}c^{i} C_{f_{h,1/2}}(\psi_i , \psi_j)\, \overline{M}^{-1}_{j\ell}[g_0,g_1]\overline{\frac{\partial S}{\partial g^{\ell}}}[g_0,g_1]=0,\end{aligned}$$ which follows from the property that the basis $\{\psi_i\}_{i\in I}$ can present the functions $\phi=\{|\bm{v}|^2,\bm{v},1\}$ exactly, and due to the null space of the form $C_f(u,w)$, which together lead to $$\begin{aligned}
\sum_{i\in I}c^{i} C_{f_{h}}(\psi_i , \psi_j)=0, \quad \forall\, j\in I.\end{aligned}$$
The final step is to verify the existence of a unique equilibrium sate. For an equilibrium state to exist, one must have $g_1=g_0=g_{\text{eq}}$. This requirement, and the evolution equation , provides $$\begin{aligned}
\sum_{\ell\in I}\overline{G}_{k\ell}[g_{\text{eq}},g_{\text{eq}}]\overline{\frac{\partial S}{\partial g^{\ell}}}[g_{\text{eq}},g_{\text{eq}}]=0, \quad \forall\, k\in I.\end{aligned}$$ Next, using the defining properties $\overline{\partial S/\partial g^{\ell}}[g_{\text{eq}},g_{\text{eq}}]=\partial S/\partial g^{\ell}(g_{\text{eq}})$ and $\overline{G}_{k\ell}[g_{\text{eq}},g_{\text{eq}}]=G_{k\ell}(g_{\text{eq}})$, we obtain $$\begin{aligned}
\sum_{\ell\in I}G_{k\ell}(g_{\text{eq}})\frac{\partial S}{\partial g^{\ell}}(g_{\text{eq}})=0, \quad \forall\, k\in I.\end{aligned}$$ From here the uniqueness of the equilibrium state follows trivially after using and the null-space argument, leading to the observation that the numerical equilibrium state is given by $$\begin{aligned}
f_{h,\text{eq}}(\bm{v})=\exp(a|\bm{v}|^2+\bm{b}\cdot\bm{v}+c),\end{aligned}$$ as expected.
The sparse matrix system
========================
To conclude our derivations, we choose a convenient discrete gradient method, and provide explicit expressions for all necessary terms, simultaneously converting into a sparse form suitable for iterative inversion techniques.
We will use the second order $\mathcal{O}(\delta t^2)$, so-called average discrete gradient [@HartenLaxLeer:1983], that is defined according to $$\begin{aligned}
\overline{\frac{\partial A}{\partial g^{\ell}}}[g_0,g_1]=\int\limits_0^1\frac{\partial A}{\partial g^{\ell}}((1-\xi)g_0+\xi g_1)d\xi.\end{aligned}$$ The explicit expression for the matrix $\overline{M}_{ij}[g_0,g_1]$ then becomes $$\begin{aligned}
\overline{M}_{ij}[g_0,g_1]=\int \psi_i\frac{\exp\left(g_{h0}\right)-\exp\left(g_{h1}\right)}{g_{h0}-g_{h1}}\psi_j\,d\bm{v},\end{aligned}$$ and the average discrete gradient of the entropy is given by $$\begin{aligned}
\overline{\frac{\partial S}{\partial g^{\ell}}}[g_0,g_1]=\overline{\frac{\partial S - 1}{\partial g^{\ell}}}[g_0,g_1]-\overline{M}_{\ell j}[g_0,g_1]\,1^j,\end{aligned}$$ where the vector $\overline{\partial S - 1/\partial g^{\ell}}[g_0,g_1]$ is defined as $$\begin{aligned}
\overline{\frac{\partial S - 1}{\partial g^{\ell}}}[g_0,g_1]=-\int \psi_{\ell}\frac{(g_{h0}-1)\exp\left(g_{h0}\right)-(g_{h1}-1)\exp\left(g_{h1}\right)}{g_{h0}-g_{h1}}\,d\bm{v}.\end{aligned}$$ For practical reasons, we have introduced the short notation $$\begin{aligned}
g_{h0}=\sum_{k\in I}g_0^k\psi_k, && g_{h1}=\sum_{k\in I}g_1^k\psi_k.\end{aligned}$$ Putting everything together, and using the null-space condition once more to remove the coefficients $1^j$ in the gradient of entropy, we obtain a coupled sparse-matrix system, $$\begin{aligned}
\label{eq:sparse-system}
\sum_{k\in I}\overline{M}_{ik}[g_0,g_1]\frac{g^k_1-g^k_0}{\delta t}&=-\sum_{j\in I}C_{f_{h,1/2}}(\psi_i , \psi_j)F_{j}, \quad \forall\, i\in I,\\
\sum_{j\in I}\overline{M}_{ij}[g_0,g_1]F_{j}&=\overline{\frac{\partial S - 1}{\partial g^{i}}}[g_0,g_1], \quad \forall\, i\in I,\end{aligned}$$ where $f_{h,1/2}=\exp((g_{h0}+g_{h1})/2)$. This system provides a nonlinear but sparse equation for solving $g_1$ in terms of $g_0$ (the dense matrix–vector product $C_{f_{h,1/2}}(\psi_i , \psi_j)F_{j}$ can be factored into sparse matrix–vector products as in [@Kraus-Hirvijoki-2017]) and, if solved to machine precision, will provide all the desired properties to machine precision as well.
Conclusions {#sec:discussion}
===========
In this article we have explored spatio-temporal discretizations for the nonlinear Landau collision operator. Our results consisted of three milestones: (i) the infinite-dimensional Landau collision operator was succesfully converted to an ODE with desired physical properties, (ii) this ODE was shown to have a metriplectic structure, and (iii) a successful temporal discretization of the metriplectic structure was achieved retaining all of the properties of the original infinite-dimensional system. Perhaps the most important observation was (ii), as without it, it would have been quite difficult to guess a successful temporal discretization scheme. Hence the milestone (ii) stresses the value of discretization methods that are based on the infinite-dimensional metriplectic formulation and, by preserving it’s structure, annihilate the guesswork. Overall, the presented results constitute an important step on the path towards ever more refined structure-preserving discretization methods for the full Vlasov-Maxwell-Landau system.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present UV images of M 31 and M 32, as observed by HST with the refurbished FOC. The galaxies were observed through the F175W and F275W filters, allowing the construction of color magnitude diagrams (CMDs) for the hundreds of detected sources found in each image. Comparison of these data with the stellar evolutionary tracks of horizontal branch stars and their progeny shows that for the first time outside of our own Galaxy, we may be measuring the colors of individual stars that are evolving along post asymptotic giant branch (PAGB), post-early AGB, and AGB-Manqu$\acute{\rm e}$ paths. Searching to the 6$\sigma$ detection limit, we find 1349 stars in M 31 and 183 stars in M 32. We compare the distribution of stars in the CMDs with the expectations from theory.'
address: |
$^{ \dagger}$NASA/GSFC, Code 681, Greenbelt, MD 20771\
$^{ \dagger\dagger}$STScI, 3700 San Martin Drive, Baltimore, MD 21218\
$^{\dagger\dagger\dagger}$IGPP, LLNL, 7000 East Avenue, Livermore, CA 94550\
$^*$Laboratoire d’Astronomie Spatiale du CNRS, Traverse du Siphon, Les Trois Lucs, F-13012, Marseille, France\
$^{**}$Dpt. of Physics & Astronomy, JHU, 3400 N. Charles Street, Baltimore, MD 21218
author:
- 'Thomas M. Brown $^{ \dagger}$, Henry C. Ferguson$^{ \dagger\dagger}$, S. Adam Stanford$^{\dagger\dagger\dagger}$, Jean-Michel Deharveng$^*$, and Arthur F. Davidsen$^{**}$'
title: Hot Evolved Stars in the Centers of M 31 and M 32
---
Motivation {#motivation .unnumbered}
==========
The spectra of elliptical and spiral galaxy bulges exhibit a strong upturn shortward of 2700 Å, dubbed the “UV upturn.” Characterized by the $m_{1550}-V$ color, the UV upturn shows strong variation (ranging from 2.05–4.50 mag) in nearby quiescent galaxies. The UV upturn is thought to vary with the composition of an evolved stellar population, because the lifetimes of stars leaving the horizontal branch vary by orders of magnitude. Stars that evolve along the asymptotic giant branch (AGB) and post-AGB evolutionary paths are short-lived in comparison to the AGB-Manqu$\acute{\rm e}$ stars that spend 10$^7$ yr at high luminosity (100 L$_{\odot}$) and high temperature (T$_{eff} > 30,000$K). The longer-lived stars are more efficient UV emitters, and thus it is thought that they comprise a greater proportion of the populations with the stronger UV upturns (see Brown et al. 1997 for a review[@brown97]).
Many groups have tried to characterize the stellar populations in early-type galaxies by fitting their composite spectral energy distributions (SEDs) with the synthetic spectra of model stellar populations (cf. Brown et al. 1997[@brown97]). However, the considerable ambiguity in this process has not determined unequivocally the populations producing the UV flux. Our goal in this work was to directly characterize these stellar populations through UV photometry of the individual stars in two galaxies: M 31 and M 32. To pursue this goal, we have taken deep exposures of these galaxies with the F175W and F275W filters on the Faint Object Camera (FOC). The resulting color-magnitude diagrams (CMDs) that can be constructed from this photometry can be compared to the predictions of stellar evolutionary theory.
Observations {#observations .unnumbered}
============
The observations for this work are summarized in Table 1. Our photometry was performed with the IRAF package apphotx, which we used to find stars that were detected at 6$\sigma$ above the diffuse background. This search found 1349 sources in M 31 and 183 in M 32. We then simulated the data in order to determine our completeness limits and spurious source contamination. These simulations show that our detection capability drops below 50% for stars dimmer than 24$^{th}$ magnitude in each individual filter for a given galaxy. However, our actual detection capability is somewhat better than that estimated from these simulations, since we generate our source list by searching in both filters. Our magnitudes use the STMAG system; $m_o$ in Table 1 corresponds to 1 count/sec in the filter. We have constructed 3-color images of the galaxies (see Ferguson & Brown, this volume) using our own FOC data and archival WFPC2 data.
[|l||c|c||c|c|]{} & &\
R.A.$_{2000}$&&\
Dec.$_{2000}$ & &\
$D$ (kpc) & &\
$E(B-V)$ (mag) & &\
$m_{1550}-V$ (mag)& &\
& F175 & F275 & F175 & F275\
Exp. (sec) & 19773 & 8390 & 16179 & 8990\
PSF FWHM (pix) & 4.6 & 4.0 & 4.3 & 3.9\
$m_o$ & 19.43 & 21.47 & 19.43 & 21.47\
=4.75in
=4.75in
Comparison with Theory {#comparison-with-theory .unnumbered}
======================
In Figs. 1 and 2, we have produced CMDs for these sources, and compared them with stellar evolutionary tracks for HB and post-HB stars. The evolutionary tracks were placed in the CMDs by finding the closest Kurucz (1993[@kurucz93]) synthetic spectrum for each step along the track, normalizing it to agree with the distance and reddening for the galaxies (Table 1), and then folding it through the appropriate instrument response curves using the IRAF task calcphot. We have also performed Monte Carlo simulations of the CMD that would arise from a population of extreme horizontal branch (EHB) stars, chosen to reproduce the M 31 F175W countrate. In our simulations, this population is assumed to originate from a flat mass distribution on the blue end of the HB, i.e., the number of stars per unit mass is uniform across the EHB. Our simulations show that there should be a gap in the distribution of stars in each CMD (see Figs. 1 & 2). The gap is present because of the rapid evolution between two evolutionary phases: the core He burning on the HB, and the later shell He burning in the “slow blue phase.”
Although the exposure times on both galaxies are approximately equal, the M 31 luminosity function peaks at $m_{275} = 26$ mag and $m_{175} = 25.5$ mag, while the M 32 luminosity function peaks at $m_{275} = 25$ mag and $m_{175} = 24$ mag. This is in line with theories of the UV upturn, since one would expect brighter (and thus short-lived) stars in the galaxy with the weaker UV upturn.
Puzzles {#puzzles .unnumbered}
=======
Since the distribution of stars in our CMDs does not agree with the expectations from our simulations, we compared our data to that from other investigations, in order to investigate possible problems with our data. These investigations turned up a few more discrepancies:
> $\bullet$ The F175W countrate in both of our images is approximately 60% higher than that predicted from IUE observations (private communication Calzetti) of M 31 and M 32.
>
> $\bullet$ Comparison of common sources in our data and that of pre- COSTAR F175W data (King et al. 1992[@king92]) shows our sources to be 1.1 mag fainter.
>
> $\bullet$ Comparison of common sources in our data and that of archival WFPC2 F300W data shows that the mean $m_{275}-m_{300}$ color for these sources is 0.23 mag, when it should be between 0.0 and -0.20 mag for hot UV sources at 10,000 K $\leq T_{eff} \leq$ 50,000 K.
Because of these discrepancies, we cannot as yet characterize the stellar populations in M 31 and M 32. We are attempting checks of the data calibration and our own calculations. This work is still in progress.
Brown, T. M., Ferguson, H. C., Davidsen, A. F., & Dorman, B. 1997, ApJ, 482, 685.
Dorman, B., Rood, R. T., & O’Connell, R. W. 1993, ApJ, 419, 596.
Ferguson, H. C., & Davidsen, A. F. 1993, ApJ, 408, 92.
King, I. R., et al. 1992, ApJ, 397, L35.
Kurucz, R. L. 1993, CD-ROM 13, ATLAS9 Stellar Atmosphere Programs and 2 km/s Grid (Cambridge: Smithsonian Astrophys. Obs.).
Vassiliadis, E., & Wood, P. R. 1994, ApJS, 92, 125.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The ground-state spins and magnetic moments of $^{49,51}$K have been measured using bunched-beam high-resolution collinear laser spectroscopy at ISOLDE-CERN. For $^{49}$K a ground-state spin $I = 1/2$ was firmly established. The observed hyperfine structure of $^{51}$K requires a spin $I > 1/2$ and from its magnetic moment $\mu(^{51}\text{K})= +0.5129(22)\, \mu_N$ a spin/parity $I^\pi=3/2^+$ with a dominant $\pi 1d_{3/2}^{-1}$ hole configuration was deduced. This establishes for the first time the re-inversion of the single-particle levels and illustrates the prominent role of the residual monopole interaction for single-particle levels and shell evolution.'
author:
- 'J. Papuga'
- 'M. L. Bissell'
- 'K. Kreim'
- 'K. Blaum'
- 'B.A. Brown'
- 'M. De Rydt'
- 'R. F. Garcia Ruiz'
- 'H. Heylen'
- 'M. Kowalska'
- 'R. Neugart'
- 'G. Neyens'
- 'W. Nörtershäuser'
- 'T. Otsuka'
- 'M. M. Rajabali'
- 'R. Sánchez'
- 'Y. Utsuno'
- 'D. T. Yordanov'
bibliography:
- 'References-G-short.bib'
title: 'Spins and Magnetic Moments of $^{49}$K and $^{51}$K: establishing the 1/2$^+$ and 3/2$^+$ level ordering beyond $N$ = 28.'
---
The nuclear shell-model forms the basis for our understanding of atomic nuclei and since the very beginning spins and magnetic moments of ground states have played a crucial role [@Mayer48; @Mayer49]. Single-particle as well as collective degrees of freedom in atomic nuclei can be described with modern large-scale shell-models [@Caurier2005]. However, the interplay between theory and experiment is indispensable for further improving the shell-model effective interactions as new regions of the nuclear chart are being explored.
Since more and more rare-isotope beams became available, strong modifications to the well-known shell structure were required in several regions of the nuclear chart. Examples are the unexpected level ordering in the “island of inversion” isotopes $^{31,33}$Mg [@Neyens2011], the weakening of the $N=28$ shell gap below Ca [@Bastin2007], the monopole migration of proton single-particle levels towards $N=28$ [@Gade2006] and $N=50$ [@Flanagan2009]. The origin of the changes in shell structure has been discussed in several theoretical papers [@Otsuka2001; @Otsuka2005; @Otsuka2006; @Smirnova2010; @Otsuka2010]. As experimental evidence is growing, effective shell-model interactions as well as mean-field models are being modified to account for the new observations [@Grasso2007; @Nowacki2009; @Zalewski2009; @Kaneko2011; @Wang2011].
In the region below Ca ($Z=20)$, where protons occupy the $sd$ shell and neutrons occupy the $pf$ shell, the gradual filling of the $\nu 1f_{7/2}$ orbit from $N = 20$ to $N = 28$ leads to a strong decrease of the first excited $1/2^+$ energy in the $_{19}$K and $_{17}$Cl isotopes. This level becomes the ground state in $^{47}$K [@Touchard1982] and in the Cl chain the inversion begins at $^{41}$Cl [@Ollier2003]. Following the explanation given by Otsuka *et al.* [@Otsuka2005], occupation of the $\nu 1f_{7/2}$ orbit (having $j=l+1/2$) leads to an increased binding of the $\pi 1d_{3/2}$ orbital (with $j^{'}=l^{'}-1/2$), and thus to a near-degeneracy (even inversion) with the $\pi 2s_{1/2}$ orbit as the $\nu 1f_{7/2}$ orbit is fully occupied at $N=28$.
This raises the question as to what happens beyond $N=28$, when the higher $pf$ orbits are being filled. Using a Woods-Saxon potential e.g. the monotonic lowering of the $\pi 1d_{3/2}$ energy with respect to the $\pi 2s_{1/2}$ energy is predicted to continue well beyond $N=28$. However, it is by now also well-known that the residual monopole interaction, which changes with isospin and thus with neutron number, plays an important role in the evolution of single-particle levels. This residual monopole interaction has a central, spin-orbit and tensor term, all of which need to be determined from experimental data. Furthermore, these monopole interactions are strongest between orbits with the same number of nodes, and thus in the region beyond $N=28$ it is the $\pi 2s_{1/2} - \nu 2p_{3/2}$ interaction strength that is the dominant one [@Otsuka2010]. Until now, very little experimental data are available which probe this part of the nucleon-nucleon residual interaction. A rather extended level scheme is available for $^{49}$K [@Broda2010] but all spins are tentative as long as the ground-state spin is not established. These recent in-beam data favor a ground-state spin $I=1/2$ while earlier $\beta$-decay work suggests $I=3/2$ [@Carraz1982]. For $^{51}$K no excited states are known. From $\beta$-decay studies the ground-state spin is tentatively assigned to be $I=3/2$ with a dominant $\pi 1d_{3/2}$ hole structure [@Perrot2006].
This letter presents the measured ground-state spins and magnetic moments of $^{49,51}$K with two and four neutrons beyond $N=28$ respectively, thus gradually filling the $\nu 2p_{3/2}$ orbit. The K$^{+}$ beams were produced at the ISOLDE facility at CERN where 1.4 GeV protons impinged on a thick $\text {UC}_{x}$ target. Atoms diffused out of the target and were surface ionized in a heated tube at $\sim$2000 $^{\circ}$C. The ions were accelerated up to 40 kV, mass separated and bunched by the gas-filled Paul trap (ISCOOL) [@Franberg2008; @Flanagan2009]. Collinear laser spectroscopy [@Mueller1983] was performed using the 769.9 nm $4s\;{^2S_{1/2}} \rightarrow 4p\;{^2P_{1/2}}$ atomic transition in K, after the beam was neutralized via collisions with neutral K atoms in a charge exchange cell (CEC). To scan the hyperfine structure the Doppler-shifted laser frequency, as experienced by the fast atoms, was modified by applying a tunable scanning voltage of $\pm 500$ V on the CEC. The fluorescence light emitted from the resonantly laser-excited atoms was subsequently detected using photomultiplier tubes (PMTs). By gating the PMT signals such that photons were only collected during the $\sim$6 $\mu$s period when the bunch of atoms pass in front of the PMT, the background photon count rate was reduced by a factor of 15000 (= ISCOOL accumulation time/ion bunch temporal width) compared to a continuous beam detection.
The relatively slow decay rate of the atomic transition ($3.8 \times 10^7$ s$^{-1}$), low quantum efficiency (2.5%) and high heat-related dark counts of PMTs operating at these wavelengths hinder optical spectroscopy of exotic K isotopes. In order to perform the measurements on a $^{51}$K beam of $\sim$4000 ions/s, a new optical detection station was developed. Details on this will be presented in a forthcoming paper [@Kreim2013].
{width="80.00000%"}
Typical hyperfine structure (hfs) spectra for $^{49}$K and $^{51}$K are shown in Fig. \[HFS-49-51K\]. To convert the recorded scanning voltage into frequency, AME 2012 [@Wang2012] masses were used with the recently measured masses of $^{49}$K [@Lapierre2012] and $^{51}$K [@Gallant2012] already included. The hfs spectra have been fitted using the $\chi^2$ minimization procedure MINUIT. The peak positions were determined by the hyperfine parameters $A(^{2}S_{1/2})$ and $A(^{2}P_{1/2})$, the nuclear spin $I$ and the center of gravity of the structure defining the isotope shift. The $A$-parameters are directly related to the nuclear magnetic moment through the relation $ A = {\mu B_{0}}/{I J}$, where $B_{0}$ is the magnetic field generated by electrons at the site of a nucleus. Voigt profiles with common line widths have been used for all peaks in the spectrum. The nuclear spin determines the number of allowed transitions. For $I=1/2$ only three transitions are allowed (left of Fig. \[HFS-49-51K\]) and the ground-state spin of $^{49}$K can be unambiguously assigned as $I=1/2$. For all other spins four resonances appear in the spectrum, as illustrated in the right part of Fig. \[HFS-49-51K\]. We have therefore fitted the $^{51}$K spectra assuming a spin 3/2, 5/2 and 7/2. In the case of a $J=1/2$ to $J=1/2$ transition the ratio of the upper-to-lower hyperfine parameter cannot be used to exclude a particular spin (method used e.g. to assign the $^{72,74}$Cu spins [@Flanagan2010]), as the *relative* hyperfine structure separations are unaffected by the nuclear spin in this case. There is, however, an indirect way to determine the $^{51}$K spin, namely by comparing the intensities of the different hyperfine structure components. From the intensity ratios observed for $^{51}$K we can exclude a spin 5/2 (or even higher) at a confidence level of 95$\%$. To further confirm the measured spin $I = 3/2$, the extracted magnetic moment for different spin assumptions can be compared to single-particle moments and predictions by shell-model calculations.
![\[g-mu\] (Color online) Experimental $g$ factors (upper) and magnetic moments (lower) from [@Beckmann1974; @Touchard1982] (dots) and the new data (stars) for $^{49}$K (1/2) and $^{51}$K (3/2). Effective single-particle $g$ factors are calculated with $g_{s}^{\rm{eff}}$=0.85$g_{s}^{\rm{free}}$ and $g_l^\pi$=1.15, $g_l^\nu$=-0.15. See text for details on calculations.](g-mu){width="40.00000%"}
[c x e f]{}
Isotope & & &\
\
$^{39}$K & 3/2 & +231.0 (3) & +27.8(2)\
$^{47}$K & 1/2 & +3413.2 (2) & +411.8(2)\
$^{49}$K & 1/2 & +2368.2(14) & +285.6(7)\
$^{51}$K & 3/2 & +302.5(13) & +36.6(9)\
$^{51}$K & (5/2) & +201.6 (9) & +24.4(6)\
$^{51}$K & (7/2) & +151.3 (7) & +18.3(4)\
\
The magnetic moments are deduced from the $A(^2S_{1/2})$ values relative to that of $^{39}$K, for which a very precise measurement of $A_{\rm{ref}}(^{2}S_{1/2})$ = 230.8598601(7) MHz and $\mu_{\rm{{ref}}} = +0.3914662(3)\, \mu_{N} $ is available from an atomic beam magnetic resonance measurement [@Beckmann1974]. The results are presented in Table \[moments\]. The statistical errors are smaller than the hyperfine anomalies measured for $^{40,41,42}$K relative to $^{39}$K [@Beckmann1974; @Eisinger1952; @Chan1969]. Following the approach of Bohr [@Bohr1951] these hyperfine anomalies are well reproduced and those for $^{47,49,51}$K are predicted to be less than $0.3\%$. This is included as an additional error on the magnetic moments (in square brackets).
[c c y y z z z]{} Isotope & $I^{\pi}$ & & & & &\
\
$^{39}$K & $3/2^+$ & +0.2611 (3) \[8\] &+0.3917 (5)\[12\] & +0.63 & +0.63 & +0.65\
$^{47}$K & $1/2^+$ & +3.8584 (2)\[116\] &+1.9292 (1)\[58\] & +1.87 & +1.91 & +1.91\
$^{49}$K & $1/2^+$ & +2.6772(16) \[80\] & +1.3386 (8)\[40\] & +1.61 & +1.81 & +1.72\
$^{51}$K & $3/2^+$ & +0.3420(15) \[10\] & +0.5129(22)\[15\] & +0.60 ([E]{}=0) & +0.63 ([E]{}=0) & +0.51 ([E]{}=0)\
$^{51}$K & ($5/2^+$) & +0.2279(10) \[7\] &+0.5698(25)\[17\] & +0.84 ([E]{}=1725) & +1.40 ([E]{}=2018) & +0.75 ([E]{}=2264)\
$^{51}$K & ($7/2^+$) & +0.1710 (8) \[5\] &+0.5986(28)\[18\] & -0.13 ([E]{}=1793) & -0.04 ([E]{}=2040) & -0.10 ([E]{}=2048)\
The ground-state wave function of the odd-$A$ potassium isotopes is dominated by one proton hole in the $Z = 20$ shell. The Schmidt moments and thus also the free-nucleon $g$ factors of the relevant single-particle orbits $\pi 1d_{3/2}$, $\pi 2s_{1/2}$ and $\pi 1f_{7/2}$ orbits are very different from each other (respectively +0.08, +5.58 and +1.65). Therefore their $g$ factors are an excellent probe to monitor in which orbital the unpaired proton occurs. Because the single nucleon does not appear as a free particle, the experimental values are compared to effective single-nucleon $g$ factors, with typical values for the $sd$ shell [@Richter2008]. The upper part of Fig. \[g-mu\] illustrates that the experimental $g$ factors of $^{39-45}$K are all close to the effective value for a hole in the $\pi 1d_{3/2}$ orbit ($g_{\rm{{eff}}} = +0.25$). Also the $^{51}$K $g$ factors (assuming different spins, given in column 3 of Table \[moments\]) agree very well with this value. This confirms the dominant $\pi 1d_{3/2}^{-1}$ component in the ground-state wave function of each of these isotopes, including $^{51}$K. The ground-state spins are known to be $3/2^{+}$ up to $^{45}$K and this result supports the $3/2^{+}$ assignment for the $^{51}$K ground state. The $^{47}$K $g$ factor is close to the effective single-particle value for a $\pi 2s_{1/2}^{-1}$ hole configuration and its spin/parity is known to be 1/2$^+$ [@Touchard1982]. The ground-state spin of $^{49}$K is established to be also $I=1/2$, but its $g$ factor suggests a rather mixed wave function.
In what follows we focus on further establishing the ground-state spin of $^{51}$K. For this we compare the experimental magnetic moments to values calculated with different effective shell-model interactions. Calculations have been performed with the SDPF-NR [@Retamosa1997; @Nummela2001] and its recently upgraded SDPF-U interaction [@Nowacki2009] as well as with the recently developed SDPF-MU interaction [@UtsunoArXiv] (Table \[moments\]). The SDPF-NR and SDPF-U interactions have monopole matrix elements that were tuned by fitting experimental spectra of isotopes in this region, from O to Ca. The SPDF-MU interaction is based on the recently developed monopole-based universal interaction $V_{\text{MU}}$ [@Otsuka2010] and involves therefore fewer fitted parameters. Calculations are performed with protons restricted to the $sd$ shell (thus only positive parity levels are calculated) and neutrons in the full $pf$ shell. The experimental magnetic moments for the $3/2^+$ ground states of $^{39-45}$K are well reproduced by all effective interactions (lower part of Fig. \[g-mu\]), and the agreement for the $3/2^+$ state in $^{51}$K is even better than 0.1 $\mu_N$ for all of them. The lowest positive parity states (5/2$^{+}$, 7/2$^{+}$) appear around 2 MeV and have magnetic moments that deviate significantly more from the experimental values (Table \[moments\]). Negative parity $3/2^{-}$, $5/2^{-}$, $7/2^{-}$ states, due to a proton excited in the $pf$ shell, all have a magnetic moment that is larger than +3.3 $\mu_{\rm{N}}$, incompatible with the observed small value around +0.5 $\mu_{\rm{N}}$. All arguments together allow to conclude that the ground-state spin/parity of $^{51}$K is $3/2^{+}$. The magnetic moment of the $1/2^+$ ground state in $^{47}$K is also reproduced very well, while the experimental moment of the $I=1/2^+$ ground state of $^{49}$K is somewhat overestimated by all calculations, suggesting that some particular mixing in the wave function is not well taken into account. A simple two-level mixing calculation shows that 25% mixing of a \[$\pi d_{3/2}^{-1}(\nu fp)_{2^{+}}$\]$_{{1/2}^{+}}$ allows to reproduce the observed $1/2^{+}$ moment [@NeyenProc].
By establishing the ground-state spins of $^{49}$K and $^{51}$K, we have demonstrated that the gradual reduction of the energy gap between the proton $\pi 2s_{1/2}- \pi 1d_{3/2}$ orbits reaches a minimum around $N=29$ and again increases towards the more neutron rich isotopes. It is the first time that such a ’re-inversion’ of single-particle levels is observed and it illustrates how the residual monopole interaction dominates their evolution.
In Fig. \[energies\] we compare the experimental $3/2^{+}$ and $1/2^{+}$ levels to those calculated with the different shell-model effective interactions. The SDPF-NR and SDPF-U interactions show the best overall agreement, which is not surprising because their monopole matrix elements were tuned by fitting to experimental spectra, including that of $^{47}$K [@Nowacki2009]. With the recently developed SDPF-MU interaction [@UtsunoArXiv] a reasonable agreement with the data is found, considering that its cross-shell interaction is described in a functional form using the simple tensor-subtracted monopole evolution as described in [@Otsuka2010], with only six parameters.
{width="100.00000%"}
In conclusion, the hyperfine structures of atomic $^{49,51}$K isotopes were measured for the first time. The data establish a ground-state spin $I=1/2$ for $^{49}$K and $I=3/2$ for $^{51}$K. The magnetic moments $\mu(^{49}\text{K})= +1.3386(8)[40]\, \mu_N$ and $\mu(^{51}\text{K})= +0.5129(22)[15]\, \mu_N$ reveal a mixed configuration for $^{49}$K and a rather pure $\pi 1d_{3/2}^{-1}$ configuration for $^{51}$K. Comparison with shell-model calculations shows good agreement for $^{51}$K, but none of the interactions reproduces the low experimental value of $^{49}$K. Best overall agreement with the ground-state moment and energy levels in $^{49}$K is observed for the SDPF-NR interaction, which predicts the highest mixing with $\pi 1d_{3/2}$ components in its wave function. The experimentally observed evolution of the $1/2^{+}$ and $3/2^{+}$ levels is now established up to $^{51}$K. Different effective interactions predict very different energy gaps between the 3/2$^{+}$ and the first exited 1/2$^+$ level in $^{51}$K. Along with the current results, spectroscopy of the excited states in $^{51}$K is required to further improve the effective interactions in this region.
This work was supported by the IAP-project P6/23, the FWO-Vlaanderen, NSF grant PHY-1068217, BMBF (05 P12 RDCIC), Max-Planck Society and EU FP7 via ENSAR (no. 262010). We would like to thank to the ISOLDE technical group for their support and assistance.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Hydrodynamical modeling of heavy ion collisions at RHIC suggests that the quark-gluon plasma (QGP) “thermalizes" in a remarkably short time scale, about 0.6 fm/c. We argue that this should be viewed as indicating fast isotropization, but not necessarily complete thermalization, of the non-equilibrium QGP. Non-Abelian plasma instabilities can drive local isotropization of an anisotropic QGP on a time scale which is faster than ordinary perturbative scattering processes. As a result, we argue that theoretical expectations based on weak coupling analysis are not necessarily in conflict with hydrodynamic modeling of the early part of RHIC collisions, provided one recognizes the key role of non-Abelian plasma instabilities.'
author:
- Peter Arnold and Jonathan Lenaghan
- 'Guy D. Moore'
- 'Laurence G. Yaffe'
title: 'Apparent Thermalization due to Plasma Instabilities in Quark-Gluon Plasma'
---
Hydrodynamic models of RHIC collisions (based on near-ideal fluids) provide a good description of a wide range of experimental data, including radial and elliptic flow measurements, provided one assumes that the initial partons thermalize in about 0.6 fm/c [@RHIC-hydro]. However, theoretical estimates based on perturbative scattering processes yield expected thermalization times in the range of 2.5 fm/c or above [@thermalization; @times]. What is the significance of this discrepancy? Are weak-coupling analyses, which should be valid for asymptotically high energy densities (and asymptotically large nuclei), inapplicable at RHIC energies? Perhaps so. Or have dynamical processes which may be responsible for this fast apparent thermalization not been correctly identified? We will argue this is the case [@other; @efforts]. Estimates based on perturbative scattering neglect essential dynamics: the collective behavior associated with non-Abelian plasma instabilities. Such instabilities can produce large non-perturbative effects, including apparent thermalization. We discuss two qualitative lessons which emerge from a weak-coupling analysis: ([*i*]{}) hydrodynamic behavior does [*not*]{} require full thermalization — isotropization of parton momenta in local fluid rest frames suffices; and ([*ii*]{}) plasma instabilities can drive isotropization at rates which are parametrically faster than perturbative scattering rates.
Apparent Thermalization
=======================
The thermalization time scale in a quark-gluon plasma, defined as the inverse relaxation rate of arbitrarily small departures from equilibrium, depends on the rate of large-angle scattering (and near-collinear splitting/joining) processes among quarks and gluons [@Arnold:2002zm]. Parametrically, this time scale is [@foot:transport] $\sim 26 \, [g^4 T \ln (2.4/g)]^{-1}$, and for plausible values of RHIC parameters it is hard to reconcile this time scale with the fast apparent thermalization observed in RHIC collisions. However, this time scale characterizing relaxation of asymptotically small perturbations is [*irrelevant*]{} to the question of when hydrodynamic models can be a good approximation to the dynamics of a non-equilibrium quark-gluon plasma. The essential assumption of ideal fluid hydrodynamic models is that the stress tensor, in the local rest frame at some point in the system, is nearly diagonal,$$T_{ij} \approx p \> \delta_{ij} \,,
\label {eq:stress}$$ with some equation of state relating the pressure $p$ to the energy density. But relation (\[eq:stress\]) is just a statement of [*isotropy*]{} (in the local fluid rest frame), and is automatically true if typical excitations have random directions — even if their energy distribution is far from thermal, or if the pressure $p$ differs from the equilibrium pressure for a given energy density. Consequently, understanding when a hydrodynamic model can first provide a good approximation to the plasma dynamics is the same question as understanding what dynamics drives isotropization.
Plasma Instabilities
====================
To begin, we summarize known results concerning gauge field instabilities in anisotropic non-Abelian plasmas. Further details may be found in Refs. [@Arnold:2002zm; @Arnold:2003rq; @instabilities; @instabilities2].
Let $\phard$ denote the characteristic momenta of typical excitations in a non-equilibrium quark-gluon plasma. (For example, in the saturation scenario [@saturation], $\phard$ equals the saturation scale $\Qs$ at time $\Qs^{-1}$.) We assume that $\phard$ is sufficiently large that these excitations act like highly relativistic particles. For time scales short compared to the mean free time between large-angle scatterings of typical excitations (and large compared to $\phard^{-1}$), the natural framework for describing the dynamics is collisionless kinetic theory. One splits the degrees of freedom into short wavelength (or “hard” momentum) excitations which may be characterized by a phase space distribution function $f(\p,\x,t)$, and long wavelength (or “soft”) gauge field modes which may be regarded as forming a classical field. For a non-Abelian theory, the resulting Boltzmann-Vlasov equation has the form [@Mrowczynski:1989np; @foot:vlasov] $$(D_t + \v \cdot D_\x) \, f
+
\half g \, \{ (\E + \v \times \B)_i , \, \nabla_{p_i} f \}
=
0 \,.
\label {eq:BV}$$ The corresponding Maxwell equations are $$(D_\nu \, F^{\mu\nu})_a = j^\mu_a
\equiv
g \, {\textstyle \int_\p} \> v^\mu \> \tr (t_a \, f) \,,
\label {eq:Max}$$ with $\int_\p \equiv \int {d^3\p \over (2\pi)^3}$, $v^\mu \equiv (1,\hat \p)$, and $t_a$ a color generator.
Any distribution which is homogeneous (in space) and colorless, combined with vanishing soft gauge field, gives a static solution to Eqs. (\[eq:BV\],\[eq:Max\]). Perturbations about such solutions obey a linearized equation of motion \[obtained by linearizing Eq. (\[eq:BV\]) in deviations from the static solution, solving for $\delta f$, and plugging the result into Eq. (\[eq:Max\])\] which (after a space-time Fourier transform) has the form $$\left\{
K^2 \, g^{\mu\nu} - K^\mu K^\nu + \Pi^{\mu\nu}(K)
\right\}
A_\nu(K) = 0 \,,
\label {eq:linear}$$ where the wavevector $K^\mu \equiv (\omega,\k)$ [@foot:metric]. The retarded gauge-field self-energy, generated by hard excitations, is $$\Pi^{\mu\nu}(K)
=
g^2 \int_\p {\partial f(\p) \over \partial p^l}
\left[
-v^\mu g^{l\nu} + {v^\mu v^\nu K^l \over v \cdot K - i\epsilon}
\right] .
\label {eq:Pi}$$ The zero-frequency spatial self-energy $\Pi_{ij}(0,\hat\k)$ depends on the direction but not the magnitude of the spatial wavevector $\k$. If $f(\p)$ is anisotropic but parity invariant, then the self-energy matrix $\Pi(0,\hat\k)$ has a negative eigenvalue for some directions of $\hat\k$. This implies that there are unstable solutions to the small fluctuation equation (\[eq:linear\]), [*i.e.*]{}, solutions for which $\omega$ has a positive imaginary part [@Arnold:2002zm; @Arnold:2003rq]. These are non-Abelian versions of Weibel instabilities in ordinary plasma physics [@Weibel].
Let $-\mu^2$ denote the most negative eigenvalue of $\Pi(0,\hat\k)$ (for any $\hat\k$). Unstable modes have $|\k| < \mu$. Let $\gamma$ denote the maximal growth rate of unstable modes. If the hard particle distribution has $O(1)$ anisotropy [@foot:anisotropy] then the maximum unstable wavevector $\mu$ and the maximum growth rate $\gamma$ are both comparable to the effective mass $m_\infty$ of hard gluons, $$%\textstyle
\mu^2 \sim \gamma^2 \sim m_\infty^2 = g^2 \int_\p \; {f(\p) \over |\p|} \,.
\label {eq:musq}$$ If $\phard$ is the momentum scale which dominates the integral (\[eq:musq\]), and $n \equiv \int_\p f(\p)$ is the spatial density of hard excitations, then $m_\infty \sim g \, \sqrt {n/\phard}$.
To compare to perturbative scattering rates consider, for example, a system with $n=O(\phard^3)$ — the same parametric relation as in equilibrium, where $p \sim T$ and $n=O(T^3)$. In this case $m_\infty$, and hence the instability growth rate $\gamma$ for $O(1)$ (or larger) anisotropy, is $O(g\,\phard)$. This rate is parametrically faster than the $O(g^4 \, \phard)$ rates for large-angle scattering or near-collinear splitting, or even the $O(g^2 \, \phard)$ rate of small-angle scattering [@Arnold:2002zm]. More generally, for $O(1)$ anisotropy $\gamma$ is faster than the large-angle scattering rate whenever $n \ll \phard^3/g^2$ [@foot:compare]. This inequality is satisfied parametrically unless there is saturation, and even in saturation scenarios, it is satisfied for $t \gg \Qs^{-1}$ [@saturation].
Numerical values depend, of course, on the specific form of the anisotropic phase space distribution. A simple example [@foot:example] involving a typical particle energy of 1 GeV, plasma energy density of 27 GeV/fm${}^3$, a phase space distribution proportional to $(\p\cdot\hat {\bf z})^4$, and $\alpha_s=0.5$ yields $m_\infty \simeq 740$ MeV, and $\gamma \simeq 280 \; \rm MeV = (0.7\; fm/c)^{-1}$ for $k \simeq 575$ MeV. With more extreme anisotropy, the growth rate $\gamma$ can approach $m_\infty$ itself [@Arnold:2003rq]. Yet other angular distributions can give slower growth rates.
Instabilities will grow exponentially until some dynamics comes into play which causes the amplitudes of unstable modes to saturate. There are two natural possibilities for when this might happen [@foot:expansion]. If the unstable modes with wavenumbers of order $\mu$ grow until the soft gauge field has an $O(\mu/g)$ amplitude \[or the field strength is $O(\mu^2/g)$\], then non-Abelian corrections to the linearized equation of motion (\[eq:linear\]) will become important and could substantially affect the further evolution [@foot:gauge]. In particular, one might expect these non-linearities to lead to efficient transfer of energy from the unstable modes to stable modes (with comparable wavenumber).
Alternatively, if instabilities do not saturate at $O(\mu/g)$ amplitudes, then they may continue growing until their amplitudes reach the scale $\phard/g$ \[and field strengths are $O(\mu \, \phard/g)$\]. This is the point where the soft gauge field no longer acts as a small perturbation on the motion of hard excitations. To see this, note that for this amplitude, the gauge field part of a covariant derivative is just as large as the ordinary derivative when acting on fluctuations with $O(\phard)$ momenta. This is also the point where the energy density in the soft gauge field becomes an $O(1)$ fraction of the total energy density, $
(F_{\rm soft}^{\mu\nu})^2
\sim (\mu \, \phard/g)^2
\sim n \, \phard
$.
There are reasons to believe the second alternative, not the first, is correct. The generalization to anisotropic plasmas of the HTL (“hard thermal loop”) effective action is [@Mrowczynski:2004kv; @foot:validity] $$\begin{aligned}
\label {eq:S-HTL}
S_{\rm eff}
&=&
-\int d^4x \>
\biggl[
{\textstyle {1\over4}}
F^a_{\mu\nu} F^{a \mu\nu}
\\ && \qquad {}
+
g^2 \int_\p
{f(\p) \over |\p|} \>
F^a_{\alpha\mu}
\left({v^\mu v^\nu \over (v\cdot D)^2}\right)_{ab}
F_\nu^{b\alpha}
\biggr] \,.
\nonumber\end{aligned}$$ Evaluating this, explicitly, for arbitrary static fields in order to examine the corresponding effective potential is not feasible. But in the special case of fields which vary in only one spatial direction, the effective action reduces to a simple local form. Let $\hat\n$ denote the direction of the wavevector of the most unstable mode. For gauge fields which depend only on $\hat\n \cdot \x$, one finds that the effective potential is $$V[\A(\hat\n \cdot \x)] = \int d^3x \>
\left[
{\textstyle {1\over4}} F^a_{ij} F^a_{ij}
+
\half A^a_i \, \Pi_{ij}(0,\hat \n) \, A^a_j
\right] .$$ When $\Pi(0,\hat\n)$ has a negative eigenvalue this potential is unbounded below. The runaway directions of steepest-descent correspond to Abelian field configurations where the commutator terms in the field strength $F_{ij}^a$ vanish. This suggests that non-Abelian non-linearities may [*not*]{} cause growing instabilities to saturate at the scale $\mu/g$, provided the field configuration evolves toward an effectively Abelian form which can continue rolling down the potential energy landscape. This behavior has been seen in time-dependent numerical simulations in 1+1 dimensions — the instability locally “Abelianizes” and continues growing. It is important to perform full 3+1 dimensional simulations of the collisionless kinetic theory (\[eq:BV\],\[eq:Max\]) to verify this conclusion. Such simulations are in progress [@inprogress]. Here, we shall assume that growth of instabilities, beyond the soft scale $\mu/g$, will be confirmed.
Isotropization
==============
Growing instabilities imply that the stress tensor of the non-equilibrium system will receive growing contributions from the soft gauge field. The fastest growing linearized modes tend to decrease the anisotropy in the total stress tensor [@instabilities]. For example, if the anisotropic hard particle distribution has a prolate form, so that $T^{\rm hard}_{zz} \gg T^{\rm hard}_{xx}, \, T^{\rm hard}_{yy}$, then the wavevectors of the fastest growing unstable modes lie in the equatorial plane and the growth of these modes produces a soft gauge field contribution to the stress tensor which is oblate, $T^{\rm soft}_{xx} \sim T^{\rm soft}_{yy} \gg T^{\rm soft}_{zz}$. Conversely, for an oblate hard particle distribution, the fastest growing unstable mode has its wavevector along the normal direction and generates a prolate contribution to the stress. Hence, even in the linearized regime, one can see that soft gauge field instabilities push the system toward greater isotropy. However, the soft contribution to the stress tensor is small compared to the hard particle contribution, and the back-reaction of the soft gauge field on the hard particles is a tiny perturbation, as long as the soft gauge field amplitude is much less than $O(\phard/g)$.
But if the soft gauge field amplitude reaches the scale $\phard/g$ then it no longer acts as a small perturbation to the dynamics of hard excitations. Recall that the radius of curvature of an excitation of momentum $p$ and charge $g$ in a magnetic field $B$ is $R = p/(gB)$. If the radius of curvature is comparable to the magnetic field coherence length $\mu^{-1}$, which means $B \sim \mu p/g$, then excitations of momentum $p$ will undergo $O(1)$ changes in direction during traversals of any single coherence-length sized magnetic field “patch” [@foot:patch].
Therefore, if unstable soft gauge field modes with $O(\mu)$ wavevectors grow until the field strength is $O(\mu \, \phard/g)$, then typical excitations will experience $O(1)$ changes in direction in times of order $\mu^{-1}$. Excitations with differing momenta or colors will receive different deflections from a given patch of (non-Abelian) magnetic field. Excitations traversing different patches of magnetic field (separated by $O(\mu^{-1})$) will receive nearly uncorrelated deflections.
The net effect is that a soft gauge field with a non-perturbative amplitude of order $\phard/g$ can effectively drive isotropization in the distribution of typical hard excitations on a time scale which equals the coherence length $\mu^{-1}$ of the soft gauge field.And isotropization of the hard particle distribution will turn off further growth in the soft gauge field (since gauge field instabilities are absent for isotropic distributions).
As with all instabilities, the time, or number of e-foldings, required for the soft gauge field to become large depends on the size of initial “seed” amplitudes in the relevant unstable modes. The amplitude of the soft ($k \sim \mu$) gauge field generated by a random color charge distribution of the hard particles can be estimated as $A^2 \sim $ $g^2 n/\mu \sim g\sqrt{n \, \phard}$. This is the smallest the seed field could be. For densities from $n=O(\phard^3)$ up to the density limit $n=O(\phard^3/g^2)$ imposed by saturation, $A \gtrsim O(g^{1/2} \phard)$. This is only a factor of $g^{3/2}$ smaller than the nonperturbative $O(\phard/g)$ amplitude. Therefore, the number of e-foldings required for instabilities to grow to this non-perturbative size is only of order $\ln (1/g)$.
Treating logs of $g$, for simplicity, as $O(1)$, this means that if the initial anisotropy is $O(1)$ then the characteristic growth time needed for unstable modes of the soft gauge field to reach the non-perturbative amplitude $\phard/g$ is only of order $\gamma^{-1}$. The resulting soft gauge field then drives isotropization of the hard particle distribution on a comparable $\gamma^{-1}$ time scale. Therefore, (up to logs of $g$ and factors of order one), the time scale for isotropization of the hard particle distribution is the [*same*]{} as the (inverse) instability growth rate $\gamma^{-1}$ [@foot:violent].
In numerical simulations of ordinary non-relativistic plasmas, essentially the same process of instability-driven isotropization has been observed [@Califano], with the growth of magnetic instabilities driving large reductions in anisotropy once the magnetic fields reach critical strength. (These simulations allowed three-dimensional momentum space variations, but assumed translation invariance in one spatial direction.) Various QGP numerical simulations [@QGP; @numerics] have failed to see any sign of this instability-driven dynamics because they did not allow full three-dimensional variations.
Although we have focused on the ability of non-perturbative soft gauge fields to generate large changes in directions of hard excitations, it should be noted that $\mu^{-1}$ is also the characteristic time scale for $O(1)$ changes in energies of hard excitations. This is inevitable, given the fact, noted earlier, that when the soft gauge field reaches the non-perturbative amplitude $\phard/g$ its energy density is comparable to the energy density in the hard excitations. But it may also be seen directly by noting that chromoelectric fields generated during the growth of instabilities will be comparable in size to chromomagnetic fields (since the growth rate of unstable modes is comparable to their wavenumbers for $O(1)$ anisotropy). So chromoelectric fields will reach the same $O(\mu \, \phard/g)$ size as magnetic fields — which means that an excitation traveling a distance $\mu^{-1}$ will have work of order $\phard$ done on it by the soft gauge field. Of course, this time scale for $O(1)$ changes in energy may be very different (and much shorter) than the time scale for true thermalization, as defined by a near-thermal energy distribution of excitations over a parametrically large dynamic range.
Conclusions
===========
We have argued that “early thermalization” in heavy-ion collisions is more properly interpreted as evidence of fast isotropization in the distribution of excitations. And we have argued that non-Abelian plasma instabilities can drive isotropization at a rate which is parametrically fast compared to perturbative scattering rates. Consequently, we see no reason to view the fast onset of hydrodynamic behavior in RHIC collisions as necessarily in conflict with theoretical expectations based on weak-coupling analysis of a quark-gluon plasma, provided one properly accounts for the effects of non-perturbative plasma instabilities. Further study of the scenario we have sketched is certainly needed; in particular full three-dimensional non-Abelian Boltzmann-Vlasov simulations with appropriate initial conditions should be conducted.
See, for example, U. W. Heinz, nucl-th/0407067. R. Baier, A. H. Mueller, D. Schiff and D. T. Son, Phys. Lett. B [**539**]{}, 46 (2002); D. Molnar and M. Gyulassy, Nucl. Phys. A [**697**]{}, 495 (2002) \[Erratum-ibid. A [**703**]{}, 893 (2002)\]. For other efforts to address this puzzle, see E. V. Shuryak and I. Zahed, hep-ph/0307267; E. Shuryak, Prog. Part. Nucl. Phys. [**53**]{}, 273 (2004); Z. Xu and C. Greiner, hep-ph/0406278; F. Gastineau, E. Blanquier and J. Aichelin, hep-ph/0404207. P. Arnold, G. D. Moore and L. G. Yaffe, JHEP [**0301**]{}, 030 (2003). This is the collision rate, in the relaxation-time approximation, which reproduces the leading-order shear viscosity in $N_{\rm f}{=}3$ QCD. For the latter, see P. Arnold, G. D. Moore and L. G. Yaffe, JHEP [**0305**]{}, 051 (2003). P. Arnold, J. Lenaghan and G. D. Moore, JHEP [**0308**]{}, 002 (2003). S. Mrówczyński, Phys. Lett. B [**314**]{}, 118 (1993); Phys. Rev. C [**49**]{}, 2191 (1994); Phys. Lett. B [**393**]{}, 26 (1997). P. Romatschke and M. Strickland, Phys. Rev. D [**68**]{}, 036004 (2003). See, for example, the discussion in the context of bottom-up thermalization in R. Baier, A. H. Mueller, D. Schiff and D. T. Son, Phys. Lett. B [**502**]{}, 51 (2001). S. Mrówczyński, Phys. Rev. D [**39**]{}, 1940 (1989).
We suppress species labels (quark [*vs.*]{} gluon) on the distribution $f$, as well as associated species sums and color Casimirs which should appear in Eqs. (\[eq:Max\]) and (\[eq:Pi\]–\[eq:S-HTL\]). See Ref. [@Arnold:2002zm] for details. For excitations in color representation $\cal R$, the distribution $f$ is a color density matrix transforming in the ${\cal R} \times \bar{\cal R}$ representation; the field strengths in Eq. (\[eq:BV\]) are matrices in the representation $\cal R$. We use a $({-}{+}{+}{+})$ metric convention, so $K^2 \equiv -\omega^2 + \k^2$.
E. S. Weibel, Phys. Rev. Lett. [**2**]{}, 83 (1959).
By $O(1)$ anisotropy, we mean generic distributions whose first and second angular derivatives, when integrated wrt. $d(\ln p)$, are $O(1)$ relative to $f$. In contrast, for nearly isotropic distributions, both $\mu$ and $\gamma$ are small compared to $m_\infty$. Also in contrast, the examples of extreme anisotropy studied in Ref. [@Arnold:2003rq] (nearly planar or nearly linear distributions) give $\mu \gg m_\infty$ while $\gamma$ remains $\sim m_\infty$. The large-angle scattering rate is $n\sigma (1{+}f)$, where $\sigma \sim g^4/\phard^2$, and $1{+}f$ is a Bose enhancement factor, with $f \sim n/\phard^3$ for $O(1)$ anisotropy. Comparison to $m_\infty$ gives the stated inequality.
Here, $f(\p)$ is taken as thermal times $5 (\hat\p\cdot\hat{\bf z})^4$.
Expansion of the plasma may also cut off instability growth, if the expansion rate is faster than the rate of the processes under discussion. See J. Randrup and S. Mrówczyński, Phys. Rev. C [**68**]{}, 034909 (2003). Estimates of gauge field amplitudes should be understood as applying in gauges, such as Coulomb gauge, which make the gauge field as smooth as possible.
S. Mrówczyński, A. Rebhan and M. Strickland, Phys. Rev. D [**70**]{}, 025004 (2004). The derivation of the effective action (\[eq:S-HTL\]) assumes that the gauge field is perturbative on the scale of the hard excitations. It ceases to be valid if the gauge field grows to a non-perturbative amplitude of order $\phard/g$.
J. P. Blaizot and E. Iancu, Phys. Lett. B [**326**]{}, 138 (1994). P. Arnold and J. Lenaghan, Phys. Rev. D [**70**]{}, 114007 (2004). A. Rebhan, P. Romatschke and M. Strickland, hep-ph/0412016. P. Arnold, G. D. Moore and L. G. Yaffe, work in progress.
If the coherence length $\mu^{-1}$ were less than the radius of curvature, then a hard excitation crossing any particular coherence-length sized patch of magnetic field will receive a small $O(gB/\mu p)$ deflection. Deflections from different patches will add randomly, so traversal of $O[(\mu p/gB)^2]\gg 1$ patches will be required to receive an $O(1)$ change in direction. This is the usual case where the soft gauge field is perturbatively small on the scale of hard momenta.
This is analogous to “violent relaxation” in gravitational dynamics. See, for example, D. Lynden-Bell, Mon. Not. Roy. Astron. Soc. [**136**]{}, 101 (1967); M. Trenti and G. Bertin, astro-ph/0406236. F. Califano, and N. Attico, F. Pegoraro, G. Bertin and S. V. Bulanov, Phys. Rev. Lett [**86**]{}, 5293 (2001).
A. Krasnitz and R. Venugopalan, Nucl. Phys. B [**557**]{}, 237 (1999); A. Krasnitz, Y. Nara and R. Venugopalan, Nucl. Phys. A [**717**]{}, 268 (2003); [*ibid.*]{} [**727**]{}, 427 (2003); T. Lappi, Phys. Rev. C [**67**]{}, 054903 (2003).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Pollution has become a major threat in almost all metropolitan cities around the world. Currently atmospheric scientists are working on various models that could help us understand air pollution. In this paper we have formulated a new metric tool called Delhi Similarity Index (DSI). The DSI is defined as the geometrical mean of the trace gases such as: ozone, sulfur-dioxide and carbon-monoxide, which ranges from 0 (dissimilar to Delhi) to 0.9-1 (similar to Delhi). The limitation of the tool concerning the result of the nitrous-di-oxide data set is also analyzed. Also, the GIS projections of PM 2.5 role for Indian cities are graphically represented. The DSI results from 2011 to 2014 data show that Bengaluru is in the threshold of becoming as polluted like Delhi with values varying from 0.8 to 0.9 (i.e. 80-90%) and Jungfraujoch with a 0.65 to 0.7 (i.e. 65-70%).'
author:
- Purusharth Saxena
- Madhu Kashyap Jagadeesh
title: 'Similarity indexing & GIS analysis of air pollution'
---
Introduction
============
The pollution in metropolitan cities is leading towards numerous modeling and analysis studies. The research about air pollution, and controlling techniques are suggested to avoid the rise of pollution \[[@Ratti] and [@Beelen]\]. But New Delhi the capital city of India is infamous for its pollution, which has risen to such an extent that it is posing a non-negligible health hazard for the permanent residents of the city \[[@Chhabra], [@Cropper], [@Agarwal], [@Chhabra], [@Cropper] & [@Kumar]\]. Despite constant research and concerns to government, the air quality in New Delhi is degrading and has reached a critical point \[[@Kumar] and [@Agarwal]\]. According to the Air Quality Index (AQI), the urban areas are facing major threat, consequently the trend study and analysis of AQI is going on at various places (Ex: [@Lan]). In this research we use the evolved similarity index formula of *Bray-Crutis* of 1957, which is applied in various studies such as ecology ([@Orloci]), marine ([@Schulz]), and astrophysics ([@Schulze-Makuch] and [@Madhu]). A new indexing measure to check the similarity of air pollution in different cities (such as: Bengaluru, New Delhi and Jungfraujoch) with respect to the New Delhi air pollution is introduced, which is called Delhi Similarity Index (DSI). DSI is basically defined as the geometrical mean of the trace gases such as: ozone, sulfur-dioxide and carbon-monoxide, which ranges from 0 (dissimilar to Delhi) to 0.9-1 (similar to Delhi). For past several years, Delhi’s PM 2.5 (Particulate Matter $\leq 2.5$ microns) level has crossed the threshold of WHO standards atleast a few times in an year, as studied by [@Tiw]. Various studies have been performed over the atmospheric and surface pollution of Delhi: [@Ghude-Jain], [@Srivastava]. Geographic Information System (GIS) has also been employed to study the spread and compare the pollution over time within a city. [@Matejicek], [@Jerrett]. Hence in this paper we have dedicated a section to understand PM 2.5 trends in New Delhi and Bengaluru along DSI analysis.\
The paper structure is as follows: Section 2 contains the formulation of DSI, Section 3 is about data collection along with the geo-political description of stations, section 4 deals with the Results of DSI, section 5 has the PM 2.5 analysis of Bengaluru and New Delhi, and section 6 is concluded by the discussion and conclusion.
Formulation of Delhi Similarity Index (DSI)
===========================================
The Bary-Curtis’ similarity index from 1957 is used in this formulation with the indices division of 0.2 intervals [@Bloom], the classifications are as follows: very low, low, moderate, high and very high similarity regions. Here, we define the threshold to very high similarity region with the threshold of $V=0.8$. Defining the physical limits $x_a$ and $x_b$ of the permissible variation of a variable with respect to $x_0$ (i.e. $x_a<x_0<x_b$), we calculate the weight exponents for the lower $w_a$ and upper $w_b$ limits, $$w_a = \frac
{\ln{V}}
{\ln\left[1-\left|\frac{x_0-x_a}
{x_0+x_a}\right|\right]}
\,,\quad
w_b = \frac
{\ln{V}}
{\ln\left[1-\left|\frac{x_b-x_0}{x_b+x_0}\right|\right]}
\,,\quad
\label{eq:weight_exponent}$$ The average weight is found by taking the geometric mean of $w_a$ and $w_b$, $$w_x=\sqrt{{w_a}\times {w_b}}\,.
\label{eq:geom_mean}$$
The Delhi Similarity Index is defined, as (for the entire abstract derivation refer [@Madhu]\], $$DSI_x = {\left[1-\Big|
\frac{x-x_0}{x+x_0}\Big| \right]^{w_x}}\,,
\label{eq:esi}$$ where $x$ is the concentration of the trace gases under observation, and $x_0$ is the standard for the respective trace gases as advised by Central Pollution Control Board (CPCB), India .\
The yearly Delhi Similarity Index is the geometric mean of the DSI value of the trace gases (\[eq:esi\]). $$DSI ={[{DSI_{O_3}} \times {DSI_{SO_2}} \times {DSI_{CO}}]^{1/3}}$$
Where, $DSI_{O_3}$, $DSI_{SO_2}$, and $DSI_{CO}$ are the Delhi Similarity Index values of ozone, sulfur-dioxide and carbon-monoxide, which ranges from 0 (dissimilar to Delhi) to 0.9-1 (similar to Delhi). \[sec:headings\]
Data Collection and Site Description
====================================
New Delhi
---------
The average altitude of Delhi is about 216m and the data for New Delhi is obtained from Central Pollution Control Board (CPCB) [^1], and this AQI data station is located at Shadipur ($28\degree 39'05.8"N, 77\degree 09'29.5"E$) India. New Delhi is one of the most populated metropolitan region in India. According to the 2011 census done by the Indian Government, the population density of Delhi is about 11,000 people per square kilometer, The last census was taken 8 years ago, and given the exponential rise in population being observed in Delhi, the United Nations estimates that Delhi will become the most populated city by 2028. The main reasons are the number of vehicles in Delhi has grown to a whooping 10 million (Gude et al. 2006). The geopolitical disadvantage of New Delhi is; it is surrounded by industrial & agricultural cities, which annually experiences a massive intake of trace gases as a direct effect of crop burning in neighbouring states.
Bengaluru
---------
The altitude of Bengaluru is about 920 m and the data for Bengaluru is also obtained from CPCB website, the data station is located at BTM layout ($12\degree 54'47.5"N, 77\degree 36'33.3"E$) which is a residential area in Bengaluru. Unlike New Delhi, Bengaluru experiences a mild temperature throughout the year with almost no extreme variations; the annual mean temperature is around 25$^\circ$ C, April and May being the hottest. Bengaluru is the target center of our study, which acts as the neutral point between New Delhi (major air pollution) and Jungfraujoch (minor air pollution).
Jungfraujoch
------------
Jungfraujoch is a station in the European Alps of Switzerland, it is located almost at the center of Europe, at an altitude of around 3460 m. It is a significant contributor to atmospheric data due to lack of pollutants at a good elevation and the nearby industrial areas located at lower altitudes. Data for Jungfraujoch was collected from the World Data Center for Greenhouse Gases \[WDCGG\] [^2] from the station monitored by Swiss Federal Laboratories for Materials Science and Technology (empa). Daily data collected, for all the three cities, from 2011 to 2014, the annual average is taken separately for all the trace gases and tabulated (Table: 2).
Results
=======
The calculated weight exponent parameters for New Delhi data are tabulated in Table \[table:dsi\_gas\]. The weight exponents calculations are done using Eq \[eq:weight\_exponent\], and the lower and upper limit of the trace gas data set is selected (here the minimum and maximum data recorded from 2011 - 2014).
[l\*[4]{}[c]{}r]{} Location & Gas & $x_0$ ($\mu gm^-3$) & Weight Exponents\
New Delhi & $O_3$ & 100 & 0.12\
New Delhi & $SO_2$ & 50 & 0.10\
New Delhi & $CO$ & 2000 & 0.081\
Based on the exponents in Table \[table:1\], we calculate the Delhi similarity index for the the three trace gases using Equation \[eq:esi\]
------ ---------- ------------ ----------- ---------- ------------ ----------- ---------- ------------ -----------
**$CO$** **$SO_2$** **$O_3$** **$CO$** **$SO_2$** **$O_3$** **$CO$** **$SO_2$** **$O_3$**
2011 0.983 0.901 0.922 0.885 0.815 0.946 0.085 0.561 0.998
2012 0.985 0.935 0.921 0.826 0.985 0.909 0.063 0.544 0.999
2013 0.961 0.912 0.909 0.893 0.972 0.654 0.065 0.546 0.998
2014 0.867 0.911 0.927 0.828 0.819 0.871 0.098 0.569 0.998
------ ---------- ------------ ----------- ---------- ------------ ----------- ---------- ------------ -----------
: Delhi Similarity Index values of individual trace gases
\[table:dsi\_gas\]
Table \[Tab:overall\_dsi\] shows the global Delhi Similarity Index estimated from the geometrical mean of $CO$, $SO_2$, and $O_3$ for the three cities across 2011-14 data.
[l\*[4]{}[c]{}r]{} Year & Delhi & Bengaluru & Jungfraujoch\
2011 & 0.934 & 0.880 & 0.691\
2012 & 0.947 & 0.904 & 0.689\
2013 & 0.927 & 0.828 & 0.689\
2014 & 0.901 & 0.839 & 0.697\
\[Tab:overall\_dsi\]
As evident from Table \[Tab:overall\_dsi\], the DSI values for Delhi is maximum for all the years, whereas it is least for Jungfraujoch. The variation in Jungfraujoch is because of high $O_3$ content observed can be attributed to its high altitude. The increase in the DSI value for Bengaluru is because of the increase in $SO_2$ concentration in 2013-14.\
The concentration variations of trace gases from 2011 to 2014 data is graphically represented in figure 1. The large variations of data is observed in $CO$ and $O_3$, than $SO_2$. This could be because of large quantity of carbon emissions from vehicles, which in turn deplete the ozone layer.
![Concentration of trace gases $CO$, $SO_2$ and $O_3$ []{data-label="fig:1"}](Figure1){width="13cm"}
The DSI results of Table\[Tab:overall\_dsi\] is graphically obtained as figure 2.
![DSI values for Delhi, Bengaluru, and Jungfraujoch (2011-14)[]{data-label="fig:2"}](Figure2){width="10cm"}
From figure 2 an increase is seen in Jungfraujoch and Bengaluru DSI values for the year 2013-14, while a downward trend is observed in Delhi DSI value. Nevertheless, the index value of Delhi remains prominent.
$NO_2$ data deviation
---------------------
The boundary conditions of similarity index observed as a limit, that the data obtained from the CPCB database doesn’t work on specific forms of data sets. The data set of $NO_2$ falls under the following condition: $x_{max} > x_{0}$ Where $x_{max}$ is the maximum value of x, and $x_{0}$ is the reference value. The weight exponent calculated for this form of condition does not yield any result. Hence $NO_2$ data is not applied in the similarity index tool, but the tool is valid for the above-mentioned trace gases except $NO_2$.\
The Daily $NO_2$ concentration for Delhi is given in Figure \[fig:3\], a pattern is detected throughout 2011-17, with a high concentration of $NO_2$ occurring during November each year, which in turn can be linked to the burning of crops in the nearby cities.
![Daily concentration of $NO_2$ from 2011-2017.[]{data-label="fig:3"}](Figure3){width="10cm"}
The 2015 data corresponds to extremely high $NO_2$ values observed in first week of November in Delhi. Subsequently, the rest of the years follow the same pattern in November.
PM 2.5 GIS analysis for Bengaluru and New Delhi
===============================================
Suspended particle matter in the air of size 2.5 is the basic full form of PM 2.5. The PM 2.5 data of the year 2017 from CPCB data is plotted using a Geographical Information system (GIS). The sites used for the analysis are Bengaluru and New Delhi, as data for Jungfraujoch, is not available. However, the ozone depletion is observed to be more significant in Jungfraujoch, than the PM 2.5 pollutant. GIS plots are used to study PM 2.5 effects and proliferation (as done in [@Jerrett], [@Matejicek]).
Bengaluru
---------
![Daily PM 2.5 Value for three CPCB stations in Bengaluru[]{data-label="fig:cbcb_stations_bangalore"}](Figure4)
Peenya, being an industrial district, harbours higher concentration of PM 2.5 when compared to residential areas (BTM, BWSSB). An increase in the relative content of PM 2.5 content in each district is observed in the month of April-May (Peak Summer months in Bengaluru). While a dip is observed in the month of June-July (onset of monsoon in Bengaluru).
![GIS interpolation of Bengaluru’s annual PM 2.5 concentration from three CPCB stations across the cities (left). Relative scatter plot of Bengaluru for the same stations (right).[]{data-label="fig:5"}](Figure5){width="13cm"}
The GIS plot (interpolation and scatter) for PM 2.5 content in Bengaluru, agrees with the above mentioned trends, like Peenya has higher concentration. The 3 stations are marked in green dots and the concentration intensity scale is shown in the right side of the GIS interpolation plot.
New Delhi
---------
For the New Delhi PM 2.5 sample of 2017, the available data is taken from nine CPCB stations, (namely: ITO, Mandir Marg, Punjabi Marg, Dwarka, DTU, IHBAS, RK Puram, Shadipur, Sirifort) and the corresponding results are obtained as shown below.
![Daily PM 2.5 concentration plot for New Delhi, 2017 for nine CPCB stations[]{data-label="fig:delhi_cpcb_stations"}](Figure6){width="8cm"}
The maximum PM 2.5 content is observed in the first week of November, which can be attributed to the burning of the crops in the neighbouring cities. The same pattern is observed in all the stations throughout the year. The PM 2.5 concentration is so high, that temperature does not seem to be the major factor in militating or mitigating the PM 2.5 content, as such seasonal variation was observed in Bengaluru (Figure \[fig:cbcb\_stations\_bangalore\]).
The GIS plot (interpolation and scatter plot) for New Delhi is given below in figure \[delhi\_interpolation\].
![GIS interpolation of New Delhi’s annual PM 2.5 concentration from around 17 CPCB stations across the city (left). Relative Scatter Plot of annual PM 2.5 average (right).[]{data-label="delhi_interpolation"}](Figure9){width="13cm"}
The green dots in figure 7 indication CPCB stations and the variations of PM 2.5 intensity is high in almost all chosen areas. The Bhuvan satellite vegetation images gives the size of built-up area (in red) of New Delhi and Bengaluru.
![New Delhi and Bengaluru Vegetation with Bhuvan satellite view[]{data-label="fig:bhuvan_satellite"}](blr_del_veg){width="15cm"}
Even though the built up area of Bengaluru seems to be larger, the air pollution constraints seems to be follow the similar trend. The green and the blue, indicating the need of clean air and water has to be protected and improved.
Discussion and Conclusion
=========================
As shown in Table \[table:dsi\_gas\], different cities have different pollution levels for different gases, which is a result of human intervention, geographic position - altitude, and proximity from the coastline of the city. Therefore, a direct comparison of the concentration of trace gases is an inefficient measure of the level of pollution in a city. In other words, DSI is an index that can be used to compare the atmospheric deterioration (due to trace/greenhouse gases) for two or more cities. A succinct mathematical model, called Delhi Similarity Index (DSI), gives a reference value for comparison, interpolation and extrapolation. The cities that are chosen for the model are New Delhi, Bengaluru and Jungfraujoch. Wherein New Delhi is in severe pollution state, Jungfraujoch is in an ideal location in the Bernese Alps, and Bengaluru is taken as a neutral point between New Delhi and Jungfraujoch. Figure \[fig:1\] and Figure \[fig:2\] shows the variation of Concentration of the respective gases, and the overall DSI comparison of the three cities, respectively (from 2011-2014). As seen from Figure \[fig:2\], the year 2012 was worst, in terms of pollution, for the Indian cities of Bengaluru and New Delhi; after which both the cities started improving. The summary for all the 3 cities as given in the Table \[Tab:overall\_dsi\]. In turn implies that in 2011, the trace gases concentration Jungfraujoch was 69.1% of New Delhi’s on the same year, while that of Bengaluru’s was 88% of New Delhi’s in the same year. Same can be stated about 2012, where a spike in New Delhi and Bengaluru DSI values are observed. similarly, the maximum value for Jungfraujoch was observed in 2014 (as seen in Figure \[fig:2\]). Although Jungfraujoch’s Carbon mono-oxide and Sulfur Dioxide content is of significantly lower amount when compared to New Delhi and Bengaluru, but the indexing gives 69% similarity to New Delhi because of soaring surface ozone concentration in Jungfraujoch. The variations the level of PM 2.5 due to seasonal changes (in Bengaluru) is an intriguing anomaly which requires further investigation, moreover New Delhi does not show such variation. Future work can be done with many more cities across a wide spectrum (coastal area, islands, tundra & equatorial region, etc), with more trace / greenhouse gases data, with similarity index technique. A comprehensive comparison can then be made for the chosen cities.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank CPCB for there online database for data on Indian cities and the World Data Centre for Greenhouse Gases (WDCGG) database for Switzerland data. Also we would like to thank NRSC of ISRO for Bhuvan satellite data on vegetation.
[17]{} natexlab\#1[\#1]{}url \#1[`#1`]{}urlprefix
Agarwal, K., Mughal, M., Upadhyay, P., Berry, J., Mawer, E. and Puliyel, J. (2002) The impact of atmospheric pollution on vitamin d status of infants and toddlers in delhi, india. *Archives of disease in childhood*, **87**, 111–113.
Beelen, R., Hoek, G., Vienneau, D., Eeftens, M., Dimakopoulou, K., Pedeli, X., Tsai, M.-Y., K[ü]{}nzli, N., Schikowski, T., Marcon, A. et al. (2013) Development of no2 and nox land use regression models for estimating air pollution exposure in 36 study areas in europe–the escape project. *Atmospheric Environment*, **72**, 10–23.
Bloom, S. A. (1981) Similarity indices in community studies: potential pitfalls. *Mar. Ecol. Prog. Ser*, **5**, 125–128.
Chhabra, S. K., Chhabra, P., Rajpal, S. and Gupta, R. K. (2001) Ambient air pollution and chronic respiratory morbidity in delhi. *Archives of Environmental Health: An International Journal*, **56**, 58–64.
Cropper, M., Simon, N. B., Alberini, A. and Sharma, P. (1997) The health effects of air pollution in delhi, india. *World Bank Policy Research*.
Ghude, S. D., Jain, S., Arya, B., Beig, G., Ahammed, Y., Kumar, A. and Tyagi, B. (2008) Ozone in ambient air at a tropical megacity, delhi: characteristics, trends and cumulative ozone exposure indices. *Journal of Atmospheric Chemistry*, **60**, 237–252.
Jerrett, M., Burnett, R. T., Kanaroglou, P., Eyles, J., Finkelstein, N., Giovis, C. and Brook, J. R. (2001) A gis–environmental justice analysis of particulate air pollution in hamilton, canada. *Environment and Planning A*, **33**, 955–973.
Kashyap, J. M., Gudennavar, S. B., Doshi, U. and Safonova, M. (2017) Indexing of exoplanets in search for potential habitability: application to mars-like worlds. *Astrophysics and Space Science*, **362**, 146.
Kumar, R., Nagar, J. K., Kumar, H., Kushwah, A. S., Meena, M., Kumar, P., Raj, N., Singhal, M. and Gaur, S. (2008) Indoor air pollution and respiratory function of children in ashok vihar, delhi: an exposure-response study. *Asia Pacific Journal of Public Health*, **20**, 36–48.
Lanzafame, R., Monforte, P., Patan[è]{}, G. and Strano, S. (2015) Trend analysis of air quality index in catania from 2010 to 2014. *Energy Procedia*, **82**, 708–715.
Mat[ě]{}j[í]{}[č]{}ek, L., Engst, P. and Ja[ň]{}our, Z. (2006) A gis-based approach to spatio-temporal analysis of environmental pollution in urban areas: A case study of prague’s environment extended by lidar data. *Ecological Modelling*, **199**, 261–277.
Onofri, S., de Vera, J.-P., Zucconi, L., Selbmann, L., Scalzi, G., Venkateswaran, K. J., Rabbow, E., de la Torre, R. and Horneck, G. (2015) Survival of antarctic cryptoendolithic fungi in simulated martian conditions on board the international space station. *Astrobiology*, **15**, 1052–1059.
Ratti, C., Di Sabatino, S., Britter, R., Brown, M., Caton, F. and Burian, S. (2002) Analysis of 3-d urban databases with respect to pollution dispersion for a number of european and american cities. *Water, Air and Soil Pollution: Focus*, **2**, 459–469.
Schulz, J. (2016) Bray-curtis dissimilarity. algorithms—similarity. alfred-wegener-institute for polar and marine research, bremerhaven, germany.
Schulze-Makuch, D., M[é]{}ndez, A., Fair[é]{}n, A. G., Von Paris, P., Turse, C., Boyer, G., Davila, A. F., Ant[ó]{}nio, M. R. d. S., Catling, D. and Irwin, L. N. (2011) A two-tiered approach to assessing the habitability of exoplanets. *Astrobiology*, **11**, 1041–1052.
Srivastava, A., Gupta, S. and Jain, V. (2008) Source apportionment of total suspended particulate matter in coarse and fine size ranges over delhi. *Aerosol Air Qual. Res*, **8**, 188–200.
Tiwari, S., Srivastava, A., Bisht, D., Parmita, P., Srivastava, M. K. and Attri, S. (2013) Diurnal and seasonal variations of black carbon and pm2. 5 over new delhi, india: influence of meteorology. *Atmospheric Research*, **125**, 50–62.
[^1]: (<http://www.cpcb.gov.in/CAAQM/frmUserAvgReportCriteria.aspx>)
[^2]: (<https://ds.data.jma.go.jp/>)
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We build on :[^1] a method for model and proposal learning based on maximizing the lower bound to the log marginal likelihood in a broad family of structured probabilistic models. Our approach relies on the efficiency of for performing inference in structured probabilistic models and the flexibility of deep neural networks to model complex conditional probability distributions. We develop additional theoretical insights and experiment with a new training procedure which can improve both model and proposal learning. We demonstrate that our approach provides a fast, easy-to-implement and scalable means for simultaneous model learning and proposal adaptation in deep generative models.'
author:
- |
Tuan Anh Le$^\dag$, Maximilian Igl$^\dag$, Tom Rainforth$^\ddag$, Tom Jin$^{\dag,\S}$, Frank Wood$^\dag$\
$^\dag$Department of Engineering Science, University of Oxford\
$^\ddag$Department of Statistics, University of Oxford\
$^\S$Department of Statistics, University of Warwick\
`{tuananh,igl,jin,fwood}@robots.ox.ac.uk, [email protected]`
bibliography:
- 'main.bib'
title: 'Auto-Encoding Sequential Monte Carlo'
---
Introduction
============
We build upon [@le2017auto], a method for model learning that itself builds on [@kingma2014auto; @rezende2014stochastic] and [@burda2016importance]. is similarly based on maximizing a lower bound to the log marginal likelihood, but uses [@doucet2009tutorial] as the underlying marginal likelihood estimator instead of . For a very wide array of models, particularly those with sequential structure, forms a substantially more powerful inference method than , typically returning lower variance estimates for the marginal likelihood. Consequently, by using for its marginal likelihood estimation, often leads to improvements in model learning compared with and . We provide experiments on structured time-series data that show that based learning was able to learn useful representations of the latent space for both reconstruction and prediction more effectively than the counterpart.
was introduced in an earlier preprint [@le2017auto] concurrently with the closely related methods of @maddison2017filtering [@naesseth2017variational]. In this work we take these ideas further by providing new theoretical insights for the resulting , extending these to explore the relative efficiency of different approaches to proposal learning, and using our results to develop a new and improved training procedure. In particular, we introduce a method for expressing the gap between an and the log marginal likelihood as a divergence between two distributions on an extended sampling space. Doing so allows us to investigate the behavior of this family of algorithms when the objective is maximized perfectly, which occurs only if the divergence becomes zero. In the case, this implies that the proposal distributions are equal to the posterior distributions under the learned model. In the case, it has implications for both the proposal distributions and the intermediate set of targets that are learned. We demonstrate that, somewhat counter-intuitively, using lower variance estimates for the marginal likelihood can actually be harmful to proposal learning. Using these insights, we experiment with an adaptation to the algorithm, which we call *alternating* , that uses different lower bounds for updating the model parameters and proposal parameters. We observe that this adaptation can, in some cases, improve model learning and proposal adaptation.
Background
==========
State-Space Models
------------------
are probabilistic models over a set of latent variables $x_{1:T}$ and observed variables $y_{1:T}$. Given parameters $\theta$, a is characterized by an initial density $\mu_{\theta}(x_1)$, a series of transition densities $f_{t, \theta}(x_t {\lvert}x_{1:t - 1})$, and a series of emission densities $g_{t, \theta}(y_t {\lvert}x_{1:t})$ with the joint density being $p_{\theta}(x_{1:T}, y_{1:T}) = \mu_{\theta}(x_1) \prod_{t = 2}^T f_{t,\theta}(x_t {\lvert}x_{1:t - 1}) \prod_{t = 1}^T g_{t,\theta}(y_t {\lvert}x_{1:t})$.
We are usually interested in approximating the posterior $p_{\theta}(x_{1:T} {\lvert}y_{1:T})$ or the expectation of some test function $\varphi$ under this posterior $I(\varphi) := \int \varphi(x_{1:T}) p_{\theta}(x_{1:T} {\lvert}y_{1:T}) \,\mathrm dx_{1:T}$. We refer to these two tasks as inference. Inference in models which are non-linear, non-discrete, and non-Gaussian is difficult and one must resort to approximate methods, for which has been shown to be one of the most powerful approaches [@doucet2009tutorial].
We will consider model learning as a problem of maximizing the marginal likelihood $p_{\theta}(y_{1:T}) = \int p_{\theta}(x_{1:T}, y_{1:T}) \,\mathrm dx_{1:T}$ in the family of models parameterized by $\theta$.
Sequential Monte Carlo
----------------------
performs approximate inference on a sequence of target distributions $(\pi_t(x_{1:t}))_{t = 1}^T$. In the context of , the target distributions are often taken to be $(p_{\theta}(x_{1:t} {\lvert}y_{1:t}))_{t = 1}^T$. Given a parameter $\phi$ and proposal distributions $q_{1, \phi}(x_1 {\lvert}y_1)$ and $(q_{t, \phi}(x_t {\lvert}y_{1:t}, x_{1:t - 1}))_{t = 2}^T$ from which we can sample and whose densities we can evaluate, is described in Algorithm \[alg:background/smc\].
Using the set of weighted particles $(\tilde x_{1:T}^k, w_T^k)_{k = 1}^K$ at the last time step, we can approximate the posterior as $\sum_{k = 1}^K \bar{w}_T^k \delta_{\tilde{x}_{1:T}^k}(x_{1:T})$ and the integral $I_{\varphi}$ as $\sum_{k = 1}^K \bar{w}_T^k \varphi(\tilde{x}_{1:T}^k)$, where $\bar{w}_T^k := w_T^k / \sum_j w_T^j$ is the normalized weight and $\delta_z$ is a Dirac measure centered on $z$. Furthermore, one can obtain an unbiased estimator of the marginal likelihood $p_{\theta}(y_{1:T})$ using the intermediate particle weights: $$\begin{aligned}
\hat Z_{\text{SMC}} := \prod_{t = 1}^T \left[\frac{1}{K} \sum_{k = 1}^K w_t^k \right]. \label{background:Z_SMC}\end{aligned}$$
The sequential nature of and the resampling step are crucial in making scalable to large $T$. The former makes it easier to design efficient proposal distributions as each step need only target the next set of variables $x_t$. The resampling step allows the algorithm to focus on promising particles in light of new observations, avoiding the exponential divergence between the weights of different samples that occurs for importance sampling as $T$ increases. This can be demonstrated both empirically and theoretically [@del2004feynman Chapter 9]. We refer the reader to [@doucet2009tutorial] for an in-depth treatment of .
Importance Weighted Auto-Encoders
---------------------------------
Given a dataset of observations $(y^{(n)})_{n = 1}^N$, a generative network $p_{\theta}(x, y)$ and an inference network $q_{\phi}(x {\lvert}y)$, [@burda2016importance] maximize $\frac{1}{N} \sum_{n = 1}^N \operatorname{\acrshort{ELBO}}_{\text{IS}}(\theta, \phi, y^{(n)})$ where, for a given observation $y$, the $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ (with $K$ particles) is a lower bound on $\log p_{\theta}(y)$ by Jensen’s inequality: $$\begin{aligned}
&\operatorname{\acrshort{ELBO}}_{\text{IS}}(\theta, \phi, y) = \int Q_{\text{IS}}(x^{1:K}) \log \hat Z_{\text{IS}}(x^{1:K}) \,\mathrm dx^{1:K}
\leq \log p_{\theta}(y), \text{ where } \label{eq:background/elbo-is} \\
&Q_{\text{IS}}(x^{1:K}) = \prod_{k = 1}^K q_{\phi}(x^k {\lvert}y),
\,\,\,\, \hat Z_{\text{IS}}(x^{1:K}) = \frac{1}{K} \sum_{k = 1}^K \frac{p_{\theta}(x^k, y)}{q_{\phi}(x^k {\lvert}y)}. \label{eq:background/q_is_z_is}\end{aligned}$$ Note that for $K = 1$ particle, this objective reduces to a [@kingma2014auto; @rezende2014stochastic] objective we will refer to as $$\begin{aligned}
\operatorname{\acrshort{ELBO}}_{\text{VAE}}(\theta, \phi, y) = \int q_{\phi}(x {\lvert}y) (\log p_{\theta}(x, y) - \log q_{\phi}(x {\lvert}y)) \,\mathrm dx. \label{eqn:background/elbo_vae}\end{aligned}$$
The optimization is performed using where a sample from $\left(\prod_{k = 1}^K q_{\phi}(x^k {\lvert}y^{(n)})\right)$ is obtained using the reparameterization trick [@kingma2014auto] and the gradient $\frac{1}{N}\sum_{n=1}^{N}\nabla_{\theta, \phi} \log \left(\sum_{k = 1}^K \frac{p_{\theta}(x^k, y^{(n)})}{q_{\phi}(x^k {\lvert}y^{(n)})}\right)$ is used to perform an optimization step.
Auto-Encoding Sequential Monte Carlo
====================================
implements model learning, proposal adaptation, and inference amortization in a similar manner to the and the : it uses on an empirical average of the over observations. However, it varies in the form of this . In this section, we will introduce the , explain how gradients of it can be estimated, and discuss the implications of these changes.
Objective Function
------------------
Consider a family of $\{p_{\theta}(x_{1:T}, y_{1:T}): \theta \in \Theta\}$ and a family of proposal distributions $\{q_{\phi}(x_{1:T} {\lvert}y_{1:T}) = q_{1, \phi}(x_1 {\lvert}y_1) \prod_{t = 2}^T q_{t, \phi}(x_t {\lvert}x_{1:t - 1}, y_{1:t}): \phi \in \Phi\}$. uses an objective based on the marginal likelihood estimator . In particular, for a given $y_{1:T}$, the objective is defined as $$\begin{aligned}
\operatorname{\acrshort{ELBO}}_{\text{SMC}}(\theta, \phi, y_{1:T}) &:= \int Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K},\end{aligned}$$ where $\hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ is defined in and $Q_{\text{SMC}}$ is the sampling distribution of , $$\begin{aligned}
Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) &= \left(\prod_{k = 1}^K q_{1, \phi}(x_1^k)\right) \left( \prod_{t = 2}^T \prod_{k = 1}^K q_{t, \phi}(x_t^k {\lvert}\tilde x_{1:t - 1}^{a_{t - 1}^k}) \cdot \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K})\right). \label{eqn:aesmc/q_smc}\end{aligned}$$
$\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ forms a lower bound to the log marginal likelihood $\log p_{\theta}(y_{1:T})$ due to Jensen’s inequality and the unbiasedness of the marginal likelihood estimator. Hence, given a dataset $(y_{1:T}^{(n)})_{n = 1}^N$, we can perform model learning based on maximizing the lower bound of $\frac{1}{N} \sum_{n = 1}^N \log p_{\theta}(y_{1:T}^{(n)})$ as a surrogate target, namely by maximizing $$\begin{aligned}
\mathcal J(\theta, \phi) &:= \frac{1}{N} \sum_{n = 1}^N \operatorname{\acrshort{ELBO}}_{\text{SMC}}(\theta, \phi, y_{1:T}^{(n)}). \label{eqn:aesmc/objective}\end{aligned}$$ For notational convenience, we will talk about optimizing in the rest of this section. However, we note that the main intended use of is to amortize over datasets, for which the is replaced by the dataset average $\mathcal J(\theta, \phi)$ in the optimization target. Nonetheless, rather than using the full dataset for each gradient update, will we instead use minibatches, noting that this forms unbiased estimator.
Gradient Estimation
-------------------
We describe a gradient estimator used for optimizing $\operatorname{\acrshort{ELBO}}_{\text{SMC}}(\theta, \phi, y_{1:T})$ using . The sampler in Algorithm \[alg:background/smc\] proceeds by sampling $x_1^{1:K}, a_1^{1:K}, x_2^{1:K}, \dotsc$ sequentially from their respective distributions $\prod_{k = 1}^K q_1(x_1^k), \; \prod_{k = 1}^K
\mathrm{Discrete}(a_1^k {\lvert}w_1^{1:K}), \; \prod_{k = 1}^K q_2(x_2^k {\lvert}x_1^{a_1^k}), \dotsc$ until the whole particle-weight trajectory $(x_{1:K}^{1:T}, a_{1:T - 1}^{1:K})$ is sampled. From this trajectory, using equation , we can obtain an estimator for the marginal likelihood.
Assuming that the sampling of latent variables $x_{1:T}^{1:K}$ is reparameterizable, we can make their sampling independent of $(\theta, \phi)$. In particular, assume that there exists a set of auxiliary random variables $\epsilon_{1:T}^{1:K}$ where $\epsilon_t^k \sim s_t$ and a set of reparameterization functions $r_t$. We can simulate the sampler by first sampling $\epsilon_1^{1:K} \sim \prod_{k = 1}^K s_1$ and setting $x_1^k = r_1(\epsilon_1^k)$ and $\tilde{x}_1^k = x_1^k$, then for $t=2,\dots,T$ cycling through sampling $a_{t-1}^{1:K} \sim \prod_{k = 1}^K \mathrm{Discrete}(a_{t-1}^k {\lvert}w_{t-1}^{1:K})$ and $\epsilon_{t}^{1:K} \sim \prod_{k = 1}^K s_{t}$, and setting $x_{t}^k = r_{t}(\epsilon_{t}^k, \tilde{x}_{1:t-1}^{a_{t-1}^k})$ and $\tilde{x}_{1:t}^k = (\tilde{x}_{1:t - 1}^{a_{t - 1}^k}, x_t^k)$. We use the resulting reparameterized sample of $(x_{1:K}^{1:T}, a_{1:T - 1}^{1:K})$ to evaluate the gradient estimator $\nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$.
To account for the discrete choices of ancestor indices $a_t^k$ one could additionally use the [@williams1992simple] trick, however in practice, we found that the additional term in the estimator has problematically high variance. We explore various other possible gradient estimators and empirical assessments of their variances in Appendix \[sec:gradients\]. This exploration confirms that including the additional terms leads to problematically high variance, justifying our decision to omit them, despite introducing a small bias into the gradient estimates.
Bias & Implications on the Proposals
------------------------------------
In this section, we express the gap between and the log marginal likelihood as a divergence and study implications on the proposal distributions. We present a set of claims and propositions whose full proofs are in Appendix \[sec:bias-proofs\]. These give insight into the behavior of and show the advantages, and disadvantages, of using our different . This insight motivates Section \[sec:improving\_proposal\_learning\] which proposes an algorithm for improving proposal learning.
\[claim:aesmc/bias/elbo\] Given an *unnormalized target density* $\tilde P: \mathcal X \to [0, \infty)$ with *normalizing constant* $Z_P > 0$, $P := \tilde P / Z_P$, and a *proposal density* $Q: \mathcal X \to [0, \infty)$, then $$\begin{aligned}
\operatorname{\acrshort{ELBO}}:= \int Q(x) \log \frac{\tilde P(x)}{Q(x)} \,\mathrm dx, \label{eqn:aesmc/bias/elbo}
\end{aligned}$$ is a lower bound on $\log Z_P$ and satisfies $$\begin{aligned}
\operatorname{\acrshort{ELBO}}= \log Z_P - {\acrshort{KL}\left(Q \middle| \middle| P\right)}. \label{eqn:aesmc/bias/elbo2}
\end{aligned}$$
This is a standard identity used in variational inference and . In the case of , applying Definition \[claim:aesmc/bias/elbo\] with $P$ being $p_{\theta}(x {\lvert}y)$, $\tilde{P}$ being $p_{\theta}(x, y)$, $Z_P$ being $p_{\theta}(y)$, and $Q$ being $q_{\phi}(x {\lvert}y)$, we can directly rewrite as $\operatorname{\acrshort{ELBO}}_{\text{VAE}}(\theta, \phi, y) = \log p_{\theta}(y) - {\acrshort{KL}\left(q_{\phi}(x {\lvert}y) \middle| \middle| p_{\theta}(x {\lvert}y)\right)}$.
The key observation for expressing such a bound for general such as $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ and $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ is that the target density $P$ and the proposal density $Q$ need not directly correspond to $p_{\theta}(x {\lvert}y)$ and $q_{\phi}(x {\lvert}y)$. This allows us to view the underlying sampling distributions of the marginal likelihood Monte Carlo estimators such as $Q_{\text{IS}}$ in and $Q_{\text{SMC}}$ in as proposal distributions on an extended space $\mathcal X$. The following claim uses this observation to express the bound between a general and the log marginal likelihood as $\gls{KL}$ divergence from the extended space sampling distribution to a corresponding target distribution.
\[claim:aesmc/bias/estimator\] Given a non-negative unbiased estimator $\hat Z_P(x) \geq 0$ of the normalizing constant $Z_P$ where $x$ is distributed according to the proposal distribution $Q(x)$, the following holds: $$\begin{aligned}
\operatorname{\acrshort{ELBO}}&= \int Q(x) \log \hat Z_P(x) \,\mathrm dx = \log Z_P - {\acrshort{KL}\left(Q \middle| \middle| P\right)},
\label{eqn:aesmc/bias/elbo3} \\
\text{where} \quad P(x) &= \frac{Q(x) \hat Z_P(x)}{Z_P} \label{eqn:aesmc/bias/P}
\end{aligned}$$ is the implied normalized target density.
In the case of , we can apply Claim \[claim:aesmc/bias/estimator\] with $Q$ and $\hat Z_P$ being $Q_{\text{IS}}$ and $\hat Z_{\text{IS}}$ respectively as defined in and $Z_P$ being $p_{\theta}(y)$. This yields $$\begin{aligned}
\operatorname{\acrshort{ELBO}}_{\text{IS}}(\theta, \phi, y) &= \log p_{\theta}(y) - {\acrshort{KL}\left(Q_{\text{IS}} \middle| \middle| P_{\text{IS}}\right)}, \text{ where } \label{eqn:aesmc/bias/elbo_is} \\
P_{\text{IS}}(x^{1:K}) &= \frac{1}{K} \sum_{k = 1}^K \left(q_\phi(x^1 {\lvert}y) \cdots q_\phi(x^{k - 1} {\lvert}y) p_\theta(x^k {\lvert}y) q_\phi(x^{k + 1} {\lvert}y) \cdots q_\phi(x^K {\lvert}y) \right).\end{aligned}$$ Similarly, in the case of , we obtain $$\begin{aligned}
\operatorname{\acrshort{ELBO}}_{\text{SMC}}(\theta, \phi, y_{1:T}) &= \log p_{\theta}(y_{1:T}) - {\acrshort{KL}\left(Q_{\text{SMC}} \middle| \middle| P_{\text{SMC}}\right)}, \text{ where } \label{eqn:aesmc/bias/elbo_smc} \\
P_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) &= Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) / p_{\theta}(y_{1:T}).\end{aligned}$$
Having expressions for the target distribution $P$ and the sampling distribution $Q$ for a given allows us to investigate what happens when we maximize that , remembering that the term is strictly non-negative and zero if and only if $P = Q$. For the and cases then, provided the proposal is sufficiently flexible, one can always perfectly maximize the by setting $p_{\theta}(x {\lvert}y) = q_{\phi}(x {\lvert}y)$ for all $x$. The reverse implication also holds: if $\operatorname{\acrshort{ELBO}}_{\text{VAE}}=\log Z_P$ then it must be the case that $p_{\theta}(x {\lvert}y) = q_{\phi}(x {\lvert}y)$. However, for , achieving $\operatorname{\acrshort{ELBO}}=\log Z_P$ is only possible when one also has sufficient flexibility to learn a particular series of intermediate target distributions, namely the marginals of the final target distribution. In other words, it is necessary to learn a particular factorization of the generative model, not just the correct individual proposals, to achieve $P=Q$ and thus $\operatorname{\acrshort{ELBO}}_{\text{SMC}} = Z_P$. These observations are formalized in Propositions \[proposition:aesmc/bias/elbo\_is\] and \[proposition:aesmc/bias/elbo\_smc\] below.
\[proposition:aesmc/bias/elbo\_is\] $Q_{\text{IS}}(x^{1:K}) = P_{\text{IS}}(x^{1:K})$ for all $x^{1:K}$ if and only if $q(x {\lvert}y) = p(x {\lvert}y)$ for all $x$.
\[proposition:aesmc/bias/elbo\_smc\] If $K > 1$, then $P_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) = Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ for all $(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ if and only if
1. $\pi_t (x_{1:t})= \int p(x_{1:T} {\lvert}y_{1:T}) \,\mathrm dx_{t + 1:T} = p(x_{1:t} {\lvert}y_{1:T})$ for all $x_{1:t}$ and $t = 1, \dotsc, T$, and
2. $q_1(x_1 {\lvert}y_1) = p(x_1 {\lvert}y_{1:T})$ for all $x_1$ and $q_t(x_t {\lvert}x_{1:t - 1}, y_{1:t}) = p(x_{1:t} {\lvert}y_{1:T}) / p(x_{1:t - 1} {\lvert}y_{1:T})$ for $t = 2, \dotsc, T$ for all $x_{1:t}$,
where $\pi_t (x_{1:t})$ are the intermediate targets used by .
Proposition \[proposition:aesmc/bias/elbo\_smc\] has the consequence that if the family of generative models is such that the first condition does not hold, we will not be able to make the bound tight. This means that, except for a very small class of models, then, for most convenient parameterizations, it will be impossible to learn a perfect proposal that gives a tight bound, i.e. there will be no $\theta$ and $\phi$ such that the above conditions can be satisfied. However, it also means that $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ encodes important additional information about the implications the factorization of the generative model has on the inference—the model depends only on the final target $\pi_T(x_{1:T}) = p_{\theta}(x_{1:T} | y_{1:T})$, but some choices of the intermediate targets $\pi_t(x_{1:t})$ will lead to much more efficient inference than others. Perhaps more importantly, is usually a far more powerful inference algorithm than importance sampling and so the setup allows for more ambitious model learning problems to be effectively tackled than the or . After all, even though it is well known in the literature that, unlike for , most problems have no perfect set of proposals which will generate exact samples from the posterior [@doucet2009tutorial], still gives superior performance on most problems with more than a few dimensions. These intuitions are backed up by our experiments that show that using $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ regularly learns better models than using $\operatorname{\acrshort{ELBO}}_{\text{IS}}$.
Improving Proposal Learning {#sec:improving_proposal_learning}
===========================
In practice, one is rarely able to perfectly drive the divergence to zero and achieve a perfect proposal. In addition to the implications of the previous section, this occurs because $q_{\phi}(x_{1:T} {\lvert}y_{1:T})$ may not be sufficiently expressive to represent $p_{\theta}(x_{1:T} {\lvert}y_{1:T})$ exactly and because of the inevitable sub-optimality of the optimization process, remembering that we are aiming to learn an amortized inference artifact, rather than a single posterior representation. Consequently, to accurately assess the merits of different for proposal learning, it is necessary to consider their finite-time performance. We therefore now consider the effect the number of particles $K$ has on the gradient estimators for $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ and $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$.
[r]{}[0.24]{}
Counter-intuitively, it transpires that the tighter bounds implied by using a larger $K$ is often harmful to proposal learning for both and . At a high-level, this is because an accurate estimate for $\hat{Z}_P$ can be achieved for a wide range of proposal parameters $\phi$ and so the magnitude of $\nabla_{\phi} \operatorname{\acrshort{ELBO}}$ reduces as $K$ increases. Typically, this shrinkage happens faster than increasing $K$ reduces the standard deviation of the estimate and so the standard deviation of the gradient estimate relative to the problem scaling (i.e. as a ratio of true gradient $\nabla_{\phi} \operatorname{\acrshort{ELBO}}$) actually increases. This effect is demonstrated in Figure \[fig:kde\] which shows a kernel density estimator for the distribution of the gradient estimate for different $K$ and the model given in Section \[sec:experiments/proposal\_learning\]. Here we see that as we increase $K$, both the expected gradient estimate (which is equal to the true gradient by unbiasedness) and standard deviation of the estimate decrease. However, the former decreases faster and so the relative standard deviation increases. This is perhaps easiest to appreciate by noting that for $K > 10$, there is a roughly equal probability of the estimate being positive or negative, such that we are equally likely to increase or decrease the parameter value at the next iteration, inevitably leading to poor performance. On the other hand, when $K=1$, it is far more likely that the gradient estimate is positive than negative, and so there is clear drift to the gradient steps. We add to the empirical evidence for this behavior in Section \[sec:exp\]. Note the critical difference for model learning is that $\nabla_{\theta} \operatorname{\acrshort{ELBO}}$ does not, in general, decrease in magnitude as $K$ increases. Note also that using a larger $K$ should always give better performance at test time; it may though be better to learn $\phi$ using a smaller $K$.
In simultaneously developed work [@rainforth2017tighter], we formalized this intuition in the setting by showing that the estimator of $\nabla_\phi \operatorname{\acrshort{ELBO}}_{\text{IS}}(\theta, \phi, x)$ with $K$ particles, denoted by $I_K$, has the following : $$\begin{aligned}
\operatorname{\acrshort{SNR}}:= \frac{\operatorname{\mathbb{E}}[I_K]}{\sqrt{\operatorname{\mathrm{Var}}[I_K]}} = O\left(\sqrt{\frac{1}{K}}\right).\end{aligned}$$ We thus see that increasing $K$ reduces the and so the gradient updates for the proposal will degrade towards pure noise if $K$ is set too high.
Alternating ELBOs {#sec:aesmc/alt}
-----------------
To address these issues, we suggest and investigate the algorithm which updates $(\theta, \phi)$ in a coordinate descent fashion using different , and thus gradient estimates, for each. We pick a $\theta$-optimizing pair and a $\phi$-optimizing pair $(A_{\theta}, K_{\theta}), (A_{\phi}, K_{\phi}) \in \{\gls{IS}, \gls{SMC}\} \times \{1, 2, \dotsc\}$, corresponding to an inference type and number of particles. In an optimization step, we obtain an estimator for $\nabla_{\theta} \operatorname{\acrshort{ELBO}}_{A_{\theta}}$ with $K_{\theta}$ particles and an estimator for $\nabla_{\phi} \operatorname{\acrshort{ELBO}}_{A_{\phi}}$ with $K_{\phi}$ particles which we call $g_{\theta}$ and $g_{\phi}$ respectively. We use $g_{\theta}$ to update the current $\theta$ and $g_{\phi}$ to update the current $\phi$. The results from the previous sections suggest that using $A_{\theta} = \gls{SMC}$ and $A_{\phi} = \gls{IS}$ with a large $K_{\theta}$ and a small $K_{\phi}$ may perform better model and proposal learning than just fixing $(A_{\theta}, K_{\theta}) = (A_{\phi}, K_{\phi})$ to $(\gls{SMC}, \text{large})$ since using $A_{\phi} = \gls{IS}$ with small $K_{\phi}$ helps learning $\phi$ (at least in terms of the ) and using $A_{\theta} = \gls{SMC}$ with large $K_{\theta}$ helps learning $\theta$. We experimentally observe that this procedure can in some cases improve both model and proposal learning.
Experiments {#sec:exp}
===========
We now present a series of experiments designed to answer the following questions: 1) Does tightening the bound by using either more particles or a better inference procedure lead to an adverse effect on proposal learning? 2) Can , despite this effect, outperform ? 3) Can we further improve the learned model and proposal by using ?
First we investigate a for model learning and a latent variable model for proposal adaptation. This allows us to compare the learned parameters to the optimal ones. Doing so, we confirm our conclusions for this simple problem.
We then extend those results to more complex, high dimensional observation spaces that require models and proposals parameterized by neural networks. We do so by investigating the *Moving Agents* dataset, a set of partially occluded video sequences.
Linear Gaussian State Space Model {#sec:experiments/lgssm}
---------------------------------
Given the following $$\begin{aligned}
p(x_1) &= \mathrm{Normal}\left(x_1; 0, 1^2\right), \\
p(x_t {\lvert}x_{t - 1}) &= \mathrm{Normal}\left(x_t; \theta_1 x_{t - 1}, 1^2\right), && t = 2, \dotsc T, \\
p(y_t {\lvert}x_t) &= \mathrm{Normal}\left(y_t; \theta_2 x_t, \sqrt{0.1}^2\right), && t = 1, \dotsc, T,\end{aligned}$$ we find that optimizing $\operatorname{\acrshort{ELBO}}_{\text{SMC}}(\theta, \phi, y_{1:T})$ w.r.t. $\theta$ leads to better generative models than optimizing $\operatorname{\acrshort{ELBO}}_{\text{IS}}(\theta, \phi, y_{1:T})$. The same is true for using more particles.
We generate a sequence $y_{1:T}$ for $T = 200$ by sampling from the model with $\theta = (\theta_1, \theta_2) = (0.9, 1.0)$. We then optimize the different $\gls{ELBO}$s w.r.t. $\theta$ using the bootstrap proposal $q_1(x_1 {\lvert}y_1) = \mu_{\theta}(x_1)$ and $q_t(x_t {\lvert}x_{1:t - 1}, y_{1:t}) = f_{t, \theta}(x_t {\lvert}x_{1:t - 1})$. Because we use the bootstrap proposal, gradients w.r.t. to $\theta$ are not backpropagated through $q$.
We use a fixed learning rate of $0.01$ and optimize for $500$ steps using . Figure \[fig:experiments/lgssm/logz\_and\_params\] shows that the convergence of both $\log p_{\theta}(y_{1:T})$ to $\max_{\theta} \log p_{\theta}(y_{1:T})$ and $\theta$ to $\operatorname*{arg\,\!max}_{\theta} \log p_{\theta}(y_{1:T})$ is faster when $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ and more particles are used.
[.49]{} ![(Left) Log marginal likelihood analytically evaluated at every $\theta$ during optimization; the black line indicates $\max_{\theta} \log p_{\theta}(y_{1:T})$ obtained by the algorithm. (Right) learning of model parameters; the black line indicates $\operatorname*{arg\,\!max}_{\theta} \log p_{\theta}(y_{1:T})$ obtained by the algorithm.[]{data-label="fig:experiments/lgssm/logz_and_params"}](log_marginal_likelihoods.pdf "fig:")
[.49]{} ![(Left) Log marginal likelihood analytically evaluated at every $\theta$ during optimization; the black line indicates $\max_{\theta} \log p_{\theta}(y_{1:T})$ obtained by the algorithm. (Right) learning of model parameters; the black line indicates $\operatorname*{arg\,\!max}_{\theta} \log p_{\theta}(y_{1:T})$ obtained by the algorithm.[]{data-label="fig:experiments/lgssm/logz_and_params"}](model_parameters.pdf "fig:")
Proposal Learning
-----------------
We now investigate how learning $\phi$, i.e. the proposal, is affected by the the choice of and the number of particles.
\[sec:experiments/proposal\_learning\] Consider a simple, fixed generative model $p(\mu)p(x {\lvert}\mu) = \mathrm{Normal}(\mu; 0, 1^2)\mathrm{Normal}(x; \mu, 1^2)$ where $\mu$ and $x$ are the latent and observed variables respectively and a family of proposal distributions $q_{\phi}(\mu) = \mathrm{Normal}(\mu; \mu_q, \sigma_q^2)$ parameterized by $\phi = (\mu_q, \log \sigma_q^2)$. For a fixed observation $x = 2.3$, we initialize $\phi = (0.01, 0.01)$ and optimize $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ with respect to $\phi$. We investigate the quality of the learned parameter $\phi$ as we increase the number of particles $K$ during training. Figure \[fig:experiments/proposals/lgssm\] (left) clearly demonstrates that the quality of $\phi$ compared to the analytic posterior decreases as we increase $K$.
Similar behavior is observed in Figure \[fig:experiments/proposals/lgssm\] (middle, right) where we optimize $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ with respect to both $\theta$ and $\phi$ for the described in Section \[sec:experiments/lgssm\]. We see that using more particles helps model learning but makes proposal learning worse. Using our algorithm alleviates this problem and at the same time makes model learning faster as it profits from a more accurate proposal distribution. We provide more extensive experiments exploring proposal learning with different and number of particles in Appendix \[sec:additional-experiments\].
[.32]{} ![*(Left)* Optimizing $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ for the Gaussian unknown mean model with respect to $\phi$ results in worse $\phi$ as we increase number of particles $K$. *(Middle, right)* Optimizing $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ with respect to $(\theta, \phi)$ for and using the algorithm for updating $(\theta, \phi)$ with $(A_{\theta}, K_{\theta}) = (\gls{SMC}, 1000)$ and $(A_{\phi}, K_{\phi}) = (\gls{IS}, 10)$. *Right* measures the quality of $\phi$ by showing $\sqrt{\sum_{t = 1}^T (\mu_t^{\text{kalman}} - \mu_t^{\text{approx}})^2}$ where $\mu_t^{\text{kalman}}$ is the marginal mean obtained from the Kalman smoothing algorithm under the model with -optimized parameters and $\mu_t^{\text{approx}}$ is an marginal mean obtained from the set of $10$ particles with learned/bootstrap proposal.[]{data-label="fig:experiments/proposals/lgssm"}](gaussian_proposals.pdf "fig:")
[.64]{} ![*(Left)* Optimizing $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ for the Gaussian unknown mean model with respect to $\phi$ results in worse $\phi$ as we increase number of particles $K$. *(Middle, right)* Optimizing $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ with respect to $(\theta, \phi)$ for and using the algorithm for updating $(\theta, \phi)$ with $(A_{\theta}, K_{\theta}) = (\gls{SMC}, 1000)$ and $(A_{\phi}, K_{\phi}) = (\gls{IS}, 10)$. *Right* measures the quality of $\phi$ by showing $\sqrt{\sum_{t = 1}^T (\mu_t^{\text{kalman}} - \mu_t^{\text{approx}})^2}$ where $\mu_t^{\text{kalman}}$ is the marginal mean obtained from the Kalman smoothing algorithm under the model with -optimized parameters and $\mu_t^{\text{approx}}$ is an marginal mean obtained from the set of $10$ particles with learned/bootstrap proposal.[]{data-label="fig:experiments/proposals/lgssm"}](lgssm_proposal_3.pdf "fig:")
Moving Agents {#sec:moving-agents}
-------------
To show that our results are applicable to complex, high dimensional data we compare and on stochastic, partially observable video sequences. Figure \[fig:experiments/max/visualisation\] in Appendix \[sec:appendix\_moving\_agents\] shows an example of such a sequence.
The dataset consists of $N = 5000$ sequences of images $(y_{1:T}^{(n)})_{n = 1}^{N}$ of which 1000 are randomly held out as test set. Each sequence contains $T = 40$ images represented as a 2 dimensional array of size $32\times32$. In each sequence there is one agent, represented as circle, whose starting position is sampled randomly along the top and bottom of the image. The dataset is inspired by [@ondruska2016deep], however with the crucial difference that the movement of the agent is *stochastic*. The agent performs a directed random walk through the image. At each timestep, it moves according to $$\begin{split}
y_{t+1} & \sim \mathrm{Normal}(y_{t+1}; y_t+0.15, 0.02^2)\\
x_{t+1} & \sim \mathrm{Normal}(x_{t+1}; 0, 0.02^2)
\end{split}$$ where $(x_t, y_t)$ are the coordinates in frame $t$ in a unit square that is then projected onto $32\times32$ pixels. In addition to the stochasticity of the movement, half of the image is occluded, preventing the agent from being observed.
For the generative model and proposal distribution we use a [@chung2015recurrent]. It extends by introducing a stochastic latent state $x_t$ at each timestep $t$. Together with the observation $y_t$, this state conditions the deterministic transition of the . By introducing this unobserved stochastic state, the is able to better model complex long range variability in stochastic sequences. Architecture and hyperparameter details are given in Appendix \[sec:appendix\_vrnn\].
Figure \[fig:experiments/max/elbos\] shows $\max(\operatorname{\acrshort{ELBO}}_{\text{IS}},\operatorname{\acrshort{ELBO}}_{\text{SMC}})$ for models trained with and for different particle numbers. The lines correspond to the mean over three different random seeds and the shaded areas indicate the standard deviation. The same number of particles was used for training and testing, additional hyperparameter settings are given in the appendix. One can see that models trained using outperform and using more particles improves the for both. In Appendix \[sec:appendix\_moving\_agents\], we inspect different learned generative models by using them for prediction, confirming the results presented here. We also tested on this task, but found that while it did occasionally improve performance, it was much less stable than and .
[.49]{} ![*(Left)* Rolling mean over 5 epochs of $\max(\operatorname{\acrshort{ELBO}}_{\text{SMC}},
\operatorname{\acrshort{ELBO}}_{\text{IS}})$ on the test set, lines indicate the average over 3 random seeds and shaded areas indicate standard deviation. The color indicates the number of particles, the line style the used algorithm. *(Right)* The table shows the final $\max(\operatorname{\acrshort{ELBO}}_{\text{SMC}}, \operatorname{\acrshort{ELBO}}_{\text{IS}})$ for each learned model.[]{data-label="fig:experiments/max/elbos"}](ELBO_Moving-Agents "fig:"){width=".95\linewidth"}
[.49]{}
Particles Method Moving Agents
----------- -------- --------------- -- -- --
-357.3
**-356.7**
-356.6
**-356.1**
-356.2
**-356.1**
Conclusions
===========
We have developed —a method for performing model learning using a new objective which is based on the marginal likelihood estimator. This objective is optimized using and the reparameterization trick. Our approach utilizes the efficiency of in models with intermediate observations and hence is suitable for highly structured models. We experimentally demonstrated that this objective leads to better generative model training than the objective for structured problems, due to the superior inference and tighter bound provided by using instead of importance sampling.
Additionally, in Claim \[claim:aesmc/bias/estimator\], we provide a simple way to express the bias of objectives induced by log of marginal likelihood estimators as a divergence on an extended space. In Propositions \[proposition:aesmc/bias/elbo\_is\] and \[proposition:aesmc/bias/elbo\_smc\], we investigate the implications of these being zero in the case of and . In the latter case, we find that we can achieve zero only if we are able to learn intermediate target distributions corresponding to marginals of the target distribution. Using our assertion that tighter variational bounds are not necessarily better, we then introduce and test a new method, alternating , that addresses some of these issues and observe that, in some cases, this improves both model and proposal learning.
### Acknowledgments {#acknowledgments .unnumbered}
TAL is supported by EPSRC DTA and Google (project code DF6700) studentships. MI is supported by the UK EPSRC CDT in Autonomous Intelligent Machines and Systems. TR is supported by the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) ERC grant agreement no. 617071; majority of TR’s work was undertaken while he was in the Department of Engineering Science, University of Oxford, and was supported by a BP industrial grant. TJ is supported by the UK EPSRC and MRC CDT in Statistical Science. FW is supported by The Alan Turing Institute under the EPSRC grant EP/N510129/1; DARPA PPAML through the U.S. AFRL under Cooperative Agreement FA8750-14-2-0006; Intel and DARPA D3M, under Cooperative Agreement FA8750-17-2-0093.
Gradients {#sec:gradients}
=========
The goal is to obtain an unbiased estimator for the gradient $$\begin{aligned}
\nabla_{\theta, \phi} \int Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K}.\end{aligned}$$
Full Reinforce
--------------
We express the required quantity as $$\begin{aligned}
&\nabla_{\theta, \phi} \int Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K} \\
&= \int \nabla_{\theta, \phi} Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) + \\
& \hspace{3em} Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K} \\
&= \int Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \left[\nabla_{\theta, \phi} \log Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) + \right. \\
&\left. \hspace{3em} \nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})\right] \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K},\end{aligned}$$ which we can estimate by sampling $(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ directly from $Q_{\text{SMC}}$ and evaluating $\left[\nabla_{\theta, \phi} \log Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) + \nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})\right]$.
Reinforce & Reparameterization {#sec:gradients/reinforce-reparam}
------------------------------
We express the required quantity as $$\begin{aligned}
\nabla_{\theta, \phi} &\int Q_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K} \\
&= \nabla_{\theta, \phi} \int \left(\prod_{k = 1}^K q_1(x_1^k)\right) \left( \prod_{t = 2}^T \prod_{k = 1}^K q_t(x_t^k {\lvert}x_{t - 1}^{a_{t - 1}^k}) \cdot \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K})\right) \nonumber\\
& \hspace{3em} \log \hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) \,\mathrm dx_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K} \\
&= \nabla_{\theta, \phi} \int \left(\prod_{k = 1}^K s_1(\epsilon_1^k)\right) \left( \prod_{t = 2}^T \prod_{k = 1}^K s_t(\epsilon_t^k) \cdot \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K})\right) \nonumber\\
& \hspace{3em} \log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K}) \,\mathrm d\epsilon_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K} \\
&= \int \left( \prod_{t = 1}^T \prod_{k = 1}^K s_t(\epsilon_t^k) \right) \left[\nabla_{\theta, \phi} \prod_{t = 2}^T \prod_{k = 1}^K \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K}) \log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K}) + \right. \nonumber\\
&\hspace{3em} \left. \left( \prod_{t = 2}^T \prod_{k = 1}^K \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K}) \right) \nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K}) \right] \,\mathrm d\epsilon_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K} \\
&= \int \left( \prod_{t = 1}^T \prod_{k = 1}^K s_t(\epsilon_t^k) \right) \left( \prod_{t = 2}^T \prod_{k = 1}^K \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K}) \right) \cdot \nonumber\\
& \hspace{3em} \left[
\nabla_{\theta, \phi} \log\left(
\prod_{t = 2}^T \prod_{k = 1}^K \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K})
\right)
\log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K}) + \right. \nonumber\\
& \hspace{3em} \left.\nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K}) \label{eq:gradients/reinforce-reparam}
\right] \,\mathrm d\epsilon_{1:T}^{1:K} \,\mathrm da_{1:T - 1}^{1:K},\end{aligned}$$ where $r(\epsilon_{1:T}^{1:K})$ denotes a sample with identical distribution as $x_{1:T}^{1:K}$ obtained by passing the auxiliary samples $\epsilon_{1:T}^{1:K}$ through the reparameterization function. We can thus estimate the gradient by sampling $\epsilon_{1:T}^{1:K}$ from the auxiliary distribution, reparameterizing and evaluating $\left[
\nabla_{\theta, \phi} \log\left(
\prod_{t = 2}^T \prod_{k = 1}^K \mathrm{Discrete}(a_{t - 1}^k {\lvert}w_{t - 1}^{1:K})
\right)
\log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K}) + \nabla_{\theta, \phi} \log \hat Z_{\text{SMC}}(r(\epsilon_{1:T}^{1:K}), a_{1:T - 1}^{1:K})
\right]$.
In Figure \[fig:gradients/reinforce\_vs\_ignore\], we demonstrate that the estimator in has much higher variance if we include the first term.
![$T = 200$ model described in Section \[sec:experiments/lgssm\]. of $\nabla_{\theta_1} \operatorname{\acrshort{ELBO}}_{\text{SMC}}$ evaluated at $\theta_1 = 0.1$ with $K = 16$ using $100$ samples.[]{data-label="fig:gradients/reinforce_vs_ignore"}](smc_across_gradient_estimators_0.pdf){width="50.00000%"}
Proofs for Bias & Implications on the Proposals {#sec:bias-proofs}
===============================================
$$\begin{aligned}
\operatorname{\acrshort{ELBO}}&= \int Q(x) \log \frac{Z_P P(x)}{Q(x)} \,\mathrm dx \\
&= \int Q(x) \log Z_P \,\mathrm dx - \int Q(x) \log\frac{Q(x)}{P(x)} \,\mathrm dx \\
&= \log Z_P - {\acrshort{KL}\left(Q \middle| \middle| P\right)}. \label{eqn:bias-proofs/basic_elbo}
\end{aligned}$$
Since $\hat Z_P(x) \geq 0$, $Q(x) \geq 0$ and $\int Q(x) \hat Z_P(x) \,\mathrm dx = Z_P$, we can let the unnormalized target density in Definition \[claim:aesmc/bias/elbo\] be $\tilde P(x) = Q(x) \hat Z_P(x)$. Hence, the normalized target density is $P(x) = Q(x) \hat Z_P(x) / Z_P$. Substituting these quantities into and yields the two equalities in .
($\implies$) Substituting for $Q_{\text{IS}}(x^{1:K}) = P_{\text{IS}}(x^{1:K})$, we obtain $$\begin{aligned}
\prod_{k = 1}^K q(x^k {\lvert}y) &= \frac{1}{K} \sum_{k = 1}^K \frac{\prod_{\ell = 1}^K q(x^\ell {\lvert}y)}{q(x^k {\lvert}y)} p(x^k {\lvert}y) \\ &= \frac{1}{K} \sum_{k = 1}^K \left[q(x^1 {\lvert}y) \cdots q(x^{k - 1} {\lvert}y) p(x^k {\lvert}y) q(x^{k + 1} {\lvert}y) \cdots q(x^K {\lvert}y)\right].
\end{aligned}$$ Integrating both sides with respect to $(x^2, \dotsc, x^K)$ over the whole support (i.e. marginalizing out everything except $x^1$), we obtain: $$\begin{aligned}
q(x^1 {\lvert}y) = \frac{1}{K} \left[p(x^1 {\lvert}y) + \sum_{k = 2}^K q(x^1 {\lvert}y)\right].
\end{aligned}$$ Rearranging gives us $q(x^1 {\lvert}y) = p(x^1 {\lvert}y)$ for all $x^1$.
($\impliedby$) Substituting $p(x^k {\lvert}y) = q(x^k {\lvert}y)$, we obtain $$\begin{aligned}
P_{\text{IS}}(x^{1:K}) &= \frac{1}{K} \sum_{k = 1}^K \frac{Q_{\text{IS}}(x^{1:K})}{q(x^k {\lvert}y)} p(x^k {\lvert}y) \\
&= \frac{1}{K} \sum_{k = 1}^K Q_{\text{IS}}(x^{1:K}) \\
&= Q_{\text{IS}}(x^{1:K}).
\end{aligned}$$
We consider the general sequence of target distributions $\pi_t(x_{1:t})$ ($p_{\theta}(x_{1:t} {\lvert}y_{1:t})$ in the case of ), their unnormalized versions $\gamma_t(x_{1:t})$ ($p_{\theta}(x_{1:t}, y_{1:t})$ in the case of ), their normalizing constants $Z_t = \int \gamma_t(x_{1:t}) \,\mathrm dx_{1:t}$ ($p_{\theta}(y_{1:t})$ in the case of ), where $Z = Z_T = p(y_{1:T})$.
($\implies$) It suffices to show that $\hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) = Z$ for all $(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ implies 1 and 2 in Proposal \[proposition:aesmc/bias/elbo\_smc\] due to equation .
We first prove that $\hat Z_{\text{SMC}}(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) = Z$ for all $(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ implies that the weights $$\begin{aligned}
w_1(x_1) &:= \frac{\gamma_1(x_1)}{q_1(x_1)} \label{eqn:weight_1} \\
w_t(x_{1:t}) &:= \frac{\gamma_t(x_{1:t})}{\gamma_{t - 1}(x_{1:t - 1}) q_t(x_t {\lvert}x_{1:t - 1})} && \text{ for } t = 2, \dotsc, T \label{eqn:weight_t}
\end{aligned}$$ are constant with respect to $x_{1:t}$.
Pick $t \in \{1, \dotsc, T\}$ and distinct $k, \ell \in \{1, \dotsc, K\}$. Also, pick $x_{1:t}$ and ${x'}_{1:t}$. Now, consider two sets of particle sets $(\bar x_{1:T}^{1:K}, \bar a_{1:T - 1}^{1:K})$ and $(\tilde x_{1:T}^{1:K}, \tilde a_{1:T - 1}^{1:K})$, illustrated in Figure \[fig:bias-proofs/particle\_sets\], such that $$\begin{aligned}
\bar x_\tau^\kappa &=
\begin{cases}
{x'}_\tau & \text{ if } \kappa = \ell \text{ and } \tau < t \\
{x'}_\tau & \text{ if } (\kappa, \tau) = (k, t) \\
x_\tau & \text{ if } \kappa = k \text{ and } \tau < t \\
x_\tau^\kappa & \text{ otherwise}
\end{cases} && \text{ for } \tau = 1, \dotsc, T, ~\kappa = 1, \dotsc, K, \\
\bar a_\tau^\kappa &=
\begin{cases}
\ell & \text{ if } (\kappa, \tau) = (k, t - 1) \text{ or } (k, t) \\
\kappa & \text{ otherwise}
\end{cases} && \text{ for } \tau = 1, \dotsc, T - 1, ~\kappa = 1, \dotsc, K, \\
\tilde x_\tau^\kappa &=
\begin{cases}
{x'}_\tau & \text{ if } \kappa = \ell \text{ and } \tau < t \\
x_\tau & \text{ if } (\kappa, \tau) = (k, t) \\
x_\tau & \text{ if } \kappa = k \text{ and } \tau < t \\
x_\tau^\kappa & \text{ otherwise}
\end{cases} && \text{ for } \tau = 1, \dotsc, T, ~\kappa = 1, \dotsc, K, \\
\tilde a_\tau^\kappa &=
\begin{cases}
\ell & \text{ if } (\kappa, \tau) = (k, t) \\
\kappa & \text{ otherwise}
\end{cases} && \text{ for } \tau = 1, \dotsc, T - 1, ~\kappa = 1, \dotsc, K.
\end{aligned}$$
[.4]{} ![(Left) particle set $(\bar x_{1:T}^{1:K}, \bar a_{1:T - 1}^{1:K})$ and (right) particle set $(\tilde x_{1:T}^{1:K}, \tilde a_{1:T - 1}^{1:K})$. Lines indicate ancestor indices.[]{data-label="fig:bias-proofs/particle_sets"}](particle_set_1.pdf "fig:"){width="\textwidth"}
[.4]{} ![(Left) particle set $(\bar x_{1:T}^{1:K}, \bar a_{1:T - 1}^{1:K})$ and (right) particle set $(\tilde x_{1:T}^{1:K}, \tilde a_{1:T - 1}^{1:K})$. Lines indicate ancestor indices.[]{data-label="fig:bias-proofs/particle_sets"}](particle_set_2.pdf "fig:"){width="\textwidth"}
The weights $\bar w_\tau^\kappa$ and $\tilde w_\tau^\kappa$ for the respective particle sets are identical except when $(\tau, \kappa) = (t, k)$ where $$\begin{aligned}
\bar w_t^k = w_t({x'}_{1:t}), \\
\tilde w_t^k = w_t(x_{1:t}).
\end{aligned}$$ Since $\hat Z(\bar x_{1:T}^{1:K}, \bar a_{1:T - 1}^{1:K}) = \hat Z(\tilde x_{1:T}^{1:K}, \tilde a_{1:T - 1}^{1:K})$, we have $w_t({x'}_{1:t}) = w_t(x_{1:t})$. As this holds for any arbitrary $t$ and $x_{1:t}$, it follows that $w_t(x_{1:t})$ must be constant with respect to $x_{1:t}$ for all $t = 1, \dotsc, T$.
Now, for $x_{1:t}$, consider the implied proposal by rearranging and $$\begin{aligned}
q_1(x_1) &= \frac{\gamma_1(x_1)}{w_1} \\
q_t(x_t {\lvert}x_{1:t - 1}) &= \frac{\gamma_t(x_{1:t})}{\gamma_{t - 1}(x_{1:t - 1}) w_t} && \text{ for } t = 2, \dotsc, T,
\end{aligned}$$ where $w_t := w_t(x_{1:t})$ is constant from our previous results. For this to be a normalized density with respect to $x_t$, we must have $$\begin{aligned}
w_1 = \int \gamma_1(x_1) \,\mathrm dx_1 &= Z_1,
\end{aligned}$$ and for $t = 2, \dotsc, T$: $$\begin{aligned}
w_t &= \int \frac{\gamma_t(x_{1:t})}{\gamma_{t - 1}(x_{1:t - 1})} \,\mathrm dx_t \\
&= \frac{\int \gamma_t(x_{1:t}) \,\mathrm dx_t}{\gamma_{t - 1}(x_{1:t - 1})} \\
&= \frac{Z_t}{Z_{t - 1}} \cdot \frac{\int \pi_t(x_{1:t}) \,\mathrm dx_t}{\pi_{t - 1}(x_{1:t - 1})}.
\end{aligned}$$ Since $\int \pi_{t + 1}(x_{1:t + 1}) \,\mathrm dx_{t + 1}$ and $\pi_t(x_{1:t})$ are both normalized densities, we must have $\pi_t(x_{1:t}) = \int \pi_{t + 1}(x_{1:t + 1}) \,\mathrm dx_{t + 1}$ for all $t = 1, \dotsc, T - 1$ for all $x_{1:t}$. For a given $t \in \{1, \dotsc, T - 1\}$ and $x_{1:t}$, applying this repeatedly yields $$\begin{aligned}
\pi_t(x_{1:t}) = \int \pi_{t + 1}(x_{1:t + 1}) \,\mathrm dx_{t + 1}
= \int \int \pi_{t + 2}(x_{1:t + 2}) \,\mathrm dx_{t + 2} \,\mathrm dx_{t + 1}
= \cdots
= \int \pi_T(x_{1:T}) \,\mathrm dx_{t + 1:T}
\end{aligned}$$ such that each $\pi_t(x_{1:t})$ must be the corresponding marginal of the final target. We also have $$\begin{aligned}
w_1(x_1) &= Z_1, \\
w_t(x_{1:t}) &= \frac{Z_t}{Z_{t - 1}}, && t = 2, \dotsc, T, \\
q_1(x_1) &= \pi_1(x_1) = \pi_T(x_1), \\
q_t(x_t {\lvert}x_{1:t - 1}) &= \frac{\pi_t(x_{1:t})}{\pi_{t - 1}(x_{1:t - 1})} = \frac{\pi_T(x_{1:t})}{\pi_T(x_{1:t - 1})}, && t = 2, \dotsc, T.
\end{aligned}$$
($\impliedby$) To complete the proof, we now simply substitute identities in 1 and 2 of Proposal \[proposition:aesmc/bias/elbo\_smc\] back to the expression of $\hat Z(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K})$ to obtain $\hat Z(x_{1:T}^{1:K}, a_{1:T - 1}^{1:K}) = Z$.
Experiments {#experiments}
===========
VRNN {#sec:appendix_vrnn}
----
In the following we give the details of our architecture. The generative model is given by:
$$p(x_{1:T}, h_{0:T}, y_{1:T}) = p(h_0) \prod_t p(x_t|h_{t-1})p(y_t|h_{t-1}, x_t)p(h_t|h_{t-1},x_t,y_t)$$
where $$\begin{split}
p(h_0) & = \mathrm{Normal}(h_0; 0, I)\\
p(x_t|h_{t-1}) & = \mathrm{Normal}(x_t;\mu^x_\theta(h_{t-1}), {\sigma^x_\theta}(h_{t-1})^2)\\
p(y_t|h_{t-1}, x_t) & = \mathrm{Bernoulli}(y_t;\mu^y_\theta(\varphi^x_\theta(x_t), h_{t-1}))\\
p(h_t|h_{t-1},x_t,y_t) & = \delta_{f(h_{t-1},\varphi^x_\theta(x_t),\varphi^y_\theta(y_t))}(h_t)
\end{split}$$
and the proposal distribution is given by
$$p(x_t|y_t, h_{t-1}) = \mathrm{Normal}(x_t;\mu^p_\phi(\varphi^y_\phi(y_t), h_{t-1}), {\sigma^p_\phi}^2(\varphi^y_\phi(y_t), h_{t-1}))$$
The functions $\mu^x_\theta$ and $\sigma^x_\theta$ are computed by networks with two fully connected layers of size 128 whose first layer is shared. $\varphi^x_\theta$ is one fully connected layer of size 128.
For visual input, the encoding $\varphi^y_\theta$ is a convolutional network with conv-4x4-2-1-32, conv-4x4-2-1-64, conv-4x4-2-1-128 where conv-wxh-s-p-n denotes a convolutional network with $n$ filters of size $w\times h$, stride $s$, padding $p$. Between convolutions we use leaky ReLUs with slope 0.2 as nonlinearity and batch norms. The decoding $\mu^y_\theta$ uses transposed convolutions of the same dimensions but in reversed order, however with stride $s=1$ and padding $p=0$ for the first layer.
A Gated Recurrent Unit (GRU) is used as RNN and if not stated otherwise ReLUs are used in between fully connected layers.
For the proposal distribution, the functions $\mu^p_\phi$ and $\sigma^p_\phi$ are neural networks with three fully connected layers of size 128 that are sharing the first two layers. Sigmoid and softplus functions are used where values in $(0,1)$ or $\mathbb{R}^{+}$ are required. We use a minibatch size of 25.
For the moving agents dataset we use with a learning rate of $10^{-3}$.
A specific feature of the architecture is that the proposal and the generative model share the component $\varphi^y_{\phi,\theta}$. Consequently, we set $\phi=\theta$ for the parameters belonging to this module and train it using gradients for both $\theta$ and $\phi$.
Moving Agents {#sec:appendix_moving_agents}
-------------
In Figure \[fig:experiments/max/visualisation\] we investigate the quality of the generative model by comparing visual predictions. We do so for models learned by *(top)* and *(bottom)*. The models were learned using ten particles but for easier visualization we only predict using five particles.
The first row in each graphic shows the ground truth. The second row shows the averaged predictions of all five particles. The next five rows show the predictions made by each particle individually.
The observations (i.e. the top row) up to $t=19$ are shown to the model. Up to this timestep the latent values $x_{0:19}$ are drawn from the proposal distribution $q(x_t|y_t, h_{t-1})$. From $t=20$ onwards the latent values $x_{20:37}$ are drawn from the generative model $p(x_t|x_{t-1})$. Consequently, the model predicts the partially occluded, stochastic movement over 17 timesteps into the future.
We note that most particles predict a viable future trajectory. However, the model learned by is not as consistent in the quality of its predictions, often ’forgetting’ the particle. This does not happen in every predicted sequence but the behavior shown here is very typical. Models learned by are much more consistent in the quality of their predictions.
![Visualisation of the learned model. Ground truth observations (top row in each sub figure) are only revealed to the algorithm up until t=19 inclusive. The second row shows the prediction averaged over all particles, all following rows show the prediction made by a single particle. *(Top)* . *(Bottom)* .[]{data-label="fig:experiments/max/visualisation"}](agents_visual_is.png){width=".95\linewidth"}
![Visualisation of the learned model. Ground truth observations (top row in each sub figure) are only revealed to the algorithm up until t=19 inclusive. The second row shows the prediction averaged over all particles, all following rows show the prediction made by a single particle. *(Top)* . *(Bottom)* .[]{data-label="fig:experiments/max/visualisation"}](agents_visual_smc.png){width=".95\linewidth"}
Optimizing Only Proposal Parameters {#sec:additional-experiments}
-----------------------------------
[.49]{} ![*(Left)* Optimizing $\operatorname{\acrshort{ELBO}}$ with respect to $\phi$ for . *(Right)* The lengths of the squares are proportional (with a constant factor) to $\sqrt{\sum_{t = 1}^T (\mu_t^{\text{kalman}} - \mu_t^{\text{approx}})^2}$ which is a proxy for inference quality of $\phi$ described in the main text. The larger the square, the worse the inference.[]{data-label="fig:additional_experiments/lgssm/inference_test_and_elbos"}](elbos.pdf "fig:"){width="\textwidth"}
[.49]{} ![*(Left)* Optimizing $\operatorname{\acrshort{ELBO}}$ with respect to $\phi$ for . *(Right)* The lengths of the squares are proportional (with a constant factor) to $\sqrt{\sum_{t = 1}^T (\mu_t^{\text{kalman}} - \mu_t^{\text{approx}})^2}$ which is a proxy for inference quality of $\phi$ described in the main text. The larger the square, the worse the inference.[]{data-label="fig:additional_experiments/lgssm/inference_test_and_elbos"}](inference_2.pdf "fig:"){width="\textwidth"}
We have run experiments where we optimize various $\operatorname{\acrshort{ELBO}}$ objectives with respect to $\phi$ with $\theta$ fixed in order to see how various objectives have an effect on proposal learning. In particular, we train $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ and $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ with number of particles $K \in \{10, 100, 1000\}$. Once the training is done, we use the trained proposal network to perform inference using both and with number of particles $K_{\text{test}} \in \{10, 100, 1000\}$.
In Figure \[fig:additional\_experiments/lgssm/inference\_test\_and\_elbos\], we see experimental results for the described in Section \[sec:experiments/lgssm\]. We measure the quality of the inference network using a proxy $\sqrt{\sum_{t = 1}^T (\mu_t^{\text{kalman}} - \mu_t^{\text{approx}})^2}$ where $\mu_t^{\text{kalman}}$ is the true marginal mean $\operatorname{\mathbb{E}}_{p(x_{1:T} {\lvert}y_{1:T})}[x_t]$ obtained from the Kalman smoothing algorithm and $\mu_t^{\text{approx}} = \left(\sum_{k = 1}^K w_T^k x_t\right) / \left(\sum_{k = 1}^K w_T^k\right) $ is an approximate marginal mean obtained from the proposal parameterized by $\phi$.
We see that if we train using $\operatorname{\acrshort{ELBO}}_{\text{SMC}}$ with $K_{\text{train}} = 1000$, the performance for inference using (with whichever $K_{\text{test}} \in \{10, 100, 1000\}$) is worse than if we train with $\operatorname{\acrshort{ELBO}}_{\text{IS}}$ with any number of particles $K_{\text{train}} \in \{10, 100, 1000\}$. Examining the other axes of variation:
- Increasing $K_{\text{test}}$ (moving up in Figure \[fig:additional\_experiments/lgssm/inference\_test\_and\_elbos\] (Right)) improves inference.
- Increasing $K_{\text{train}}$ (moving to the right in Figure \[fig:additional\_experiments/lgssm/inference\_test\_and\_elbos\] (Right)) worsens inference.
- Among different possible combinations of (training algorithm, testing algorithm), (, ) $\succ$ (, ) $\succ$ (, ) $\succ$ (, ), where we use “$a \succ b$” to denote that the combination $a$ results in better inference than combination $b$.
[^1]: This work builds upon an earlier preprint [@le2017auto] along with the independent, simultaneously developed, closely related, work of @maddison2017filtering and @naesseth2017variational.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Recently we have shown that there are crucial similarities in the physics of cosmic microwave background radiation (CMBR) anisotropies and the flow anisotropies in relativistic heavy-ion collision experiments (RHICE). We also argued that, following CMBR anisotropy analysis, a plot of root-mean square values of the flow coefficients, calculated in a lab fixed frame for RHICE, can yield important information about the nature of initial state anisotropies and their evolution. Here we demonstrate the strength of this technique by showing that elliptic flow for non-central collisions can be directly determined from such a plot without any need for the determination of event-plane.'
author:
- 'Ananta P. Mishra'
- 'Ranjita K. Mohapatra'
- 'P. S. Saumia'
- 'Ajit M. Srivastava'
title: 'Using CMBR analysis tools for flow anisotropies in relativistic heavy-ion collisions'
---
Introduction
============
Relativistic heavy-ion collision experiments (RHICE) are often termed as [*Little Bangs*]{} in analogy of the [*Big Bang*]{} representing the initial stage of the Universe. Indeed, there are tempting similarities between the early universe and these experiments. For example, it is often mentioned that the surface of last scattering for the cosmic microwave background radiation (CMBR) is similar to the freezeout surface in RHICE in the sense that one has to learn about the early stages of the system from hadrons coming out from the freezeout surface, just as CMBR encodes information about the early universe. However, these correspondences and analogies have been primarily made from a motivational view point. It has recently been shown by us [@cmbhic] that the physics of RHICE may have deeper connections, not just with the above mentioned aspects of CMBR, but even to the most celebrated aspect of inflationary physics of the universe namely the presence of superhorizon fluctuations.
Following the techniques used for CMBR anisotropy analysis, it was proposed in ref.[@cmbhic] that instead of focusing on the average values of the flow coefficients $v_n$, one should calculate root-mean square values of the flow coefficients $v_n^{rms}$. Further, these calculations should be performed in a lab fixed frame, which eliminates the difficulties associated with determination of event plane for conventional elliptic flow analysis for non-central collisions. To distinguish from the conventional flow coefficients $v_n$ which are defined with respect to the event plane, we will denote, in this paper, the flow coefficients defined with respect to the lab fixed frame as ${\tilde v}_n$. The root mean square value is then denoted as ${\tilde v}_n^{rms}$. Plots of ${\tilde v}_n^{rms}$ (denoted as $v_n^{rms}$ in [@cmbhic]) using model estimates based on initial energy density anisotropies from HIJING [@hijing] were given there for large range of values of $n$ (with $n$ ranging from 1 to 30) for central collisions and it was argued that such plots can yield important information about the nature of initial state anisotropies and their evolution for RHICE.
In this paper we demonstrate the strength of this analysis technique by showing that a plot of values of ${\tilde v}_n^{rms}$ vs. $n$ can be used for directly probing various flow coefficients, in particular, the elliptic flow for non-central collisions [@flow0], without any need for the determination of event-plane. For non-central collisions, a plot of ${\tilde v}_n^{rms}$ (using model estimates from the spatial distribution of initial parton energy density), shows a prominent peak at $n = 2$, as we will see below. Further, the height of the peak increases with increasing value of the impact parameter $b$, thus allowing direct determination of elliptic flow (in terms of ${\tilde v}_2^{rms}$). This can be very useful for analyzing huge wealth of data as the determination of event plane for each event becomes redundant.
The correspondence between the CMBR physics and RHICE explored at a deeper level as discussed in [@cmbhic] may lead to new set of interesting phenomena and new techniques for RHICE. We mention that such a connection between physics of RHICE and that of inflationary universe was never anticipated earlier, and indeed, at first sight, it looks surprising that a concept like superhorizon fluctuation which arises from highly non-trivial, superluminal expansion phase of the very early universe could have any relevance for relativistic heavy-ion collision experiments in laboratory. Such superhorizon fluctuations in RHICE originate from the fact that in the center of mass frame the thermalization happens rather quickly, within about 1 fm. This time is too short for any fluctuations (inhomogeneities) in the transverse direction to disappear, by processes of homogenization, which have wavelengths larger than 1 fm. Initial parton energy density distribution from HIJING show that transverse fluctuations with wavelengths significantly larger than 1 fm are necessarily present at the time 1 fm even in central collisions. These arise from localization of partons inside initial nucleons, as well as from the fluctuations in nucleon coordinates. Analyzing the development of flow from such initial large wavelength fluctuations, we had argued that the anisotropies in the final particle momenta may show characteristic features of acoustic oscillations of subhorizon modes as well as suppression of modes which remain superhorizon at the freezeout stage. We will show below that such features do not mask the peak at $n = 2$ in non-central collisions.
The paper is organized in the following manner. In section II, we recall the adoption of CMBR anisotropy analysis method for the case of RHICE. Here we also show that calculations of ${\tilde v}_n^{rms}$ in the lab fixed frame can directly yield information about various flow coefficients defined with respect to the event plane. In particular, the conventional elliptic flow can be determined in this manner. Section III presents the results of the application of these analysis methods for the determination of ${\tilde v}_n^{rms}$ for the case of non-central collisions. In section IV we discuss the $p_T$ dependence, and in section V we discuss the issue of anisotropic detector acceptance. Section VI presents conclusions and discussions.
Using CMBR analysis tools for RHICE
===================================
Traditional analysis of elliptic flow in RHICE [@flow0] is based on measuring the average value of the 2nd Fourier coefficient of particle transverse momentum anisotropy. The average value is determined by carrying out the Fourier expansion in the event-plane reference frame. Even for the so-called ’central collisions’ one determines event plane for very small impact parameter collisions and then carries out the traditional analysis of determination of average flow [@cntrl]. However, determination of event plane is highly non-trivial making determination of higher flow coefficients very difficult. Indeed, flow coefficients have been determined only upto $v_6$.
It is here that the use of techniques of CMBR analysis can make a significant difference. For CMBR, the temperature anisotropies are analyzed using spherical harmonics, as appropriate for the surface of 2-sphere (the CMBR sky) [@cmbr]. The coefficients of the expansion (denoted as $a_{lm}$ corresponding to the spherical harmonic $Y_{lm}$) are degenerate in the argument $m$ and when averaged over different values of $m$ yield zero averages due to isotropy of the universe (when suitably corrected for local velocities etc.). What one plots is the variance of $a_{lm}$ denoted by $C_l$ and this leads to the celebrated power spectrum of CMBR anisotropies [@cmbr].
We propose using the same technique [@cmbhic] for analyzing particle momentum anisotropies, using lab fixed frame, in RHICE to probe the generation and evolution of flow. For RHICE, focusing on central rapidity region, one will be analyzing momentum anisotropies on a circle, requiring use of the Fourier coefficients, which we denote as ${\tilde v}_n$ to distinguish from the conventional flow coefficients $v_n$ which are defined with respect to the event plane. With a fixed lab frame, the event average values of these ${\tilde v}_n$s will all be zero due to rotational symmetry. We then propose to use the variance of ${\tilde v}_n$, i.e. ${\tilde v}_n^{rms}$ in analogy with $C_l$ for CMBR.
We first establish the relation between ${\tilde v}_n^{rms}$ calculated in the lab fixed frame and the conventional flow coefficients $v_n$ which are defined with respect to the event plane as follows
$$\rho(\phi^\prime) = \sum_{n=-\infty}^\infty v_n e^{in\phi^\prime}$$
Where the angle $\phi^\prime$ is measured with respect to the event plane. Assuming that the event plane is oriented by angle $\psi$ with respect to the lab fixed frame, we can write the above equation in the lab fixed frame as,
$$\rho_\psi(\phi) = \sum_{n=-\infty}^\infty v_n e^{in(\phi-\psi)}$$
where now the angle $\phi$ is measured with respect to the lab fixed frame. The subscript for $\rho_\psi(\phi)$ remind us that the above Fourier series expansion uses the flow coefficients defined with respect to the event plane which is oriented at an angle $\psi$ in the lab frame. Without the knowledge of $\psi$, we will calculate the lab fixed flow coefficients ${\tilde v}_n(\psi)$ as
$${\tilde v}_n(\psi) = \frac{1}{2\pi} \int_{-\pi}^{\pi}
\rho_\psi(\phi) e^{-in\phi} d\phi$$
Here the argument $\psi$ in ${\tilde v}_n(\psi)$ is used to remind that these quantities are calculated in the lab fixed frame, and are different from the values of conventional flow coefficients $v_n$ defined in Eq.(1). The event average value $\overline {{\tilde v}_n}$ of these flow coefficients ${\tilde v}_n$, is obtained by averaging ${\tilde v}_n$ over a large number of events (of similar type, e.g. centrality etc.). $\overline {{\tilde v}_n}$ is determined by the following expression
$$\overline {{\tilde v}_n} = \frac{1}{2\pi} \int_{-\pi}^\pi (\frac{1}{2\pi}
\int_{-\pi}^\pi \rho_\psi(\phi) e^{-in\phi} d\phi ) d\psi$$
Integration over $\psi$ reflects the random variation of the orientation of the event plane in the lab fixed frame from event to event. By changing the order of integration for $d\phi$ and $d\psi$, and using Eq.(2), we immediately see that the $\overline {{\tilde v}_n} = 0$ for all $n > 0$ as mentioned above.
We next calculate the event averaged values of $|{\tilde v}_n(\psi)|^2$ as
$$({\tilde v}_n^{rms})^2 = \overline {{\tilde v}_n^2} = \frac{1}{2\pi} \int_{-\pi}^\pi d\psi
{\tilde v_n(\psi)} {\tilde v_n^*(\psi)}$$
Straightforward calculation using Eqs.(2)-(5) gives
$${\tilde v}_n^{rms} = |v_n|$$
This is an important result. It shows that various conventional flow coefficients $v_n$, defined with respect to the event plane, can be directly calculated in the lab fixed frame in terms of ${\tilde v}_n^{rms}$. However, one has to be careful in interpreting the above equation. The above equation has been derived for the event averaged values. Thus the right hand side $|v_n|^2$ also represents the event averaged value of the conventional flow coefficients. Even if we assume that each event is almost identical in terms of centrality selection etc. (which it is not), still there will be random fluctuations in the fluid especially due to initial state fluctuations [@cmbhic]. If the contributions of such fluctuations becomes dominant then the right hand side of the above equation will not have a neat interpretation of relating to $|v_n|$ defined with respect to the event plane.
Fortunately this problem does not exist for the elliptic flow $v_2$. Below we will see, with model calculations of ${\tilde v}_2^{rms}$ for non-central events from HIJING, that the contribution arising from the elliptic shape of the produced partons in the overlap region is very large compared to the contribution from the random, initial state fluctuations. This will be seen as a prominent peak in the plot of ${\tilde v}_n^{rms}$ vs. $n$ at $n = 2$. Thus, as far as determination of the elliptic flow is concerned, Eq.(6) shows that calculation of ${\tilde v}_2^{rms}$ in the lab fixed frame can be used to determine the value of $|v_2|$. Neglecting the contributions of random fluctuations, symmetry of the elliptical shape of the overlap region for non-central collisions will then imply that $|v_2|$ is the same as $v_2$ conventionally defined as the coefficient of the $cos2\phi$ term.
For other values of $n$ the Eq.(6) has to be suitably interpreted with proper account of event by event random fluctuations. As we have emphasized, values of ${\tilde v}_n^{rms}$ determined by averaging over events should be directly used to probe the statistical properties of fluctuations and anisotropies of the initial plasma region (as is done for CMBR). Hydrodynamical simulations, with proper incorporation of such fluctuations can then be used to predict the values of ${\tilde v}_n^{rms}$ whose comparison with data can then be used to constrain/determine various physical inputs such as equation of state, viscosity etc.
We now describe our method for calculating ${\tilde v}_n^{rms}$ using HIJING [@hijingp]. We first recall from ref. [@cmbhic] the estimates of initial spatial anisotropies. We start with the initial transverse energy density distribution for Au-Au collision at 200 GeV/A center of mass energy from HIJING. For parton positions we use random locations inside the parent nucleon, (similarly, for partons produced from the string systems, random position are used along the line joining the two corresponding nucleons). The transverse energy density at a given transverse position ${\vec x}$, at proper time $\tau = \tau_{eq}$, is taken as [@hijing],
$$\epsilon_{tr}({\vec x},\tau_{eq}) = {1 \over \Delta A}
\sum_i E_{tr}^i ~F(\tau_{eq},p_{tr})~
\delta^2({\vec x} - {\vec x}^i_0 - {\vec v}^i \tau_{eq}) ~\Delta(y^i)$$
where ${\vec x}^i_0$ denotes the initial transverse coordinates of the $i_{th}$ parton (determined using the coordinates of the parent nucleon in HIJING as discussed above), $E^i_{tr}$ is its transverse energy, $p_{tr}$ the transverse momentum, and ${\vec v^i}$ is its transverse velocity. For the rapidity window we take $\Delta(y^i) = 1$ centered at $y = 0$, [@hijing]. The sum over $i$ includes all partons in a small transverse area element $\Delta A (\simeq 0.5$ fm$^2$) at position ${\vec x}$. We have included a factor $F(\tau_{eq},p_{tr}) \equiv 1/(1 +
1/(p_{tr} \tau_{eq})^2)$ to account for the probability of formation of partons with zero rapidity [@hijing].
We assume that the hydrodynamic description becomes applicable by $\tau =\tau_{eq}$, which we take to be 1 fm and calculate the anisotropies in the fluctuations in the spatial extent $R(\phi)$ at this stage, where $R(\phi)$ represents $\epsilon_{tr}$ weighted average of the transverse radial coordinate in the angular bin at azimuthal coordinate $\phi$ [@cmbhic]. As emphasized above, angle $\phi$ is taken in a lab fixed coordinate frame. We divide the region in 50 - 100 bins of azimuthal angle $\phi$, and calculate the Fourier coefficients of the anisotropies in ${\delta R}/R
\equiv (R(\phi) - {\bar R})/{\bar R}$ where $\bar R$ is the angular average of $R(\phi)$. Note that in this way we are representing all fluctuations essentially in terms of fluctuations in the boundary of the initial region. We use $F_n$ to denote Fourier coefficients for these spatial anisotropies, and use ${\tilde v}_n$ to denote $n_{th}$ Fourier coefficient of expected momentum anisotropy in ${\delta p}/p$ defined in the lab frame. Here $\delta p$ represents fluctuation in the momentum $p$ of the final particles from the average momentum, in a given azimuthal angle bin.
We should clarify that in the conventional analysis, $v_2$ is directly related to the elliptical shape of the fireball, and normally one does not call it a fluctuation. However, in our language, the elliptic shape of the fireball itself is taken to represent a fluctuation (of large wavelength) from isotropic case. As our analysis is carried out in the Lab fixed frame, all shape fluctuations (including the elliptical ones) are treated uniformly by calculating different Fourier coefficients $F_n$.
Again we emphasize the most important difference between the conventional discussions of the elliptic flow and our analysis; here one does not try to determine any special reaction plane on event-by-event basis. A fixed coordinate system is used for calculating azimuthal anisotropies. This is why, as discussed in Sect.II and as is seen in the plots in ref.[@cmbhic], averages of $F_n$s (and hence of ${\tilde v}_n$s) vanish when large number of events are included in the analysis. However, the root mean square values of $F_n$s, and hence of ${\tilde v}_n$s, are non-zero in general and contain non-trivial information. In fact, it is the same as the standard deviation for the distribution of $F_n$s since the average value of $F_n$s is zero. This is what is exactly done for the CMBR case also [@cmbr].
One important difference from CMBR, which may be of crucial importance in finding any non-trivial features in the plot of ${\tilde v}_n^{rms}$ for RHICE, is the fact that for the CMBR case, for each $l$ mode of the spherical harmonic, there are only $2l+1$ independent measurements available, as there is only one CMBR sky to observe. For small $l$ values this becomes dominant source of uncertainty, leading to the accuracy limited by the so called cosmic variance [@cmbr]. In contrast, for RHICE, each nucleus-nucleus collision (with same parameters like collision energy, centrality etc.) provides a new sample event. Also, in the absence of any special reflection symmetry here (which was present in the traditional elliptic flow analysis) all flow coefficients give non-zero contributions to ${\tilde v}_n^{rms}$, and the $``sin"$ terms give same values as the $``cos"$ terms. In the plots of ${\tilde v}_n^{rms}$ below, we show the sum of these two contributions, i.e. square root of the sum of the squares of the $``sin"$ term and the $``cos"$ term
Results for non-central collisions
==================================
We have generated events using HIJING and we present sample results for Au-Au collision at 200 GeV/A center of mass energy. In all the plots, the averages are taken over 10000 events, and the root mean square values ${\tilde v}_n^{rms}$ of the flow Fourier coefficients are obtained from spatial $F_n$s simply by using proportionality factor of 0.2. The relation between the Fourier coefficients of the spatial anisotropy and resulting momentum anisotropy in our model can only be obtained using a full hydrodynamical simulation, with proper accounts of any surface tension, as well as factors such as horizon crossing etc. to properly account for the physics discussed here. In the absence of such a simulation, we make a strong assumption here that all Fourier coefficients for momentum anisotropy are related to the corresponding coefficients for spatial anisotropy by roughly the same proportionality factor, which we take to be 0.2 for definiteness. (As mentioned in ref.[@cmbhic], this choice also gave reasonably good agreement with the results for ${\tilde v}_2^{rms}$ in the literature for (almost) central events [@cntrl].) Although overall shape of the plot of ${\tilde v}_n^{rms}$ will crucially depend on the values of these proportionality constants, and hence may change completely from what is shown here, the qualitative features such as the presence of peaks may remain unaffected. (Though, the peak positions may shift depending on these proportionality constants.) $F_n$s are calculated directly from the parton energy distribution at $t = 1$ fm for each event, within unit central rapidity window.
Fig.1 presents plots of ${\tilde v}_n^{rms}$ for different values of impact parameters. Solid, dashed, and dotted plots correspond to impact parameter $b$ = 0, 5 fm, and 8 fm, respectively with the spread in $b$ taken to be 1 fm for non-zero $b$ cases. The solid plot for central collisions shows monotonically decreasing values of ${\tilde v}_n^{rms}$. We emphasize that even here the plot has non-trivial structure compared to the case when partons are randomly distributed inside the nuclear volume with uniform probability, which yields a flat plot for corresponding ${\tilde v}_n^{rms}$. (Though, the rise of $b = 0$ plot for small $n$ does not appear prominent here due to larger vertical scale of the plot.) Errors in these plots are very small.
Important thing to note is the prominent peak in non-central collisions for $b$ = 5, and 8 fm at $n = 2$. The peak height is significantly larger for $b$ = 8 fm. Important thing to realize is that this feature (which within our model simply represents elliptical distribution of partons in position space at the initial time) is present even when the ${\tilde v}_n^{rms}$ have been calculated in a lab fixed reference frame. Average values of ${\tilde v}_n$ for all values of impact parameters, calculated for large number of events, continue to be zero as shown in ref. [@cmbhic] for the central collision case. This is expected from rotational symmetry in lab fixed frame. Larger overall values of ${\tilde v}_n^{rms}$ for larger values of $b$ are simply due to smaller number of initial partons leading to larger fluctuations. We mention that the solid plot in Fig.1 is given exactly for $b = 0$ to show the difference between zero and non-zero $b$ cases. However, with experimental data, $b$ = 0 will inevitably include events with small but non-zero values of $b$. Thus even for $b = 0$ case one should expect to observe a (small) peak at $n = 2$.
-1.5in -1.5in
We have also calculated the root mean square values of the Fourier coefficients of the momentum anisotropy directly using the momenta of final particles from HIJING. These plots are all similar to each other for all values of the impact parameter $b$ and hence similar to the central case $b = 0$ (apart from larger statistical fluctuations for larger values of $b$ due to smaller number of partons as discussed above for Fig.1.) This is expected because, despite initial spatial anisotropy of parton production for non-central collisions, momentum distribution remains isotropic in the absence of any hydrodynamic evolution. We do not show these momentum plots here because such plots have been given in ref.[@cmbhic] for the central collision case.
Fig.1 shows the strength of this technique of plotting root-mean-square values of flow coefficients in lab fixed coordinate frame as compared to the traditional analysis of elliptic flow. This technique is simple to implement (apart from the issue of non-flow correlations [@nonflow; @accpt1], which one has to resolve in any case), allowing larger statistics to be generated. What one needs is predictions from hydrodynamics about ${\tilde v}_n^{rms}$ calculated in lab fixed frame and then compare with data analyzed similarly. As the plots in Fig.1 show, the values of ${\tilde v}_n^{rms}$ even for very large values of $n$ continue to be significantly non-zero and with the possibility of sufficient statistics, one may be able to determine any important features in these plots.
In [@cmbhic] we had argued that certain important aspects of inflationary density fluctuations, such as suppression of superhorizon modes and acoustic oscillations may also be present in these plots of ${\tilde v}_n^{rms}$ in RHICE. Basic ideas can be stated as follows. Acoustic peaks in CMBR primarily result from the coherence and acoustic oscillations of the inflationary density fluctuations with coherence originating from the superhorizon nature of fluctuations, essentially freezing the fluctuations on superhorizon scales. This should be reasonably true for RHICE as the transverse velocity to begin with is expected to be zero and becomes non-zero only due to pressure gradients (which does not become fully effective on superhorizon scales). The oscillatory behavior for the fluctuations in the universe results from attractive forces of gravity and counter balancing forces from radiation pressure (with the coupling of baryons to the radiation). The oscillatory behavior of the fluctuations in RHICE is expected due to flow development with non-zero pressure gradients, at least for sufficiently small wavelength modes. These arguments suggest that some features like acoustic peaks may be present in the plots of ${\tilde v}_n^{rms}$ for RHICE also. Further, modes with wavelengths larger than the acoustic horizon size $H^{fr}_s$ at freezeout should be naturally suppressed as pressure gradients do not become fully effective on those scales to generate full flow anisotropy before freezeout occurs. This leads to the following suppression factor
$$({\tilde v}_n)_{observed} = {2H^{fr}_s \over \lambda} ({\tilde v}_n)_{max}$$
where $\lambda \sim 2\pi {\bar R}^{fr}/n, ~~~(n \ge 1)$, is the measure of wavelength of the anisotropy corresponding to the $n_{th}$ Fourier coefficient. Here $\bar R^{fr}$ represents the transverse radius at the freezeout time $\tau_{fr}$. Using the rough estimate of the rate of change of the transverse velocity to be about 0.1 fm$^{-1}$ at the early stages at these energies [@vtr], we can estimate $\bar R^{fr} \simeq {\bar R} + 0.05 (\tau_{fr} -
\tau_{eq})^2 = {\bar R} (1 + 0.05 {\bar R}/c_s^2)$. Here ${\bar R} \equiv
{\bar R}(\tau_{eq}) = c_s (\tau_{fr} - \tau_{eq})$. The largest wavelength $\lambda_{max}$ of spatial anisotropy which will have chance to develop to its maximum hydrodynamic value is, therefore, $\lambda_{max} \simeq 2 H^{fr}_s = 2 c_s (\tau_{fr} - \tau_{eq})
= 2 {\bar R}(\tau_{eq})$. This gives us the corresponding minimum value $n_{min}$ of $n$ below which flow coefficients should show suppression due to being superhorizon,
$$n_{min} = \pi (1 + {0.05 {\bar R}(\tau_{eq}) \over c_s^2} )$$
The possibilities of the superhorizon suppression and acoustic oscillations lead to new peak structures in the plot of ${\tilde v}_n^{rms}$ as was shown in [@cmbhic] for central collisions. It is therefore important to know whether the peak in ${\tilde v}_n^{rms}$ due to elliptic flow for non-central collisions (as in Fig.1) survives if these new features are also present, especially with the suppression of superhorizon modes at smaller $n$. For this purpose, we show in Fig.2, plots (for different values of impact parameter $b$ as in Fig.1), when superhorizon suppression is included and when acoustic oscillations are also included in modeling the values of ${\tilde v}_n^{rms}$. It is clearly seen that the peak at $n = 2$ corresponding to the elliptic flow is present in all the plots and remains most prominent for $b = 8$ fm.
Here we mention that recently Sorensen has proposed [@srnsn] a novel explanation of azimuthal correlations observed at the Relativistic Heavy-ion Collider (RHIC) at BNL in terms of the suppression of the superhorizon fluctuations discussed in [@cmbhic]. One of the plots (Fig.4) in [@srnsn] (where suitable subtraction of $v_2$ has been made) shows possible suppression of values of the variance for small $n$, as predicted in [@cmbhic].
-0.4in -0.3in
We mention that we calculate ${\tilde v}_n^{rms}$ by direct calculation of variances of the distributions of ${\tilde v}_n$ (actually $F_n$s) in the laboratory fixed frame. Equivalently one can calculate ${\tilde v}_n^{rms}$ by calculating the two-particle azimuthal correlation functions, as is done for the elliptic flow calculations. To extract elliptic flow coefficient from two-particle azimuthal correlations one has to separate non-flow contributions [@nonflow; @accpt1]. Two particle azimuthal correlations, which are experimentally measured, contain contributions from non-flow effects such as jets, resonance decays, HBT correlations, final state interactions etc. Various methods have been discussed to separate out the non-flow contributions to the azimuthal correlations [@nonflow; @accpt1]. Our estimates of ${\tilde v}_n^{rms}$ will also contain such non-flow contributions which have to be properly accounted for.
$p_T$ dependence
================
In the conventional analysis of elliptic flow, $v_2$ has a non-trivial dependence on $p_T$ of hadrons [@v2pt]. $v_2$ increases with $p_T$, saturates, and then decreases. Initial increase of $v_2$ with $p_T$ is easily determined [@sv; @oltr; @oltr1]. Large $p_T$ behavior of flow is governed by viscous effects, parton energy loss, as well as by parton coalescence [@v2pt]. The same physics will be applicable in our model as well. For simplicity, we will follow the approach in ref. [@oltr] where the linear increase of the elliptic flow $v_2$ with $p_T$ is determined to be
$$v_2 = {\alpha \over T}(p_T - v m_T)$$
where $\alpha$ characterizes the magnitude of elliptic flow, $v$ is the flow velocity (see below) and $m_T = \sqrt{p_T^2 + m^2}$ is the transverse mass. We mention that this linear dependence of $v_2$ on $p_T$ is expected for the intermediate range of $p_T$. The behavior for small $p_T$ may be more non-trivial [@sv]. However, recall from Eq.(6) (and discussion in Sect.II) that in the context of our model, ${\tilde v}_2^{rms}$ is simply given by the conventional $v_2$, hence whatever be the dependence of $v_2$ on $p_T$, ${\tilde v}_2^{rms}$ should show the same dependence. Still, it is useful to directly check the expected $p_T$ dependence of ${\tilde v}_2^{rms}$ in our model by adopting the analysis of $v_2$ in ref. [@oltr] for our case. This is for two reasons. First, the relevance of random fluctuations needs to be assessed at different stages of the calculations to see the effects of event averaging and to see whether similar statements can be made about other flow coefficients. (For example, see the discussion following Eq.(12) below about the form of the fluid 4-velocity.) Secondly, the analysis for $v_2$ in ref. [@oltr] is carried out to linear order in $\alpha$. Since we calculate the root mean square values, we need to keep terms of order $\alpha^2$. Though, as we will see below, the $\alpha^2$ term will turn out to be only relevant for $v_4$ and will not affect the analysis of ${\tilde v}_2^{rms}$.
We start by noting that the expression for $v_2$ in Eq.(10) corresponds to the conventional analysis with the determination of event plane. In our approach, the analysis is carried out in the lab fixed frame. For each non-central event the magnitude of the flow anisotropy will still be characterized by (say) the above equation. However, the maximum flow direction will now vary randomly from one event to another because the event plane is left undetermined. Thus, it looks reasonable to expect that ${\tilde v}_2^{rms}$ should show the same dependence on $p_T$ as $v_2$ in Eq.(10), as is indeed shown by Eq.(6). We will check this directly by adopting the analysis of $v_2$ in ref. [@oltr] for our case for the estimation of ${\tilde v}_2^{rms}$ in the lab fixed frame (and keeping terms of order $\alpha^2$). Let us start with the transverse momentum distribution of emitted particles [@oltr]
$${dN \over p_T dp_T dp_z d\phi} \equiv \rho(\phi) \propto
exp({-m_T u^0(\phi) + p_T u(\phi) \over T})$$
Here, $u(\phi)$ is the space component of the fluid flow 4-velocity $u_\mu$ in $\phi$ direction, and $u^0(\phi)$ is the time component.
For non-central collisions, the anisotropic flow can be parameterized as
$$u(\phi) = u + 2\alpha cos 2(\phi - \psi)$$
Here $u$ is the angular average of maximum fluid 4-velocity, and $\alpha > 0$ characterizes the magnitude of elliptic flow. Note, that we have introduced here the angle $\psi$ characterizing the orientation of the event plane in the lab frame so that the angle $\phi$ is measured in the lab frame. Also, it is important to realize that here we are only characterizing elliptic flow anisotropy of the fluid, without worrying about the presence of general fluctuations. The justification for this is that, as can be seen from Fig.1, for non-central collisions in our method it is only ${\tilde v}_2^{rms}$ which has a significantly larger value than the values of ${\tilde v}_n^{rms}$ for other $n$. Thus, the presence of general fluctuations (as accounted for in Fig.1) will play an important role in determination of these other values of ${\tilde v}_n^{rms}$ even for non-central collisions. In contrast, for ${\tilde v}_2^{rms}$ the elliptic anisotropy resulting from non-central collision seems to play dominant role over the random fluctuations. Hence the parametrization of Eq.(12) may capture important aspects of the dependence of ${\tilde v}_2^{rms}$ on $p_T$ etc.
We are interested in calculating the root mean square value (instead of $v_2$ as in ref. [@oltr]). Thus, we expand $u^0(\phi) = \sqrt{u(\phi)^2 + 1}$ to second order in $\alpha$, and get (with $v = u/u^0$),
$$u^0(\phi) = u^0 + 2\alpha v cos2(\phi - \psi) + {2 \alpha^2 \over (u^0)^3}
cos^22(\phi - \psi)$$
Using this equation, and Eq.(11), we get particle transverse momentum distribution as
$$\rho_\psi(\phi) = A[1 + 2 \alpha cos2(\phi-\psi) {(p_T - v m_T) \over T}
+ \alpha^2 (1 + cos4(\phi - \psi)) ({(p_T - v m_T)^2 \over T^2}
- {m_T \over T (u^0)^3} ) ]$$
where $A$ is independent of $\phi, \psi$. We see that this equation is in the same form as Eq.(2). It is important to note here that the $\alpha^2$ term here only comes for $v_4$ (requiring higher powers of $\alpha$ for the calculation of ${\tilde v}_4^{rms}$) and $v_2$ remains linear in $\alpha$.
Following the discussion in Sect.II following Eq.(2), and using Eq.(6), we immediately see that
$${\tilde v}_2^{rms} = {\alpha \over T} (p_T - v m_T)$$
Which is the same dependence on $p_T$ as in Eq.(10) for the conventional analysis of $v_2$ using event plane determination.
We again emphasize that in our picture, the randomness in the orientation of the event plane w.r.t. the lab fixed frame is easy to model by introducing angle $\psi$. The other intrinsic sources of fluctuations will also contribute to ${\tilde v}_2^{rms}$ in our method, but as shown in Fig.1, their contributions appear to remain small (as seen by the plot for $b$ = 0). (Note again, overall larger scale of plots of ${\tilde v}_n^{rms}$ for larger values of $b$ is primarily due to smaller number of particles leading to larger statistical fluctuations.) As we had discussed earlier, ${\tilde v}_n^{rms}$ for $n \ne 2$ may be dominated by these intrinsic sources of fluctuations. If the contribution of such fluctuations remains sub dominant, then our results show that direct determination of $v_2$, in particular its $p_T$ dependence ([@sv]) is possible by measuring ${\tilde v}_2^{rms}$ in lab fixed frame completely avoiding the determination of event plane.
detector acceptance
===================
One of the main advantages of using the techniques proposed here is that it makes it much easier to extract the information about flow without any need for event plane determination. Thus large amount of data can be analyzed improving accuracy. It then becomes important to address the issue of being able to incorporate data with varying azimuthal acceptance, as has been emphasized for the conventional flow analysis. Several methods have been proposed for accounting for azimuthal dependence of detector acceptance in the conventional flow analysis [@accpt1; @accpt2]. These can be suitably adopted to the method proposed here. As mentioned above, due to fixed lab frame being used for analysis, the average values of all flow coefficients will be zero here due to azimuthal symmetry. Clearly, this will no longer be true in the presence of azimuthal anisotropy in the detector acceptance. For example, let $A(\phi)$ be the acceptance function characterizing the probability that a hadron is detected at azimuthal angle $\phi$, with $A(\phi)$ being normalized as [@accpt1; @accpt2]
$$\int^{\pi}_{-\pi} A(\phi) d\phi = 2\pi$$
In the lab fixed frame, the values of ${\tilde v}_n$ in our method may be calculated from Eq.(3) (Sect.II). Note that for anisotropic detector acceptance there is an implicit factor of $A(\phi)$ in the integrand in Eq.(3). One can then directly calculate the event average value $\overline {{\tilde v}_n}$ of these flow coefficients ${\tilde v}_n$, by averaging ${\tilde v}_n$ over a large number of events (of similar type, e.g. centrality etc.). The actual value of $\overline {{\tilde v}_n}$, for non-central collisions, is determined by the following expression
$$\overline {{\tilde v}_n} = \frac{1}{2\pi} \int_{-\pi}^\pi (\frac{1}{2\pi}
\int_{-\pi}^\pi A(\phi)
\rho_\psi(\phi) e^{-in\phi} d\phi ) d\psi$$
Here, $\psi$ denotes the orientation of the event plane in the lab fixed frame (see, Eq.(4) in Sect.II), which varies from event to event. We have made here explicit the acceptance function $A(\phi)$ in the above expression. (For simplicity we assume that $A(\phi)$ does not depend on rapidity, $p_T$ etc.). For ideal detector with uniform acceptance $A(\phi) = 1$ and by changing the order of $d\phi$ and $d\psi$ integrations one can easily see that $\overline {{\tilde v}_n} = 0$ as discussed in Sect.II (following Eq.(4)). However, for anisotropic acceptance, with $A(\phi)
\ne 1$, $\overline {{\tilde v}_n}$ need not be zero. In fact, apart from a normalization (relating to angular average of $\rho$), $\overline
{{\tilde v}_n}$ gives the Fourier coefficients for the series expansion of the function $A(\phi)$, thus allowing us to determine the acceptance function $A(\phi)$ directly from experimental data [@accpt1; @accpt2].
For the simple case when $A(\phi)$ is non-zero for every $\phi$, one can directly use $A(\phi)$, determined from $\overline {{\tilde v}_n}$ as explained above, to compensate for the effect of anisotropic detector acceptance by including additional factor of $1/A(\phi)$ in the integrand of Eq.(3) for calculation of ${\tilde v}_n$.
$${\tilde v}_n = \frac{1}{2\pi} \int_{-\pi}^\pi
{\rho_\psi(\phi) \over A(\phi)} e^{-in\phi} d\phi$$
With this modification, Eq.(17) shows that the event average value $\overline {{\tilde v}_n}$ will again be zero, just as for the case of isotropic detector acceptance. The above Eq.(18) can then be used to determine the values of ${\tilde v}_n^{rms}$ as discussed in earlier sections.
However, for more general case, e.g. for incomplete detector coverage, where $A(\phi) = 0$ for a range of $\phi$ the above simple method does not help. For such cases one can adopt the techniques discussed in [@accpt1; @accpt2] for the present case. Though, due to the important role played by the event plane determination, these techniques need to be suitably adopted for our case where a fixed lab frame is used for the analysis. In fact, it is more helpful to directly adopt the techniques used for CMBR case where one uses a fixed lab frame (say, galactic coordinates) for writing down temperature anisotropies in the sky in terms of spherical harmonics. For CMBR analysis also one needs to compensate for the effects of partial coverage of the sky. This happens either due to limited coverage of the sky by the detector, or due to galactic foreground as for WMAP [@cmbrprtl]. Adopting that approach for our case, we start with the experimentally determined flow coefficients ${\rm v}_n$ for each event (which, following [@cmbrprtl] we call [*pseudo*]{} flow coefficients).
$${\rm v_n} = \frac{1}{2\pi} \int_{-\pi}^\pi e^{-in\phi} A(\phi)
\rho_\psi(\phi) d\phi$$
where, now, $A(\phi)$ may be zero for a range of $\phi$ values. Using Fourier coefficients $a_m$ for the acceptance function $A(\phi)$, we get,
$${\rm v_n} = \frac{1}{2\pi} \int_{-\pi}^\pi \sum_m a_m e^{-i(n-m)\phi}
\rho_\psi(\phi) d\phi ~~ = \sum_l a_{(n-l)}~ {\tilde v}_l$$
where ${\tilde v}_l$ is the true flow coefficient (defined with respect to the lab fixed frame). The above equation shows the mode-mode coupling of flow coefficients resulting from incomplete detector coverage [@cmbrprtl; @accpt1; @accpt2]. For an approximate determination of ${\tilde v}_l$ we can use appropriate truncation of the above matrix equation to a finite set of linear equations and solve for ${\tilde v}_l$. This analysis can be done for each event, and with the values of ${\tilde v}_l$ thus determined one can directly get the estimate of the true variance ${\tilde v}_n^{rms}$ for a set of events.
Conclusions
===========
In summary we emphasize the important lessons from CMBR analysis techniques. We have argued that important information about initial anisotropies of the system and their evolution in relativistic heavy-ion collisions can be obtained by plotting the root mean square values of the Fourier coefficients ${\tilde v}_n^{rms}$ of the anisotropies in the fluctuations $\delta p/p$ of the particle momenta, calculated in a fixed laboratory frame, starting from $n = 1$ upto large values of $n \simeq 30$. (which corresponds to $\lambda \sim 1$ fm). Note that $n = 30$ almost corresponds to wavelength of fluctuation $\lambda$ at the surface of the region, at $\tau_{fr}$, being of order 1 fm. Fluctuations with wavelengths smaller than 1 fm presumably cannot be treated within hydrodynamical framework, so we restrict attention within this range of $n$. One will expect that beyond a critical value of $n$ the nature of the curve should change in some qualitative manner indicating breakdown of underlying hydrodynamical description for smaller modes. The wavelength corresponding to that critical value of $n$ will determine the smallest scale below which hydrodynamical description is not valid. The plot of ${\tilde v}_n^{rms}$ for $n$ larger than this critical value will probe fluctuations in parton density at even smaller length scales and may provide a bridge with the perturbative regime.
For non-central collisions our technique provides a direct method to probe elliptic flow by determining ${\tilde v}_2^{rms}$ in lab fixed frame. If random fluctuations do not dominate then our results show that ${\tilde v}_n^{rms}$ can directly probe $|v_n|$ for $n \ne 2$ also. This may be important for collisions of deformed nuclei. It is important to appreciate that a plot of ${\tilde v}_n^{rms}$ tells us about the statistical nature of fluctuations and anisotropies of the initial plasma region (just as in the CMBR case), hence it will always have valuable information irrespective of its shape (e.g. a flat curve). Hydrodynamical simulations, with proper incorporation of such fluctuations can be used to predict directly the values of ${\tilde v}_n^{rms}$ whose comparison with data will constrain/determine various physical inputs such as equation of state, viscosity etc. Further, in analogy with CMBR analysis, one can determine higher moments of momentum anisotropies (again in lab fixed frame, as we have discussed above) probing detailed nature of initial fluctuations, e.g. non-Gaussianity etc. We hope to present a more detailed analysis of such issues in a future work. We also intend to explore various aspects of [*horizon entering*]{} of density fluctuations which can be probed under laboratory conditions using RHICE. This may help in bringing at least some aspects of inflationary physics (e.g. causal aspects, excluding those features which relate directly to gravity) of the universe under some experimental control.
Acknowledgments {#acknowledgments .unnumbered}
===============
We are very grateful to Raju Venugopalan, Rajeev Bhalerao, and Sergei Voloshin for very useful comments and suggestions. We also thank Sanatan Digal, Abhishek Atreya and Anjishnu Sarkar for useful comments.
[99]{}
A. P. Mishra, R. K. Mohapatra, P. S. Saumia, and A. M. Srivastava, Phys. Rev. [**C 77**]{}, 064902 (2008).
M. Gyulassy, D. H. Rischke, and B. Zhang, Nucl. Phys. [**A 613**]{}, 397 (1997).
J.-Y. Ollitrault, Phys. Rev. [**D 46**]{}, 229 (1992); S. Voloshin and Y. Zhang, Z. Physik [**C70**]{}, 665 (1996); S.A. Volosin, A.M. Poskanzer, and R. Snellings, arxiv:0809.2949.
B. M. Tavares, H. J. Drescher, and T. Kodama, Braz. J. Phys. [**37**]{}, 41 (2007); P. Sorensen (for the STAR collaboration), J. Phys. [**G 34**]{}, S897 (2007); S. Manly et al. (for PHOBOS collaboration), nucl-ex/0702029; H. Sorge, Phys. Rev. Lett. [**82**]{}, 2048 (1999).
S. Dodelson, “Modern Cosmology”, (Academic Press, California, 2003).
X.N. Wang and M. Gyulassy, Phys. Rev. [**D 44**]{}, 3501 (1991); Comput. Phys. Commun. [**83**]{}, 307 (1994).
N. Borghini and J.-Y. Ollitrault, Phys. Rev. [**C 70**]{}, 064905 (2004); X. Dong, S. Esumi, P. Sorensen, N.Xu, and Z. Xu, Phys. Lett. [**B 597**]{}, 328 (2004)
N. Borghini, P. M. Dinh, and J.-Y. Ollitrault, Phys. Rev. [**C 64**]{}, 054901 (2001).
P.F. Kolb, J. Sollfrank, and U. Heinz, Phys. Rev. [**C 62**]{}, 054909 (2000).
P. Sorensen, Proc. 24th Winter Workshop on Nuclear Dynamics, 2008, arXiv:0808.0503.
M. Gyulassy, I. Vitev, X.N. Wang, Phys.Rev.Lett. [**86**]{}, 2537 (2001); Y. Bai (for STAR collaboration), J.Phys. [**G34**]{}, S903 (2007); B.I. Abelev et al. (for STAR Collaboration), Phys.Rev. [**C77**]{}, 054901 (2008).
J.-Y. Ollitrault, Eur. J. Phys. [**29**]{}, 275 (2008).
N. Borghini, and J.-Y. Ollitrault, Phys. Lett. [**B 642**]{}, 227 (2006).
P. Huovinen, P.F. Kolb, U. Heinz, P.V. Ruuskanen, and S.A. Voloshin, Phsy. Lett. [**503**]{}, 58 (2001).
I. Selyuzhenkov and S. Voloshin, Phys. Rev. [**C 77**]{}, 034904 (2008); R. S. Bhalerao, N. Borghini, and J.-Y. Ollitrault, Nucl. Phys. [**A 727**]{}, 373 (2003).
E. Hivon, et al. Astrophys. J. [**567**]{},2 (2002).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
bibliography:
- 'unifyref.bib'
---
=10000 CALT-TH-2017-22
[ Unifying Relations for Scattering Amplitudes ]{}\
\
[*${}^a$Walter Burke Institute for Theoretical Physics,\
California Institute of Technology, Pasadena, CA 91125\
${}^b$Mani L. Bhaumik Institute for Theoretical Physics,\
Department of Physics and Astronomy, UCLA, Los Angeles, CA 90095* ]{}[^1]\
**Abstract**
> We derive new amplitudes relations revealing a hidden unity among a wide-ranging variety of theories in arbitrary spacetime dimensions. Our results rely on a set of Lorentz invariant differential operators which transmute physical tree-level scattering amplitudes into new ones. By transmuting the amplitudes of gravity coupled to a dilaton and two-form, we generate all the amplitudes of Einstein-Yang-Mills theory, Dirac-Born-Infield theory, special Galileon, nonlinear sigma model, and biadjoint scalar theory. Transmutation also relates amplitudes in string theory and its variants. As a corollary, celebrated aspects of gluon and graviton scattering like color-kinematics duality, the KLT relations, and the CHY construction are inherited traits of the transmuted amplitudes. Transmutation recasts the Adler zero as a trivial consequence of the Weinberg soft theorem and implies new subleading soft theorems for certain scalar theories.
Introduction and Summary
========================
The modern S-matrix program has exposed marvelous structures long hidden in plain sight within gauge theory and gravity. The fact that these theories are endowed with exceptional properties is perhaps unsurprising given that their form is uniquely dictated by locality and gauge invariance. Nevertheless, in recent years it has become abundantly clear that many of these structures are actually commonplace. For example, the Kawai-Lewellen-Tye (KLT) relations [@Kawai:1985xq], Bern-Carrasco-Johansson (BCJ) color-kinematics duality [@Bern:2008qj], and Cachazo-He-Yuan (CHY) construction [@Cachazo:2013hca; @Cachazo:2013iea; @Cachazo:2014xea] all apply across a tremendous range of theories, including Einstein-Yang-Mills (EYM) theory, Dirac-Born-Infeld (DBI) theory, the special Galileon (SG), the nonlinear sigma model (NLSM), and the biadjoint scalar (BS) theory. A natural question then arises: why [*these*]{} theories? Imbued with symmetries of disparate origin and character, these theories do not obviously conform to any cohesive organizing principle that would place them on equal footing in the eyes of the S-matrix.
In this paper, we offer an explanation for the peculiar universality of these structures. Our results follow from a set of simple unifying relations which “transmute” the tree-level scattering amplitudes of certain theories into those of others. By transmuting the S-matrix of the mother of all theories—gravity coupled to a dilaton and two-form, [*i.e.*]{}, “extended gravity”—we beget the S-matrices of all the theories previously mentioned, thus revealing their many shared traits as congenital. Transmutation is by design independent of the particular form in which an amplitude happens to be represented. To achieve this, we study amplitudes as formal functions of Lorentz invariant products of the external on-shell kinematic data: $e_i e_j$, $p_i e_j$, and $p_i p_j$ where $e_i$ and $p_i$ denotes polarization and momentum of particle $i$, respectively, and $i\neq j$.[^2] [^3] In this language, it is straightforward to construct a basis of gauge invariant operators which preserve on-shell kinematics. Perhaps unsurprisingly, these operators automatically transmute gauge invariant scattering amplitudes into new ones. Our roster of transmutation operators is summarized as follows:
- The trace operator ${\cal T}_{ij}=\partial_{e_i e_j}$ reduces the spin of particles $i$ and $j$ by one unit and places them within a new color trace structure. This operator transmutes gravitons into photons, gluons into biadjoint scalars, and BI photons into DBI scalars.
- The insertion operator ${\cal T}_{ijk} = \partial_{p_i e_j}- \partial_{p_ke_j}$ reduces the spin of particle $j$ by one unit and inserts it between particles $i$ and $k$ within a color trace structure. This operator transmutes gravitons into gluons, gluons into biadjoint scalars, and BI photons into pions, at the level of color-ordered amplitudes of the resulting theories, which we will properly define in [Sec. \[sec:web\]]{}.
- The longitudinal operator ${\cal L}_i = \sum_j p_i p_j \partial_{p_j e_i}$ reduces the spin of particle $i$ by one unit while converting it to a longitudinal mode. Necessarily applied to all particles at once, this operator transmutes gravitons into BI photons, gluons into pions, and BI photons into SG scalars.
These transmutation operators form the building blocks of the unified web of theories depicted in [Fig. \[fig:unified\_web\]]{}. Theory space is partitioned into three triangular regions corresponding to the nearest of kin to extended gravity, BI, and YM. The corners of the web denote all of the theories previously discussed, while the edge and bulk regions denote hybrid theories. Remarkably, these triangular regions are not independent—due to KLT [@Kawai:1985xq] and BCJ [@Bern:2008qj], extended gravity can be recast as the double copy of YM, and BI as the product of YM and the NLSM. Hence, one can actually construct the entire web purely from the middle triangle corresponding to YM and its descendants.
Armed with the notion of transmutation, it is then straightforward to import familiar structures in gauge theory and gravity into their descendants. For example, transmutation trivially explains the near-universal applicability of the KLT, BCJ, and CHY constructions. Likewise, the Weinberg soft theorems transmute into the Adler zero conditions on pions and Galileons, as first observed in [@Cheung:2016prv]. At subleading order, we can also derive new soft theorems in the spirit of [@Cachazo:2016njl].
The remainder of this paper is organized as follows. In Sec. \[sec:setup\], we systematically construct a basis of transmutation operators. We define the notion of color-ordering which is crucial for insertion and discuss the web of theories related by transmutation in Sec. \[sec:web\]. Examples are illustrated in Sec. \[sec:example\]. A proof by induction of our claims is presented in Sec. \[sec:proof\], followed by a discussion of the infrared structure in Sec. \[sec:soft\]. Finally, we conclude with a discussion of outlook and future directions in Sec. \[sec:outlook\].
{width="75.00000%"}
Transmutation {#sec:setup}
=============
Physical Constraints
--------------------
In an abstract sense, the purpose of a transmutation operator ${\cal T}$ is to convert a gauge invariant object $A$ into a new gauge invariant object ${\cal T} \cdot A$. At the very minimum, ${\cal T}$ should conform to the following baseline physical criteria:
- $\cal T$ preserves on-shell kinematics.
- $\cal T$ preserves gauge invariance.
Let us consider each of these conditions in turn.
### On-shell Kinematics {#on-shell-kinematics .unnumbered}
A physical scattering amplitude of massless particles is well-defined on the support of the on-shell conditions, $p_i p_i = p_i e_i =0$, and momentum conservation, $\sum_i p_i = 0$. Representations of the amplitude that differ by terms which vanish on-shell are physically equivalent. It is then crucial that our construction be agnostic to such differences.
To preserve the on-shell conditions, we simply define the physical scattering amplitude $A$ to be a function of $p_i p_j$, $p_i e_j$, and $e_i e_j$ for $i\neq j$. For momentum conservation, we define the total momentum operator [$$\begin{aligned}
{\cal P}_v \equiv \sum_i p_i v \, ,\end{aligned}$$]{} where $i$ runs over all external legs and and $v$ labels any momentum or polarization vector. Due to the implicit momentum-conserving delta function in $A$, we have that [$$\begin{aligned}
{\cal P}_v \cdot A=0 \,.\end{aligned}$$]{} To ensure that ${\cal T}$ conserves momentum it is then sufficient to require that [$$\begin{aligned}
[\, {\cal P}_v \, , \, {\cal T} \,] \cdot A =0 \,,
\label{eq:mom_con}\end{aligned}$$]{} so ${\cal P}_v \cdot {\cal T} \cdot A ={\cal T} \cdot {\cal P}_v \cdot A= 0$, [*i.e.*]{} the transmuted amplitude also conserves momentum. This does not trivially hold for differential operators involving $p_i p_j$ and $p_i e_j$ as we will see explicitly in next section. Since $[\, {\cal P}_v \, , \, {\cal P}_w \,]=0$, this operator itself conserves momentum.
### Gauge Invariance {#gauge-invariance .unnumbered}
A physical scattering amplitude should also be gauge invariant. To incorporate this constraint we define a differential operator corresponding to the Ward identity on particle $i$, [$$\begin{aligned}
{\cal W}_i \equiv \sum_v p_i v \, \partial_{ v e_i} \, .\end{aligned}$$]{} Here $v$ runs over all momentum and polarization vectors in the amplitude. Any gauge invariant amplitude is annihilated by this operator, [$$\begin{aligned}
{\cal W}_i \cdot A = 0 \, .\end{aligned}$$]{} Perhaps unsurprisingly, ${\cal W}_i$ is itself momentum-conserving and gauge invariant. To establish that ${\cal W}_i$ conserves momentum we compute the commutator $$\begin{split}
&[\, {\cal W}_i \, , \, {\cal P}_v \, ]= \sum_{j,w} [\, p_i w \, \partial_{w e_i} \, , \, p_j v \, ]
= \delta_{v e_i} {\cal P}_{p_i} \, ,
\end{split}$$ which vanishes when acting on a physical amplitude $A$, so $[\, {\cal W}_i \, , \, {\cal P}_v \, ] \cdot A=0$. Meanwhile, to verify that ${\cal W}_i$ is gauge invariant, we compute the commutator $$\begin{split}
&[\, {\cal W}_i \, , \, {\cal W}_j \, ]= \sum_{v,w} [\, p_i v \, \partial_{v e_i } \, , \, p_j w \, \partial_{w e_j } \, ]
=0 \, ,
\end{split}$$ which vanishes automatically.
Transmutation Operators
-----------------------
We are now equipped to derive a systematic basis of operators which preserve gauge invariance and on-shell kinematics. To begin, we define an ansatz expressed in a basis of first order differentials, [$$\begin{aligned}
{\cal T} \equiv \sum\limits_{i,j } {\cal A}_{ij} \partial_{p_i p_j} + {\cal B}_{ij} \partial_{p_i e_j} +{\cal C}_{ij} \partial_{e_i e_j}\, ,
\end{aligned}$$]{} where ${\cal A}_{ij}$, ${\cal B}_{ij}$, and ${\cal C}_{ij}$ are general functions of the external kinematic data. For later notational convenience we choose ${\cal A}_{ii} = {\cal B}_{ii} = {\cal C}_{ii}=0$, which is consistent with the on-shell conditions. Let us now constrain this ansatz with momentum conservation and gauge invariance.
First, requiring that ${\cal T}$ preserves conservation of momentum implies that $$\begin{aligned}
[ \, {\cal T} \, , \, {\cal P}_v \, ]
= \sum_{i,j ,k} [ \, {\cal A}_{ij} \partial_{p_i p_j} + {\cal B}_{ij} \partial_{p_i e_j} \, , \, p_k v \, ]
=0 \, .\end{aligned}$$ Choosing $v$ to be a momentum vector or polarization vector, respectively, we obtain [$$\begin{aligned}
\sum_{i} {\cal A}_{ij} + {\cal A}_{ji} =\sum_{i } {\cal B}_{ij} = 0 \, ,\end{aligned}$$]{} for all $j$. So momentum conservation constrains every row and column of ${\cal A}_{ij}$ sums to zero, every column of ${\cal B}_{ij}$ sums to zero, and leaves ${\cal C}_{ij}$ unfixed. This leads us to a simple basis of mutually commuting operators which span the solution set of these constraints, $$\begin{split}
{\cal T}_{ij} &\equiv \partial_{e_i e_j} \\
{\cal T}_{ijk} &\equiv \partial_{p_i e_j} - \partial_{p_k e_j} \\
{\cal T}_{ijkl} &\equiv \partial_{p_i p_j}- \partial_{p_k p_j} +\partial_{p_k p_l}- \partial_{p_i p_l} \, ,
\end{split}$$ where all indices are distinct with the symmetry properties $$\begin{split}
{\cal T}_{ij} &= {\cal T}_{ji} \\
{\cal T}_{ijk} &= -{\cal T}_{kji} \\
{\cal T}_{ijkl} &= -{\cal T}_{kjil}= {\cal T}_{klij} = -{\cal T}_{ilkj} \, .
\end{split}$$ Thus we obtain a basis of momentum-conserving operators.
Second, we consider the gauge invariance of these basis operators. A simple calculation yields $$\begin{aligned}
\label{eq:commutators}
[\, {\cal T}_{ij} \, , \, {\cal W}_k \, ] &= 0 \nonumber \\
[\, {\cal T}_{ijk} \, , \, {\cal W}_l \, ] &=\delta_{il} {\cal T}_{ij}- \delta_{kl} {\cal T}_{jk} \\
[\, {\cal T}_{ijkl} \, , \, {\cal W}_m \, ] &= (\delta_{im}-\delta_{km}){\cal T}_{jml} +(\delta_{jm}-\delta_{lm}) {\cal T}_{imk} \, , \nonumber\end{aligned}$$ so ${\cal T}_{ij}$ is intrinsically gauge invariant while ${\cal T}_{ijk}$ and ${\cal T}_{ijkl}$ are not. As we will later see, operators which are not intrinsically gauge invariant can still be [*effectively*]{} gauge invariant if the right-hand side of [Eq. ]{} annihilates the amplitude. Note that ${\cal T}_{ijkl}$ is especially peculiar because it is a differential operator in $p_i p_j$, and so yields generic double poles when applied to a physical amplitude. As a result, ${\cal T}_{ijkl}$ is not necessarily a generator of amplitudes, although we will see later that it can appear in certain soft limits. In any case, hereafter we focus on ${\cal T}_{ij}$ and ${\cal T}_{ijk}$.
### Trace Operators {#trace-operators .unnumbered}
Since ${\cal T}_{ij}$ is intrinsically gauge invariant we can apply it with impunity to any gauge invariant object to produce a new gauge invariant object. As a differential operator, ${\cal T}_{ij}$ eliminates all appearances of $e_i$ and $e_j$ except when appearing in the combination $e_i e_j$. Hence ${\cal T}_{ij}$ is equivalent to a dimensional reduction in which the polarization vectors $e_i$ and $e_j$ are chosen to be extra-dimensional, [*i.e.*]{} orthogonal to the directions spanned by the external gluons. This kinematic configuration sets $e_i e_j=1$ and $v e_i = v e_j=0$ for any momentum or polarization vector $v$ associated with a gluon, which is mathematically equivalent to applying ${\cal T}_{ij}$. From this perspective it is obvious why ${\cal T}_{ij}$ conserves momentum while preserving gauge invariance on the remaining particles—it is simply an implementation of dimensional reduction.
At the level of scattering amplitudes, ${\cal T}_{ij}$ then transmutes a pair of gluons into a pair of scalars, [$$\begin{aligned}
{\cal T}_{ij} \cdot A(\cdots, g_i , g_j,\cdots) = A(\cdots, \phi_i \phi_j,\cdots) \, ,\end{aligned}$$]{} where the ellipses denote spectator particles and the transmuted scalars should be interpreted as biadjoint scalars carrying the original color index as well as an additional dual color index. Here and throughout this paper, we use a notation where commas separate sets of particles which are ordered according to the dual color. Since the original gluons do not carry dual color, each gluon in the original amplitude is separated by a comma. In contrast, the transmuted biadjoint scalars carry a dual color ordering $i,j$ within a new dual color trace, so they are not separated by a comma. Note that for the simple case of transmuting a pair of gluons, the ordering $i,j$ is irrelevant because ${\cal T}_{ij}$ is symmetric. Since ${\cal T}_{ij}$ initializes a new dual color trace, we refer to it as a “trace operator”.
### Insertion Operators {#insertion-operators .unnumbered}
Unfortunately, ${\cal T}_{ijk}$ is not intrinsically gauge invariant because it acts nontrivially on the momenta $p_i$ and $p_k$ generated by gauge transformations on $e_i$ and $e_k$. If, however, $e_i$ and $e_k$ have already been eliminated prior to acting with ${\cal T}_{ijk}$, then there is no issue with gauge invariance. In other words, ${\cal T}_{ijk}$ is effectively gauge invariant when appearing in tandem with enough supplemental transmutation operators. For example, [$$\begin{aligned}
[\, {\cal T}_{ik} \cdot {\cal T}_{ijk} \, , \, {\cal W}_l \, ] = \delta_{il} \, {\cal T}_{ik}\cdot {\cal T}_{ij}- \delta_{kl} \, {\cal T}_{ik}\cdot {\cal T}_{jk} \, ,\end{aligned}$$]{} where the right-hand side vanishes on any physical amplitude which is necessarily multi-linear in $e_i$ and $e_k$. Said another way, ${\cal T}_{ijk}$ is effectively gauge invariant if its commutator in [Eq. ]{} annihilates the amplitude. Putting this all together at the level of scattering amplitudes, we find that [$$\begin{aligned}
{\cal T}_{ijk} \cdot A(\cdots \phi_i \phi_k \cdots, g_j,\cdots)
=A(\cdots \phi_i \phi_j \phi_k \cdots ,\cdots) \, . \label{eq:insertion_claim}\end{aligned}$$]{} As before, the transmuted states are biadjoint scalars carrying the original color as well as a new dual color. Since the indices of ${\cal T}_{ijk}$ are ordered, these scalars have the ordering $i,j,k$ within the dual color trace. As noted earlier, the resulting biadjoint scalars are not separated by commas because they reside in the same dual color trace. Because ${\cal T}_{ijk}$ transmutes a gluon into a biadjoint scalar and inserts it between two existing biadjoint scalars, we will refer to it as an “insertion operator”.
### Longitudinal Operators {#longitudinal-operators .unnumbered}
Let us now define a set of “longitudinal operators” which transmute particles into derivatively coupled longitudinal modes, [*e.g.*]{} pions and Galileons. These operators are defined as $${\cal L}_i \equiv \sum_j p_i p_j \partial_{p_j e_i} \qquad \textrm{ and } \qquad
{\cal L}_{ij} \equiv - p_i p_j \partial_{e_i e_j} \, , \label{eq:long_def}
$$ where $j$ runs over all momenta in the amplitude and we have included a relative sign for later convenience. The longitudinal operators preserve on-shell kinematics because they are linear combinations of the trace and insertion operators: ${\cal L}_i = \sum_{j\neq k} p_i p_j {\cal T}_{jik}$ and ${\cal L}_{ij} = - p_i p_j {\cal T}_{ij}$ where $k$ is an arbitrary reference leg. By the same logic, ${\cal L}_{ij}$ is intrinsically gauge invariant but ${\cal L}_{i}$ is not, since [$$\begin{aligned}
[\, {\cal L}_i \, , \, {\cal W}_j \, ] &= -{\cal L}_{ij} \, . \label{eq:LW}\end{aligned}$$]{} Just as before, however, ${\cal L}_i$ can still be effectively gauge invariant if the right-hand side of the commutator annihilates the amplitude.
It is clear from [Eq. ]{} that ${\cal L}_i$ and ${\cal L}_{ij} $ are closely related by gauge invariance. We can exploit this kinship to derive a simple formula relating products of ${\cal L}_i$ to products of ${\cal L}_{ij}$. To do this we write the Ward identity in the form [$$\begin{aligned}
{\cal W}_i = {\cal L}_i + \Delta_i \, , \label{eq:Wi_to_Li}\end{aligned}$$]{} where $\Delta_i = \sum_j p_i e_j \partial_{e_j e_i}$ is the difference between the Ward identity operator and the longitudinal operator. Starting from any product of ${\cal L}_i$ operators, we insert ${\cal L}_i = {\cal W}_i - \Delta_i$ and commute factors of ${\cal W}_i$ to the right until they annihilate the amplitude. Iterating this procedure, we find that $$\begin{split}
{\cal L}_i \cdot {\cal L}_j &= {\cal L}_{ij} + \cdots \\
{\cal L}_i \cdot {\cal L}_j \cdot {\cal L}_k \cdot {\cal L}_l &= {\cal L}_{ij}\cdot { \cal L}_{kl}+{\cal L}_{ik} \cdot {\cal L}_{jl}+{\cal L}_{il} \cdot {\cal L}_{jk} + \cdots \, ,\label{eq:Lrelation}
\end{split}$$ where the ellipses denote terms which have ${\cal W}_i$ commuted all the way to the right or which involve $\Delta_i$. In order to generalize [Eq. ]{} we define a shorthand for a product of longitudinal operators, [$$\begin{aligned}
{\cal L} \equiv \prod_{i} {\cal L}_{i} = \sum_{ \rho } \prod_{i,j\,\in \textrm{pairs}} {\cal L}_{ij} +\cdots, \label{eq:Ldef}\end{aligned}$$]{} where in the first equality $i$ runs over all the gluons in the scattering amplitude and in the second equality $\rho$ runs over all partitions of the external gluons into pairs and $i,j$ run over each pair in a partition. Again, the ellipses denote terms with ${\cal W}_i$ on the right of involving $\Delta_i$. The former terms identically annihilate the physical amplitude while the latter terms vanish due to the multi-linearity of the amplitude in polarizations.
Since ${\cal L}_i$ and ${\cal L}_{ij}$ carry explicit factors of $p_i p_j$, they reduce the spin of states while also increasing the number of derivatives per particle. As we will see, these operators transmutes states into longitudinal modes, [*e.g.*]{} pions or Galileons. Note that ${\cal L}_{ij}$ is equivalent to the “compactify” operation for CHY integrands presented in [@Cachazo:2014xea], here generalized to any representation of the amplitude.
Unified Web of Theories {#sec:web}
=======================
Applied in various combinations, the transmutation operators form an interlocking web of scattering amplitudes relations across a wide range of theories. Remarkably, all S-matrices descend from the S-matrix of extended gravitons. As discussed in the introduction, transmutation can be understood for YM and then straightforwardly generalized to extended gravity and BI via double copy construction. Therefore, we focus here mainly on the transmutation on YM amplitudes. Let us briefly review some of the basics of color structure in YM theory. The full color-dressed YM amplitude can be expressed in terms of traces of color generators [$$\begin{aligned}
A_{\rm YM}^{\rm full}(g_1, g_2,\cdots, g_n) = \sum_{\alpha \in S_n/Z_n} \textrm{Tr}(T^{\alpha_1}T^{\alpha_2}\cdots T^{\alpha_n}) \, A_{\rm YM}(g_{\alpha_1} g_{\alpha_2} \cdots g_{\alpha_n}),
\label{eq:color_ordering}\end{aligned}$$]{} where $\alpha$ sums over all orderings modulo cyclic permutations and the right-hand side is comprised of so-called color-ordered amplitudes. As before, we use commas to delineate sets of particles with no relative ordering. For this reason the states in the color-dressed amplitude in [Eq. ]{} are separated by commas, while those in the color-ordered amplitudes are not. Similarly, we would label a non-planar amplitude with commas separating sets of particles in difference traces.
This decomposition not only works for YM but also for the flavor group of pions. Also, in the case of BS theory, there are two color groups so we can apply the above color decomposition twice [$$\begin{aligned}
A_{\rm BS}^{\rm full}(\phi_1, \phi_2,\cdots, \phi_n) = \sum_{\alpha,\beta \in S_n/Z_n} \textrm{Tr}(T^{\alpha_1}T^{\alpha_2}\cdots T^{\alpha_n}) \textrm{Tr}(T^{\beta_1}T^{\beta_2}\cdots T^{\beta_n})\, A_{\rm BS}(\phi_{\alpha_1} \cdots \phi_{\alpha_n} | \phi_{\beta_1} \cdots \phi_{\beta_n}) \nonumber
\\
\label{eq:double_color}\end{aligned}$$]{} where $A_{\rm BS}(\phi_{\alpha_1} \cdots \phi_{\alpha_n} | \phi_{\beta_1} \cdots \phi_{\beta_n}) $ is the doubly color-ordered amplitude encoding a color and dual color ordering, often abbreviated as $A_{\rm BS}(\alpha|\beta)$. For notational simplicity, we will often keep the dual color ordering arising from transmutation explicit, while suppressing the original color ordering inherited from starting amplitude.
We stress here that the insertion operators, which transmute graviton to gluons, gluons to biadjoint scalars, or BI photons to pions, produce color-ordered amplitudes. The purpose of insertion in this paper is to reduce the spins but project to an ordering of the particles, which we will show concretely in Sec. \[sec:BCJ\]. The mapping of the full amplitudes, however, requires summing all orderings dressed with color factors put by hands. Let us discuss the resulting single trace and multiple trace scattering amplitudes from transmutation in turn.
Single Trace Amplitudes
-----------------------
To begin, it will be convenient to define a shorthand notation denoting a single trace operator followed by a sequence of insertion operators, [$$\begin{aligned}
{\cal T}[\alpha] \equiv {\cal T}_{\alpha_1 \alpha_n} \cdot \prod_{i =2}^{n-1} {\cal T}_{\alpha_{i-1} \alpha_i \alpha_n} \, , \label{eq:Odef}\end{aligned}$$]{} where $\alpha$ is an ordered set. Applying ${\cal T}[\alpha] $ to a scattering amplitude of gluons, we obtain $${\cal T}[i_1 \cdots i_n] \cdot A(g_{i_1},\cdots ,g_{i_n},\cdots) = A( \phi_{i_1} \cdots \phi_{i_n},\cdots) \, ,
\label{eq:master}$$ which is an amplitude of biadjoint scalars coupled to the remaining nontransmuted gluons denoted by the ellipses after the last comma. If $\alpha$ is the set of all gluons, then no gluons remain after transmutation and the resulting amplitude is that of the BS theory. If, on the other hand, $\alpha$ is a subset of gluons, then we obtain an amplitude of the gauged BS theory, which describes biadjoint scalars interacting with gluons through the original color index. Since the dual color ordering $\alpha$ of the biadjoint scalars is defined modulo cyclic permutations, ${\cal T}[\alpha]$ is actually cyclically invariant, though not manifestly so.
Remarkably, the above construction is quite general. For example, by transmuting a extended gravity amplitude, we obtain [$$\begin{aligned}
{\cal T} [i_1 \cdots i_n] \cdot A(h_{i_1},\cdots ,h_{i_n},\cdots) =A(g_{i_1} \cdots g_{i_n}, \cdots) \, ,\end{aligned}$$]{} which is an amplitude of gluons coupled to extended gravitons, [*i.e.*]{} an EYM amplitude. Applied to photon scattering amplitudes in BI theory, transmutation yields [$$\begin{aligned}
{\cal T}[i_1 \cdots i_n] \cdot A(\gamma_{i_1}, \cdots, \gamma_{i_n},\cdots) = A( \pi_{i_1} \cdots \pi_{i_n}, \cdots) \, ,\end{aligned}$$]{} which is an amplitude of pions coupled to BI photons.
Meanwhile, the longitudinal operator transmutes all the gluons in an amplitude to pions, [$$\begin{aligned}
{\cal L} \cdot A(g_{i_1}, \cdots, g_{i_n}, \cdots)
=A(\pi_{i_1}, \cdots, \pi_{i_n}, \cdots) \, , \label{eq:long_claim}\end{aligned}$$]{} where the ellipses after the last comma denote any spectator biadjoint scalars in the amplitude. If the initial amplitude does not contain any biadjoint scalars, then the transmuted object actually vanishes. If, on the other hand, the initial amplitude starts with precisely two biadjoint scalars, then the resulting amplitude is actually equal to an amplitude of all external pions. That is, irrespective of which pair of external legs are initially chosen to be biadjoint scalars, the longitudinal operator produces a permutation invariant amplitude of the NLSM. This obfuscation of permutation invariance is reminiscent of the cubic NLSM action proposed in [@Cheung:2016prv]. Finally, if the initial amplitude has three or more biadjoint scalars, then the resulting amplitude is that of the extended NLSM [@Cachazo:2016njl], describing pions coupled to biadjoint scalars.
As before, this construction also applies to other theories. In particular, the longitudinal operator transmutes an amplitude of extended gravitons coupled to gluons into $${\cal L} \cdot A(h_{i_1}, \cdots , h_{i_n}, \cdots)
=A(\gamma_{i_1} , \cdots , \gamma_{i_n}, \cdots) \, ,$$ which is an amplitude of BI photons coupled to gluons. Similarly, the longitudinal operator transmutes an amplitude of BI photons coupled to pions into $${\cal L} \cdot A(\gamma_{i_1}, \cdots , \gamma_{i_n}, \cdots)
= A(\phi_{i_1}, \cdots, \phi_{i_n}, \cdots) \, ,$$ which is an amplitude of SG scalars coupled to pions.
Multiple Trace Amplitudes
-------------------------
It is straightforward to generalize our results to include multiple trace structures. To start, let us consider the simplest multiple trace amplitudes where each trace is comprised of a pair of states. Starting from a gluon amplitude, we obtain [$$\begin{aligned}
{\cal T}[i_1 j_1] \cdots {\cal T}[i_m j_m] \cdot A(g_{i_1} , g_{j_1} , \cdots , g_{i_m} , g_{j_m},\cdots)
= A(\phi_{i_1 }\phi_{j_1}, \cdots , \phi_{i_m }\phi_{j_m},\cdots) \, , \label{eq:trace_claim} \end{aligned}$$]{} where the ellipses after the last comma again denote nontransmuted spectator gluons. The resulting amplitude describes biadjoint scalars coupled to gluons via gauge interactions. This operation also transmutes extended gravity into EM theory and BI theory into the DBI scalar theory.
To discuss more complex trace structures it will be convenient to make the dependence on the original color ordering explicit. In particular, if we start with a color ordered gluon amplitude $A(\beta)$, then [Eq. ]{} can be written in the shorthand [$$\begin{aligned}
{\cal T}[\alpha] \cdot A(\beta) = A( \alpha |\beta ) \, ,\end{aligned}$$]{} in the notation of [Eq. ]{}. The $A( \alpha |\beta )$ on the right-hand side is the doubly color-ordered amplitude for BS theory in [Eq. ]{}. Applying this operator multiple times, we obtain [$$\begin{aligned}
{\cal T}[\alpha_1] \cdots {\cal T}[\alpha_m] \cdot A(\beta) &= A(\alpha_1, \cdots, \alpha_m |\beta) \, ,\end{aligned}$$]{} which is a multiple trace amplitude in gauged BS theory.
In contrast, since the amplitudes of extended gravity and BI theory are not color-ordered to start, [$$\begin{aligned}
{\cal T}[\alpha] \cdot A = A( \alpha) \,.\end{aligned}$$]{} In this case the multiple trace amplitude is [$$\begin{aligned}
{\cal T}[\alpha_1] \cdots {\cal T}[\alpha_m] \cdot A &= A(\alpha_1, \cdots, \alpha_m ) \, ,\end{aligned}$$]{} corresponding to the amplitudes of the NLSM coupled to BI photons and EYM theory.
Summary of Unifying Relations
-----------------------------
Let us briefly summarize the unifying relations associated with some better-known theories, corresponding to the corners of the web depicted in [Fig. \[fig:unified\_web\]]{}. From YM amplitudes we obtain, $$\begin{split}
A_{\rm YMS} &= {\cal T}[i_1 j_1]\cdots {\cal T}[ i_m j_m] \cdot A_{\rm YM} \\
A_{\rm BS} &= {\cal T}[i_1 \cdots i_n]\cdot A_{\rm YM} \\
A_{\rm NLSM} &={\cal L} \cdot {\cal T}[i_1 i_n] \cdot A_{\rm YM} \, .
\end{split}
\label{eq:YM_sum}$$ From BI amplitudes we obtain $$\begin{split}
A_{\rm DBI} &= {\cal T}[i_1 j_1]\cdots {\cal T}[ i_m j_m] \cdot A_{\rm BI} \\
A_{\rm NLSM} &= {\cal T}[i_1 \cdots i_n]\cdot A_{\rm BI} \\
A_{\rm SG} &={\cal L} \cdot {\cal T}[i_1 i_n] \cdot A_{\rm BI} \,.
\end{split}
\label{eq:BI_sum}$$ From extended gravity amplitudes we obtain $$\begin{split}
A_{\rm EM} &= {\cal T}[i_1 j_1]\cdots {\cal T}[ i_m j_m] \cdot A_{\rm G} \\
A_{\rm YM} &= {\cal T}[i_1 \cdots i_n]\cdot A_{\rm G} \\
A_{\rm BI} &={\cal L} \cdot {\cal T}[i_1 i_n] \cdot A_{\rm G} \,.
\end{split}
\label{eq:GR_sum}$$ Since the above amplitudes are all ultimately derived from extended gravity, they share the kindred relations depicted in [Fig. \[fig:unified\_web\]]{}.
A trivial corollary of these unifying relations is the universality of the KLT, BCJ, and CHY constructions. Because these structures generalize to arbitrary spacetime dimension, the associated scattering amplitudes can be represented as functions of Lorentz invariant products of momentum and polarization vectors—so transmutation applies. Given YM theory and gravity as input, it is then straightforward to import the KLT and BCJ relations into all the theories in [Fig. \[fig:unified\_web\]]{} via transmutation. Similarly, the CHY integrand for gravity can be transmuted into new CHY integrands for the descendant theories. We leave a full analysis of these ideas for future work.
Ultraviolet Completion {#sec:more}
----------------------
Transmutation is based on the dual assumptions of gauge invariance and on-shell kinematics—for massless external particles. Notably, these stipulations do not preclude the presence of massive particles exchanged as intermediate states, as would arise from ultraviolet completion. Hence, transmutation also defines a web of unifying relations for the [*ultraviolet completions*]{} of the theories in [Fig. \[fig:unified\_web\]]{}. Whether the resulting amplitudes are fully consistent is not guaranteed—after all, our factorization proof assumes a massless spectrum—but it is still illuminating to see the variety of objects that arise from this procedure.
To begin, consider an ultraviolet completion of YM theory in which new massive degrees of freedom arrive at a physical scale suggestively denoted by $\alpha'$. Since the amplitude is a gauge invariant object, transmutation generates a new gauge invariant object, [$$\begin{aligned}
{\cal T}[i_1 \cdots i_n] \cdot A_{\rm YM+ \alpha'}(g_{i_1}, \cdots, g_{i_n}) = A_{\rm BS + \alpha'}(\phi_{i_1} \cdots \phi_{i_n} ) \, ,\end{aligned}$$]{} corresponding to a possible ultraviolet completion of the BS theory with $\alpha'$ corrections. A natural candidate for an ultraviolet completion of YM theory is open superstring theory, $A_{\rm YM + \alpha'} \sim A_{\rm open}$, in which case the transmutation equation becomes [$$\begin{aligned}
\label{eq:stringrelation1}
{\cal T}[i_1 \cdots i_n] \cdot A_{\rm open}(g_{i_1}, \cdots, g_{i_n}) = A_{\rm Z}(\phi_{i_1} \cdots \phi_{i_n} ) \, ,\end{aligned}$$]{} where $ A_{\rm BS + \alpha'} \sim A_{\rm Z}$ is the amplitude of biadjoint scalars obtained from the Z-theories of Carrasco, Mafra, and Schlotterer [@Carrasco:2016ldy; @Mafra:2016mcc; @Carrasco:2016ygv]. Explicitly, this amplitude is equal to the world sheet integral [$$\begin{aligned}
A_{\rm Z}(\phi_{i_1} \cdots \phi_{i_n})
=
(\alpha')^{n-3} \int_{D } \, { dz_1 dz_2 \cdots dz_n \over {\rm vol} (SL(2,R)) }
{\prod^n_{i<j} |z_{ij}|^{\alpha' p_i p_j} \over z_{i_1 i_2} z_{i_2 i_3} \cdots
z_{i_n i_1} } \,, \end{aligned}$$]{} where $z_{ij} = z_i - z_j$, and the integration domain $D$ is along the real line, [$$\begin{aligned}
D = \{ - \infty < z_1 < z_2< \cdots <z_n < + \infty\} \, .\end{aligned}$$]{} An analogous argument for the ultraviolet completions of BI theory and the NLSM imply that [$$\begin{aligned}
\label{eq:stringrelation2}
{\cal T}[i_1 \cdots i_n] \cdot A_{\rm open}( \gamma_{i_1}, \ldots , \gamma_{i_n}) = A_{ \rm Z }( \pi_{i_1} \cdots \pi_{i_n} ) \, ,\end{aligned}$$]{} corresponding to the BI photon amplitudes of open superstring theory and the pion amplitudes of abelian Z-theory [@Carrasco:2016ldy], respectively.
The relations in [Eq. ]{} and [Eq. ]{} can be trivially derived from the KLT relations. Based on earlier results in [@Mafra:2011nv; @Mafra:2011nw], the open superstring amplitudes can be recast as a KLT product of amplitudes in Z-theory and YM theory [@Broedel:2013tta], [$$\begin{aligned}
\label{eq:string1}
A_{\rm open} = A_{\rm Z } \otimes A_{\rm YM}\, ,\end{aligned}$$]{} where $\otimes$ denotes the field theory KLT product. Transmuting the left-hand side sends $A_{\rm YM}$ to $A_{\rm BS}$ on the right-hand side. Since a KLT product with $A_{\rm BS}$ is trivial, we obtain ${\cal T}[i_1 \cdots i_n] \cdot A_{\rm open} = A_{\rm Z } \otimes A_{\rm BS} = A_{\rm Z} $, which is equivalent to [Eq. ]{} and [Eq. ]{}. Instead applying a longitudinal operator, we obtain $ \mathcal{L} \cdot \mathcal{T}[i_1 i_n ]\cdot A_{\rm open} = A_{\rm Z} \otimes A_{\rm NLSM}$, whose scalar amplitudes coincide with the NLSM at low energies, but curiously does not correspond to abelian Z-theory.
A similar logic also holds for extended gravity and its ultraviolet completion. Transmuting the graviton scattering amplitude yields [$$\begin{aligned}
\label{eq:stringrelation3}
{\cal T}[i_1 \cdots i_n] \cdot {A}_{\rm G+ \alpha'}(h_{i_1}, \cdots, h_{i_n}) =
A_{\rm YM + \alpha' } (g_{i_1}\cdots g_{i_n}) \, .\end{aligned}$$]{} Interestingly, when taking $A_{\rm G + \alpha'} \sim A_{\rm closed}$ to be amplitudes of the closed superstring, we find that the right-hand side corresponds not to open superstring amplitudes, but to the so-called single value projection [@Brown:2013gia] [@Schlotterer:2012ny; @Stieberger:2013wea; @Stieberger:2014hba] of the open-string amplitudes, $A_{\rm YM + \alpha'} \sim A_\textrm{open}^{\rm SV}$. For a detailed discussion on the single-value projection, see for instance [@Stieberger:2014hba] and references there in. In any case, the above claim is understood by the convenient KLT form for closed superstring amplitudes [$$\begin{aligned}
\label{eq:closestringKLT2}
{A}_{\rm closed } =
A_\textrm{open }^{\rm SV} \otimes
A_{\rm YM} \,,\end{aligned}$$]{} where the KLT kernel is for field theory rather than for string theory [@Kawai:1985xq].
Examples {#sec:example}
========
To familiarize the reader with the mechanics of transmutation, let us analyze a few concrete examples. Later on, some of these will serve as base cases for a proof by induction of our claims.
Three-Particle Amplitudes
-------------------------
To begin, consider the three-particle scattering amplitude of gluons, [$$\begin{aligned}
A(g_1 ,g_2, g_3) = \frac{1}{2} e_1 e_2 (p_2 e_3 - p_1 e_3) + \textrm{cyclic} \, ,\end{aligned}$$]{} where the YM color ordering $1,2,3$ is implicitly assumed and only dual color ordering is shown. By applying the trace operator, we transmute any two gluons into biadjoint scalars, $$\begin{split}
A(\phi_1 \phi_2, g_3) &= {\cal T}[12]\cdot A(g_1, g_2, g_3) = \frac{1}{2} (p_2 e_3 - p_1 e_3) \\
A(\phi_2 \phi_3, g_1) &= {\cal T}[23]\cdot A(g_1, g_2, g_3) = \frac{1}{2}(p_3 e_1 - p_2 e_1) \\
A(\phi_3 \phi_1, g_2) &= {\cal T}[31]\cdot A(g_1, g_2, g_3) = \frac{1}{2}(p_1 e_2 - p_3 e_2) \, .
\end{split}
\label{eq:A3example}$$ Alternatively, we can transmute all three gluons simultaneously, [$$\begin{aligned}
A(\phi_1 \phi_2 \phi_3) &= {\cal T}[123]\cdot A(g_1 , g_2, g_3) = 1 \, ,
\label{eq:phi3amp}\end{aligned}$$]{} into biadjoint scalars with the dual color ordering $1,2,3$ here. Here we have dropped the dimensionful cubic coupling associated with the BS theory. For the reverse ordering $3,2,1$, we apply [$$\begin{aligned}
A(\phi_3 \phi_2 \phi_1) &= {\cal T}[321]\cdot A(g_1, g_2, g_3) = -1 \, .\end{aligned}$$]{} The relative minus sign for the second case captures the anti-symmetry of the dual color factor. Finally, applying the longitudinal operator to [Eq. ]{} yields $$\begin{split}
A(\pi_1, \pi_2, \pi_3) &= A(\phi_1 \phi_2, \pi_3) = {\cal L}\cdot {\cal T}[12]\cdot A(g_1, g_2, g_3) \\
&= A(\phi_2 \phi_3, \pi_1) = {\cal L}\cdot {\cal T}[23]\cdot A(g_1, g_2, g_3) \\
&= A(\phi_3 \phi_1, \pi_2) = {\cal L}\cdot {\cal T}[31]\cdot A(g_1, g_2, g_3) = 0 \, ,
\end{split}$$ where as noted earlier, we can make any choice for the pair of biadjoint scalars to obtain the permutation invariant pion amplitude. As expected, the three-particle amplitude for pions is zero.
Four-Particle Amplitudes
------------------------
Next, let us apply transmutation to the four-particle scattering amplitude of gluons. The expression from color-ordered Feynman diagrams yields[^4] [$$\begin{aligned}
A(g_1,g_2,g_3,g_4) &= \frac{n_{[12][34]}}{p_1 p_2}+\frac{n_{[23][41]}}{p_2 p_3}
-\frac{1}{2}\left[ ({e_{1}e_{2}}) ({e_{3}e_{4}})+({e_{2}e_{3}}) ({e_{4}e_{1}}) \right]
+ ({e_{1}e_{3}}) ({e_{2}e_{4}}) \, ,\end{aligned}$$]{} where the numerator of the cubic diagram is $n_{[12][34]} = n_{1234} - n_{2134} - n_{1243} + n_{2143} $ and [$$\begin{aligned}
n_{1234} =&
\frac{1}{4} ({e_{1}e_{2}}) ({e_{3}e_{4}}) ({p_{1}p_{3}})
-({e_{2}e_{3}}) ({p_{3}e_{4}}) ({p_{2}e_{1}})
-\frac{1}{2} ({e_{1}e_{2}}) ({p_{2}e_{3}})({p_{1}e_{4}})
-\frac{1}{2} ({e_{3}e_{4}}) ({p_{4}e_{1}}) ({p_{3}e_{2}}) \, , \end{aligned}$$]{} and similarly for $n_{[23][41]}$ up to relabelling. Again, the YM color ordering $1,2,3,4$ is being suppressed and we only display the ordering of the dual color group. By transmuting two pairs of gluons, we obtain the scattering amplitudes for biadjoint scalars, $$\begin{split}
A(\phi_1 \phi_2, \phi_3 \phi_4) &= {\cal T}[12]\cdot {\cal T}[34]\cdot A(g_1,g_2,g_3,g_4) = \frac{p_1 p_3}{p_1 p_2} \\
A(\phi_1 \phi_3, \phi_2 \phi_4) &= {\cal T}[13]\cdot {\cal T}[24]\cdot A(g_1,g_2,g_3,g_4) = 1 \, .
\end{split}$$ Note the presence of a contact quartic scalar interaction in both cases.
Meanwhile, the transmuted single trace four-particle amplitudes are $$\begin{split}
A(\phi_1 \phi_2, g_3, g_4) & = {\cal T}[12]\cdot A(g_1, g_2, g_3, g_4) =
\frac{(p_1 e_3) (p_2 e_4) + (p_1 p_3) (e_3 e_4) - (p_1 e_4) (p_2 e_3) }{p_1 p_2}
-\frac{ (p_1 e_4) (p_2 e_3) }{p_2 p_3} \\ A(\phi_1 \phi_2 \phi_3, g_4) & = {\cal T}[123]\cdot A(g_1, g_2 ,g_3, g_4) =\frac{p_3 e_4}{p_1 p_2}-\frac{p_1 e_4}{p_1 p_4} \\
A(\phi_1 \phi_2 \phi_3 \phi_4) & = {\cal T}[1234]\cdot A(g_1, g_2, g_3, g_4) = \frac{1}{p_1 p_2} + \frac{1}{p_2 p_3} \, .
\end{split}
\label{eq:4pt_example}$$ Note that the ordering in the transmutation operator is crucial because it dictates the order in the dual color trace. In in particular, the relative ordering of the original and dual color traces will affect the resulting scattering amplitude. So for example, we find that $$\begin{split}
A(\phi_1 \phi_3 \phi_2 \phi_4) &= {\cal T}[1324]\cdot A(g_1, g_2, g_3, g_4) = -\frac{1}{p_2 p_3} \\
A(\phi_1 \phi_3 \phi_4 \phi_2) &= {\cal T}[1342]\cdot A(g_1, g_2, g_3, g_4)= -\frac{1}{p_1 p_2} \, ,
\end{split}$$ so as expected, these amplitudes exhibit factorization channels for particles which are adjacent in both the original color and the dual color. Furthermore, we have, [$$\begin{aligned}
A(\phi_1 \phi_2 \phi_3 \phi_4) + A(\phi_1 \phi_3 \phi_2 \phi_4) + A(\phi_1 \phi_3 \phi_4 \phi_2) = 0 \, ,\end{aligned}$$]{} which is the transmuted form of the $U(1)$ decoupling relation.
Lastly, let us consider transmutation via the longitudinal operators. Again applying the freedom of choosing an initial pair of biadjoint scalars, we obtain $$\begin{split}
A(\pi_1, \pi_2, \pi_3, \pi_4) &= A(\phi_1 \phi_2, \pi_3, \pi_4) =
{\cal L}\cdot {\cal T}[12]\cdot A(g_1, g_2, g_3, g_4) \\
&= A(\phi_1 \phi_3, \pi_2, \pi_4) = {\cal L} \cdot {\cal T}[13]\cdot A(g_1, g_2, g_3, g_4) = p_1 p_3 \, , \end{split}$$ which is the correct four-particle color-ordered pion amplitude.
Gluon and Graviton Amplitudes {#sec:BCJ}
-----------------------------
Last but not least, consider amplitudes comprised of all gluons or all extended gravitons. We start by proving [Eq. ]{} in the special case where all the external states are gluons. A tree amplitude in YM can be expressed as, [$$\begin{aligned}
A^{\rm full}_{\rm YM} = \sum_i \frac{N_i C_i}{D_i} \, ,\end{aligned}$$]{} where $C_i$ are the color factors and $N_i$ are respectively the color and kinematic BCJ numerators associated with a cubic graph $i$. The color and kinematic numerators can be expanded in the Del Duca-Dixon-Maltoni half-ladder bases [@DelDuca:1999rs], [$$\begin{aligned}
N_i = \sum_\alpha \sigma_{i}(\alpha) N(\alpha) \qquad \textrm{and} \qquad C_i = \sum_\beta \sigma_{i}(\beta) C(\beta) \, .\end{aligned}$$]{} Projecting the full amplitude into the color ordering $\beta$, [$$\begin{aligned}
A_{\rm YM}(\beta) = \sum_\alpha N(\alpha) A_{\rm BS}(\alpha |\beta) \, ,\end{aligned}$$]{} where $A_{\rm BS}(\alpha | \beta) $ is actually the double-trace biadjoint scalar amplitude [$$\begin{aligned}
A_{\rm BS}(\alpha | \beta) = \sum_i \frac{\sigma_{i}(\alpha) \sigma_{i}(\beta) }{D_i} \, ,\end{aligned}$$]{} which forms a basis of doubly color-ordered biadjoint scalar amplitudes. As we will show shortly, the operator in [Eq. ]{} is constructed so that [$$\begin{aligned}
{\cal T}[\alpha] \cdot N(\beta) = \delta_{\alpha \beta} \, , \label{eq:inverse}\end{aligned}$$]{} which is one if $\alpha$ and $\beta$ are the same permutation and zero otherwise. Applying this to the color-ordered gluon amplitude yields precisely [Eq. ]{}. Thus, this operator extracts precisely the doubly color-ordered biadjoint scalar amplitude from the color-ordered YM amplitude.
Meanwhile, we can lift this construction to extended gravity by again utilizing the double copy, [$$\begin{aligned}
A_{\rm G} = \sum_i \frac{\tilde N_i N_i }{D_i} = \sum_{\alpha, \beta} \tilde N(\alpha) A_{\rm BS}(\alpha|\beta)
N(\beta) = \sum_{ \beta} N(\beta) A_{\rm YM}(\beta) \, .\end{aligned}$$]{} Using the relation in [Eq. ]{} we find that [$$\begin{aligned}
{\cal T}[\alpha] \cdot A_{\rm G} = A_{\rm YM}(\alpha) \, , \label{eq:GRtoYM}\end{aligned}$$]{} so we obtain color-ordered gluon amplitudes from graviton amplitudes.
To derive [Eq. ]{} we will use the fact that the BCJ numerator for the gluon amplitude takes the schematic form [$$\begin{aligned}
N(\beta) \sim (e e)(pe)^{n-2} + (e e)^2(pe)^{n-4}(pp) + \ldots \, ,
\label{eq:schematic}\end{aligned}$$]{} where the ellipses denote terms with higher and higher powers of $(ee)$. The above schematic expression of $N(\beta)$ is determined simply by the power counting. By construction, the operator ${\cal T}[\alpha]$ eliminates a single factor of $(ee)$ and $n-2$ factors of $(pe)$, so it projects out all but the first term. Because this term carries no powers of $(pp)$, this term in the numerator can never cancel propagator poles, so it must correspond to the pure cubic Feynman diagram contributions to the gluon amplitude. It is trivial to compute the cubic Feynman diagram contribution, yielding [$$\begin{aligned}
N(\alpha) = e_{\beta_1} e_{\beta_n} \prod_{i=2}^{n-1} \sum_{j=1}^{i-1} p_{\beta_j} e_{\beta_i} + \ldots \, ,\end{aligned}$$]{} where the ellipses denote numerous other permutations of terms that are all annihilated by ${\cal T}[\alpha]$. It finishes the proof of Eq. (\[eq:inverse\]). Finally we remark that Eq. (\[eq:GRtoYM\]) may also be argued applying the statement of [@Arkani-Hamed:2016rak], which states that scattering amplitudes in YM theories are uniquely fixed by the requirement of simple poles, power counting and gauge invariance, and the operator ${\cal T}[\alpha]$ clearly preserves all those properties.
Proof by Induction {#sec:proof}
==================
We now present an induction proof for the unifying relations summarized in [Sec. \[sec:web\]]{} and depicted diagrammatically in [Fig. \[fig:unified\_web\]]{}. In particular, we show that ${\cal T}_{ij}$, ${\cal T}_{ijk}$, and ${\cal L}$ transmute physical amplitudes into other physical amplitudes. In our earlier discussion, we observed that ${\cal T}_{ijk}$ and ${\cal L}$ are only gauge invariant when applied to an amplitude already containing states transmuted by ${\cal T}_{ij}$. For this reason we must initialize the induction with a base case: the amplitude in [Eq. ]{}, which describes any number of pairs of biadjoint scalars in distinct dual color traces interacting with gluons. Any single or multiple trace amplitude is obtained by applying ${\cal T}_{ijk}$ and ${\cal L}$ to this amplitude. Other useful base cases are the three- and four-particle amplitudes derived in [Sec. \[sec:example\]]{}.
In order to establish the unifying relations we demonstrate that a transmuted amplitude satisfies necessary and sufficient conditions that define a physical amplitude. This proof occurs in three steps. First, we verify that the transmuted amplitude is on-shell constructible, [*i.e.*]{} determined by the residues on each factorization channel and hence defined via on-shell recursion relations. As we will see, it will suffice that the amplitude vanish either at large momentum transfer or in the soft limit. Second, we check explicitly that the transmuted amplitude factorizes properly into products of physical amplitudes, using crucially that the same is true of the original amplitude by the induction hypothesis. In this way we prove the unifying relations for single and multiple trace scattering amplitudes of gluons, biadjoint scalars, and pions summarized in [Eq. ]{} and portrayed collectively by the lower triangle in Fig. \[fig:unified\_web\]. Third, we use the double copy construction to extend our proof to the remaining unifying relations listed in [Eq. ]{} and [Eq. ]{} and depicted in [Fig. \[fig:unified\_web\]]{}.
On-shell Constructibility
-------------------------
Consider an initial amplitude $A$ comprised of gluons and biadjoint scalars. Using the results of [@Britto:2004ap; @Britto:2005fq; @Cohen:2010mi; @Cheung:2015cba], $A$ is is on-shell constructible by an all-line recursion relation. In particular, for greater than four particles, the deformed amplitude $A$ falls off at large $z$ for an all-line momentum shift parameterized by $z$. This result follows straightforwardly from dimensional analysis and does not require detailed information about the amplitude [@Cohen:2010mi; @Cheung:2015cba]. Applying the insertion operator to $A$ yields a new object ${\cal T}_{ijk} \cdot A$. However, since ${\cal T}_{ijk} \cdot A$ carries strictly fewer derivatives than $A$, we can apply the exact same power counting argument from [@Cohen:2010mi; @Cheung:2015cba] to establish vanishing large $z$ behavior of ${\cal T}_{ijk} \cdot A$. Thus, we determine that this object is also on-shell constructible.
Instead applying the longitudinal operator, we obtain a transmuted object ${\cal L}\cdot A$ with more derivatives than $A$, so large $z$ fall-off is not guaranteed. Fortunately, an amplitude without large $z$ fall-off can still be on-shell constructible if it has compensating infrared properties like the Adler zero or its generalizations [@Cheung:2015ota; @Luo:2015tat]. As we will show in [Sec. \[sec:soft\]]{}, the pions in ${\cal L} \cdot A$ indeed manifest the Adler zero as a transmutation of the Weinberg soft theorems of gauge theory. This applies even in the presence of spectator biadjoint scalars. In particular, the growth in large $z$ behavior incurred from the derivatives in ${\cal L}$ is always balanced by the Adler zeros arising from the resulting pions.
Factorization
-------------
Since the transmuted amplitude has either large $z$ fall-off or vanishing soft limits, our remaining task is to verify proper factorization. By the assumption of the induction hypothesis, the original initial $A$ factorizes according to [$$\begin{aligned}
A(\cdots) &\sim \sum_{I } A_L(\cdots I_L) \, A_R(I_R \cdots) \, ,\end{aligned}$$]{} where $\sim$ denotes the residue on the factorization channel of the singular propagator, $1/p_L^2 = 1/p_R^2$. Here the sum on $I$ runs over all possible gluon and scalar internal states. In what follows, we certify the proper factorization of ${\cal T}_{ijk} \cdot A$ and ${\cal L} \cdot A$, respectively.
### Insertion Operator {#insertion-operator .unnumbered}
As shown in [Eq. ]{}, the insertion operator ${\cal T}_{ijk}$ transmutes the gluon $g_j$ into a biadjoint scalar, inserting it between the two biadjoint scalars $\phi_i$ and $\phi_k$ which are adjacent in the dual color. There are a handful of possible arrangements of these states relative to the factorization channel. Let us consider each case in turn, demonstrating that the resulting amplitude factorizes correctly.
[*Gluon on Same Side as Scalars.*]{} If the gluon and both biadjoint scalars are on the same side of the factorization channel, then initial scattering amplitude factorizes according to $$A(\cdots\phi_i \phi_k\cdots, g_j, \cdots) \sim \sum_I A_L(\cdots\phi_i \phi_k\cdots, g_j,\cdots, I_L) \, A_R(I_R,\cdots) \, .$$ Applying an insertion operator to both sides, we otain $$\begin{split}
&{\cal T}_{ijk} \cdot A(\cdots\phi_i \phi_k\cdots, g_j, \cdots) \sim \sum_I A_L(\cdots\phi_i \phi_j \phi_k\cdots, \cdots, I_L) \, A_R(I_R, \cdots) \, ,
\end{split}$$ where in the second line we have used the induction hypothesis, which is that the claim in [Eq. ]{} holds for all lower point amplitudes. Hence, in this case the left-hand side factorizes correctly, [*i.e.*]{} as the transmuted $A(\cdots \phi_i \phi_j\phi_k \cdots, \cdots)$ should.
[*Gluon on Opposite Side as Scalars.*]{} If the gluon is on the opposite side of the factorization channel from the biadjoint scalars, then $$\begin{split}
& A(\cdots\phi_i \phi_k\cdots , g_j, \cdots) \sim \sum_I A_L(\cdots\phi_i \phi_k\cdots, I_L) \, A_R(I_R, g_j \cdots) \\
&\sim A_L(\cdots\phi_i \phi_k\cdots \phi_L) \, A_R(\phi_R, g_j \cdots) +A_L(\cdots\phi_i \phi_k\cdots , g_L) \, A_R(g_R, g_j \cdots) \, ,
\end{split}$$ where the first and second terms correspond to intermediate scalar and gluon exchange, respectively. Next, consider ${\cal T}_{ijk}$ applied to each of these terms.
The first term is annihilated by ${\cal T}_{ijk}$ because the gluon polarization $e_j$ can only dot into the combination $p_i + p_k$ because the intermediate particle is a scalar and cannot transmit differences of momenta across the channel. This agrees with the factorization of $A(\cdots \phi_i \phi_j \phi_k\cdots, \cdots)$, which should vanish on this factorization channel since it requires $\phi_j$ to be adjacent to other external scalars in $A_R$ but also to be inserted between $\phi_i$ and $\phi_k$, which is impossible.
The second term is not annihilated by ${\cal T}_{ijk}$, but since the intermediate particle is a gluon we can apply the completeness relation, [$$\begin{aligned}
\partial_{p_i e_j } - \partial_{p_k e_j} \rightarrow (\partial_{p_i e_{L} } - \partial_{p_k e_{L}}) \partial_{e_R e_{j}} \, ,\end{aligned}$$]{} which expressed in terms of transmutation operators is [$$\begin{aligned}
{\cal T}_{ijk} \rightarrow {\cal T}_{iLk} {\cal T}_{Rj} \, .\end{aligned}$$]{} Applying this to the factorization equation, we obtain $${\cal T}_{ijk} \cdot A(\cdots\phi_i \phi_k\cdots, g_j,\cdots) \sim A_L(\cdots \phi_i \phi_L \phi_k\cdots,\cdots) \, A_R(\phi_R \phi_j, \cdots) \, ,$$ which again factorizes as $A(\cdots \phi_i \phi_j \phi_k\cdots, \cdots)$ should.
[*Scalars on Opposite Sides.*]{} The final case arises when the two biadjoint scalars are on opposite sides of the factorization channel. Since $\phi_i$ and $\phi_k$ are by assumption within the same dual color trace, this means that a biadjoint scalar must be exchanged across the factorization channel $$A(\cdots\phi_i \phi_k\cdots, g_j, \cdots)
\sim A_L(\cdots\phi_i \phi_L\cdots, g_j,\cdots) \, A_R(\phi_R \phi_k,\cdots) \, .
$$ ${\cal T}_{ijk}$ only extracts the dependence on the momentum difference between $p_i$ and $p_k$. However, only the total momenta through the factorization channel, $p_L$, so $p_k$ can only enter through the combination of momenta $p_L$. This implies that on this factorization channel, [$$\begin{aligned}
\partial_{p_i e_j } - \partial_{p_k e_j} \rightarrow \partial_{p_i e_j } - \partial_{p_L e_j} \, ,\end{aligned}$$]{} which in terms of transmutation operators implies [$$\begin{aligned}
{\cal T}_{ijk}\rightarrow {\cal T}_{ijL} \, .\end{aligned}$$]{} Applying this to the factorization equation we obtain $${\cal T}_{ijk}\, A(\cdots\phi_i \phi_k\cdots, g_j,\cdots) \sim A_L(\cdots\phi_i \phi_j \phi_L\cdots,\cdots) \, A_R(\phi_R \phi_k,\cdots) \, ,$$ which factorizes in the same way as $A(\cdots \phi_i \phi_j \phi_k\cdots, \cdots)$.
### Longitudinal Operator {#longitudinal-operator .unnumbered}
As shown in [Eq. ]{}, the longitudinal operator transmutes all the gluons in a single trace amplitude into pions, leaving spectator biadjoint scalars untouched. To prove that the resulting amplitude is physical, it will be convenient to split ${\cal L}$ according to its contributions coming from each side of the factorization channel, [$$\begin{aligned}
{\cal L} = {\cal L}_L \cdot {\cal L}_R \qquad \textrm{where} \qquad {\cal L}_L = \prod_{i\in L} {\cal L}_i \qquad \textrm{and} \qquad {\cal L}_R = \prod_{i\in R} {\cal L}_i \, .
\label{eq:L_fact}\end{aligned}$$]{} Note that the ${\cal L}_i$ factors still depend on momenta on both sides of the factorization channel, given the definition in [Eq. ]{}. As before, there are several possible configurations of the biadjoint scalars in the amplitude relative to the factorization channel, which we now consider in turn.
[*Scalars on Both Sides.*]{} By assumption, the starting amplitude is single trace. Since there are biadjoint scalars on either side of the factorization channel, the exchanged particle must also be a biadjoint scalar, so [$$\begin{aligned}
A(\cdots\phi_i \phi_l\cdots, g_j, g_k,\cdots)\sim A_L(\cdots\phi_i \phi_L, g_j, \cdots) \, A_R(\phi_R \phi_l \cdots, g_k,\cdots) \, ,\end{aligned}$$]{} where $g_j$ and $g_k$ are by definition gluons on to the left and right of the factorization channel. Next, we consider a longitudinal operator ${\cal L}_j$ acting on gluon $g_j$. Since the intermediate particle is a scalar, the polarization vector $e_j$ can only couple to $p_k$ for $k\in R$ through the combination given by the total momentum flowing through the channel, $p_I =\sum_{k \in R} p_k$. On the factorization channel, ${\cal L}_j$ is effectively equivalent to [$$\begin{aligned}
{\cal L}_j &=
\sum_{l \in L} p_l p_j \partial_{p_l e_j} + \sum_{k \in R} p_k p_j \partial_{p_k e_j}
\rightarrow \sum_{l \in L} p_l p_j \partial_{p_l e_j} + p_I p_j \partial_{p_I e_j} = {\cal L}'_j \, .
\label{eq:L_fact_easy}\end{aligned}$$]{} where we effectively replace $\partial_{p_k e_j}$ with $\partial_{p_I e_j}$. We recognize ${\cal L}'_j$ as a longitudinal operator acting purely on states in the left lower-point amplitude. Applying the same logic to the right, we obtain $$\begin{split}
{\cal L} \cdot A(\cdots\phi_i \phi_l\cdots, g_j, g_k,\cdots) &\sim {\cal L}_L \cdot A_L(\cdots\phi_i \phi_L, g_j, \cdots) \, {\cal L}_R \cdot A_R(\phi_R \phi_l \cdots, g_k,\cdots) \\
&\sim A_L(\cdots\phi_i \phi_L, \pi_j,\cdots) \, A_R(\phi_R \phi_l \cdots,\pi_k,\cdots) \, ,
\end{split}$$ which is the proper factorization equation for the transmuted amplitude, $A(\cdots\phi_i \phi_l\cdots, \pi_j, \pi_k,\cdots)$.
[*Scalars on Same Side.*]{} When all of the biadjoint scalars are on the same side of factorization channel, the original amplitude factorizes with an intermediate gluon, [$$\begin{aligned}
A(\cdots\phi_i \phi_l\cdots, g_j, g_k,\cdots) \sim A_L(\cdots\phi_i \phi_l\cdots , g_j, g_L,\cdots) \, A_R(g_R, g_k,\cdots) \, .
\label{eq:pion_case2}\end{aligned}$$]{} Similar to the previous case, on the factorization channel, ${\cal L}_j$ is effectively equivalent to $$\begin{aligned}
{\cal L}_j
&\rightarrow {\cal L}'_j+ \sum_{q \in R} p_q p_j \partial_{p_q e_R} \cdot {\cal T}_{L j} \, ,
\label{eq:Lleft_expand}\end{aligned}$$ where ${\cal T}_{L j}=\partial_{e_L e_j}$ is the trace operator acting on $g_{j}$ and $g_L$. The additional term compared with Eq. accounts for the fact that the intermediate particle is now a gluon, so the polarization vector $e_j$ can couple to $p_k$ through $p_k e_R$ via the completeness relation. Likewise we have $$\begin{aligned}
{\cal L}_k
&\rightarrow {\cal L}'_k+ \sum_{l \in L} p_l p_k \partial_{p_l e_L} \cdot {\cal T}_{R k} \,,\end{aligned}$$ for $k\in R$. Plugging the above into ${\cal L}_R$, this becomes the logitudinal operators for the right sub-amplitude except for the two gluons $g_{k,R}$, whose transmutation is given by the extra term in the above equation. The trace operator contained there turns gluons into biadjoint scalars, which is exactly what we need to form pion amplitude on the right. Effectively, $${\cal L}_R \rightarrow
\sum_{k \in R, l \in L} p_k p_l \partial_{p_l e_L} \cdot ({\cal L}'_{R} \cdot {\cal T}_{R k}) = {\cal L}'_{I}\cdot ({\cal L}'_{R} \cdot {\cal T}_{R k}) \, ,$$ where $ {\cal L}'_{I}$ is the longitudinal operator for the intermediate gluon on the left and $({\cal L}'_{R} \cdot {\cal T}_{R k})$ is the operator taking YM sub-amplitude into NLSM one. To get the second equality, we use the fact that $({\cal L}'_{R} \cdot {\cal T}_{R k})$ yields the same pion amplitude for any $k$, so the summation on $k$ can be carried out using $\sum_{k \in R} p_k = p_I$, which then gives ${\cal L}'_{I}$. We find the ${\cal L}_R$ on the factorization channel simplifies into two parts: transforming the YM sub-amplitude on the right into a NLSM one as well as taking the internal gluon on the left into a pion. Applying this to the full amplitude yields $$\begin{split}
{\cal L} \cdot A(\cdots\phi_i \phi_l\cdots, g_j, g_k,\cdots)
&\sim {\cal L}_L \cdot {\cal L}_R \cdot A_L(\cdots\phi_i \phi_l\cdots , g_j, g_L,\cdots) \, A_R(g_R, g_k,\cdots) \\
&\sim {\cal L}_L \cdot {\cal L}'_I\cdot A_L(\cdots\phi_i \phi_l\cdots , g_j, g_L,\cdots) \, A_R(\pi_R, \pi_k\cdots) \\
&\sim A_L(\cdots\phi_i \phi_l\cdots, \pi_j,\pi_L ,\cdots ) \, A_R(\pi_R , \pi_k\cdots) \, ,
\end{split}$$ which factorizes correctly with an internal pion. Going from second to third line, the second term in Eq. is irrelevant because $g_L$ has been transmuted. The rest of operators in ${\cal L}_L {\cal L}'_I$ transmute gluons in $A_L$ into pions, yielding the desired factorization.
Double Copy Construction
------------------------
Our induction proof establishes the unifying relations which are summarized in [Eq. ]{} and depicted in the lower triangle of Fig. \[fig:unified\_web\]. To derive additional unifying relations, we simply employ the double copy construction in the spirit of [@Cachazo:2014xea]. Crucially, since transmutation is independent of the particular representation chosen for the amplitude, this procedure commutes with the KLT relations. In particular, recall that the KLT construction dictates that $$\begin{aligned}
A_{\rm BI} = A_{\rm YM} \otimes \tilde{A}_{\rm NLSM} \qquad \textrm{and} \qquad
A_{\rm G} = A_{\rm YM} \otimes \tilde{A}_{\rm YM} \, .\end{aligned}$$ By taking the KLT product of the amplitudes in [Eq. ]{} with those of YM theory and the NLSM, we obtain the unifying relations for extended gravity and BI shown in [Eq. ]{} and [Eq. ]{}, respectively, and represented by the other triangles in Fig. \[fig:unified\_web\]. This completes our proof.
Infrared Structure {#sec:soft}
==================
Since transmutation is a functional relationship between scattering amplitudes, it applies in all kinematic regimes. Nevertheless, it will be enlightening to study this operation specifically in the soft limit, where certain universal structures arise. In particular, the gluon and graviton scattering amplitudes famously obey soft theorems taking the schematic form $$A^{(n)} \xrightarrow{p_i \rightarrow 0} {\cal S}^{(i)} \cdot A^{(n-1)} \, ,
\label{eq:soft_original}$$ in the limit where particle $i$ is soft. Here we have defined the soft operator, ${\cal S}^{(i)}$, which acts on the lower-point amplitude $A^{(n-1)}$ to yield $A^{(n)}$ in the soft limit. At leading order in the soft limit, ${\cal S}^{(i)}$ is a simple multiplicative factor, but at higher order it acts as a differential operator on the hard external kinematic data.
Next, let us study the image of these soft theorems for gluons and gravitons under transmutation. Consider a generic product of transmutation operators ${\cal T}^{(n)}$ acting on $A^{(n)}$. For later convenience, we factorize ${\cal T}^{(n)}$ into [$$\begin{aligned}
{\cal T}^{(n)}={\cal T}^{(i)}\cdot {\cal T}^{(n-1)} \, ,\label{eq:Tn}\end{aligned}$$]{} where ${\cal T}^{(i)}$ transmutes the soft particle and ${\cal T}^{(n-1)}$ transmutes the remaining hard particles. Since ${\cal T}^{(n-1)}$ and $A^{(n-1)}$ are by definition independent of the soft particle, they do not contain $e_i$ and so $[\, {\cal T}^{(i)} \, ,\, {\cal T}^{(n-1)} \, ]=0$ and ${\cal T}^{(i)} \cdot A^{(n-1)}=0$. Using this observation, we apply the transmutation operator in [Eq. ]{} to the right-hand side of Eq. , yielding $$\begin{split}
{\cal T}^{(n)} \cdot A^{(n)} \xrightarrow{p_i \rightarrow 0} {\cal T}^{(i)} \cdot {\cal T}^{(n-1)} \cdot {\cal S}^{(i)} \cdot A^{(n-1)}
= [ \, {\cal T}^{(i)} \cdot {\cal T}^{(n-1)}\, ,\, {\cal S}^{(i)} \, ]\cdot A^{(n-1)} \, .
\end{split}$$ The hard trace and insertion operators commute with the soft operator, so $[\, {\cal T}^{(n-1)}\, , \, {\cal S}^{(i)} \, ]=0$ and the above equation further simplifies to $$\begin{split}
{\cal T}^{(n)} \cdot A^{(n)} \xrightarrow{p_i \rightarrow 0} [ \, {\cal T}^{(i)} \, ,\, {\cal S}^{(i)} \, ] \cdot {\cal T}^{(n-1)}\cdot A^{(n-1)} \, .
\end{split}
\label{eq:soft_transmutation}$$ Since ${\cal T}^{(n-1)}\cdot A^{(n-1)}$ is by definition the transmuted lower-point amplitude, this implies that the transmuted soft operator is simply $ [ \, {\cal T}^{(i)} \, ,\, {\cal S}^{(i)} \, ]$. In what follows let us discuss the transmuted cousins of the leading and subleading soft theorems.
Soft Theorems at Leading Order
------------------------------
Our jumping-off point is the leading order soft graviton factor of Weinberg [@Weinberg:1965nx], [$$\begin{aligned}
{\cal S}_{\rm G}^{(i)}= \sum_{ j \neq i} \frac{ p_i e_j \, p_i \tilde e_j }{ p_{i} p_j } \, ,\end{aligned}$$]{} where we have maintained a factorized form for the polarizations to emphasize our focus on extended gravity. We obtain the leading soft gluon factor by taking the commutator of a transmutation operator and the leading soft graviton factor.
Applying the trace operator, we find a vanishing commutator, [$$\begin{aligned}
{\cal S}_{{\rm EM}}^{(ij)} &= [\, {\cal T}_{ij} \, ,\, {\cal S}_{{\rm G}}^{(i)} \,] = 0 \, ,\end{aligned}$$]{} which correctly establishes the vanishing of soft photon amplitudes in EM theory at leading order. For the insertion operator, this reproduces the known leading soft gluon factor, [$$\begin{aligned}
{\cal S}_{{\rm YM}}^{(ijk)} &= [ \, {\cal T}_{ijk} \, ,\, {\cal S}_{{\rm G}}^{(j)} \, ] = \frac{ p_i e_j}{ p_i p_j } -\frac{ p_k e_j}{ p_k p_j } \, ,\end{aligned}$$]{} corresponding to a soft particle $j$ adjacent to particles $i$ and $k$. The differential operators here act implicitly on the tilded polarizations. Note that the appearance of color ordering is given by the definition of the insertion operator.
Applying the trace operator to the soft gluon factor, we obtain [$$\begin{aligned}
{\cal S}_{{\rm YMS}}^{(ijk|jl)} &= [ \, {\cal T}_{jl} \, ,\, {\cal S}_{{\rm YM}}^{(ijk)} \, ] = 0 \, ,\end{aligned}$$]{} where the vertical bar in the superscript separates the original and dual color orderings, so particle $j$ in the same dual color trace as particle $l$. As expected, we find that soft scalar amplitudes of YMS also vanish at leading order. Instead, if we apply the insertion operator again, we obtain the leading soft biadjoint scalar soft factor, [$$\begin{aligned}
S_{{\rm BS}}^{(ijk|ljm)} &= [ \, {\cal T}_{ljm} \, ,\, {\cal S}_{{\rm YM}}^{(ijk)} \, ] = \frac{\Delta_{lim}}{ p_i p_j } -\frac{\Delta_{lkm}}{ p_k p_j } \, ,\end{aligned}$$]{} where $\Delta_{ijk}=\delta_{ij} - \delta_{jk}$ and the differential operators act implicitly on the nontilded polarizations. Here the soft factor is non-zero only when $i$ or $k$ are still adjacent to $j$ in the dual color ordering. Just like the leading soft factors for gluons and gravitons, the leading soft factor for biadjoint scalars can also be trivially derived from Feynman diagrams.
Last of all, let us consider the action of the longitudinal operator on the leading soft theorems. As shown in [Eq. ]{}, the longitudinal operator acting on a particle can be written as ${\cal L}_i = {\cal W}_i + \Delta_i$, where ${\cal W}_i$ is the Ward identity operator and thus annihilates a physical amplitude. Since the soft gluon and graviton factors are independent of $e_i e_j$ factors, they are annihilated by ${\cal L}_i$, so the transmuted soft factors are all vanishing. In particular, [$$\begin{aligned}
S_{\rm BI}^{(i)} = S_{\rm NLSM}^{(ijk)} =0 \, ,\end{aligned}$$]{} which is expected since BI theory is a manifestly derivatively coupled theory and the NLSM exhibits the well-known Adler zero [@Adler:1964um]. This reconfirms the observation of [@Cheung:2016prv] that the ${\cal O}(p)$ vanishing of pion amplitudes follows directly from leading soft gluon theorem. Note that the Adler zero derived above also applies to a general amplitude of pions coupled to biadjoint scalars, which is necessary for the proof in [Sec. \[sec:proof\]]{}. Furthermore, a trivial corollary is that [$$\begin{aligned}
S_{\rm SG}^{(i)} = S_{\rm DBI}^{(ij)} =0 \, ,\end{aligned}$$]{} since the vanishing soft limits of BI and NLSM are inherited by their descendants, DBI and SG.
Soft Theorems at Subleading Order
---------------------------------
Next, let us transmute the soft gluon theorem at subleading order to learn its implications for the descendants of YM theory, which are the BS theory, YMS theory, and the NLSM. The soft gluon theorem, including leading and subleading contributions [@Casali:2014xpa; @Schwab:2014xua; @Bern:2014vva; @Broedel:2014fsa], is $${\cal S}_{\rm YM}^{(ijk)} = \frac{p_i e_j -p_j J_i e_j}{p_i p_j}
-(i \leftrightarrow k) \, ,$$ where $p_j J_i e_j = p_j^\mu J_{i\mu \nu} e_j^\nu$ and we have defined [$$\begin{aligned}
J_{i\mu \nu} = p_{i[\mu} \frac{\partial}{\partial{p_i^{\nu]}}}+ e_{i[\mu} \frac{\partial}{\partial{e_i^{\nu]}}} \, ,\end{aligned}$$]{} which is the angular momentum operator of particle $i$, split between orbital and spin components.
At subleading order, the commutator of the trace operator and the soft gluon operator is [$$\begin{aligned}
{\cal S}^{(ijk|jl)}_{\rm YMS}= [ \, {\cal T}_{jl} \, ,\, {\cal S}_{\rm YM}^{(ijk)} \, ] =- \sum_{m\neq i,k}\delta_{lm}{\cal T}_{ilk}+\left \lbrace \frac{{\cal N}_{ijkl}}{p_i p_j}
- \left( i \leftrightarrow k\right) \right \rbrace,
\label{eq:soft_YMS}\end{aligned}$$]{} where ${\cal N}_{ijkl}=\sum_{m\neq i} \left( \delta_{il} p_j e_m- \delta_{lm} p_j e_i \right) {\cal T}_{im}+
\sum_{m\neq i,k} \delta_{il} p_j p_m {\cal T}_{mik} $. This is a new subleading soft theorem for YMS theory. Instead taking the commutator of the insertion operator, we obtain [$$\begin{aligned}
{\cal S}_{\rm BS}^{(ijk|ljm)} = [\, {\cal T}_{ljm} \, , \, {\cal S}_{\rm YM}^{(ijk)} \, ]
= \left\{ \frac{\Delta_{lim} \left(1+p_j \frac{\partial}{ \partial {p_i} }
\right)}{p_i p_j }
-\frac{p_j e_i}{p_i p_j} {\cal T}_{lim}
-\left( i \leftrightarrow k\right) \right\} - {\cal T}_{ilkm} \, ,
\label{eq:soft_BS}\end{aligned}$$]{} which is a new soft theorem for a biadjoint scalar at subleading order. Note that the above expression is valid even for biadjoint scalars coupled to gluons. Finally, let us derive the subleading soft theorem for the NLSM. Here it will be convenient to use the longitudinal operator in the form ${\cal L}_{ij}=-p_i p_j {\cal T}_{ij}$. While [Eq. ]{} requires $[ \, {\cal T}^{(n-1)} \, ,\, S^{(j)} \, ]=0$, this is not strictly speaking true when the transmutation involves longitudinal operators and the soft gluon operator at subleading order. However, the nonzero contribution to the commutator comes from terms of the form $[ \, {\cal T}^{(n-1)} \, , \, \partial_{p_l p_m} \, ]$, which can only arise from the orbital angular operator in the subleading contributions to the soft operator. From the form of the orbital angular momentum operator this implies that the soft polarization $e_j$ must be dotted into a momentum vector, and so this term is automatically annihilated by ${\cal L}_{ij}$. Hence the commutator $[\, {\cal T}^{(n-1)} \, ,\, S^{(j)} \, ]$ effectively vanishes, so we can employ the results obtained for the trace and insertion operators. In particular, exploiting that ${\cal L}_{ij}$ is simply a linear combination of ${\cal T}_{ij}$ operators, then from which we can read off the subleading soft theorem of the NLSM from that of YMS theory in [Eq. ]{}, yielding [$$\begin{aligned}
{\cal S}^{(ijk)}_{\rm NLSM} = \sum_{l\neq i,k} p_j p_l {\cal T}_{ilk} \, .\end{aligned}$$]{} The presence of the insertion operator implies that the subleading soft theorem for pions involves a lower-point amplitude of a biadjoint scalar, in agreement with the results of [@Cachazo:2016njl]. Naively, it should be a straightforward exercise to extend these results to the subleading soft graviton theorem [@Cachazo:2014fwa; @Schwab:2014xua; @Afkhami-Jeddi:2014fia; @Bern:2014vva; @Broedel:2014fsa; @Zlotnikov:2014sva; @Kalousios:2014uva]. This is of particular interest given the central role of the soft graviton as the stress tensor for a CFT whose correlators are flat space scattering amplitudes [@Kapec:2016jld; @Cheung:2016iub]. Unfortunately, there is an obstruction to this path since transmutation requires a extended gravity amplitude which incorporates the graviton, dilaton, and two-form. To the best of our knowledge, no such factorized subleading soft theorem exists for a general representation of the amplitude [@DiVecchia:2015oba; @DiVecchia:2016amo]. One exception to this comes from the SG theory, where transmutation acts symmetrically on the two polarizations associated with the graviton. There the leading and subleading soft graviton operators commute with ${\cal L}_{ij}$, verifying the subleading vanishing ${\cal O}(p^2)$ soft behavior observed in [@Cheung:2014dqa]. Furthermore, as shown in [@Cachazo:2016njl], the subsubleading soft theorem can be extracted from the KLT relations.
Outlook and Future Directions {#sec:outlook}
=============================
We have derived a set of simple differential operators which transmute physical scattering amplitudes into new ones. Applied repeatedly, transmutation then spawns the family tree of theories depicted in [Fig. \[fig:unified\_web\]]{}. Because all of these theories are descendants of extended gravity, they are bred with innate structures like the KLT and BCJ relations, the CHY formulation, and soft theorems. Our results offer several avenues for future work, which we now discuss.
First and foremost is the question of the [*physical*]{} meaning of transmutation. While the trace operator is a trivial implementation of dimensional reduction, the insertion and longitudinal operators are somewhat murkier in interpretation. In principle, one should be able to derive transmutation purely at the level of the Lagrangian. If such a prescription exists, it will likely shed light on the symmetry and algebra for color-kinematic duality observed in [@Cheung:2016prv].
Clarifying the nature of transmutation might give us clue about how to include fermions. The discussion in this paper crucially relies on amplitudes valid in arbitrary spacetime dimension, so naturally only bosons are considered. On the other hand, it is well-known that supersymmetric Ward identities also relate states of different spins, including fermions, albeit they only exists in certain dimensions. It would be of interest to consider the combination of transmutation and supersymmetric Ward identities. For instance, the commutators of the operators between them may lead to a way of applying transmutation on fermions.
A second but related issue concerns what it even means to define a physical theory. Historically, theories are defined by considerations of symmetry. YM theory and gravity are dictated by gauge invariance and general covariance [@Arkani-Hamed:2016rak], while effective theories are controlled by nonlinearly realized symmetries. From the perspective of the S-matrix, these same constraints enter through Lorentz invariance, locality, and infrared properties [@Cheung:2016drk; @Rodina:2016jyz]. We have found here that transmutation maps these otherwise unrelated concepts onto each other. Nevertheless, not every transmuted theory is obviously controlled by a physical principle. In particular, it would be interesting to see what dictates the structure of the BS theory in the spirit of [@Arkani-Hamed:2016rak; @Rodina:2016jyz; @Rodina:2016mbk].
Last but not least is the question of generalizing our results to loop-level amplitudes. Obviously, transmutation applies to integrands on the unitarity cuts where the integrands reduce to a product of tree-level amplitudes. Whether some vestige of transmutation persists for the integrand in generic loop momenta, or even further at the level of full amplitudes remains to be seen but clearly deserves further study.
Acknowledgment {#acknowledgment .unnumbered}
==============
We would like to thank Henrik Johansson and Ellis Yuan for discussions which inspired this work, as well as Bo Feng, Song He, Yu-tin Huang, Jan Plefka, Oliver Schlotterer, Stephen Stieberger, and Wadim Wormsbecher for helpful discussions. CHS and CW are grateful to the KITP, Santa Barbara, for hospitality during the final stages of this work. CC is supported by a Sloan Research Fellowship and CC, CHS, and CW are supported in part by a DOE Early Career Award under Grant No. DE-SC0010255 and by the NSF under Grant No. NSF PHY-1125915.
[^1]: e-mail: [[email protected]]([email protected]), [[email protected]]([email protected]), [[email protected] ]([email protected] )
[^2]: So as not to complicate our notation we omit the usual “$\cdot$” denoting Lorentz-invariant contractions.
[^3]: In the extended gravity, all the degrees of freedom can be incorporated into a single polarization tensor $e_{\mu\nu}$. It can written as a product of two copies of YM polarizations, $e_{\mu\nu}=e_{\mu}\tilde{e}_{\nu}$. The usual pure graviton corresponds to the symmetric and traceless component of $e_{\mu\nu}$. Since the transmutation only acts on one of the copies, we focus on the un-tilded copy without loss of generality.
[^4]: Our normalization convention for amplitudes effectively sets all propagator denominators to $\sum_{i\neq j} p_i p_j$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present results obtained from a series of observations of the supernova remnant RX J1713.7$-$3946 by the [*Suzaku*]{} satellite which cover about two-thirds of the remnant surface. Hard X-rays have been detected from each pointing up to $\sim 40~{\rm keV}$. The hard X-ray spectra are described by power-law functions with photon indices of $\sim 3.0$, which are larger than those in the energy region below 10 keV. The combination of the spatially-integrated XIS and HXD spectra clearly reveals a spectral cutoff in the X-ray spectrum which is linked to the maximum energy of accelerated electrons emitting synchrotron radiation. The broad-band coverage of [*Suzaku*]{} observations from 0.4 keV to 40 keV allows us to derive, for the first time, the energy spectrum of parent electrons in the cutoff region. The inferred cutoff energy in the spatially-integrated X-ray spectrum indicates that the electron acceleration in the remnant proceeds close to the Bohm-diffusion limit. We discuss implications of the spectral and morphological properties of [*Suzaku*]{} data in the context of the origin of nonthermal emission. The [*Suzaku*]{} X-ray and the H.E.S.S. TeV gamma-ray data together hardly can be explained within a pure leptonic scenario, unless we introduce an additional component of relativistic electrons with softer energy spectrum. Moreover, the leptonic models require very weak magnetic field which does not agree with the recently discovered filamentary structure and short-term variability features of the X-ray emitting region. The hadronic models with strong magnetic field provide perfect fits to the observed X-ray and TeV gamma-ray spectra through the synchrotron radiation of electrons and [*p-p*]{} interactions of protons, but require special arrangements of model parameters to explain the lack of thermal component of X-ray emission. For the morphology studies, we compare the X-ray an TeV gama-ray surface brightness maps using the [*Suzaku*]{} XIS and the H.E.S.S. data. We confirm the previously reported strong correlation between X-ray and TeV gamma-ray emission components. At the same time the [*Suzaku*]{} data reveal a deviation from the general tendency, namely, the X-ray emission in the western rim regions appears brighter than expected from the average X-ray to gamma-ray ratio.'
author:
- 'Takaaki Tanaka, Yasunobu Uchiyama, Felix A. Aharonian, Tadayuki Takahashi, Aya Bamba, Junko S. Hiraga, Jun Kataoka, Tetsuichi Kishishita, Motohide Kokubun, Koji Mori, Kazuhiro Nakazawa, Robert Petre, Hiroyasu Tajima, and Shin Watanabe'
title: 'Study of Nonthermal Emission from SNR RX J1713.7$-$3946 with [*Suzaku*]{}'
---
INTRODUCTION
============
Supernova remnants (SNRs) have long been considered to be likely acceleration sites of cosmic-ray particles below the energy of the [*knee*]{}, $\sim 10^{15}~{\rm eV}$. The energy supply to explain the energy density of comic rays is satisfied if $\sim 1$–10% of the energy of each supernova is transferred to accelerated particles. Also, the well developed theory of diffusive shock acceleration nicely explains the universal power-law spectrum of cosmic rays (e.g. @BE87 [@malkov01]). Although synchrotron emission detected in the radio band supports this idea observationally, no evidence of acceleration to TeV energy had been observed until recently. During the last decade, such evidence was revealed through observations of X-rays and TeV gamma rays from several shell-type SNRs. [@koyama95] discovered synchrotron X-rays from the shell of SN 1006, which indicates electrons are accelerated up to multi-TeV energies. This finding was followed by detections of synchrotron X-rays from other SNRs, including RX J1713.7$-$3946 (e.g. @koyama97 [@slane01]). Further evidence for multi-TeV particles (electrons and/or protons) has been provided by discovery of TeV gamma rays from some SNRs, such as Cassiopeia A [@aha01] or RX J1713.7$-$3946 [@muraishi00], although their spectral parameters and morphologies were not well determined due to the limited sensitivity of TeV observatories. Subsequently, high quality morphological and spectral studies have been performed by H.E.S.S. (e.g. @aha04 [@aha06; @aha07]). These pioneering measurements by the H.E.S.S. telescope, together with the high resolution X-ray data, have enabled direct comparison of X-ray and TeV gamma-ray data.
The shell-type SNR RX J1713.7$-$3946 (also known as G347.3$-$0.5), is one of the best-studied SNRs from which both non-thermal X-rays and TeV gamma rays are detected. This SNR was discovered in soft X-rays during the [*ROSAT*]{} All-Sky Survey [@pfe96]. The [*ASCA*]{} satellite, with wider energy coverage than that of [*ROSAT*]{}, revealed that the X-ray spectrum is featureless and can be best interpreted as synchrotron emission from very high energy electrons in the TeV regime [@koyama97; @slane99]. The X-ray spectrum was well fitted with a power-law function of photon index $\Gamma = 2.2$–$2.4$ and interstellar absorption column density $N_{\rm H} = 0.6$–$0.8 \times 10^{22}~{\rm cm}^{-2}$ without any observable evidence for a thermal emission component. Subsequent observations by [*Chandra*]{} and [*XMM-Newton*]{} have unveiled structure with a complex network of bright filaments and knots, in the western part of the SNR [@uchi03; @laz04; @cassam04; @hiraga05].
TeV gamma-ray emission from RX J1713.7$-$3946 was first reported by the CANGAROO collaboration in 1998 [@muraishi00], and confirmed by the subsequent observations with CANGAROO-II in 2000 and 2001 [@enomoto02]. Later, the H.E.S.S. collaboration obtained a resolved image of the source in TeV gamma rays [@aha04] showing that the gamma-ray emission from RX J1713.7$-$3946 arises mainly in the shell. These observations revealed a striking correlation between the X-ray and the gamma-ray images, which indicates a strong connection between the physical processes responsible for X-ray and TeV gamma-ray emission components [@aha06]. Based on the spectral and morphological information, they discussed two possible gamma-ray emission scenarios, one where gamma rays are generated by inverse Compton scattering of accelerated electrons with diffuse radiation fields (the so-called leptonic scenario) and the other where the decay of secondary $\pi^{0}$-mesons is responsible for gamma rays (hadronic scenario). The later observations with H.E.S.S. revealed that the flux extends to 30 TeV and, likely, beyond, which implies particle acceleration up to energies well above 100 TeV for either model [@aha07] .
Most recently, our X-ray observations using [*Chandra*]{} and [*Suzaku*]{} have provided important clues for understanding the acceleration process in the SNR. From a series of observations of the northwest part of the SNR with [*Chandra*]{} in 2000, 2005 and 2006, we discovered that compact regions of the northwest (NW) shell are variable in flux on a one-year time scale [@uchi07]. The fast variability was interpreted as one-year scale acceleration and synchrotron cooling of electrons with amplified magnetic fields of order of 1 mG. Such a large magnetic field in compact regions strongly favors $\pi^0$-decay emission as the origin of TeV gamma rays. Also, thanks to the wide-band coverage of [*Suzaku*]{} and its low background level, we were able to measure a hard X-ray spectrum up to 40 keV from the southwest portion of RX J1713.7$-$3946 with a clear indication of a high-energy cutoff in the synchrotron spectrum (@takahashi07, hereafter Paper I). Combined with the upper limit on a shock speed of $4500~{\rm km}~{\rm s}^{-1}$ placed by [*Chandra*]{}, the cutoff energy determined by the [*Suzaku*]{} observation of the southwest part indicates that particle acceleration within the SNR shock is so efficient that it approaches the theoretical limit corresponding to the so-called Bohm diffusion regime (e.g. @malkov01).
In this paper, we present results of mapping observations of RX J1713.7$-$3946 with [*Suzaku*]{}, which covers about two-thirds of the SNR region with 11 pointings. The low background level of the Hard X-ray Detector (HXD) enables us to detect hard X-ray emission up to $\sim 40~{\rm keV}$ from each of the pointings. At the same time, its small field-of-view (FoV) of $\sim 25^{\prime} \times 25^{\prime}$ FWHM gives us information about the spatial distribution of hard X-ray emission and spectral differences from region to region. Thanks to its low instrumental background and large effective area, the other detector system aboard [*Suzaku*]{}, the X-ray Imaging Spectrometer (XIS), also uncovers new observational facts, such as spectral features below 10 keV and the morphology of relatively dim regions left unclear in previous studies by [*ASCA*]{}, [*Chandra*]{}, and [*XMM-Newton*]{}. By combining the XIS and HXD spectra summed over the data from all the pointings, we show a wide-band X-ray spectrum (0.4–40 keV) with quite high statistics, with which we investigate not only the existence of a cutoff, but also its shape. We then compare the cutoff shape obtained with theoretical predictions.
In Section 2, we describe our [*Suzaku*]{} observations and the data reduction procedures. Analysis and results of HXD and XIS data are shown in §3.1 and §3.2, respectively. We present the wide-band spectrum by connecting the XIS and the HXD data in §3.3. A detailed study regarding the cutoff structure is also given there. Section 4 is devoted to multi-wavelength spectral and morphological studies. The results obtained are discussed in the following section, and the results are finally summarized. Throughout this paper, we assume that the distance to RX J1713.7$-$3946 is close to 1 kpc as proposed by [@koyama97] based on the $N_{\rm H}$ value. A similar distance has been claimed based on the NANTEN CO data [@fukui03; @moriguchi05]. The typical age of the remnant for such a distance is estimated of order of 1000 yr, which can be an indication of association of RX J1713.7$-$3946 with an explosion in A.D. 393 as proposed by [@wang97].
SUZAKU OBSERVATIONS AND DATA REDUCTION
======================================
The [*Suzaku*]{} observatory [@mitsuda07] is the fifth Japanese X-ray astronomy satellite, jointly developed by Japan and the US. Its scientific payload consists of two co-aligned detector systems, the XIS [@koyama_xis07] and the HXD [@takahashi_hxd07; @kokubun07]. The XIS consists of four X-ray CCD cameras which are located in the foci of X-ray telescopes (XRT; @serlem07). Three of the XIS sensors are front-illuminated (FI; 0.4–12 keV) CCDs and the other is back-illuminated (BI 0.2–12 keV). The non-imaging, collimated hard X-ray instrument, the HXD, covers the 10–600 keV bandpass. Two main detector units, silicon PIN diodes and GSO scintillators, are buried at the bottom of well-type active shield of BGO. The former covers the lower energy band of 10–60 keV, while the latter detects higher energy photons of 40–600 keV.
We performed 11 pointing observations of RX J1713.7$-$3946 with [*Suzaku*]{}. The observation log is summarized in Table 1, and the pointing position of each observation is shown in Figure \[fig:obs\_fov\]. The southwest part of the SNR, labeled as Pointing 0, was observed in 2005 during the Performance Verification phase, while the other ten observations were performed in 2006 during the [*Suzaku*]{} AO1 phase. Since the SNR is located on the Galacic plane, it is of importance to check the hard X-ray background associated with the Milky Way. We therefore observed two nearby background regions containing no bright X-ray point sources. The pointing positions of these “OFF” observations are shown with red squares in Figure \[fig:obs\_fov\]. The XIS was operated in the normal full-frame clocking mode without spaced-row charge injection during all the observations. Since the results from Pointing 0, together with those from the OFF observations, are already reported in Paper I, we do not give a detailed description on the analysis and results for these data.
We used data products from the pipeline processing version 1.2. For the XIS analysis, we retrieved “cleaned event files” which are screened using standard event selection criteria. For the 2006 data, we recalculated the values of pulse invariant (PI) and the grade values since incorrect CALDB is applied to the pipeline processing of these data as announced by the [*Suzaku*]{} instrument teams. We further screened the cleaned events with following criteria as recommended by the [*Suzaku*]{} instrument teams – (1) cut-off rigidity larger than 6 GV and (2) elevation angle from the Earth rim larger than $10^{\circ}$. For the HXD data, “uncleaned event files” were screened using standard event screening criteria. The exposure times after these screenings are shown in Table 1. Due to unstable operation of 16 PIN diodes installed in Well-counter units, W00–W03; hereafter W0, the bias voltage for these diodes was reduced to 400 V from the nominal voltage of 500 V on 26 May 2006. On 2006 October 4, the bias voltage of 16 more PIN diodes (the PIN diodes in the Well-counter units of W10–W13; W1) was reduced to 400 V for the same reason. The reduction of the PIN diode bias voltage leads to a decrease of their effective area and also affects their energy response. Since the current response matrices do not include these effects, only the PIN diodes with a bias voltage of 500 V are utilized in the following analysis. Throughout this paper, the data reduction and analysis are performed using HEADAS 6.2 and the spectral fitting is done with XSPEC 11.3.2.
[ccccc]{} \[tab:obs\] 0 & 100026010 & ($17^{\rm h}12^{\rm m}17^{\rm s}.0$, $-39^{\rm d}56^{\rm m}11^{\rm s}$) & 55/48 & 26/9/2005\
1 & 501063010 & ($17^{\rm h} 11^{\rm m} 51^{\rm s}.5$, $-39^{\rm d} 31^{\rm m} 13^{\rm s}$) & 17/17 & 11/9/2006\
2 & 501064010 & ($17^{\rm h} 12^{\rm m} 38^{\rm s}.0$, $-39^{\rm d} 40^{\rm m} 14^{\rm s}$) & 18/22 & 11/9/2006\
3 & 501065010 & ($17^{\rm h} 12^{\rm m} 38^{\rm s}.2$, $-39^{\rm d} 22^{\rm m} 15^{\rm s}$) & 19/18 & 11/9/2006\
4 & 501066010 & ($17^{\rm h} 11^{\rm m} 04^{\rm s}.5$, $-39^{\rm d} 40^{\rm m} 10^{\rm s}$) & 19/21 & 12/9/2006\
5 & 501067010 & ($17^{\rm h}11^{\rm m} 05^{\rm s}.1$, $-39^{\rm d} 22^{\rm m} 10^{\rm s}$) & 16/19 & 12/9/2006\
6 & 501068010 & ($17^{\rm h} 14^{\rm m} 11^{\rm s}.6$, $-39^{\rm d} 40^{\rm m} 14^{\rm s}$) & 20/19 & 13/9/2006\
7 & 501069010 & ($17^{\rm h} 14^{\rm m} 11^{\rm s}.4$, $-39^{\rm d} 22^{\rm m} 15^{\rm s}$) & 12/11 & 19/9/2006\
8 & 501070010 & ($17^{\rm h} 14^{\rm m} 11^{\rm s}.8$, $-39^{\rm d} 58^{\rm m} 14^{\rm s}$) & 19/20 & 19/9/2006\
9 & 501071010 & ($17^{\rm h} 12^{\rm m} 17^{\rm s}.6$, $-39^{\rm d} 18^{\rm m} 50^{\rm s}$) & 16/15 & 20/9/2006\
10 & 501072010 & ($17^{\rm h} 15^{\rm m} 44^{\rm s}.5$, $-39^{\rm d} 40^{\rm m} 10^{\rm s}$) & 15/15 & 5/10/2006\
OFF1 & 100026020 & ($17^{\rm h} 09^{\rm m} 31^{\rm s}.9$, $-38^{\rm d} 49^{\rm m} 24^{\rm s}$) & 28/24 & 25/9/2005\
OFF2 & 100026030 & ($17^{\rm h} 09^{\rm m} 05^{\rm s}.1$, $-41^{\rm d} 02^{\rm m} 07^{\rm s}$) & 30/28 & 28/9/2005
ANALYSIS AND RESULTS
====================
HXD Data Analysis {#subsec:hxd_ana}
-----------------
### Spectral Analysis {#subsec:rxj1713_hxd_spec}
The HXD PIN spectrum from each pointing was constructed and compared with the background model estimated for the each observation period. In the analysis below, the non-Xray background (NXB) model [@watanabe07] provided by the HXD team is used for the background generation. Since the NXB model does not include the contributions from the cosmic X-ray background (CXB), a simulated CXB spectrum was added to the NXB model. Specifically, based on the reanalysis of the data from [*HEAO-1*]{} observations in the 1970’s [@gruber99], the CXB spectrum was modeled as $$\begin{aligned}
\frac{dN}{d\varepsilon} = 7.9~{\varepsilon_{\rm keV}}^{-1.29} \exp \left( -\frac{\varepsilon_{\rm keV}}{\varepsilon_{\rm p}} \right)~{\rm ph}~{\rm s}^{-1}~{\rm keV}^{-1}~{\rm cm}^{-2}~{\rm str}^{-1}, \end{aligned}$$ where $\varepsilon_{\rm keV} = \varepsilon/1~{\rm keV}$ and $\varepsilon_{\rm p} = 41.13$. We estimated the CXB signal in each HXD PIN spectrum using the response matrix for spatially uniform emission, [ae\_hxd\_pinflat\_20060809.rsp]{}, [ae\_hxd\_pinflat123\_20060809.rsp]{}, and [ae\_hxd\_pinflat23\_20060809.rsp]{}. The contribution from the CXB flux to the detected count rate is estimated to be $\sim 5$% of the NXB.
Figure \[fig:hxd\_spec\_p8\] shows the HXD PIN spectrum obtained from Pointing 8, where a clear detection of hard X-rays can be seen. Likewise, the HXD PIN detected signals from all other pointings with a count rate of 20–50% of the NXB. We fitted the background-subtracted spectra with a simple power law model: $dN/d\varepsilon \propto \varepsilon^{-\Gamma}$. Since the extended nature of the source does not cause any spectral steepening or flattening (Paper I), we used the point-source response matrix at the XIS-nominal position, [ae\_hxd\_pinxinom\_20060814.rsp]{}, [ae\_hxd\_pinxinom123\_20060814.rsp]{}, and [ae\_hxd\_pinxinom23\_20060814.rsp]{}. Table 2 gives the best-fit parameters with the statistical errors at 90% confidence level. The obtained photon indices are generally larger than those obtained from the corresponding XIS spectra (see below). This difference indicates that a spectral cutoff is not unique to the SW region (Paper I), but a common feature throughout the remnant.
In order to confirm the results obtained above, we evaluated the systematic errors due to uncertainties in the NXB modeling. As described in [@mizuno06], the current reproducibility of the NXB model is $\sim 5$%. Therefore, we examined how much the values of the photon index change by increasing or decreasing the background model by 5%. The systematic errors were found to be smaller than the statistical errors indicated in Table 2.
The systematic errors due to the misestimation of NXB were examined in another way. Considering the physical size of this target, the emission from the remnant should be constant during the observations. Therefore, background-subtracted lightcurves should be constant during a observation. Although the lightcurves shown are almost constant within statistical errors, the background-subtracted count rate becomes higher when the total count rate increases for Pointings 0 and 8. The light curves for Pointings 0 and 8 are shown in Paper I and Figure \[fig:hxd\_lc\_p8\], respectively. Since this behavior is thought to be caused by misestimation of the NXB, we examined how much the photon indices change with and without those time regions. When we discard the time region which corresponds to the last bin of Figure \[fig:hxd\_lc\_p8\], the photon index changes by $\Delta \Gamma \simeq 0.2$ from the values in Table 2.
Diffuse emission from the Galactic plane also can affect the spectra. However, no emission above the 5% level of the NXB was detected from the OFF pointings (Paper I). Moreover, when the excess counts marginally detected from the OFF pointings are added to the background spectrum for each pointing, the fitting result agrees within the statistical errors.
[ccccc]{} \[tab:hxd\_fit\] 0 & $3.2\pm0.2$ & $2.5\pm0.1$ & 1.15 (36) & W0–3\
1 & $3.3\pm0.2$ & $3.2\pm0.2$ & 0.91 (18) & W1–3\
2 & $3.0\pm0.3$ & $3.2\pm0.2$ & 0.84 (21) & W1–3\
3 & $3.4\pm0.5$ & $2.0\pm0.2$ & 0.58 (17) & W1–3\
4 & $2.9\pm0.3$ & $3.5\pm0.2$ & 0.77 (23) & W1–3\
5 & $3.2\pm0.5$ & $1.7\pm0.2$ & 0.76 (17) & W1–3\
6 & $2.9\pm0.3$ & $2.9\pm0.2$ & 0.72 (20) & W1–3\
7 & $3.9^{+1.4}_{-1.1}$ & $1.2^{+0.3}_{-0.2}$ & 0.42 (9) & W1–3\
8 & $2.6^{+0.5}_{-0.4}$ & $2.2\pm0.2$ & 0.87 (19) & W1–3\
9 & $3.0^{+1.2}_{-1.0}$ & $1.1\pm0.2$ & 0.86 (12) & W1–3\
10 & $4.4^{+1.6}_{-1.2}$ & $1.0\pm0.2$ & 0.20 (8) & W2, 3
### Spatial Distribution of Hard X-ray Emission {#subsec:hxd_spatial}
Using the spectral parameters obtained for the 11, pointings we attempt to reconstruct the spatial distribution of the hard X-ray emission. Since we expect spatial variation of both brightness and spectral index, we use a Monte Carlo simulator, [simHXD]{} [@terada05], for the estimation. We input the spatial distribution of brightness and spectral index into [simHXD]{}, and compare the simulated spectra with observed ones shown in §\[subsec:rxj1713\_hxd\_spec\].
Firstly we simulated by taking a simple model in which the brightness distribution is same as the [*ASCA*]{} image shown in Figure \[fig:sim\_dist4\] and the spectral index is constant at $\Gamma = 3.0$ throughout the SNR ([*simulation1*]{}). By using the [*ASCA*]{} image as an input to [simHXD]{}, we simulated hard X-ray spectra in 10–40 keV, which are expected to be observed by the PIN in each pointing. Figure \[fig:norm\_hikaku\] compares the flux from the observations and the simulations. In this figure, each value is normalized to that of Pointing 0. Here, the systematic error for each observational data point is $\sim 20$% if we use the 5% of the background as systematic errors. The observations and the simulations appear consistent with each other. Therefore, we expect the brightness distribution above 10 keV is not drastically different from the distribution below 10 keV.
The largest discrepancy between the observation and the simulation is found for Pointing 4. In this pointing, there is a known point source at the corner of the HXD-PIN field of view. The source is listed in the Second IBIS/ISGRI Soft Gamma-Ray Survey Catalog [@bird06] as IGR J17088$-$4008. According to [@bird06], the average flux of this source is $1.1\pm0.2$ mCrab in 20–40 keV and $2.2\pm0.3$ mCrab in 40–100 keV. It is noted that the source is not bright enough for IBIS/ISGRI to determine spectral parameters. The transmission of fine collimators of the HXD for the source is estimated to be $\sim 0.05$. The estimated count rate of this source corresponds to $\sim 2$% of the detected signals from Pointing 4. However, the variability of this source could increase the count rate to a non-negligible level. The observation and the simulation become consistent if the point source was 10 times brighter than its average value during the [*Suzaku*]{} observation. Also, the angular response of the HXD near the edge of the FoV can have some uncertainties since calibrations for a source with such a large offset angle is difficult. The uncertainties can affect our estimate of the contamination from IGR J17088$-$4008. At this moment, we cannot conclude that the large difference between the observation and the simulation is due to the spatial distribution of the SNR emission or the point source in the FoV.
Next we tried another simulation taking into account the spatial distribution of spectral indices ([*simulation2*]{}). We adopt a “toy model” shown in Figure \[fig:sim\_dist4\] since the results of the spectral analysis presented in Table 2 suggest that the hard X-ray spectrum may be flatter in the inner region of the SNR than near the rim. In the model, the photon index of the inner region is set to $\Gamma = 2.6$ and that of the rim region is set to $\Gamma = 3.5$. The [*ASCA*]{} image was used to provide the brightness distribution at 10 keV.
Comparisons of the data and the simulation results are shown in Figure \[fig:norm\_hikaku\] and Figure \[fig:index\_hikaku\] for the flux and photon index, respectively. The detected flux obtained from [*simulation2*]{} is somewhat similar to that from [*simulation1*]{}, and the simulation data generally follow the observational data. As for the distribution of photon indices, the toy model gives a similar distribution to the observational results for the western portion of the SNR. Obtaining a better fit for the eastern region (Pointing 7, 8, and 10) may require more complex assumptions than the toy model for [*simulation2*]{}.
XIS Data Analysis {#subsec:xis_ana}
-----------------
### Image Analysis
Figure \[fig:xis\_mosaic\_image\] shows mosaic images of RX J1713.7$-$3946, constructed using the data from XIS0, 2, and 3 (FI-CCDs). The upper panel shows the soft band image in 1–5 keV and the lower panel shows the hard band image in 5–10 keV. Both images are smoothed with a Gaussian of $\sigma = 0^{\prime}.3$. Instrumental background signals are subtracted from both images. The signal to background ratio in the hard band is smaller than that in the soft band by one order of magnitude. Thus, the background must be carefully subtracted. We utilized the Night Earth Background Database consisting of event data obtained when the satellite is looking at the night earth and the non X-ray background becomes dominant. After subtracting the background, the vignetting effects of the XRTs were corrected by means of the simulation program called [xissim]{} [@ishisaki07]. In the program, an image from a flat field can be simulated by a Monte Carlo method.
As is seen in Figure \[fig:xis\_mosaic\_image\], the [*Suzaku*]{} XIS has covered most of the remnant. Thanks to little stray-light contamination of the XRTs and low background level of the XIS, high quality images are obtained even in the high energy band above 5 keV. The double shell structure revealed by [*XMM-Newton*]{} is clearly seen in the XIS images. In addition to the bright structures of the western part, the XIS revealed detailed morphology of the dim parts of the remnant. The dim structures are highlighted in Figure \[fig:xis\_hess\_image\], which is the same as Figure \[fig:xis\_mosaic\_image\] (a) but displayed with a different color scale. In this figure, the contours of the H.E.S.S. gamma-ray image (from @aha07) are overlaid for comparison. As is clearly seen, not only the bright rims but also the eastern portion shows a striking similarity between the two energy regimes. This correlation is discussed more quantitatively in §\[subsec:mws\].
Two point sources seen in the soft band image are listed in the [*ROSAT*]{} bright source catalogue. One is 1WGA J1714.4$-$3945, which was associated with a Wolf-Rayet star by [@pfe96]. The other is 1WGA J1713.4$-$3949, which is located between the two FoVs of Pointing 2 and 6. This source has been suggested to be the neutron star associated with the SNR by [@laz03].
Comparing the soft-band and hard-band images provides information about the spatial variations of the spectral properties. At first sight, the bright structures are very similar in the soft band and hard band images. In order to compare the images in more detail, a radial profile around the center of SNR ($\alpha_{\rm J2000} = 17^{\rm h}13^{\rm m}33^{\rm s}.6$, $\delta_{\rm J2000} = -39^{\rm d}45^{\rm m} 36^{\rm s}$) is presented in Figure \[fig:xis\_radial\], where circular regions of $2^{\prime}.1$ radius centered on the two point sources are excluded. The hard band to soft band ratio is significantly different between the bright outer region part and the interior. This difference suggests a corresponding difference in spectral properties.
### Spectral Analysis {#subsec:xis_spec}
As already reported in Paper I, the XIS data reveal a cutoff below 10 keV in the spectrum of the SW rim. The XIS spectra from the various regions should reveal whether the cutoff varies across the remnant. We extracted spectra from the regions shown in Figure \[fig:xis\_mosaic\_image\] with the green squares. The region labeled as “1’ (Region 1) is the same one that we used for the spectral analysis in Paper I. The background spectrum needs to be accumulated from nearby regions since RX J1713.7$-$3946 is located almost on the Galactic plane. The OFF observation data cannot be used for background estimation other than for Region 1. This is because the contamination on the optical blocking filters of the XIS [@koyama_xis07] significantly changed the detector response to low-energy X-rays between 2005 and 2006. For Region 1–10 we therefore accumulated a background spectrum from the regions indicated with the magenta polygons in Figure \[fig:xis\_mosaic\_image\]. In the spectral fitting discussed below, the standard RMF files version 2006-02-13 were used, whereas ARF files were produced using [xissimarfgen]{}. Each spectrum is binned so that each bin contains at least 300 counts. After the binning, we ignore those bins whose energy is smaller than 0.4 keV or larger than 12 keV. We also excluded bins between 1.7 and 1.9 keV because there exist large systematic uncertainties in the response matrices. Finally, we co-added the spectra, RMF files, and ARF files from the FI chips to produce a single data set for the XIS.
First, we fitted all the spectra with a simple power law absorbed by the interstellar medium. The results are summarized in Table 3. This model yields acceptable fits for most of the spectra. However, the fits to the spectra of Region 1, 2, and 6 are not acceptable even at the 99% confidence level. We show the spectrum of Region 2 with the best-fit power-law and the residuals in Figure \[fig:xis\_spec\_reg2\]. Although not so drastic as Region 1 which is shown in Paper I, a correlated pattern to the residuals can be seen for the Region 2. A similar pattern is also seen in the data of Region 6.
We the fitted all the spectra with a cutoff power law. This function gives acceptable fits for all regions, and consistently gives better values of $\chi_{\nu}^2$. In Figure \[fig:xis\_spec\_reg2\], we plot the residuals for the spectrum of Region 2. These results, together with the steeper spectra detected with HXD, indicate the existence of a cutoff somewhere between the bandpasses of the XIS and HXD.
Although statistically rejected for some regions, the results obtained with the power-law fits are consistent with previous studies with [*ASCA*]{}, [*Chandra*]{}, and [*XMM-Newton*]{} [@koyama97; @slane99; @uchi03; @cassam04; @hiraga05]. The results with [*Suzaku*]{} suggest that the photon index is larger and the absorption is smaller for the inner regions than for the outer regions. This tendency agrees with the radial profile shown in Figure \[fig:xis\_radial\]. The difference of $N_{\rm H}$ between the western bright spots and the inner region is $\Delta N_{\rm H} \simeq 0.3$–$0.4 \times 10^{22}~{\rm cm}^{-2}$, which is consistent with the results from [*XMM-Newton*]{}. According to the discussion by [@hiraga05], there could be a correlation between the difference of $N_{\rm H}$ and the presence of the molecular clouds in the western part of the SNR detected with the NANTEN telescope [@fukui03; @moriguchi05].
[cccccccccccc]{} \[tab:xis\_fit\] 1 & 108 & $0.87\pm0.01$ & $2.39\pm 0.01$& $6.06\pm0.03$ & 1.38 (718) & & $0.77\pm0.01$ & $1.96\pm0.05$ & $9\pm1$ & $5.7\pm0.1$ & 1.04 (717)\
2 & 77.2 & $0.83\pm0.02$ & $2.38\pm0.02$ & $5.22\pm0.05$ & 1.23 (242) & & $0.73\pm0.03$ & $2.00\pm0.09$ & $10^{+3}_{-2}$ & $4.9\pm0.2$ & 1.01 (241)\
3 & 86.6 & $0.55\pm0.03$ & $2.62\pm0.04$ & $1.41\pm0.03$ & 1.14 (94) & & $0.47\pm0.04$ & $2.2\pm0.2$ & $8^{+7}_{-3}$ & $1.3\pm0.1$ & 1.02 (93)\
4 & 42.2 & $0.54\pm0.04$ & $2.72\pm0.07$ & $0.66\pm0.02$ & 1.38 (44) & & $0.44\pm0.07$ & $2.2\pm0.3$ & $6^{+8}_{-2}$ & $0.60^{+0.06}_{-0.05}$ & 1.23 (43)\
5 & 116 & $0.78\pm0.03$ & $2.39\pm0.04$ & $1.88\pm0.04$ & 1.18 (120) & & $0.70\pm0.05$ & $2.1\pm0.2$ & $11^{+13}_{-4}$ & $1.8\pm0.1$ & 1.12 (119)\
6 & 51.4 & $0.88^{+0.06}_{-0.05}$ & $2.43\pm0.06$ & $1.27\pm0.04$ & 1.63 (48) & & $0.68\pm0.09$ & $1.7\pm0.3$ & $5^{+3}_{-2}$ & $1.1^{+0.2}_{-0.1}$ & 1.24 (47)\
7 & 59.6 & $0.71\pm0.04$ & $2.38\pm0.05$ & $1.18\pm0.03$ & 1.15 (56) & & $0.62\pm0.07$ & $2.0\pm0.2$ & $10^{+14}_{-4}$ & $1.1\pm0.1$ & 1.03 (55)\
8 & 72.4 & $0.78\pm0.03$ & $2.55\pm0.04$ & $1.91\pm0.04$ & 1.10 (110) & & $0.68\pm0.05$ & $2.1\pm0.2$ & $8^{+5}_{-2}$ & $1.8\pm0.1$ & 0.92 (109)\
9 & 65.9 & $0.81\pm0.03$ & $2.49^{+0.04}_{-0.03}$ & $2.03\pm0.04$ & 1.11 (109) & & $0.74^{+0.04}_{-0.05}$ & $2.2\pm0.2$ & $14^{+16}_{-5}$ & $1.9^{+0.2}_{-0.1}$ & 1.03 (108)\
10 & 54.0 & $0.60\pm0.04$ & $2.23\pm0.05$ & $1.49\pm0.03$ & 0.88 (75) & & $0.52\pm0.06$ & $1.9\pm0.2$ & $12^{+19}_{-5}$ & $1.4\pm0.1$ & 0.79 (74)
Broad-Band X-Ray Spectral Analysis {#subsec:broad}
----------------------------------
In this section, we connect the XIS and HXD spectra, which is crucial for a quantitative study of the cutoff structure and also for multi-wavelength study described in §\[subsec:mws\]. In order to study the general characteristics of the emission from RX J1713.7$-$3946, the XIS and HXD spectra obtained in §\[subsec:xis\_ana\]/§\[subsec:hxd\_ana\] were co-added. Then, the summed XIS and HXD spectra were scaled to account for the flux from the whole remnant, which makes it easier not only to connect the XIS and HXD spectra but also to compare the combined spectrum directly to those of other wavelengths in §\[subsec:mws\]. In scaling the spectra to the whole remnant, we assumed the surface brightness of the [*ASCA*]{} GIS image (1–5 keV) shown in Figure \[fig:obs\_fov\], the angular response of the XIS/XRT system (ARFs), and that of the HXD PIN presented in Figure 6 of Paper I. The relative normalization factor between XIS and HXD derived from Crab observations [@ishida06] is included in the scaling. Therefore, the scaled XIS and HXD can be connected to each other without additional scaling if no systematic errors are considered. However, we estimate that the normalization of the XIS and HXD spectra obtained by the procedures above should contain systematic errors of 10–20%. In order to account for the systematic errors, we include a constant normalization factor and deal with this factor as a free parameter when fitting the XIS+HXD spectrum below. In this procedure, we discarded the HXD data from Pointing 4 since they seem to be contaminated by a nearby hard X-ray source as described in §\[subsec:hxd\_spatial\].
Before the XIS and HXD spectra were jointly analyzed, each co-added spectrum was independently fitted using a power-law function. Table 4 summarizes the fit results. The HXD spectrum gives an acceptable fit. In contrast, the XIS does not, with a large chi-squared value of ${\chi_\nu}^2 = 1.57$ for 711 degrees of freedom and residuals that begin to become large at $\sim 6$ keV. This fact is clearly seen in Figure \[fig:joint\_spec\_pl\], where the XIS and HXD spectra are plotted together with a power-law function which represents the XIS spectrum. The spectral steepening begins in the XIS band and continues smoothly into the HXD band. This plot strongly suggests that spectral steepening occurs around 10 keV. We have already reported such a spectral feature in Paper I for Pointing 0 data. The same kind of feature has been revealed in this spectrum averaged over the SNR, which suggests that the spectral steepening is common in the entire region.
We then quantitatively evaluated the spectral steepening by fitting a model with a cutoff structure to the combined XIS/HXD spectrum. Below we use our numerically calculated synchrotron spectrum which was embedded into the XSPEC package. As an electron distribution, we adopt a generalized form, namely, $$\begin{aligned}
\frac{dN_e}{dE} \propto E^{-s} \exp \left[ - \left( \frac{E}{E_0} \right)^\beta \right]\label{eq:e_spec},\end{aligned}$$ instead of taking a simple [*exponential cutoff*]{} often used in the literature. Here, $s$ represents index of electron spectrum and $\beta (> 0)$ determines rapidity of high-energy cutoff. The photon spectrum can be calculated as $$\begin{aligned}
\frac{dN}{d\varepsilon} \propto \varepsilon^{-1} \int F \left(\frac{\varepsilon}{\varepsilon_c}\right) \frac{dN_e}{dE}~dE. \end{aligned}$$ Here the function $F(x)$ is defined as $$\begin{aligned}
F(x) \equiv x \int^{\infty}_{x} K_{5/3} (\xi)~d\xi,\end{aligned}$$ where $K_{5/3}$ is the modified Bessel function of 5/3 order. When pitch angles are isotropic, the characteristic photon energy $\varepsilon_c$ is given as $$\begin{aligned}
\varepsilon_c = 0.543 \left( \frac{B}{100~\mu {\rm G}} \right) \left( \frac{E}{10~{\rm TeV}} \right)^2~{\rm keV}. \end{aligned}$$ The model consists of four parameters: $s$, $\Pi \equiv E_0B^{1/2}$, $\beta$, and the flux normalization. Photon spectra were calculated and tabulated for reasonable ranges of the four parameters, to be used as an XSPEC table model. The parameters characterizing the electron energy distribution can be obtained directly from the fit to the X-ray data.
For RX J1713.7$-$3946, the most probable value for $s$ is 3.0 rather than 2.0 due to significant synchrotron cooling of electrons during the lifetime of the SNR $t_0$ ($\sim 1000~{\rm yr}$). The electron spectrum becomes steeper by a factor of 1 ($s \rightarrow s +1$), when the injection of electrons is constant and the lifetime $t_0$ is smaller than the time scale of synchrotron cooling $t_{\rm sync}$. Since $t_{\rm sync}$ can be written as $$\begin{aligned}
t_{\rm sync} = 28 \left( \frac{B}{100~\mu{\rm G}} \right)^{-3/2} \left( \frac{\varepsilon}{3~{\rm keV}} \right)^{-1/2}~{\rm yr}, \label{eq:sync_cooling_time}\end{aligned}$$ the condition above is satisfied in the energy range, $$\begin{aligned}
\varepsilon > \varepsilon_b \equiv 2.3 \left( \frac{B}{100~\mu{\rm G}} \right)^{-3} \left( \frac{t_0}{10^3~{\rm yr}} \right)^{-2}~{\rm eV}. \label{eq:ebreak}\end{aligned}$$ To explain the variability of X-ray emission reported on a year timescale from compact regions of the shell of RX J1713.7$-$3946, [@uchi07] proposed that the magnetic field in these compact regions is amplified to 1 mG. The average large-scale magnetic field in the remnant should be significantly lower, but even for $B \simeq 100~{\mu}{\rm G}$ the above condition in the X-ray domain is safely satisfied. Therefore, for a strong shock with a compression ratio of 4.0, the index of X-ray emitting electrons should be close to $s = 3.0$.
We fitted the [*Suzaku*]{} spectrum with the synchrotron spectrum described above. In the fitting procedure, we fixed the electron index $s$. Table 5 summarizes the result in the case of $s = 3.0$, which corresponds to the most probable case following the discussion above. We also present results when $s = 2.0$ for comparison. This case can be realized if the magnetic field and/or the age of the remnant are much smaller than we expect, as is seen in equation (\[eq:ebreak\]). The spectral fitting with the electron index of $s = 3.0$ yields rather rapid steepening with $\beta = 3.4^{+0.7}_{-0.5}$ compared to the conventionally used [*exponential cutoff*]{}. However, one should be careful with the physical interpretation of the fit based on equation (\[eq:e\_spec\]). Indeed, while the electron spectrum is derived from the broad-band [*Suzaku*]{} X-ray data assuming a specific spectral form given by equation (\[eq:e\_spec\]), the curve 1 in Figure \[fig:cutoff\_hikaku\] has a more general meaning. It does not depend on the assumed analytical presentation and, in fact, relates [*uniquely*]{} the energy spectrum of electrons to the measured X-ray spectrum. Indeed, the curve 1 can be presented in different mathematical forms. In particular, the electron spectrum shown by curve 1 in Figure \[fig:cutoff\_hikaku\] is quite close to the theoretical prediction for the spectrum of shock accelerated electrons in a young SNR [@zira07], $$\begin{aligned}
\frac{dN_e}{dE} \propto E^{-3} \left[ 1+ 0.66 \left( \frac{E}{E_0} \right)^{5/2} \right]^{9/5} \exp \left[ -\left( \frac{E}{E_0} \right)^2 \right]. \label{za07_e_spec}\end{aligned}$$ This spectrum is derived under the assumption that electrons are accelerated by a strong shock in the Bohm diffusion regime and that the energy losses of electrons are dominated by synchrotron cooling. It is seen that the energy spectrum below the cutoff is described by a function which deviates from a pure power-law form, by an exponential term with $\beta=2$. The latter has a simple physical interpretation and is a result of combination of two effects — acceleration in the Bohm diffusion regime, and energy losses in the synchrotron regime. The curve 2 shown in Figure 12 is calculated for the parameter $\Pi=E_0 B^{1/2} = 201~{\rm TeV}~\mu{\rm G^{1/2}}$.
The energy spectrum of synchrotron radiation corresponding to the electron spectrum given by equation (\[za07\_e\_spec\]) has a form [@zira07] $$\begin{aligned}
\frac{dN}{d\varepsilon} \propto \varepsilon^{-2} \left[ 1+ 0.46 \left( \frac{\varepsilon}{\varepsilon_0}\right)^{0.6} \right]^{2.29} \exp \left[ -\left( \frac{\varepsilon}{\varepsilon_0} \right)^{1/2} \right]. \label{za07_sync_spec}\end{aligned}$$ This spectrum fits well the broad-band [*Suzaku*]{} X-ray data for a single parameter $\varepsilon_0=0.67 \pm 0.02$ keV. Note that the exponential term in the synchrotron spectrum is a weak function of energy ($\propto \exp[-(\varepsilon/\varepsilon_0)^{1/2}]$); therefore the characteristic energy $\varepsilon_0$ only formally can be considered as the cutoff energy. In fact, for the shape given by equation (\[za07\_sync\_spec\]), the break in the spectrum starts at much higher energies ($\varepsilon \sim 10 \varepsilon_0$).
[ccccc]{} \[tab:xishxd\_all\_plfit\] XIS (0.4–12 keV) & $0.79\pm0.01$ & $2.39\pm0.01$ & $7.65\pm0.03$ & 1.57 (711)\
HXD (12–40 keV) & — & $3.2\pm0.1$ & $1.17\pm0.03$ & 1.16 (67)
[cccccc]{} \[tab:xishxd\_all\_modelfit\] Electron Distribution & $0.71\pm0.01$& $1.03\pm0.06$ & $s=3.0$ (fixed), $\beta = 3.4^{+0.7}_{-0.5}$, $\Pi = 402\pm6~{\rm TeV}~\mu{\rm G}^{1/2}$& $7.2\pm0.1$ & 1.11 (778)\
of Equation (\[eq:e\_spec\]) & $0.68\pm0.01$ & $1.03^{+0.06}_{-0.05}$ & $s=2.0$ (fixed), $\beta = 1.5\pm0.2$, $\Pi = 207^{+21}_{-20}~{\rm TeV}~\mu{\rm G}^{1/2}$& $7.1\pm0.1$ & 1.07 (778)\
Z&A (2007) & $0.70\pm0.01$ & $1.08\pm0.04$ & $\varepsilon_0 = 0.67\pm0.02~{\rm keV}$ & $7.2\pm0.1$ & 1.11 (779)
Multi-Wavelength Study {#subsec:mws}
----------------------
### Spectral Energy Distribution {#subsubsec:sed}
The high quality data of [*Suzaku*]{} obtained over two decades in energy and combined with the TeV gamma ray data, allow definite conclusions concerning the origin of multi-TeV parent particles based on the comparison of model predictions with observations. In particular it is important that the [*Suzaku*]{} data provide unambiguous information about the shape of the energy spectrum of electrons in the cutoff region.
Figure \[fig:SED\] shows the Spectral Energy Distribution (SED), $E^2 dN/dE$, of RX J1713.7$-$3946 from radio to TeV energies. The X-ray data correspond to the entire remnant; they are reconstructed assuming the best-fit model given by equation (\[za07\_sync\_spec\]) corrected for the interstellar absorption with $N_{\rm H} = 0.70 \times 10^{22}~{\rm cm}^{-2}$. The points in the TeV gamma-ray range corresponding to the fluxes of the whole remnant, are from the latest report of the H.E.S.S. collaboration [@aha07]. The EGRET upper limit is taken from [@aha06]; it is obtained through modeling and subtracting the resolved EGRET source 3EG 1714$-$3857 [@hartman99]. The two data points in the radio band are shown as measured with the ATCA telescope [@laz04]. But it should be noted that these fluxes are detected only from the NW rim region, therefore they should be treated as lower limits when compared to model predictions for the whole remnant. The recent estimates based on observations with ATCA and the 30-m radiotelescope IRA show that the flux of the entire remnant should be a factor of few larger than the flux of the NW rim (F. Acero et al. 2008, in preparation; G. Dubner 2008, private communications).
In Figure \[fig:SED\] we show theoretical fluxes of broad-band electromagnetic radiation produced by both accelerated electrons and protons. The results depend strongly on the strength of the average magnetic field $B$. Although the magnetic field in compact filaments can be as large as $\sim 1~{\rm mG}$ [@uchi07], the field in less bright and more diffuse regions where the bulk of the synchrotron X-ray emission is produced, should be significantly weaker. Yet, a moderately large field exceeding $100~\mu{\rm G}$ is required to explain the X-ray flux ratio between the diffuse and compact zones. The integrated X-ray flux from the diffuse zone (zone 1) is larger by a factor of $\sim 4$ than that from the compact zone (zone 2): $F_1 \sim 4 F_2$ [@uchi03]. Using the ratio of the emission volumes estimated from the X-ray image ($V_1 \sim 1000 V_2$), the number density of X-ray emitting electrons, $n \propto F/(VB^2)$, in the zone 2 relative to the zone 1 is estimated as $n_2 \sim 250(B_1/B_2)^2 n_1$. For $B \gtrsim 1~{\rm mG}$, and assuming $n_2 \geq n_1$, one finds $B_1 \geq 100~\mu{\rm G}$.
In Figure \[fig:SED\] the synchrotron, inverse Compton (IC), and $\pi^0$-decay fluxes are calculated assuming strong uniform magnetic field of strength $B=200~\mu{\rm G}$. The calculations are performed with constant injection of electrons and protons over the last 1000 yr. The injection spectrum for both electrons and protons is assumed to be a power-law with an index $s=2.0$. For calculations of IC fluxes we used the diffuse photon fields proposed by [@porter06], including 2.7 K CMB, as well as interstellar radiation consisting of two, optical (starlight) and infrared (dust) emission components.
The best fits for the synchrotron radiation and $\pi^0$-decay gamma rays are achieved for cutoffs in the energy distributions of electrons and protons at 28.4 TeV and 130 TeV, respectively. The spectra of $\pi^0$-decay gamma rays are calculated using the analytical presentations by [@kelner06]. Since the highest energy tail of the observed gamma-ray spectrum is best fitted with $\exp [-(\varepsilon/\varepsilon_0)^{\beta_\gamma}]$ with $\beta_\gamma \approx 0.5$ [@aha07], the proton spectrum in the cutoff region should have an exponential behavior, $\beta_p \approx 2 \times \beta_\gamma =1$ (see @kelner06).
The electrons suffer synchrotron losses, thus the electron spectrum above a certain energy becomes steeper ($E^{-s} \rightarrow E^{-(s+1)}$). The position of the break in the electron spectrum appears at $$\begin{aligned}
E_b = 1.25 \left( \frac{B}{100~\mu{\rm G}} \right)^{-2}
\left( \frac{t_0}{10^3~{\rm yr}} \right)^{-1}~{\rm TeV}\label{eq:ebreak_electron}. \end{aligned}$$ For a magnetic field $B=200 \mu \rm G$ and age of the source $T \leq 1000$ yr, the spectral break in the synchrotron spectrum corresponding to the transition of the electrons spectrum from uncooled to cooled regime, appears around 1 eV (see Figure \[fig:SED\]). Thus, for young SNRs the detection of the synchrotron break at optical/infrared wavelengths would be an additional argument in favor of strong magnetic field. The cooling break in the electron spectrum is reflected also in the gamma-ray band. Namely, for $B=200 \mu \rm G$ the corresponding signature in the IC gamma-ray spectrum appears around 10 GeV. Unfortunately (see Figure \[fig:SED\]), for such a strong magnetic field, the IC component is suppressed, and falls well below the sensitivity of gamma-ray detectors in this energy band, including [*GLAST*]{}. At the presence of such a strong magnetic field, the only viable mechanism which can produce TeV gamma-rays at the flux level detected by H.E.S.S. is related to interactions of ultrarelativistic protons with ambient gas through production and decay of $\pi^0$-mesons. To explain the detected flux of gamma-rays protons should be accelerated to energies well beyond 100 TeV, and the parameter $A=(W_{\rm p}/10^{50} \rm erg) (n/1 \ cm^{-3}) (d/1 \rm kpc)^{-2}$ should be between 1.5–3, depending on the spectrum of protons.
In order to increase the IC flux to the level of the observed TeV gamma-ray flux, the magnetic field should be reduced down to $10~\mu{\rm G}$ and $15~\mu{\rm G}$. For the given magnetic field, the high quality X-ray data of [*Suzaku*]{} obtained over two energy decades allow derivation of the electron spectrum with high accuracy within the interval covering one energy decade: from $\sim 50(B/10~\mu{\rm G})^{-1/2}~{\rm TeV}$ to $\sim ~500(B/10~\mu{\rm G})^{-1/2}~{\rm TeV}$. This allows us to calculate the spectrum and absolute flux of IC gamma rays above a few TeV without any model assumptions, as long as the main target for the IC gamma-ray production remains the 2.7 K CMB. The contribution of the diffuse optical/infrared radiation fields generally is less, however the optical photons may provide enhanced TeV emission at low, sub-TeV energies. For calculations of the IC spectrum we used the interstellar radiation model of [@porter06], who proposed significantly larger flux of optical and infrared components compared to the generally accepted flux. However, the results presented in Figure \[fig:SED\_leptonic\] show that even this high diffuse optical and infrared radiation fails to account for the observed gamma-ray flux below a few TeV. In order to fill this gap one needs to assume an unreasonably large density of optical radiation. This is demonstrated in Figure \[fig:SED\_leptonic\_opt\], where an agreement of IC calculations with the reported gamma-ray fluxes is achieved assuming an additional, although in our view quite unrealistic, component of optical radiation with a density of $140~{\rm eV}~{\rm cm}^{-3}$.
It should be noted that the problem of explanation of low energy gamma rays in Figure \[fig:SED\_leptonic\] is related to the large energy of the break in the electron spectrum given by equation (\[eq:ebreak\_electron\]), and correspondingly to the position of the Compton peak which appears above 1 TeV in the spectral energy distribution of gamma rays. Thus, the reduction of the break energy down to 200 GeV could in principle solve the problem. However, since the magnetic field in this model cannot exceed $15~\mu{\rm G}$, the only way to shift the Compton peak to sub-TeV energies is to assume that the supernova remnant is older than $10^4$ yr.
An alternative solution for explanation of gamma-ray data within the IC models is to assume a higher density of relativistic electrons responsible for $\leq 1$ TeV gamma rays, i.e. to postulate a second, low-energy electron component in the shell. Formally this assumption does not [*a priori*]{} contradict the observations since the additional electrons are required to be present below 20 TeV, i.e. in the energy range which is not constrained by [*Suzaku*]{} observations (for production of the lowest energy X-rays detected by [*Suzaku*]{} in a magnetic field of $15~\mu{\rm G}$ the electron energy must exceed 40 TeV). The results calculated under such assumption are shown in Figure \[fig:SED\_leptonic\_2ele\]. The second electron component is assumed to have a power-law injection spectrum with the same index as the first (main) component, $s=2.0$, but with a high energy cutoff around 10 TeV. The latter is required to prevent the conflict with the observed gamma-ray spectrum above 1 TeV.
### Morphology
A comparison between the X-ray and the TeV gamma-ray morphology is expected to provide information about the acceleration/emission processes. As is already shown in Figure \[fig:xis\_hess\_image\], the [*Suzaku*]{} XIS image revealed a significant correlation with that of H.E.S.S. telescopes not only in the bright structures but also in the dim regions of the remnant. In the following, the two images are compared with each other on a more quantitative basis.
It should be taken into account that the X-ray morphology can be affected by the spatial distribution of absorption column density, $N_{\rm H}$. In order to avoid the effect, here we disregard the energy band lower than 1 keV, and use two energy intervals, 2–5 keV and 5–10 keV, for comparison. For the $N_{\rm H}$ variation as shown in Table 3, the count rate can vary by 5% and 0.7% in the 2–5 keV band and 5–10 keV band, respectively. One should also take into account the difference of point spread functions between [*Suzaku*]{} XIS and H.E.S.S. when comparing the two images. To prevent this, we compare surface brightness for each square region with a size of $10^{\prime}.8 \times 10^{\prime}.8$, which is larger than the point spread functions of either observatory. The regions are indicated with green dashed lines in Figure \[fig:mosaic\_image\_sqreg\].
Figure \[fig:plot\_keV\_vs\_TeV\] shows a scatter plot between the [*Suzaku*]{} XIS count rate ($F_{\rm keV}$) and that of H.E.S.S. ($F_{\rm TeV}$). As seen in this plot, the X-ray count maps correlate strongly with the gamma-ray count map. The correlation coefficients are calculated to be 0.85 and 0.83, for 2–5 keV band and 5–10 keV band, respectively. It is worth noting that there are some deviations (X-ray intensity excesses) in the bright regions. Figure \[fig:map\_keV-TeV\] shows a map of $F_{\rm keV} - F_{\rm TeV}$ (for 2–5 keV band) overlaid with the H.E.S.S. contours, in which one can see that the X-ray excesses are present along the NW and SW rims.
DISCUSSION
==========
Cutoff in the Synchrotron Spectrum
----------------------------------
We conducted a series of [*Suzaku*]{} observations which covers about two-thirds of the surface of SNR RX J1713.7$-$3946. Through the data analysis, we successfully detected signals up to $\sim 40$ keV from each of the pointings. The HXD spectra above 10 keV are significantly steeper than those obtained from the XIS below 10 keV, suggesting that a spectral cutoff is common throughout the remnant. By combining the XIS and HXD spectra, we obtained a wide-band spectrum with high statistics, which clearly shows a cutoff around 10 keV.
Taking advantage of the high photon statistics, we performed a detailed study of the cutoff shape and compared it with a recent theoretical prediction by [@zira07]. A sharp cutoff of the accelerated electron spectrum is needed to reproduce the cutoff shape in the synchrotron spectrum detected with [*Suzaku*]{}. The spectrum of electrons derived from [*Suzaku*]{} data is in good agreement with the analytical model of [@zira07].
The cutoff energy in the spectrum of synchrotron radiation contains an important information about the efficiency of diffusive shock acceleration. For acceleration in the Bohm diffusion regime and when energy losses of electrons are dominated by synchrotron cooling, the cutoff energy, $\varepsilon_0$ in equation (\[za07\_sync\_spec\]) is expressed as [@zira07] $$\begin{aligned}
\varepsilon_0 = 0.55 \left( \frac{v_s}{3000~{\rm km}~{\rm s}^{-1}} \right)^2 \eta^{-1}~{\rm keV}, \end{aligned}$$ where $v_s$ is the shock speed and $\eta~(\ge1)$ is the so-called “gyrofactor”. The case of $\eta = 1$ corresponds to the “Bohm limit”, and implies high level of turbulence $\delta B \sim B$. The [*Suzaku*]{} spectrum is characterized by the best-fit parameter $\varepsilon_0 = 0.67~{\rm keV}$ which gives $v_s = 3300 \eta^{1/2}~{\rm km}~{\rm s}^{-1}$. Here, we assume that the shock speed $v_s$ is uniform throughout the remnant, which is supported by the fact that the outer boundary of the X-ray morphology is nearly circular. The upper-limit of the shock speed $v_s \le 4500~{\rm km}~{\rm s}^{-1}$ derived from the [*Chandra*]{} data [@uchi07], results in $\eta \le 1.8$. This is a strong evidence of acceleration of electrons in the regime close to the Bohm limit. Note that a similar result was obtained for the SW rim of the remnant in [@uchi07]. Here we confirm this conclusion for a larger area of the remnant with higher statistics.
Multi-wavelength Spectrum
-------------------------
While there is little doubt in the synchrotron origin of broad-band X-ray emission measured by [*Suzaku*]{}, the X-ray spectrum alone does not give preference to the strength of the magnetic field in the region of production of synchrotron radiation. Formally, the field can be as small as $10~\mu{\rm G}$ and as large as $100~\mu{\rm G}$. Meanwhile, the strength of the magnetic field has dramatic impact on the origin of TeV gamma-rays. The so-called leptonic or inverse Compton models require magnetic field between $10~\mu{\rm G}$ and $15~\mu{\rm G}$. Even so, it is difficult to achieve, at least within a simple one-zone model, a satisfactory explanation of both X-ray and TeV gamma-ray spectral features, unless we invoke an extremely high diffuse radiation field of optical photons to enhance the IC gamma-radiation below 1 TeV (see Figure \[fig:SED\_leptonic\_opt\]). A more realistic approach for explanation of the broad-band TeV gamma-ray spectrum within IC models can be realized under the assumption of existence an additional, low-energy electron component in the shell (see Figure \[fig:SED\_leptonic\_2ele\]). Even so, the most serious problem for IC models remains the requirement of low magnetic field in the gamma-ray production region, in contrast to large magnetic field required to explain the fast variability of X-ray emission on small scales. Formally, one may assume that gamma-rays are mainly produced in “voids”, i.e. in regions with very low magnetic field. This would imply quite inhomogeneous distribution of the magnetic field in the shell. One the other hand, the observed strong X-ray and TeV correlation within the IC models can be explained only in the case of homogeneous distribution of magnetic field.
The large-scale magnetic fields on parsec scales with an average strength larger than $\geq 15~\mu{\rm G}$ make the IC gamma-ray production inefficient, and thus give preference to the so-called hadronic models of gamma-rays produced at interactions of accelerated protons with the ambient gas via production and decay of secondary $\pi^0$-mesons. What concerns X-rays, they are produced, as in leptonic models, by synchrotron radiation of directly accelerated electrons. This is demonstrated in Figure \[fig:SED\] for very strong magnetic field, $B=200~\mu{\rm G}$. Note that while comparing the model predictions with measurements in the radio band, one should take into account that the radio points shown correspond to measurements of NW rim, while the X-ray and gamma-ray points are for the entire remnant. If the ratio of the radio flux from the NW rim to that from the whole remnant is not much different from the corresponding ratio in X-rays, the flux from the whole SNR should be significantly larger. This would reduce the difference between the measurements and predictions. In any case, the radio flux can be significantly reduced assuming somewhat smaller magnetic field or harder electron spectrum. Indeed, in Figure \[fig:SED\_100uG\] we show model calculations performed for a magnetic field $B=100~\mu{\rm G}$. While the synchrotron X-ray flux is described perfectly as before (in Figure \[fig:SED\]), the radio flux is by a factor of four lower; at 1.4 GHz it is 34 Jy, which is close to the latest estimates of radio flux from the whole remnant based on observations with ATCA and the 30-m radio telescopes of IRA (F. Acero et al. 2008, in preparation; G. Dubner 2008, private communications).
The radio flux can be suppressed even for the ambient field larger than $100~\mu {\rm G}$, provided that the electron injection spectrum is harder than $E^{-2}$. Figure \[fig:SED\_17\] demonstrates this possibility, where we assume an electron/proton index of $s = 1.7$ which corresponds to a compression ratio of $\sigma = 5.3$. Note that $\sigma$ can exceed the adiabatic upper limit of 4 as described by [@bere06]. Note that the value of $s = 1.7$ is consistent with the conclusion of [@villa07] based on semi-analytical derivation of the parent proton spectrum from the H.E.S.S. data. In model calculations shown in Figure \[fig:SED\_17\], the spectrum of protons requires an “early” exponential cutoff at $E_{p0} = 25.0~{\rm TeV}$. Note that formally the spectral index $s=1.7$ implies shock acceleration in non-linear regime which in fact predicts some deviation from pure power-law distribution of accelerated particles (see e.g. @Ellison07 [@bere06]). This would lead to further reduction of the radio flux.
The convection of low energy electrons could be another reason for low radio flux. Note that the escape of electrons through convection has strong impact only on low energy electrons; because of fast synchrotron cooling, the effect of escape is negligible for multi-TeV electrons. A significant quantity of low-energy electrons can escape from the shell of the SNR before emitting radio photons. Therefore, the radio flux can be reduced while the X-ray flux will remain unchanged.
For calculations shown in Figure \[fig:SED\], the total energy of electrons is estimated as $W_e = 3.1 \times 10^{46}~(d/1~{\rm kpc})^2~{\rm erg}$, and the energy for protons as $W_p = 2.7 \times 10^{50}~(n/1~{\rm cm}^{-3})^{-1}(d/1~{\rm kpc})^2~{\rm erg}$. In the case of harder energy spectra with power-law index $s=1.7$ corresponding to Figure \[fig:SED\_17\], one has $W_e = 6.0 \times 10^{45}~(d/1~{\rm kpc})^2~{\rm erg}$, and $W_p = 1.6 \times 10^{50}~(n/1~{\rm cm}^{-3})^{-1}(d/1~{\rm kpc})^2~{\rm erg}$, respectively. The proton/electron ratio in either case is very small, $K_{ep} \leq 10^{-4} (n/1~{\rm cm}^{-3})$. This value is significantly smaller than that for directly observed local cosmic rays ($K_{ep} \sim 0.01$), unless a large ambient matter density of $n \sim 100~{\rm cm}^{-3}$ is assumed. [@katz08] and [@butt08] argued that the hadronic scenario for this SNR has difficulties because the $K_{ep}$ value should be consistent with local cosmic rays and other SNRs. However, the $K_{ep}$ value of one SNR at a fixed age does not necessarily need to agree with the local cosmic ray value. The low-energy electrons are likely produced in later stages of SNR evolution, when the $K_{ep}$ value can be different from the present value. A comparison with other SNRs should be performed with care, as well.
Cutoff energy in the gamma-ray spectrum should give an important hint whether SNRs are sources of cosmic rays below the [*knee*]{}, if gamma rays observed by H.E.S.S. have hadronic origins. [@plaga08] argued that the cutoff energy of the H.E.S.S. spectrum of RX J1713$-$3946 is around 18 TeV, which can be translated to proton energy more than 10 times below the energy of the [*knee*]{}. Indeed our multi-wavelength study requires a cutoff in the proton spectrum around 100 TeV or even less for hard acceleration spectra of protons. However, one should take into account that even in the case of effective acceleration the highest energy protons beyond 100 TeV escape the source in a quite short time scales, and hence do not contribute to the gamma-ray production at the present epoch [@ptuskin05; @gabici07].
A unique feature of RX J 1713.7-3946 is the lack of thermal X-ray emission. Recently, [@katz08] and [@butt08] interpreted this fact as an argument against the hadronic model for TeV gamma rays. Generally it is true that plasma in young supernova remnants is heated to high temperatures observed via thermal X-ray emission of hot electrons. However, one should take into account that we deal with a unique object, and the lack of thermal X-ray emission cannot [*a priori*]{} be invoked as an argument against the hadronic origin of the observed TeV gamma rays.
It is important to note that in SNR shocks the formation of high plasma temperatures with $kT_i=3/16~m_i v_s^2$ is relevant only to protons (ions), and that a high ion temperature does not automatically (from first principles) mean a high electron temperature. In fact, the only known heating process of thermal electrons is Coulomb collisions between electrons and protons (ions), which, however, has too long time scale to establish electron-proton equipartition. On the other hand, we do know from X-ray observations that the electrons in young SNRs are heated to keV temperatures. This can be explained by assuming that a hypothetical mechanism, most likely related to the energy exchange through excited plasma waves, is responsible for effective electron heating in SNRs. As long as the nature of this mechanism in collisionless shocks remains unknown, one cannot predict, even qualitatively, the specifics of its operation on a source by source basis.
We indeed deal with two interesting facts. First, many young SNRs, like Tycho and Cassiopeia A, with intense thermal X-ray emission and intense nonthermal radio emission emit little (or do not emit at all) TeV gamma rays. On the other hand, RX J1713.7$-$3946, with lack of (or rather very low) thermal X-ray emission and with relatively weak nonthermal radio emission, is a source of powerful TeV radiation. These two facts can be treated as a hint for low efficiency of establishing equipartition in the thermal plasma in very effective TeV particle accelerators like RX J1713.7$-$3946. It is interesting to note in this regard, that such a tendency is found also for another effective TeV accelerator — SN 1006 [@vink03]. Whether the reduction of the exchange rate between different particle species in thermal plasma has a link to the particle acceleration in high Mach number shocks, as proposed by [@vink03], is a very interesting question to be explored in future deep theoretical and phenomenological studies. In this regard, RX J1713.7$-$3946 can serve as a key “template” source for such studies.
[@hughes00] also discussed low electron temperature based on [*Chandra*]{} observations of a young SNR in the Small Magellanic Could, 1E 0102.2$-$7219. They measured a blast-wave velocity of $\sim 6000~{\rm km}~{\rm s}^{-1}$ from the expansion rate and predicted the electron temperature of $kT_e > 2.5~{\rm keV}$ by considering Coulomb heating. However, the electron temperature derived from their spectral analysis is 0.5 keV, which is far below the prediction. According to their discussion, not only electron heating but also ion heating is suppressed and substantial fraction of energy may be going into cosmic-ray production due to the non-linear effects in the shock.
The nonlinear shock acceleration in this object can convert a significant, up to $f \sim 0.5$ fraction of the kinetic energy of explosion into relativistic particles. Correspondingly the fraction of available energy which goes to the heating of the ambient plasma will be reduced $1-f \sim 0.5$. Yet, conservative estimates show that plasma in RX J1713.7$-$3946 can be heated to quite high temperatures even the heating of electrons and protons proceeds only through the Coulomb exchange. This question recently has studied by [@Ellison07] for a standard SNR of age $t_{\rm SNR}=500$ yr and energy $E_{\rm SN}=10^{51}$ erg. In particular, it has been shown that in the case of effective diffusive shock acceleration and the plasma density $n=0.1~{\rm cm}^{-3}$, the ratio of synchrotron luminosity to thermal (bremsstrahlung) luminosity can be as large as 100. This implies that in the case of RX J1713.7$-$3946, from which thermal X-ray emission is not observed, the plasma density cannot significantly exceed $0.1~{\rm cm}^{-3}$ (the luminosity of thermal bremsstrahlung is proportional to $n^2$). The low density of order of $0.1~{\rm cm}^{-3}$ reduces the parameter space for hadronic models but does not exclude it. Indeed, as mentioned above, the TeV gamma-ray flux of RX J1713.7$-$3946 can be explained by interactions of protons if the parameter $A=(W_p/10^{50}~{\rm erg}) (n/1~{\rm cm}^{-3}) (d/1~{\rm kpc})^{-2}$ exceeds 1.5 to 3, depending on the spectrum of protons. Assuming that more than $30\%$ of the explosion energy of this SNR is released in accelerated protons, and that the plasma in the gamma-ray production region is compressed by a factor of few, we find that the located of the source at a distance of about 1 kpc would marginally support the hadronic model. While closer location of the source would make the model requirements quite viable and flexible, the location of the source beyond 1 kpc hardly can be accommodated within a standard shock acceleration scenario. If the SNR is closer, the distance of $d \gtrsim 0.5~{\rm kpc}$ seems reasonable considering the upper limit on the shock speed of $4500(d/1~{\rm kpc})~{\rm km}~{\rm s}^{-1}$ by [@uchi07].
Finally we should mention the model suggested by [@Malkov05] which can naturally explain both the low synchrotron flux at radio frequencies and lack of thermal X-ray emission of RX J1713.7$-$3946. The standard scenarios of gamma-ray production in SNRs assume that radiation is produced in downstream where the density of both relativistic particles and thermal plasma is higher than in upstream. However, in the cases when the shock is expanding into a low-density wind bubble and approaching cold dense material, e.g. swept-up shell or surrounding molecular clouds, the gamma-radiation is contributed predominantly from upstream. While the energy distribution of accelerated particles downstream is coordinate-independent in both linear and nonlinear regimes, the particle distribution upstream is coordinate-dependent. Because of energy-dependent diffusion coefficient, the high-energy particles diffuse ahead of low-energy particles, thus a dense material adjacent upstream will “see” relativistic particles (protons and electrons) with low-energy exponential cutoff, $E_{\rm min}$, which depends on the location of the dense regions. This implies that the effective production of TeV gamma-rays (from [*p-p*]{} interactions) and X-rays (from synchrotron radiation of TeV electrons) will be not accompanied by low energy (GeV) gamma-rays and synchrotron radio emission. Obviously, this model is not constraint by lack of thermal emission.
Morphology
----------
In addition to the spectral information, the comparison of X-ray and TeV gamma-ray images presented in Figure \[fig:plot\_keV\_vs\_TeV\] helps us to draw the physical picture of RX J1713.7$-$3946. Let us first discuss the tight correlation observed in most parts of the remnant. Within the hypothesis of hadronic origin of gamma-rays, the gamma-ray flux is proportional to the number densities of the ambient matter and relativistic protons, while the X-ray flux is proportional to number density of electrons. If the matter distribution significantly varies throughout the SNR, we need fine parameter tuning among the matter distribution, the electron injection rate, and the proton injection rate, in order to produce the tight correlation. Therefore, more natural explanation is that the matter density is uniform and the injection rate of the electrons and that of the protons are proportional to each other.
The X-ray flux excess along the NW and SW rims provides a unique probe of recent acceleration activity. Let us consider a “toy” model and compare its predictions with the observational results. In the toy model, the injection rate of electrons and protons keeps constant but increases by a factor of 1.5 in the last 10 yr only at the NW and SW rims. What is important here is the difference of cooling time between electrons and protons. The synchrotron cooling time of an electron emitting synchrotron photons with energy of $\varepsilon$ is given as equation (\[eq:sync\_cooling\_time\]), while the cooling time of protons due to $p$-$p$ interactions is expressed as $$\begin{aligned}
t_{pp} = 5.3 \times 10^7 \left( \frac{n}{1~{\rm cm}^{-3}} \right)^{-1}~{\rm yr}\end{aligned}$$ and is almost energy-independent. For the magnetic field $B=100 \ \rm \mu G$, the cooling times of electrons emitting 2 keV and 5 keV X-rays are $12~{\rm yr}$ and $7.6~{\rm yr}$, respectively. For any reasonable density of ambient gas, the cooling time of protons is much longer than the lifetime of this SNR. While the synchrotron X-rays we observe at present are emitted by electrons accelerated during the last $\sim 10~{\rm yr}$. the flux of $\pi^0$-decay gamma rays is provided by protons accelerated throughout the lifetime of the SNR. Figure \[fig:plot\_keV\_vs\_TeV\_toymodel\] shows the scatter plot of $F_{\rm keV}$ and $F_{\rm TeV}$ expected from the toy model, which shows a similar distribution to the observational results in Figure \[fig:plot\_keV\_vs\_TeV\]. The recent active acceleration increases the X-ray flux while keeping the gamma-ray flux almost unchanged.
SUMMARY
=======
We observed SNR RX J1713.7$-$3946 with the [*Suzaku*]{} observatory. Hard X-rays up to $\sim 40~{\rm keV}$ are detected from each of the 11 pointings. The hard X-ray morphology estimated by the HXD PIN data is generally consistent with an extrapolation of that in the energy region below 10 keV. When the HXD spectra are fitted with a power law, the photon indices are larger than those obtained from the XIS data. The difference of photon indices between the XIS and the HXD varies from region to region. Although this may suggest a variation of spectral shape or cutoff energy, the FoV of the HXD PIN is not small enough to allow detailed imaging spectroscopy. Such studies will become possible with upcoming missions with hard X-ray mirrors like [*NeXT*]{}, [*NuSTAR*]{}, or [*Simbol-X*]{}. Moreover these missions will extend the effective studies up to 100 keV. If the TeV gamma-ray emission of RX J1713.7$-$3946 is of hadronic origin, this should allow detection and study of synchrotron X-ray emission of electrons produced in proton-proton interactions via decays of secondary charged pions [@aha_book]. In this regard, the above mentioned missions can provide very effective tools for deep (at the level as low as $10^{-13} ~{\rm erg}~{\rm cm}^{-2}~{\rm s}^{-1}$) probes of hadronic processes in SNRs in the $\geq 100$ TeV energy regime with unprecedented (subarcmin) angular resolution.
Using the XIS and HXD data, we obtained a synchrotron spectrum in the energy range of two decades (0.4–40 keV), which means we can probe the parent electron distribution in the energy range of one decade regardless of the magnetic field strength. The wide-band coverage enables us to see a clear high-energy cutoff in the synchrotron spectrum, and for the first time allows detailed studies of the cutoff shape and derivation of the spectrum of the parent electrons. The spectral shape in the cutoff region was quantitatively evaluated, which revealed that the cutoff shape is compatible with a theoretical prediction by [@zira07]. Based on their model, the cutoff energy is obtained as $\varepsilon_0 = 0.67 \pm 0.02~{\rm keV}$. This result, together with upper limit of the shock velocity from [*Chandra*]{} [@uchi07], indicates that acceleration efficiency in this remnant approaches the Bohm limit.
We modeled the multi-wavelength spectrum of RX J1713.7$-$3946 within both the leptonic and hadronic gamma-ray production scenarios. The hadronic model nicely fits the data while it seems difficult to explain all observational data with a simple leptonic model. The major problem with leptonic models is related to the requirement of low magnetic field, $B \leq 15~\mu {\rm G}$ which does not agree with the recent discovery of variability of local regions of the shell on year timescales. The hadronic models recently have be criticized based on the lack of thermal emission and low radio flux. However, these fact can be accommodated within a standard shock acceleration model in a young SNR with magnetic field as large as $100~\mu {\rm G}$, plasma density $n \sim 0.1~{\rm cm}^{-3}$, and distance to the source $d \leq 1~{\rm kpc}$, assuming that more than 30% of energy of explosion is released in relativistic protons with spectrum as hard as $E^{-2}$ and exponential cutoff around 100 TeV.
Finally, besides the general strong correlation between X-ray and TeV gamma-ray emission, we found an excess X-ray emission compared to gamma-ray emission in the brightest regions of the remnant. This excess can be explained by recent activity accompanied by effective acceleration of electrons in localised reagions of the shell.
The authors would like to thank all the members of the [*Suzaku*]{} Science Working Group for their help in the spacecraft operation, instrumental calibration, and data processing. The authors thank Una Hwang for carefully reading the manuscript and Stefan Funk for assisting with the comparison of the morphologies with the H.E.S.S. data and for his helpful comments on the manuscript. The authors also thank Misha Malkov and Vladimir Zirakashvili for discussions related to different aspects of the diffusive shock acceleration theory. T. Tanaka and A. Bamba are supported by research fellowships of the Japan Society for the Promotion of Science for Young Scientists.
Aharonian, F. A., et al. 2001, , 370, 112
Aharonian, F. A. 2004, Very High Energy Cosmic Gamma Radiation: A Crucial Window on the Extreme Universe (World Scientific)
Aharonian, F. A., et al. 2004, , 432, 75
Aharonian, F. A., et al. 2006, , 449, 223
Aharonian, F. A., et al. 2007, , 464, 235
Berezhko, E. G., & Völk, H. J. 2006, , 451, 981
Bird, A. J., et al. 2006, , 636, 765
Blandford, R., & Eichler, D. 1987, , 154, 1
Butt, Y., Porter, T., Katz, B., & Waxman, E. 2008, , in press
Cassam-Chenaï, G., Decourchelle, A., Ballet, J., Sauvageot, J.-L., Dubner, G., & Giacani, E. 2004, , 427, 199
Ellison, D.C., Patnaude, D.J., Slane, P., Blasi, P., Gabici, S. 2007, , 661, 879
Enomoto, R., et al. 2002, , 416, 823
Fukui, Y., et al. 2003, , 55, L61
Gabici, S., & Aharonian, F. A. 2007, , 665, L131
Gruber, D. E. Matteson, J. L.; Peterson, L. E., & Jung, G. V. 1999, , 520, 124
Hartman, R. C., et al. 1999, , 123, 79
Hiraga, J. S., Uchiyama, Y., Takahashi, T., & Aharonian, F. A. 2005, , 431, 953
Hughes, J. P., Rakowski, C. E., & Decourchelle, A. 2000, , 543, L61
Ishida, M., et al. 2006, Suzaku Memo, JX-ISAS-SUZAKU-MEMO-2006-40
Ishisaki, Y., et al. 2007, , 59, S113
Katz, B., & Waxman, E. 2008, Journal of Cosmology and Astroparticle physics, in press
Kelner, S. R., Aharonian, F. A., & Bugayov, V. V. 2006, , 74, 034018
Kokubun, M., et al. 2007, , 59, S53
Koyama, K., Petre, R., Gotthelf, E. V., Hwang, U., Matsuura, M., Ozaki, M., & Holt, S. S. 1996, , 378, 255
Koyama, K., Kinugasa, K., Matsuzaki, K., Nishiuchi, M., Sugizaki, M., Torii, K., Yamauchi, S., & Aschenbach, B. 1997, , 49, L7
Koyama, K., et al. 2007, , 59, S23
Lazendic, J. S., Slane, P. O., Gaensler, B. M., Plucinsky, P. P., Hughes, J. P., Galloway, D. K., & Crawford, F. 2003, , 593, L23
Lazendic, J. S., Slane, P. O., Gaensler, B. M., Reynolds, S. P., Plucinsky, P. P., & Hughes, J. P. 2004, , 602, L201
Malkov, M.A., Diamond, P.H., Sagdeev, R.Z. 2005, , 624, L37
Malkov, M. A., and Drury, L. O’C. 2001, Rep. Prog. Phys., 64, 429
Mitsuda, K., et al. 2007, , 59, S1
Mizuno, T., et al. 2006, Suzaku Memo, JX-ISAS-SUZAKU-MEMO-2006-42
Moriguchi, Y., Tamura, K., Tawara, Y., Sasago, H., Yamaoka, K., Onishi, T., & Fukui, Y. 2005, , 631, 947
Muraishi, H., et al. 2000, , 354, L57
Pfeffermann, E., & Aschenbach, B. 1996, in Röntgenstrahlung from the Universe, ed. H. U. Zimmermann, J. Trümper, & H. Yorke, MPE Rep., 263, 267
Plaga, R. 2008, , 453, 48
Porter, T. A., Moskalenko, I. V., & Strong, A. W. 2006, , 648, L29
Ptuskin, V. S., & Zirakashvili, V. N. 2005, , 429, 755
Serlemitsos P. J., et al. 2007, , 59, S9
Slane, P., Galensler, B. M., Dame, T. M., Hughes, J. P., Pluncinsky, P. P., & Green, A., 1999, . 525, 357
Slane, P., Hughes, J. P., Edgar, R. J., Plucinsky, P. P., Miyata, E., Tsunemi, H., & Aschenbach, B. 2001, , 548, 814
Takahashi, T., et al. 2007, , 59, S35
Takahashi, T., et al. 2008, , 60, S131 (Paper I)
Terada, Y., et al. 2005, IEEE Trans. Nucl. Sci., 52, 902
Uchiyama, Y., Aharonian, F. A., & Takahashi, T. 2003, , 400, 567
Uchiyama, Y., Aharonian, F. A., Takahashi, T., Hiraga, J. S., Moriguchi, Y., & Fukui, Y. 2005, High Energy Gamma-Ray Astronomy, 745, 305
Uchiyama, Y., Aharonian, F. A., Tanaka, T., Takahashi, T., & Maeda, Y. 2007, , 449, 576
Villante, F. L., & Vissani, F. 2007, , 76, 125019
Vink, J., Laming, J. M., Gu, M. F., Rasmussen, A., & Kaastra, J. S. 2003, , 587, L31
Wang, Z. R., Qu, Q.-Y., & Chen, Y. 1997 , 318, L59
Watanabe, S., et al. 2007, Suzaku Memo, JX-ISAS-SUZAKU-MEMO-2007-01
Zirakashvili, V. N., & Aharonian, F. 2007 , 465, 695
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'A new and universal method for implementing scale invariance, called *best matching*, is presented. It extends to scaling the method introduced by Bertotti and the author to create a fully relational dynamics that satisfies Mach’s principle. The method is illustrated here in the context of non-relativistic gravitational particle dynamics. It leads to far stronger predictions than general Newtonian dynamics. The energy and angular momentum of an ‘island universe’ must be exactly zero and its size, measured by its moment of inertia, cannot change. This constancy is enforced because the scale invariance requires all potentials to be homogeneous of degree -2. It is remarkable that one can nevertheless exactly recover the standard observed Newtonian laws and forces, which are merely accompanied by an extremely weak universal force like the one due to Einstein’s cosmological constant. In contrast to Newtonian and Einsteinian dynamics, both the gravitational constant G and the strength of the cosmological force are uniquely determined by the matter distribution of the universe. Estimates of their values in agreement with observations are obtained. Best matching implements a *dynamics of pure shape* for which the action is a dimensionless number. If the universe obeys such scale-invariant law, steadily increasing inhomogeneity, not expansion of the universe, causes the Hubble red shift. The application of best matching to geometrodynamics is treated in a companion paper.'
---
**[SCALE-INVARIANT GRAVITY: PARTICLE DYNAMICS]{}**
Electronic address: [email protected]
Introduction
============
The Euclidean symmetries of three-dimensional space (*3-space*) do not fully determine Newtonian kinematics. Absolute time and inertial frames are puzzling extra structures. They are eliminated in this paper in a dynamics that is *fully invariant* under Euclidean translations, rotations *and* dilatations. Besides results of independent interest, this paper introduces the techniques of the companion paper [@ABFO], which modifies Einstein’s general relativity (GR), making it scale invariant in a similar manner. Both papers extend the *3-space approach* of [@BOF; @AB], which derives spacetime kinematics, including the universal light cone and the gauge principle, from 3-space first principles.
Weyl [@Weyl1] made the first serious attempt to create a scale-invariant (spacetime) dynamics, but his use of a conformally transformed 4-vector field to compensate conformal transformations of the gravitational field came to grief. Einstein [@Einstein] noted that it would make atoms emit spectral lines with history-dependent frequencies. This contradicted the observed facts. Many years later, Dirac [@DiracSI] revived Weyl’s idea in a simplified form in which the compensating 4-vector field is replaced by a simpler scalar compensating field. In his theory, there is no longer the flagrant disagreement with observation associated with Einstein’s spectral argument. Nevertheless, Einstein would probably still have had reservations, since in Dirac’s theory (which is a form of Brans–Dicke theory [@BD]) the coupling constants of the various forces of nature are epoch and location dependent. The strong equivalence principle (SEP), according to which the laws of nature, including the values of coupling constants, must be exactly the same everywhere and at every epoch, is violated. Dirac regarded this as a virtue, since he hoped it would justify his Large Numbers Hypothesis, but others would prefer the rigour of the SEP. From the late eighties, numerous authors have proposed Dirac-type theories [@Wetterich]. The scalar compensating field is now widely called the dilaton, and it has been conjectured [@Wetterich] that its vacuum expectation value determines the Planck length in an emergent fashion. A dilaton field is a necessary concomitant of the Einsteinian graviton in string theory, and there is now a vast literature on related matters.
There are reasons to be unhappy about the present situation. The general principles of physics impose few restrictions on scalar fields, so the scope for ‘ad hocery’ in order to match the latest observations is great. Even in string theory, the many different compactification possibilities lead to different dilatonic properties, and there seems to be little hope at this stage of unambiguous predictions for the form of the resulting effective four-dimensional theory. In addition, quantization, which often fails to respect classical scale invariance, introduces further uncertainty in the form of anomalies and arbitrary cutoffs.
In this and the companion paper [@ABFO], a new and universal approach to scale invariance is proposed. It makes no use of compensating fields at all and leads directly and naturally to a mechanism whereby the strong equivalence principle can be maintained. Since scalar fields play no essential role in the theory, there is much less scope for fudging. Scale-invariant gravity makes at least some unambiguous predictions, and it is therefore decidedly vulnerable to experimental disproof.
The idea behind scale-invariant gravity is in fact closely related to Weyl’s original inspiration, which was to implement Descartes’s dream of explaining all of physics by geometry. In 1919, despite Einstein’s reservations, Weyl wrote [@Weyl2]: “Descartes’s dream of a purely geometrical physics is fulfilled in a remarkable way, though admittedly not one that he could have foreseen. In its concepts, physics in no way goes beyond geometry.” However, years later he admitted defeat and wrote [@Weyl3] “the facts of atomism teach us that length is not relative but absolute […]{} physics can never be reduced to geometry as Descartes had hoped.” He came to this conclusion because the failure of his theory had persuaded him that physics did not respect the symmetries, like scaling, that one would expect it to inherit from space. He attributed the failure to the effect of quantum mechanics. As will be explained in this and the companion paper [@ABFO], scale-invariant gravity exploits the symmetries that so appealed to Weyl’s intuition, but does so without compensating fields. Its realization in particle dynamics in this paper uses no concepts that, except for the Newtonian notion of point masses, go beyond those of the three-dimensional geometry axiomatized by Euclid. The geometrodynamical realization in [@ABFO] uses nothing but the concepts of modern three-dimensional differential geometry, above all Riemannian 3-spaces and scalar and 3-vector fields defined on them.
The key insight into scale-invariant gravity is the realization that the apparent breaking of scale invariance might not be due to quantum mechanics but to *an inadequate theory of inertia*. As currently described, inertia violates scale invariance. This fact seems to have escaped notice. Moreover, the defect can be remedied without compensating fields. Our approach to scale invariance is direct and, we believe, new. It is universal because it can be implemented for any spatial geometry that admits a similarity (scaling) symmetry.
The claim that the existing descriptions of inertia break scale invariance develops a criterion taken from Poincaré’s discussion of absolute and relative motion in [@Poincare1]. It demands that, for any given spatial geometry, dynamics must be *maximally predictive* in a well-defined sense. Relationists like Mach [@Mach] had argued that particle dynamics should be expressed solely using quantities invariant under Euclidean translations and rotations: the mutual separations $r_{ij}$ of the particles. But Poincaré pointed out that this can always be done through an automatic process of elimination. The real test is whether specification at an initial time of the $r_{ij}$ and their time derivatives $\dot{r}_{ij}$ (together with the particle masses and the force law) suffices to predict the future evolution (of an isolated system). Newtonian theory fails this precisely formulated prediction test because the $r_{ij}$ and $\dot{r}_{ij}$ contain no information about the total angular momentum **J**, and very different evolutions result for different values of **J**. As a result, when Newton’s equations are expressed in terms of the $r_{ij}$ they contain third derivatives with respect to the time. Clearly, such a theory is less predictive than one containing only second derivatives. For more details, see [@Barbour2001].
Poincaré did not attempt to formulate a maximally predictive dynamics. Like Weyl later in the case of scaling, he simply accepted that nature did not respect mathematical intuition. This resignation may have been premature. Maximal predictive strength is a powerful constructive principle.[^1] It imposes very characteristic structures on theories that possess the property. The papers [@BOF; @AB] have shown how it can explain some very basic known laws of nature: the universal light cone, the gauge principle, and Einstein’s gravitational field equations. Now we use it in the hope of discovering *new physics*.
The first step is to push the principle to its limit. For Poincaré still allowed Newton’s absolute time and length scale. Let us introduce configuration spaces. Let T be the one-dimensional space of Newtonian absolute times, and Q be the $3N$-dimensional Newtonian configuration space of $N$ point particles of masses $m_{i}$. Let $\textrm{Q}_{\textrm{\scriptsize
RCS}}$ (RCS stands for *relative configuration space*) be the $(3N-6)$-dimensional space obtained from Q by quotienting with respect to the Euclidean translations and rotations. Thus, a point in $\textrm{Q}_{\textrm{\scriptsize RCS}}$ is the entire 6-dimensional group orbit in Q generated by them from a point $q\in Q$. All particle configurations that can be carried into each other by translations and rotations are identified. Finally, quotienting by the dilatations, we arrive at the $(3N-7)$-dimensional space $\textrm{Q}_{0}$ of shapes of $N$-particle configurations: *shape space*.
Standard Newtonian theory is formulated in the space $\textrm{Q}\times\textrm{T}$. Given the masses and the force law, its equations predict the future given an initial point $qt$ in $\textrm{Q}\times\textrm{T}$ and initial direction in $\textrm{Q}\times\textrm{T}$ at $qt$. (The direction in Q determines the direction of the momenta, and the slope in T their magnitude, i.e., the particle speeds.) Poincaré’s criterion is met in $\textrm{Q}\times\textrm{T}$ but not in $\textrm{Q}_{\textrm{\scriptsize RCS}}\times\textrm{T}$ and, a fortiori, not in $\textrm{Q}_{0}$.
There are two known ways of constructing a maximally predictive particle dynamics. The first uses directly the particle separations $r_{ij}$ or, in a scale-invariant theory, ratios of them. Schrödinger [@Schrodinger] was one of many who explored this route in the 20th century, but it leads to anisotropic effective masses that are ruled out by sensitive experiments [@HD]. The absence of mass anisotropy is a very tough experimental test of relational theories. In [@BB], Bertotti and I proposed a maximally predictive form of variational dynamics that is appropriately called *best matching*. This method leads to isotropic masses and is presented in section 4. In particle dynamics, it leads to a small subset of the solutions allowed by Newton’s equations. The conditions that select them show how elimination of non-relational kinematics leads to a dynamics that is fully determined by its underlying geometry. In [@ABFO], it leads to a scale-invariant generalization of GR: *conformal gravity*.
Each symmetry has its own distinctive best matching and imposes a constraint on the canonical momenta $\textbf{p}_{i}$. Simultaneously, the best matching imposes further conditions on the Lagrangian and even on the solutions (in conformal gravity) that ensure propagation of the constraint. It is illuminating to exhibit the constraints and list the additional conditions.
Translational symmetry constrains the total momentum (all sums are over $i=1-N$): $$\textbf{P}=\sum\textbf{p}_{i}=0,
\label{MC}$$ and the potential $U$ must be translational invariant, so Newton’s third law is enforced. These are not new restrictions, since Galilean invariance enforces the invariance and ensures (\[MC\]) holds in the centre-of-mass (c.m) frame.
Rotational best matching leads to the non-trivial vanishing of the total angular momentum $\textbf{J}$:
$$\textbf{J}=\sum\textbf{x}_{i}
\times{\textbf{p}_{i}}=0,
\label{JC}$$
and in conjunction with translational best matching enforces $U=U(r_{ij})$. At this stage, the energy $E$ is arbitrary, and inertial motion, for which $U=\textrm{constant}$, is possible. The results (\[MC\]) and (\[JC\]) were obtained in [@BB]. The unique (vanishing) value of $\textbf{J}$ imposed by best matching eliminates the ambiguity of prediction noted by Poincaré.
The extension of best matching to dilatations is new and leads to the most interesting result. It concerns the scalar quantity $ D=\sum{\textbf{p}_{i}}\cdot{\textbf{x}_{i}}$, which has not, so far as I know, been given a name in the literature. Since it bears the same relation to expansion as angular momentum does to rotation (and has the same dimensions – action), I shall call it the *dilatational momentum*. Dilatational best matching forces it to vanish: $$\ D=\sum{\textbf{p}_{i}}\cdot{\textbf{x}_{i}}=0.
\label{VC}$$ In addition, $U$ must be homogeneous of degree -2 in $r_{ij}$, and $E$ must be zero. Thus pure inertial motion, for which $U=\textrm{constant}$ and $E>0$, is not allowed. It is in this sense that inertia violates scaling. *There is no maximally predictive inertial dynamics on shape space*. One cannot formulate a theory of pure inertial motion without introducing additional kinematic structure – an absolute scale of length – that mathematical intuition suggests one should not employ.[^2] If one wishes to have any dynamics at all on shape space that satisfies the Poincaré criterion, *it must include forces and have vanishing energy*. This casts new light on the old problem highlighted by Hertz [@Hertz; @Lanczos]: why does energy exist in two disparate and independent forms, kinetic and potential? Their separate existence is the reason why pure inertial motion is allowed in Newtonian mechanics. In scale-invariant dynamics, this cannot be. The two forms of energy must come as inseparable twins and their sum must be zero. If the universe is scale invariant, this has implications for dimensional analysis and our understanding of the constants of nature (for a review of the present understanding, see [@Barrow]). In fact, it leads to *determination of the gravitational constant* G (Secs. 6 and 7).
So far as I know, formulating a scale-invariant theory through a requirement on the form of the initial-value problem in shape space has not hitherto been considered. There is no indication of there having been work done in this direction in the often-cited review [@Fulton]. To avoid confusion, the reader should keep in mind this new criterion of scale invariance (or conformal invariance). It is used throughout both this paper and the companion [@ABFO].
Whereas $\textbf J$, like $\textbf M$, is conserved for isolated Newtonian systems, nothing forces $\textbf J=0$. Moreover, the dilatational momentum $D$ is not even conserved in general let alone zero. The best-matching conditions are therefore non-trivial, and (\[VC\]) is decidedly restrictive. Note also that (\[MC\])–(\[VC\]) and the associated conditions on $U$ and $E$ are as purely geometrical in origin as the operators grad, curl and div of vector analysis, to which they exactly correspond. This reflects the universal geometrical nature of best matching.
It might seem that dilatational best matching is academic, since gravity and electrostatics have $1/r$ potentials and not the $1/r^{2}$ potentials required by (\[VC\]). However, there exists a unique, natural and universal way to select $1/r^{2}$ potentials that are manifested as effective $1/r$ potentials indistinguishable from the Newtonian counterparts (section 5). The strong equivalence principle is then satisfied for all these forces, including gravity. The sole result of this procedure is to introduce a weak long-range force $\textbf F$ like the one generated by Einstein’s cosmological constant $\Lambda$. However, unlike $\Lambda$, the strength of $\textbf F$ is uniquely determined by the potential of the system and is therefore epoch dependent. Moreover, $\textbf
F$ forces the ‘size’ of the universe, measured by the moment of inertia $I$ (and by the spatial volume $V$ in conformal gravity), to remain exactly constant. We shall see that there is a sense in which the cosmological force introduces an weak violation of the strong equivalence principle. It is, however, a violation that is reduced to the absolute minimum that is possible.
Another important thing to note is that the entire treatment of this and the companion paper [@ABFO] is based on the premise that the universe can be treated as a self-contained closed dynamical system. In the particle model, the moment of inertia of an $N$-particle system taken to represent an ‘island universe’ in Euclidean space plays an essential physical role in the theory. In conformal gravity [@ABFO], a similar role is played by the 3-volume of the universe. Scale invariance that respects the SEP relies on the use of ratios of local separations divided by global quantities: the moment of inertia and the 3-volume, respectively. Thus, the assumption of a self-contained universe is essential.
Section 2 shows how the condition (\[VC\]) can arise as a very exceptional case within Newtonian theory and by how much Newtonian theory fails to be maximal predictive. Section 3 shows how time is eliminated by Jacobi’s principle for the dynamical orbit of a system in its configuration space. (Dynamical orbits should not be confused with group orbits. Both exist in configuration space, and both play important roles in best matching.) Section 4 introduces best matching and develops the necessary formal techniques in Lagrangian form. Section 5 formulates the rule for passing from an arbitrary set of Newtonian potentials to scale-invariant counterparts with the associated cosmological force $\textbf F$. Section 6 considers how dimensional analysis is changed by the elimination of time and an absolute scale of length. It also shows how scale invariance leads to an explicit expression for the gravitational constant G and the strength of **F**. Section 7 contains estimates of the strength of $\textbf F$ predicted by scale-invariant gravity and shows how the explicit expression for G in terms of the matter distribution of the universe permits a determination of the mass and size of the universe from the empirical value of G. Section 8 considers whether the actual universe is scale invariant. The greatest need is for an explanation of the Hubble red shift that does not rely on expansion of the universe. It is this that makes the theory vulnerable to experimental disproof. Section 9 considers possible quantum implications. Finally, the Appendix gives the Hamiltonian form of the theory.
The Lagrange–Jacobi Relation
============================
In the c.m frame, let an isolated system have Lagrangian ${\cal{L}}=T-U$ with $T={\sum}{(m_{i}/2)}{\dot{\textbf{x}}_{i}}\cdot{\dot{\textbf{x}}_{i}}$ and potential $U$. The energy $E=T+U$ is conserved. The c.m moment of inertia $I$ is $$I=\sum m_{i}\textbf{x}_{i}\cdot\textbf{x}_{i}\equiv
{1\over
M}\sum_{i<j}m_{i}m_{j}r_{ij}^{2},\hspace{.5cm}M=\sum{m_{i}}.
\label{MofI}$$ Lagrange and Jacobi noted[^3] that $
\ddot{I}=2{\sum}m_{i}{\dot{\textbf{x}}_{i}}\cdot{\dot{\textbf{x}}_{i}}
+2{\sum}m_{i}{\textbf{x}_{i}}\cdot{\ddot{\textbf{x}}_{i}}$ has important properties. By Newton’s second law $m_{i}\ddot{\textbf{x}}_{i}=-\partial{U}/\partial\textbf{x}_{i}$ and the definition of $T$, we find $
\ddot{I}=4T-2\sum\textbf{x}_{i}\cdot{\partial U\over\partial
\textbf{x}_{i}}$. If now $U$ is homogeneous of degree $k$, then, using Euler’s theorem and $T=E-U$, we find $$\ddot{I}=4(E-U)-2kU.
\label{Iddh}$$ Consider Newtonian celestial mechanics, for which $k=-1$. Then $\ddot{I}=4E-2U$, from which Lagrange deduced the first qualitative result in dynamics. Since $U<0$ for gravity, $E\geq 0$ implies $ \ddot I>0.$ Thus $I$ is concave upwards and must tend to infinity as $t\longrightarrow +\infty$ and $t\longrightarrow -\infty$. In turn, this means that at least one of the interparticle distances must increase unboundedly, so that any system with $E\geq 0$ is unstable.
Another consequence of (\[Iddh\]) is the virial theorem. For suppose that the system has virialized, so that $I\approx0$. Then $4E=(2k+4)U.$
For our purposes, the most interesting consequence of (\[Iddh\]) arises when $k=-2$. For then $$\ddot I=4E.
\label{Ih-2}$$
Thus, $I$ has the parabolic dependence $I=2Et^{2}+bt+c$ on the time and will tend rapidly to zero or infinity. Such a system is extremely unstable, either imploding or exploding.
However, suppose $E=0$. Then $\ddot I=0$ by (\[Ih-2\]), so that $\dot I=2\sum m_{i}\dot\textbf x_{i}\cdot \textbf x_{i}=2\sum
\textbf p_{i}\cdot \textbf x_{i}=\textrm{constant}.$ Thus, the dilatational momentum $D=\sum \textbf p_{i}\cdot \textbf x_{i}$ is conserved if both $U$ is homogeneous of degree -2 and $E=0$. (The arbitrary additive constant in $E$ is fixed by requiring $U\longrightarrow
0$ when all $r_{ij}\longrightarrow 0$.) Obtained thus, conservation of $D$ is a fluke. However, as we just noted, dilatational best matching requires not merely conservation but vanishing of $D$. The homogeneity of $U$ and vanishing of $E$ are enforced by the symmetry. Homogeneity is crucial for scale invariance.
Newtonian dynamics has always been seen as the paradigm of rationality. However, Poincaré’s analysis [@Poincare1] of the initial-value problem shows that rationality is ‘arena dependent’. If one treats the universe as an isolated dynamical system in shape space, Newtonian dynamics needs more initial data than seem necessary. It is worth spelling this out for the three-body problem, for which shape space Q$_{0}$ is the two-dimensional space of triangle shapes. Let it be coordinatized by two angles $\alpha$ and $\beta$ of the triangle. If they are taken as dependent and independent variables, respectively, the orbit in Q$_{0}$ for a theory that satisfies the Poincaré criterion should be described by $${\textrm{d}^{2}\alpha\over
\textrm{d}\beta^{2}}=f\left(\alpha,\beta,{\textrm{d}\alpha\over
\textrm{d}\beta}\right),
\label{Dziobek}$$ and three initial data would be needed: $\alpha,\beta,{\textrm{d}\alpha/\textrm{d}\beta}.$
Newtonian theory fails this ideal of a *dynamics of pure shape*. To determine the curve of a generic solution in Q$\times$T projected down to Q$_{0}$, *five* further data are needed. Two fix the direction of $\textbf J$ relative to the instantaneous triangle, one is the constant ratio (rotational K.E/total K.E), another is the varying ratio $T/U$, and the final one is the varying ratio (K.E in change of size/K.E in change of shape). There is a similar mismatch for the $N$-body problem. All the extra data look natural in the Newtonian arena but incongruous in shape space. Rotational best matching eliminates the first three, dilatational best matching the other two.
Jacobi’s Principle
==================
The first step to a dynamics of pure shape is the elimination of time by Jacobi’s principle [@Lanczos], which describes all Newtonian motions of one $E$ as geodesics on configuration space. This topic has already been discussed in [@BOF; @CQG94; @EOT], so the ground already covered there will be only briefly recapitulated. However, this section will also consider the important issue of dimensions in a timeless and scale-invariant theory, which has not hitherto been discussed.
For $N$ particles of masses $m_{i}$ with potential $U(\textbf{x}_{1}, \dots , \textbf{x}_{N})$ and energy $E$, the Jacobi action is [@Lanczos] $$I_{\scriptsize\textrm{Jacobi}} = 2\int\sqrt{E -
U}\sqrt{\tilde T} \textrm{d}\lambda,
\label{Jacobi}$$ where $\lambda$ labels the points on trial curves and $ \tilde{T}
= \sum {m_{i} \over
2}{\textrm{d}{\textbf{x}}_{i}\over\textrm{d}\lambda}\cdot
{\textrm{d}{\textbf{x}}_{i}\over\textrm{d}\lambda}$ is the parametrized kinetic energy. The action (\[Jacobi\]) is timeless since the label $\lambda$ could be omitted and the mere displacements $d\textbf{x}_{i}$ employed, as is reflected in the invariance of $I_\textrm{\scriptsize{Jacobi}}$ under the reparametrization $$\lambda \rightarrow f(\lambda).\label{rep}$$
In fact, it is much more illuminating to write the Jacobi action in the form $$I_{\scriptsize\textrm{Jacobi}} = 2\int\sqrt{E -
U}\sqrt{T^*},\hspace{.5cm}T^*=\sum {m_{i} \over
2}\textrm{d}{\textbf{x}}_{i}\cdot
{\textrm{d}{\textbf{x}}_{i}},
\label{Jacobi*}$$ which makes its timeless nature obvious and dispenses with the label $\lambda$.
The characteristic square roots of $I_{\scriptsize\textrm{Jacobi}}$ fix the structure of the canonical momenta: $$\textbf{p}_{i} =
{\partial {\cal L} \over \partial(\textrm{d}\textbf{x}_{i}/\textrm{d}\lambda)} =
m_{i}\sqrt{E - U \over {\tilde T}} {d \textbf{x}_{i}
\over d \lambda},
\label{CanMom}$$ which, being homogeneous of degree zero in the velocities, satisfy the constraint [@Dirac] $$\sum{{{\textbf{p}}_{i}}\cdot
{{\textbf{p}}_{i}}\over 2m_{i}} = E -
U.
\label{QuadCon}$$
The Euler–Lagrange equations are $${\textrm{d}\textbf{p}^{i} \over \textrm{d} \lambda}= {\partial
{\cal L} \over \partial \textbf{x}_{i}} = -\sqrt{{\tilde T} \over
E - U}{\partial U\over
\partial \textbf{x}_{i}} ,
\label{JacobiEL}$$ where $\lambda$ is still arbitrary. If we choose it such that $${{\tilde T} \over E - U} = 1 \Rightarrow {\tilde T} = E - U
\label{EnCon}$$ then (\[CanMom\]) and (\[JacobiEL\]) become $$\textbf{p}_{i} = m_{i}{{\textrm{d}\textbf{x}_{i}} \over \textrm{d}
\lambda},\hspace{1.0cm} {\textrm{d} \textbf{p}_{i} \over
\textrm{d}\lambda} = -{\partial U\over\partial \textbf{x}_{i}},$$ and we recover Newton’s second law w.r.t this special $\lambda$. However, (\[EnCon\]), which is usually taken to express energy conservation, becomes the *definition of time*. Indeed, this emergent time, chosen to make the equations of motion take their simplest form [@Poincare2], is the astronomers’ operational ephemeris time [@Clemence]. It is helpful to see how ‘change creates time’. The increment $\delta t$ generated by displacements $\delta \textbf{x}_{i}$ is $$\delta t={\sqrt{\sum
m_{i}\delta\textbf{x}_{i}\cdot\delta\textbf{x}_{i}}\over\sqrt{2(E-U)}}
\equiv{\delta s\over\sqrt{2(E-U)}}.
\label{clem}$$ Each particle ‘advances time’ in proportion to the square root of its mass and to its displacement, the total contribution $\delta s$ being weighted by $\sqrt{2(E-U)}$. In a scale-invariant theory, $E=0$, and $\delta t$ takes an especially suggestive form: a given kinetic $\delta s$ ‘advances time’ by an amount inversely proportional to $|U|$.
Best Matching
=============
Our goal is a theory of Jacobi-type geodesics on shape space Q$_{0}$. To do this, we might interpolate shapes $A_{0},B_{0}\in{\textrm{Q}_{0}}$ with suitably continuous curves $q_{0}(\lambda)$ and extremalize an action of the form $${I_{\scriptsize\textrm{Q}}}_{0}=\int_{\lambda_{A_{0}}}^{\lambda_{B_{0}}}
\textrm{d}\lambda
\sqrt{f_{ij}{\textrm{d}q_{0}^{i}\over
\textrm{d}\lambda}{\textrm{d}q_{0}^{j}\over
\textrm{d}\lambda}},$$ w.r.t certain shape coordinates $q_{0}^{i}, \dots,q_{0}^{3N-7}$ on Q$_{0}$. The $f_{ij}$ will be functions of the shape coordinates, $f_{ij}=f_{ij}(q_{0}^{1},\dots,q_{0}^{3N-7})$, and the usual summation is understood. But shape coordinates are awkward, and this approach is impracticable.
Next, we try to work in Q and attempt an action of similar form: $$I_{\scriptsize\textrm{Q}}=\int_{\lambda_{A}}^{\lambda_{B}}
\textrm{d}\lambda
{\cal L},\hspace{.5cm}{\cal L}=\sqrt{f_{ij}{\textrm{d}q^{i}\over
\textrm{d}\lambda}{\textrm{d}q^{j}\over
\textrm{d}\lambda}},$$ where now $i=1,\dots,3N$ and $f_{ij}=f_{ij}(q)$. Since $I_{\scriptsize\textrm Q}$ should depend only on the projected curve $q_{0}(\lambda)$ in Q$_{0}$, we seek to make $\cal L$ invariant under $\lambda$-dependent transformations that shift the $q(\lambda)$’s by arbitrary $\lambda$-dependent amounts along their group orbits. The $\lambda$-dependence is decisive.
Let us consider translations in one dimension and infinitesimally differing $A$ and $B$. Then $\lambda_B-\lambda_A=\delta\lambda$ can be infinitesimal, and to first order we shall be minimizing $$I_{\textrm{\scriptsize{TS}}}=\sqrt{f_{ij}\delta
q^{i}\delta q^{j}},\hspace{.5cm}\delta
q^{i}={\textrm{d}q^{i}\over\textrm{d}\lambda}\delta\lambda,
\label{TS}$$ where TS stands for thin sandwich. The action (\[TS\]) must be stationary to first order under $$q^{i}(\lambda)\longrightarrow
q^{i}(\lambda)+b(\lambda)e^i
\label{BM3}$$ for *all* choices of $b(\lambda)$. Here $e^i=1$ for all $i$ is the generator of translations for particle $i$. If this is so, $I_{\scriptsize{\textrm{TS}}}$ will depend only on the orbits of $A$ and $B$, since different choices of $b(\lambda)$ move $A$ and $B$ arbitrarily along their orbits. Note that under (\[BM3\]) the velocities change: $$v^{i}\longrightarrow
v^{i}+b'e^i,\hspace{.5cm}v^{i}={\textrm{d}q^{i}\over
\textrm{d}\lambda},\hspace{.5cm}b'={\textrm{d}b\over
\textrm{d}\lambda}.
\label{BM4}$$
Let us make a Taylor expansion of $b(\lambda)$: $$b(\lambda)=b(0)+
b'(0)\delta\lambda+\dots.$$
The *thin-sandwich problem* (TSP) requires us to minimize (\[TS\]) to first order w.r.t $b(0)$ and $b'(0)$. The $\lambda$-dependence of the transformations is reflected in the fact that for the TSP these are to be treated as independent Lagrange multipliers. The first shifts $A$ and $B$ together along their orbits; the second displaces one relative to the other. This twofold freedom gives rise to a universal characteristic structure of such variational problems. Provided the TSP has a unique solution,[^4] its limit as $\delta\lambda\longrightarrow 0$ will clearly define a metric on Q. For it can be used to define the ‘distance’ $I_{\textrm{\scriptsize{TS}}}$ (\[TS\]) between any two neighbouring points in Q$_{0}$. This process gets its name because the ‘distance’ is obtained when the configurations are *best matched*: brought as near to congruence, as measured by (\[TS\]), as possible.
To implement best matching in a convenient formal scheme, we introduce an *auxiliary gauge variable* $a\in\Gamma$, where $\Gamma$ for one-dimensional translations is $R^1$. However, for a general symmetry group defined by $d$ parameters, the auxiliary space $\Gamma$ will have $d$ dimensions ($d=7$ for the full Euclidean group). The space $\Gamma$ is used to extend the configuration space Q to a larger formal space Q$\times\Gamma$. [^5] Since Q contains $d$ unphysical degrees of freedom, $d$ is the dimension of the group orbits that foliate Q into Q$_{0}$ and is the *degeneracy*. Thus, the extension *doubles the degeneracy*. We shall denote the coordinates on $\Gamma$ collectively by $a$, which is a single variable in the present case.
We now use the auxiliary $a$ to define the *corrected coordinate*[^6] $$\bar{q}^{i}=q^{i}+ae^i.
\label{CoCo}$$
It is invariant under the pair of mutually compensating transformations: $$q^{i}(\lambda)\longrightarrow
q^{i}(\lambda)+b(\lambda)e^i,\hspace{.5cm}a(\lambda)\longrightarrow
a(\lambda)-b(\lambda),
\label{banal}$$ which, in view of the triviality of the invariance thereby achieved, we shall call the *banal transformation*. It induces equally trivial mutually compensating transformations of the velocities:
$$v^{i}\longrightarrow v^{i}+b'e^i,
\hspace{.5cm}a'\longrightarrow a'-b',\hspace{.5cm}a'={\textrm da\over\textrm d\lambda}.
\label{banalvelocities}$$
We now define the *corrected velocity* $\bar v^{i}$: $$\bar{v}^{i}=v^{i}+a'e^i.
\label{BM7}$$
Unlike the *naive (uncorrected) coordinate* $q^{i}$ and the *naive velocity* $v^{i}$, both $\bar{q}^{i}$ and $\bar{v}^{i}$ are invariant under the banal transformation (\[banal\]). One correction compensates the other. It is important to distinguish the transformation parameter $b$ from the auxiliary $a$. Although they seem to ‘do the same thing’ to the $q^i$’s \[cf. (\[BM3\]) and (\[CoCo\])\], their conceptual status is quite different.
Two rules now yield a fully invariant action of curves on Q$_{0}$: 1) On Q$\times\Gamma$, choose a Lagrangian $\bar{\cal L}=\bar{\cal
L}[\bar{q}^{i},\bar{v}^{i}]$ that is the square root of a functional homogeneously quadratic in the $\bar{v}^{i}$ and otherwise depends only on the $\bar q^i$; 2) ensure that the action along any curve in Q$\times\Gamma$ is invariant w.r.t *arbitrary* variations $\delta a$ of $a$. The ‘arbitrary’ means that there are *no fixed end points* of the $a$ variation. The point is that the $a$ variation is effectively being used to solve successive TSPs to determine the metric on Q$_{0}$. Were we to fix $a$ at the end points, we could not determine the metric at them. We shall say that $a$ is varied by the *free-end-point method*. The $q$ variation is different. It is being used to find the geodesics w.r.t the metric determined by the $a$ variation.
Free-end-point variation has consequences of three kinds: 1) Conditions on the $\bar{q}$ dependence of $\bar{\cal L}$; 2) constraints on the canonical momenta that the Euler–Lagrange equations propagate because $\bar{\cal L}$ satisfies conditions 1); 3) subsidiary equations that must be satisfied in addition to the constraints and the Euler–Lagrange equations. (This happens in GR and conformal gravity [@ABFO].)
The conditions on the $\bar q$ dependence of $\bar{\cal L}$ imposed under 1) can often cause the auxiliary $a$ to disappear from $\bar{\cal L}$ (its velocity $a'$ never does). This happens for translations, as we shall soon see, and rotations and in Abelian and non-Abelian gauge theories.[^7] However, $a$ appears explicitly in scale-invariant theories and also in a representation of GR [@ABFO].
The reason for this is a characteristic property of scaling transformations, which, in contrast to the additive nature of the translations (\[BM3\]), are multiplicative and change the norm. Indeed, the compensating transformations for dilatations are $$q^{i}\longrightarrow bq^{i},\hspace{.5cm}a\longrightarrow {a\over
b}.
\label{DilBanal}$$ The corrected coordinate and corrected velocity are $$\bar q^{i}=aq^{i},\hspace{.5cm}\bar v^{i}=av^{i}+q^{i}a'.
\label{DilCorr}$$
The device of replacing the ‘partly redundant’ $q^i$ by corrected $\bar q^i$ is evidently universally applicable, i.e., it is always possible to introduce a compensating transformation like the second members of (\[banal\]) and (\[DilBanal\]) (which are both ‘inverses’ – one by subtraction, the other by division). More complicated symmetries and several simultaneous symmetries are beyond the scope of this paper. However, translations and dilatations in particle mechanics are a useful and adequate preparation for the conformal symmetry in [@ABFO]. For simplicity, we take the Lagrangian $\bar{\cal L}[\bar q^i,\bar v^i]$ diagonal in the $\bar v^i$: $$I_{\scriptsize\textrm{BM}} =\int\textrm{d}\lambda\bar{\cal L}[\bar
q^i,\bar v^i]= \int{\textrm{d}}{\lambda}2\sqrt{B(\bar
q^{1},\dots,\bar q^{N})T},\hspace{.5cm} T = {\scriptsize{1\over
2}}{\sum_{i=1}^{N}}m_{i} (\bar
v^{i})^{2}.
\label{Action}$$
While we shall always insist on free-end-point variation w.r.t $a$, the remaining Euler–Lagrange part of the variation can be performed in two different ways. We can vary either w.r.t the $q^i$ or w.r.t the $\bar q^i$. The latter variation is neater, but we still use the former, since it shows how the absolute part of Newtonian kinematics, i.e., the part not dictated by three-dimensional Euclidean geometry, arises.
The physical variables $q^i$ and the auxiliary variable $a$ have the canonical momenta $$p^{i}\equiv {{\partial\bar{\cal L}}\over\partial
v^{i}}=\sqrt{B\over T}{\partial T\over\partial\bar v^i}{\partial\bar v^i\over\partial v^i},\hspace{.5cm}p^a\equiv {{\partial\bar{\cal L}}\over\partial
a'}=\sum_i\sqrt{B\over T}{\partial T\over\partial\bar v^i}{\partial\bar v^i\over\partial a'}.$$ Substituting the values of $\partial\bar v^i/\partial v^i$ and $\partial\bar v^i/\partial a'$ for translations and dilatations, we find $$p^i_{\scriptsize\textrm t}=\sqrt{B\over T}{\partial T\over\partial\bar v^i},\hspace{.5cm}p^a_{\scriptsize\textrm t}=\sum_i\sqrt{B\over T}{\partial T\over\partial\bar v^i};\hspace{.7cm}p^i_{\scriptsize\textrm d}=\sqrt{B\over T}{\partial T\over\partial\bar v^i}a,\hspace{.5cm}p^a_{\scriptsize\textrm d}=\sum_i\sqrt{B\over T}{\partial T\over\partial\bar v^i}q^i.
\label{p}$$
Therefore, in the respective cases we have the constraints $$\sum_i p^i_{\scriptsize\textrm t}\equiv
P_{\scriptsize\textrm t}=p^a_{\scriptsize\textrm t},\hspace{.5cm}\sum_i p^i_{\scriptsize\textrm d}q^i\equiv D=ap^a_{\scriptsize\textrm d},
\label{Niall}$$ which are primary [@Dirac] because they arise from the mere form of the action and not through variation (which gives rise to secondary constraints). The constraints (\[Niall\]) reflect the banal invariance of the action. They are homogeneously linear in the $p^{i}$ with coefficients solely determined by the symmetry and arise because $\bar{\cal L}$ depends on the corrected velocities $\bar v^i$. The actual form of $\bar{\cal L}$ is immaterial.
In accordance with the free-end-point method, the action must be stationary with respect to arbitrary variations of $a$. This means that at any instant both $\partial\bar{\cal L}/\partial a=0$ and $\partial\bar{\cal L}/\partial a'=0$. But $\partial\bar{\cal L}/\partial a'$ is the canonical momentum $p^a$ of the auxiliary variable, which must therefore vanish as a secondary constraint. The primary constraints (\[Niall\]) now become $$P_{\scriptsize\textrm t}=0,\hspace{.5cm} D=0,
\label{NiallBis}$$ i.e., the momentum and dilatational momentum must vanish, as promised at the end of the Introduction.
As Dirac points out [@Dirac], as soon as one has constraints in any dynamical system, one must verify that they are propagated by the Euler–Lagrange equations. An important feature of best matching is the manner in which its first condition $\partial\bar{\cal L}/\partial a=0$ automatically ensures propagation of the constraint enforced by its second condition $\partial\bar{\cal L}/\partial a'=0$. We can see this in two ways.
First, since the action must be stationary for all variations of $a$, they will include variations with fixed end points, from which we deduce the standard Euler–Lagrange equation $${\textrm dp^a\over\textrm d\lambda}={\partial\bar{\cal L}\over\partial a}.$$
Since we also require ${\partial\bar{\cal L}/\partial a}=0$, this ensures that $p^a$ will be conserved. Thus, if $p^a$ is initially zero, it will remain zero. By the primary constraints (\[Niall\]), the relations that hold for $p^a$ must also hold for the corresponding combinations $P_{\scriptsize\textrm t}$ and $D$ of the physical canonical momenta, so the constraints (\[NiallBis\]) must propagate.
Second, we can, of course, confirm the propagation of (\[NiallBis\]) directly, using the Euler–Lagrange equations of the physical variables in conjunction with the condition ${\partial\bar{\cal L}/\partial a}=0$. It is worth doing this to see the explicit form of this condition. In the case of translations $${\partial\bar{\cal L}\over\partial a}=\sum_i{\partial\bar{\cal L}\over\partial\bar q^i}{\partial\bar q^i\over\partial a}=\sum_i\sqrt{B\over T}{\partial B\over\partial\bar q^i}=0.
\label{q}$$ From the Euler–Lagrange equations $${\textrm dp^i\over\textrm d\lambda}=\sqrt{B\over T}{\partial B\over\partial\bar q^i},$$ we directly conclude that $${\textrm dP_{\scriptsize\textrm t}\over\textrm d\lambda}=\sum_i{\textrm dp^i\over\textrm d\lambda}=\sum_i\sqrt{B\over T}{\partial B\over\partial\bar q^i},$$ which does indeed vanish by virtue of (\[q\]). Of course, (\[q\]) is the condition for the potential $B$ to be translationally invariant (which ensures fulfilment of Newton’s third law). One of the easiest ways to achieve this is to choose $$B=B(|\bar q^i-\bar
q^j|)=B(|q^i-q^j|),$$ the second equation holding because the correcting $ae^i$ in the $\bar q^i$ simply drops out of the difference. Thus, for corrected coordinates formed additively the potential will not depend on the auxiliary gauge variable $a$. Also any constant $E$ can be added to the potential $B$, matching the total energy that appears in the Jacobi action (\[Jacobi\]).
In the case of dilatations, we have a much more intimate interconnection of conditions. The condition $\partial\bar{\cal L}/\partial a=0$ gives $$0={\partial\bar{\cal L}\over\partial a}\equiv\sum_i\left(\sqrt{T\over B}{\partial B\over\partial\bar q^i}{\partial\bar q^i\over\partial a}+\sqrt{B\over T}{\partial T\over\partial\bar v^i}{\partial\bar v^i\over\partial a}\right)=\sum_i\left(\sqrt{T\over B}{\partial B\over\partial\bar q^i}q^i+\sqrt{B\over T}{\partial T\over\partial\bar v^i}v^i\right).$$ We now divide by $a$ and add and subtract $Da'/a$, where $D$ is the dilatational constraint (dilatational momentum): $$0={1\over a}\sum_i\left(\sqrt{T\over B}{\partial B\over\partial\bar q^i}aq^i+\sqrt{B\over T}{\partial T\over\partial\bar v^i}(av^i+q^ia')\right)-{1\over a}\sum_i\sqrt{B\over T}{\partial T\over\partial\bar v^i}q^ia'.$$ Since $av^i+q^ia'=\bar v^i$, $(\partial T/\partial\bar v^i)\bar v^i=2T$ and the constraint is required to vanish, we conclude $$\sum_i \left(\sqrt{T\over B}{\partial B\over\partial
\bar q^i}\bar q^i+\sqrt{B\over T}2T\right)=0
\Rightarrow\sum_i \left({\partial B\over\partial
\bar q^i}\bar q^i+2B\right)=0.
\label{HomCond}$$ By Euler’s theorem, this tells us that the scheme will be consistent if $B$ is homogeneous of degree -2 in the $\bar q^i$’s. The last step in establishing the mutual consistency of our interlocking system of equations is to note that the propagation of the dilatational constraint by the Euler–Lagrange equations leads to exactly the same condition (\[HomCond\]).
To make the final connection with the special Newtonian situation summarized at the end of section 1 and derived from the Lagrange–Jacobi relation in section 2, we now exploit the freedom to make a banal transformation and so pass from a general frame of reference to the unique *distinguished representation* in which $a=1$ and $\lambda$ is chosen to make $B/T=1$. In the distinguished representation, we recover Newton’s laws exactly subject to the homogeneity condition on the potential and the dilatational constraint that the dilatational momentum vanishes. The $q^i$ are then coordinates in an inertial frame of reference, and the special $\lambda$ is indistinguishable from Newton’s absolute time. We have derived the non-geometrical (absolute) part of Newtonian kinematics from a purely relational scheme.
As mentioned in the introduction, a similar treatment for the rotation group [@BB; @LR] yields the constraint that the total c.m angular momentum vanishes (and a further restriction on the potential, which must be a function of the inter-particle separations). Note how well the relational origin of this dynamics is hidden. The constraints and the condition $E=0$ apply to the complete ‘island universe’ of $N$ particles in Euclidean space, and need not apply to subsystems like the solar system – only the sums over all subsystems need satisfy the constraints. Since it is difficult to verify the constraints and the condition $E=0$ for the complete universe, the sole hard evidence for the relational origin of the dynamics is in the form of the observed potentials. The conditions imposed by the translations and rotations are respected by the observed Newtonian approximations to relativistic physics, but there seems to be a spectacular failure to match the condition needed for a fully scale-invariant dynamics. The next section will show how it may be only an apparent failure. However, let me first conclude this section with two comments.
The first concerns the somewhat elaborate formalism of the free-end-point method. Why is it necessary? The velocity of our auxiliary variable plays exactly the same role as the scalar potential $\phi$ in the Hamiltonian formulation of Maxwell theory. Since $\phi$ has the dimensions of velocity and the action contains no variable of which it is the velocity, it should strictly be treated as an ignorable coordinate [@Lanczos]. It is, however, habitually treated as a multiplier, variation with respect to which yields the Gauss constraint, which is a stronger result than would follow from regarding $\phi$ as an ignorable coordinate. Our free-end-point method for translations shows why the multiplier treatment is correct. The Maxwell action contains the velocity of an auxiliary variable but not the variable itself. It would be possible to start with a more general action assumed to depend on both an auxiliary and its velocity. The variation with respect to the auxiliary would then lead to a condition on the Lagrangian that is, in fact, satisfied by the Maxwell Lagrangian, which, as in the case of particle translations, does not depend on the auxiliary after all. As already noted, this is the case in all gauge theories hitherto studied. However, as this and the companion paper [@ABFO] show, scale-invariant theories necessarily contain both an auxiliary variable and its velocity. The free-end-point method shows that correct results are obtained by treating both as independent multipliers even though one is the velocity of the other.
The second comment is closely related to the first and relates to Dirac’s classic study of generalized Hamiltonian dynamics [@Dirac]. He presents the consistent propagation of constraints derived from singular Lagrangians as more or less a matter of luck, which it certainly is in some cases [@BOF]. However, as this paper and [@ABFO] show, best matching provides a transparent scheme in which constraints linear in the canonical momenta are automatically propagated because they reflect the existence of a geometrically defined metric on a quotient configuration space. The case of dilatational constraints provides a new and interesting example of consistent constraint propagation. Its generalization to geometrodynamics [@ABFO], which closely follows the treatment given here, leads to decidedly nontrivial relations.
Scale-Invariant Dynamics with Specific Potentials
=================================================
We recall that the Jacobi action for a general Newtonian system is $$I_\textrm{\scriptsize{Jacobi}} = 2\int\sqrt{E - U}\sqrt{\tilde T}
\textrm{d}\lambda.$$
Best matching with respect to translations and rotations restricts the Newtonian possibilities to a relatively limited extent. The total angular momentum must be zero and the potential $U$ must be a function of the inter-particle separations. Best matching with respect to dilatations has more far reaching consequences: the energy $E$ must necessarily be exactly zero and the potential must be homogeneous of degree -2.
We first show how Newtonian gravity can be recovered to a good accuracy from a potential that is homogeneous of degree -2. There are two easy ways to do this. Let $$W=\sum_{i<j}{m_{i}m_{j}\over r_{ij}}.$$ Thus, $W$ is minus the Newtonian gravitational potential with Newton’s constant G set equal to unity. The first ‘way to Newton’ is to take as potential $$U=-{W^{2}\over 2}.
\label{U1}$$
In the distinguished frame with the special $\lambda=t$, the equations of motion are $${\textrm{d}\textbf{p}^{i}\over\textrm{d}t}=
-W{\partial W\over\partial\textbf{x}^{i}}.
\label{FirstEoM}$$
This is Newton’s law with G replaced by $W$. If the system has virialized and resembles a globular cluster, $W$ will remain effectively constant, and the motions will be essentially Newtonian. Astronomers will observe $1/r$ forces, but the hidden dynamics will be scale invariant.
An island universe described by (\[FirstEoM\]) can even mimic a Hubble-type expansion even though expansion has no physical meaning in scale-invariant dynamics. To see this, note that two processes can change $W$. The first is general expansion or contraction. Now this would change the moment of inertia (\[MofI\]), $$I={1\over M}\sum_{i<j}m_{i}m_{j}r_{ij}^{2},
\hspace{.5cm}M=\sum_{i}m_{i},$$ but the vanishing of the dilatational momentum $D$ enforces constancy of $I$. However, a changing shape of the matter distribution changes the *scale-invariant* Newtonian potential[^8] $$U_0=-M\sqrt{I}W.
\label{ScalInvPot}$$ The constancy of $I$ does not prevent the matter from ‘clumping’, which will increase both $U_0$ and $-W$, and with them the ‘gravitational constant’. Suppose a planet orbiting a sun in an island universe that is becoming clumpier. In the distinguished frame, the gravitational forces will become stronger, and the planet–sun separation will adiabatically decrease. However, if we insist that, by definition, gravity is constant, this effect can be offset by an adiabatic increase of all scales from the distinguished frame. This will mimic a Hubble-type expansion.
In fact, although this is suggestive I doubt whether the Hubble red shift can be explained this way. The analogue of squaring $W$ in the conformal setting leads to a very complicated theory that may not even be consistent [@ABFO]. However, there is the much more interesting possibility that uses the unique conserved quantity in a scale-invariant universe, namely $I$, or rather $\sqrt{MI}=\mu$. This has a simple conformal analogue, which is the volume of 3-space.
Therefore, to achieve scale invariance, and the strong equivalence principle, we shall use $$\mu=\sqrt{\sum_{i<j}m_{i}m_{j}r_{ij}^{2}}.
\label{Rho}$$
Just as one passes from special to general relativity (with gravity minimally coupled to matter) by replacing ordinary derivatives in the matter Lagrangians by covariant derivatives, Newtonian potentials can be converted into potentials that respect scale invariance. One simply multiplies by an appropriate power of $\mu$, which has the dimensions of length. This is a rather obvious mechanism. What is perhaps unexpected is that the modified potentials lead to forces $\textit{identical}$ to the originals accompanied by a universal cosmological force with minute local effects. The scale invariance is hidden because $\mu$ is conserved.
Let $U$ consist of a sum of potentials $U_{k}$ each homogeneous of degree $k$: $$U=\sum_{k=-\infty}^{\infty}a_{k}U_{k}.
\label{NewtPot}$$
The $a_{k}$ are freely disposable strength constants. The energy $E$ in the Jacobi action (\[Jacobi\]) will be treated as a constant potential ($k=0$). (It plays a role like the cosmological constant $\Lambda$ in GR).
Now replace (\[NewtPot\]) by $$\tilde{U}=\sum_{k=-\infty}^{\infty}b_{k}U_{k}\mu^{-(2+k)}.
\label{ScaledPot}$$
The equations of motion for (\[NewtPot\]) are $${\textrm{d}\textbf{p}^{i}\over\textrm{d}t}=
-\sum_{k=-\infty}^{\infty}a_{k}{\partial
U_{k}\over\partial\textbf{x}^{i}};$$ for (\[ScaledPot\]) they are $${\textrm{d}\textbf{p}^{i}\over\textrm{d}t}=
-\sum_{k=-\infty}^{\infty}b_{k}\mu^{-(2+k)}
{\partial U_{k}\over\partial\textbf{x}^{i}}+
\sum_{k=-\infty}^{\infty}(2+k)b_{k}\mu^{-(2+k)}U_{k}
{1\over\mu}{\partial\mu\over\partial\textbf{x}^{i}}.$$
Since $\mu$ is constant ‘on shell’, we can define new strength constants that are truly constant: $$b_{k}=a_{k}\mu^{2+k}.
\label{DefB}$$ The equations for the modified potential become $${\textrm{d}\textbf{p}^{i}\over\textrm{d}t}=
-\sum_{k=-\infty}^{\infty}a_{k}
{\partial U_{k}\over\partial\textbf{x}^{i}}+
\sum_{k=-\infty}^{\infty}(2+k)a_{k}U_{k}
{1\over\mu}{\partial\mu\over\partial\textbf{x}^{i}}.
\label{ModEq}$$ If we define $$C(t)={\sum_{k=-\infty}^{\infty}(2+k)a_{k}U_{k}\over
2\sum_{i<j}m_{i}m_{j}r_{ij}^{2}}
\label{DefC}$$ and express $\mu$ in terms of $r_{ij}$, then equations (\[ModEq\]) become $${\textrm{d}\textbf{p}^{i}\over\textrm{d}t}=
-\sum_{k=-\infty}^{\infty}a_{k}
{\partial U_{k}\over\partial\textbf{x}^{i}}+
C(t)\sum_{j}m_{i}m_{j}{\partial{r_{ij}^{2}}\over\partial\textbf{x}^{i}}.
\label{AbbModEq}$$
We recover the original forces exactly together with a universal force. It has an epoch-dependent strength constant $C(t)$ and gives rise to forces between all pairs of particles that, like gravitational forces, are proportional to the inertial mass but increase in strength linearly with the distance.[^9] The universal force will be attractive or repulsive depending on the sign of $C(t)$, which is an explicit function of the $r_{ij}$’s. For small enough $r_{ij}$, they will be negligible compared with Newtonian gravity. An estimate will be made in section 7, but a first conclusion is this. A Newtonian-like universe with forces that appear to violate scaling may nevertheless be scale invariant with an as yet unrecognized cosmological force. Cosmologists might be greatly deceived, but, having once got the idea, they could predict the epoch-dependent $C(t)$ from the observed matter distribution and the locally observed forces.
Moreover, their revised cosmological model will be stable: its ‘size’, measured by $\mu$, remains constant – the automatic adjustment of $C(t)$ ensures it. Scale-invariant cosmology is free of the inescapable instability of Newtonian and Einsteinian cosmologies. What is more, $C(t)$ is unambiguously fixed by a fundamental symmetry, unlike Einstein’s ad hoc $\Lambda$. It seems paradoxical that a theory constructed to eliminate size as a dynamical variable uses ‘size’ ($\mu$) to generate a force that conserves the size (in the distinguished representation). Of course, the numerical value of the size is purely nominal. Moreover, the paradox is an artefact of the formalism. Could we but use shape coordinates conveniently, cosmology would be reduced to angles and mass ratios.
Note also that the ‘size of the universe’ is not kept constant because a repulsive cosmological force balances attractive gravity, as in Einstein’s original cosmological model of 1917 [@Einstein17]. That achieved only unstable equilibrium. Scale-invariant cosmology is stable whatever the sign of $C(t)$. In Newtonian terms, the stability arises from the interplay of the potential’s homogeneity and the simultaneous vanishing of the energy $E$ and the dilatational momentum $D$ (section 2). If one retains the same forces but allows the minutest deviation from zero of $E$ or $D$, the stability is lost, parabolically for $E\neq 0$ and linearly for $D\neq 0$. This again shows how cosmologists could be tricked. If they did not realize that a symmetry enforces stability, a scale-invariant cosmology would appear to be balanced on a knife edge. This is reminiscent of the ‘flatness knife edge’ of modern relativistic cosmology.
Before we obtain estimates, two potential flaws in the device of achieving homogeneity with powers of $\mu$ should be mentioned. The first is due to the possibility of collisions, which lead to infinities in the case of point particles. In Newtonian $N$-body theory, the most intractable collisions are the so-called central collisions in which all $N$ particles collide at once [@Chenciner]. They cannot be regularized and are rather like the Big-Bang or Big-Crunch singularities in GR. Both in Newtonian theory and in GR, the cosmological singularities clearly stem from the dynamical role of scale. Such singularities are not present in scale-invariant cosmology [@ABFO]. However, lesser collisions of $n<N$ particles are possible (and some form of gravitational collapse may also occur in conformal gravity). In the particle model, this will certainly cause $C(t)$ to become infinite even at a two-particle collision, and the equations will be ill defined at such collisions. There is a similar potential problem in conformal gravity, and further research will be needed to establish whether it is serious. The second possible difficulty is related to dimensional analysis, to which we now come.
Dimensional Analysis
====================
In Newtonian dynamics – and modern physics – there are three fundamental physical dimensions: mass $[m]$, length $[l]$, and time $[t]$. As long as no cosmological assumptions are made, these are necessarily independent and absolute, in the sense that one believes that ‘they exist out there in the world’. One is forced to choose local units to measure them. If the universe is assumed to be self contained, the situation is different, because then the local unit is merely a fraction of some global total. More significantly, time can be eliminated as a dimension. One just has mass $[m]$ and length $[l]$. In scale-invariant geometrodynamics [@ABFO], dimensional analysis is even simpler and is based on length alone. The results obtained in this section are only the first step to this ideal. They are nevertheless suggestive. The method used here develops further the analysis of [@BB1].
In developing dimensional analysis, we consider two standpoints. The first is that of an imagined external observer who possesses a god’s eye view of the universe. The second is that of an internal observer forced to use locally chosen units of mass, length, and time. The internal observer has no knowledge of the universe as a whole. For both observers, length $[l]$ is the most important dimension. Indeed, time is measured by length – the distance traversed by the hand of a clock – and masses are deduced from accelerations, which involve lengths and times. Let us therefore assume both kinds of observers are equipped with rods, on which they mark units of length. These rods can be used to measure the inter-particle separations $r_{ij}$. A measure of time $[t]$ is chosen by postulating that the motion of a particular body (the clock) measures time: when the clock traverses the already chosen unit of distance, one unit of time elapses. Finally, as Mach showed [@Mach], relative masses can be deduced from the reciprocal ratio of the accelerations that bodies impart to each other when interacting in accordance with Newton’s third law. If some given mass is chosen as the unit, other masses can then be determined in this way.
An extension of the same technique permits determination of charges. Since both masses and charges are deduced from accelerations, it seems to me that they should all have the dimension mass, denoted $[m]$. We shall see that, at least for gravitational and electrostatic forces, this leads to a consistent scheme in which the action principle of the universe contains no dimensional coupling constants but only passive mass (Newton’s original inertial mass) and active charges: gravitational (which happen to be proportional to the passive mass), electrostatic and perhaps more. In standard dimensional analysis, the electric charge does not have the dimensions $[m]$ of mass, but it will in the proposed scheme. It then turns out (empirically) that the electric charge of the proton is $10^{19}$ times its gravitational charge (active gravitational mass), and both have dimension $[m]$. As we shall see at the end of this section, this assumption need not lead to conflict with the fact that the fine-structure constant $\alpha=e^2/hc$ is dimensionless because Planck’s constant $h$ and the speed of light $c$ will have different dimensions.
Let us start by considering standard dimensional analysis in Newtonian theory for the case of gravitational and electrostatic interactions (of unit charges $|e_i|=e$). The Lagrangian is $$\sum_i m_i{\textrm{d}\textbf{x}_i\over\textrm{d}t}\cdot{\textrm{d}\textbf{x}_i\over\textrm{d}t}+
\textrm{G}\sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-\sum_{i<j}{e_ie_j\over r_{ij}},\hspace{.5cm}|\bar e_i|=\bar e,$$ where bars have been added to the masses in the gravitational potential to indicate that these are charges. Since the dimensions of the kinetic energy are taken as fiducial and are $ml^2t^{-2}$, it is clear that Newton’s constant G must have dimensions $m^{-1}l^3t^{-2}$ if the gravitational charges are taken to have the dimensions of mass. It is equally clear that $e^2$ must have dimension $ml^3t^{-2}$. This can be achieved by setting $$\sum_{i<j}{e_ie_j\over r_{ij}}=\textrm{G}\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}},$$ where $\bar e/\bar m$ is the ratio of the electrostatic force generated by unit electric charge to the gravitational force generated by the unit mass $\bar m$. For definiteness, I shall assume that there are $N=2n$ particles of constant mass, $n$ of each carrying positive and negative unit electric charge. The total mass $M$ of the universe is equal to the sum of the inertial masses and equal to the sum of the gravitational and electrostatic charges, the contribution of the latter being in fact zero by the assumption of overall electric charge neutrality.
Let us now consider Jacobi’s principle for Newtonian theory under the assumption that some hypothetical principle forces the energy $E$ to be exactly zero. The action can then be written in the form $$A=2\sqrt \textrm G\int\sqrt{\sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}}}
\sqrt{\sum_i {m_{i} \over
2}\textrm{d}{\textbf{x}}_{i}\cdot
{\textrm{d}{\textbf{x}}_{i}}},$$ which has the normal dimensions $ml^2t^{-1}$ of action: $[\sqrt \textrm G]=m^{-1/2}l^{3/2}t^{-1}$ while \[integrand\]=$m^{3/2}l^{1/2}$. However, the constant $\sqrt G$ in front of the action $A$, being a common factor, can have no effect on the observed motions and can be simply omitted. Then the action $\bar A=A$/G in Jacobi’s principle has dimensions $m^{3/2}l^{1/2}$. If a non-vanishing energy $E$ is included, it will have dimensions $m^2l^{-1}$, as is evident from (\[Jacobi\*\]). To recover the normal dimensions of energy, we must take the Newtonian time to have dimensions $m^{-1/2}l^{3/2}$. This is confirmed by the explicit expression (\[clem\]) for the Newtonian time if, as above, we take the potential energy without G and therefore to have dimensions $m^2l^{-1}$.
Let us next consider the canonical momenta (\[CanMom\]). They are homogeneous of degree zero in the velocities, and therefore any attempt to introduce a dimension of time into them will achieve nothing. They have the dimensions $m^{3/2}l^{-1/2}$. It is true that if we multiply the dimensions by $m^{-1/2}l^{3/2}/m^{-1/2}l^{3/2}=m^{-1/2}l^{3/2}/t$, we recover the Newtonian $mlt^{-1}$. However, this distorts the true nature of Jacobi geodesics, which involve only distances and masses.
We must now show how Newton’s constant G is recovered as an emergent constant from the action $\bar A$. The Euler– Lagrange equations that follow from this action are $${\textrm{d}\over\textrm{d}\lambda}\left(\sqrt{\sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-
\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}}\over \sum_i {m_{i} \over
2}{\textrm{d}{\textbf{x}}_{i}\over\textrm{d}\lambda}\cdot
{\textrm{d}{\textbf{x}}_{i}\over\textrm{d}\lambda}}m_i{\textrm{d}\textbf{x}_i\over\textrm{d}\lambda}\right)
=-\sqrt{\sum_i {m_i \over
2}{\textrm d\textbf{x}_i\over\textrm d\lambda}\cdot
{\textrm d\textbf{x}_i\over\textrm d\lambda}\over \sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-
\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}}}{\partial\over\partial\textbf{x}_i}
\left(\sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-
\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}}\right)
\label{k}$$
In the discussion of Jacobi’s principle, we simplified this equation by choosing the unique $\lambda$ that makes the square root on the right-hand side unity at all $\lambda$. In this way, we recovered Newtonian time. If this method is applied to (\[k\]), obtained from the action without G, the outcome can be interpreted in Newtonian terms as due to a choice of the time unit that makes G=1. Clearly, if we make the square root equal to a constant $k$, we simply recover a different time that runs uniformly w.r.t the first time but at a different rate. With this more general simplifying choice, (\[k\]) becomes $$m_i{\textrm{d}^2\textbf{x}_i\over\textrm{d}t^2}=-k^2{\partial\over\partial\textbf{x}_i}
\left(\sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-
\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}}\right)
\label{GravEq}$$ with ‘gravitational constant’ $$k^2={\sum_i {m_i \over
2}{\textrm d\textbf{x}_i\over\textrm d t}\cdot
{\textrm d\textbf{x}_i\over\textrm d t}\over \sum_{i<j}{\bar m_i\bar m_j\over r_{ij}}-
\sum_{i<j}{\bar e_i\bar e_j\over r_{ij}}},
\label{GravCon}$$ which has the correct dimensions $m^{-1}l^3t^{-2}$. Let us now recall our internal observers within such a universe, who have chosen arbitrarily units of mass, length, and time and use standard Newtonian dimensional analysis. Making purely local observations within their solar system, they will find the the gravitational constant has some value G. Making cosmological observations, they will discover that G has the value $k^2$ (\[GravCon\]). Since they include G (=$k^2$) as a factor in the potential energy, they will interpret (\[GravCon\]) as saying that minus the Newtonian potential energy is equal to the kinetic energy. Thus, they will find that the total energy of the universe is exactly zero and will interpret this as a ‘cosmic coincidence’ since they can find no theoretical argument that enforces the condition.
From the point of view of a dynamics of pure shape, Newtonian theory in Jacobi form is clearly imperfect, this being reflected in the fact that the action depends on the length scale. Also, no fundamental principle justifies the assumption that $E=0$. However, the dependence on the mass scale is not serious, because all masses are constant and can be expressed as dimensionless ratios of the total mass $M$.
Let us now repeat the dimensional analysis and ‘calculation of G’ for scale-invariant gravity (SIG) and electrostatics with the assumption that the necessary homogeneity of degree -2 of the potential is achieved as above by means of $\mu$. This is a much more interesting theory because both the energy and the dilatational momentum must be exactly zero.
As is evident from (\[GravEq\]), the basic form of the equation of motion for Jacobi-type actions (with $E=0$ either by stipulation or enforced by scale invariance) is $$m_i{\textrm{d}^2\textbf{x}_i\over\textrm{d}t^2}=-{T\over U}{\partial U\over\partial\textbf{x}_i}$$ with $k^2=T/U$ identified as Newton’s constant. To pass from Newtonian theory to scale-invariant gravity, we simply replace the Newtonian potential $U$ by $U/\mu$. Then the equation of motion becomes $$m_i{\textrm{d}^2\textbf{x}_i\over\textrm{d}t^2}=-{T\mu\over U}{\partial \over\partial\textbf{x}_i}
\left({U\over\mu}\right)=-{T\over U}{\partial U\over\partial\textbf{x}_i}+
{T\over \mu}{\partial \mu\over\partial\textbf{x}_i}.
\label{ScaleGravEq}$$
We have the identical expression $T/U$ for G, but there is now a more subtle reason for its constancy. First, $T\mu/U$ is made constant by choosing one of the distinguished times. This is the device that made $T/U$=G constant in the Newtonian case. Second, $\mu$ is constant in scale-invariant gravity as a consequence of the dynamics. Thus, in (\[ScaleGravEq\]) the coefficient of the gravitational force is constant because of these two reasons together. The coefficient of the cosmological force is epoch dependent. Both coupling constants are emergent and uniquely determined. We shall estimate them in the next section. The strength of electrostatics relative to gravity is constant and, in the absence of unification of the forces of nature, must be determined empirically.
Two possible problems in scale-invariant gravity were noted at the end of the previous section. The one to do with dimensional analysis concerns the extension of the method presented in this section to Newtonian forces with other than $1/r$ distance dependence. Presumably, one will always want the dimensions of all terms in the potential to be the same. For gravity and electrostatics, the dimensions are $ml^{-2}$, but it seems to me that it will be difficult to find the right combination of charges to give these dimensions for other than $1/r$ Newtonian forces without introducing undetermined constants with dimensions of mass. This may not be a problem or, if it is, it might even be a virtue through ruling out such forces, which are not observed. However, I think the particle model is probably not a reliable guide for such arguments. At the least, there is no problem with gravity and electrostatics.
Continuing with the dimensional analysis, calculation of the various dimensions of the most important dynamical quantities for scale-invariant gravity (SIG) together with the dimensions for standard Newtonian theory (SNT) and ‘Jacobianized’ Newtonian theory (JNT) as calculated earlier gives the following results: $$\hspace{1cm}Action\hspace{.5cm}Length\hspace{.5cm}Momenta\hspace{.5cm}Time\hspace{.5cm}Energy$$ $$\textrm{SNT}\hspace{.3cm}ml^2l^{-1}\hspace{.7cm}l\hspace{1.4cm}mlt^{-1}
\hspace{1.1cm}t\hspace{1.1cm}ml^2t^{-2}$$ $$\textrm{JNT}\hspace{.4cm}m^{3/2}l^{1/2}\hspace{.8cm}l\hspace{1cm}m^{3/2}l^{-1/2}
\hspace{.5cm}m^{-1/2}l^{3/2}\hspace{.5cm}m^2l^{-1}$$ $$\textrm{SIG}\hspace{1cm}m\hspace{1.2cm}l\hspace{1.5cm}ml^{-1}
\hspace{1.2cm}l^2\hspace{1.2cm}ml^{-2}$$
The results for the scale-invariant case are striking and suggestive. First, since mass is a ratio, the action is effectively dimensionless. This suggests that for a scale-invariant theory the quantum of action will be a pure number. A similar calculation for scale-invariant geometrodynamics also suggests that action is a pure number. Also remarkable are the dimensions of the canonical momenta, which exactly mirror the de Broglie relation for the momentum in quantum mechanics. It is worth emphasizing that these results are non-trivial and arise from a ‘trinity’ of requirements: no time, scale invariance, and the implementation of the strong equivalence principle by the use of $\mu$.
The at first enigmatic result that in the scale-invariant case ‘time is an area’ is very natural. One of the length dimensions arises from the fact that the hand of clock must traverse a length to measure time. This length is supplied by the particle displacements in the numerator of (\[clem\]). The other length comes from the weighting by the denominator of (\[clem\]) and the fact that scale invariance and the SEP require the potential to have dimensions $ml^{-2}$.
Let us now consider what happens to the fine-structure constant by applying the rule for passing from the Newtonian dimensions $m, l, t$ to their scale-invariant timeless counterparts, which is simply to replace $t$ by $l^2$ and remember that the Newtonian potentials have been divided by $\mu$. Thus, the Newtonian dimensions of velocity are $[v]=lt^{-1}$, but now $[v]=l^{-1}$. Similarly, the Newtonian dimensions of Planck’s constant are $[h]=ml^2t^{-1}$, but now they become $[h]=m$, as we have already found. Thus, the $hc$ that appears as the denominator of the fine-structure constant no longer has its Newtonian dimensions $ml^3t^{-2}$ but $ml^{-1}$. Let us now consider the electric charge $e$. In Newtonian mechanics, it appears in the potential $\psi=\sum_{i<j}e_ie_j/r_{ij}$, but in the new scheme this is replaced by $\psi\mu^{-1}$. In Newtonian mechanics, $e^2/hc$ is dimensionless, so $[e^2]=ml^3t^{-2}$. However, if our approach is on the right track, the quantity that hitherto has been taken to be $e^2$ is actually $e^2\mu^{-1}$, and $[\mu]=ml$. Therefore, if we postulate $[e]=m$, the new dimensions of $e^2$ are $ml^{-1}$, which is what we just found for $[hc]$, so the ‘new fine-structure constant’ is still dimensionless. Even though gravitational charge does not appear to be quantized in the way electric charge is, one can still define a ‘gravitational fine-structure constant’, which will also be dimensionless.
Only if and when a scale-invariant quantum cosmology has been created will it be possible to say whether the relations found in this section have a deep meaning or are merely fortuitously suggestive. However, there must surely be an *a priori* presumption that scale invariance will overturn current ideas about the fundamental constants of nature. When Planck [@Planck] deduced his famous units of length, time, and mass, he employed $h, c, G$. He noted that the units will have fundamental significance so long as the theories from which they are deduced are correct. The analysis of this section suggests that, if the universe is self contained and scale invariant, the quantum of action is a pure number. If so, it seems inevitable that a reorientation of our ideas about dimensional analysis and the constants of nature will be necessary.
Estimates
=========
Let us make estimates for the model with gravity and electrostatics. Suppose an island universe of $N$ point particles half of which carry positive unit charges $e_{i}$ and the other half negative charges $-e_{i}=-1$ ($N$ is an even integer). With the notation of the previous section, take the Jacobi-type action to be $$I_{\scriptsize\textrm{Model}}=\int\textrm{d}\lambda\sqrt{{E\over\mu^{2}}+
{1\over\mu}\sum_{i<j}{\bar m_{i}\bar m_{j}\over r_{ij}}-
{1\over\mu}\sum_{i<j}{\bar e_{i}\bar e_{j}\over
r_{ij}}}\sqrt T =\int \textrm{d}\lambda\sqrt P\sqrt T,
\label{ModelAction}$$ where $T$ is best matched.
Although no principle forbids the ‘energy’ term $E/\mu^{2}$ in the potential, it will merely lead to a further cosmological force that adds no interest and introduces the undetermined constant $E$ (with dimension $[m]$). Let us therefore omit it, and hope that eventually some principle will rule it out – and determine the electrostatics-to-gravity strength ratio $\bar e/\bar m$, which for the moment must be fixed empirically. Let us make estimates using $\bar e^2/\bar m^2\approx 10^{40}$, reflecting the relative strengths of the electrostatic and gravitational interactions for the common elementary particles. Let us also take $N=10^{80}$ particles, since this is the estimated number of nucleons in the observed universe within the Hubble radius H. For simplicity, let us take all masses $\bar m_i=\bar m$ with $\bar mN=M$.
As cosmological model, we first assume a spherically symmetric and more or less uniform ‘plasma’ matter distribution like a globular cluster with mean radius $R$. In the estimation of (minus) the potentials $$\mu,\hspace{.5cm}\psi=\sum_{i<j}{\bar m_i\bar m_j\over r_{ij}},
\hspace{.5cm}\phi=-\sum_{i<j}{\bar e_{i}\bar e_{j}\over r_{ij}},$$ the fact that all masses are positive while half the charges are positive and half negative leads to big differences. Both $\mu$ and $\psi$ contain $\approx N^{2}$ terms formed by all pairs of particles, for which the mean separation will be $R$. Each mass ‘feels’ all the other $N-1$ masses, and the mean contribution of each pair will be $\bar m^2/R$. In contrast, each charge $\bar e_{i}$ finds itself in an almost exactly neutral cloud. In fact, the total charge outside any considered charge $\bar e_i$ will be $-\bar e_{i}$, and the ‘mean distance’ of the opposite charge will be $R$. Therefore, to a good accuracy the electrostatic potential $\phi$ consists of $N$ terms, each of magnitude $\approx \bar e^2/R$.[^10] (The minus has disappeared because the charges are opposite.)
Therefore, in the ‘plasma’ state $$\mu\approx N\bar mR,\hspace{.5cm}\psi\approx{N^2\bar m^2\over R},
\hspace{.5cm}\phi\approx{N\bar e^2\over R},
\label{Estimates}$$ and the corresponding contributions to $P$ in (\[ModelAction\]) are $$P_{\scriptsize\textrm{plasma}}={\psi\over\mu}+
{\phi\over\mu}\approx
{N\bar m\over R^{2}}+{10^{40}\over R^{2}},\hspace{.5cm}\bar m=1,
\label{PlasmaEst}$$ so that for $N\sim 10^{80}$ the (positive) electrostatic contribution to $P$ will be only $10^{-40}$ of gravity’s.
Now suppose the ‘plasma’ evolves into a ‘post-nucleosynthesis’ state in which all the charges are in $N/2$ neutral pairs with inter-charge separation $\bar r$. This pairing will change $\phi$ to $\phi\approx -N\bar e^2/\bar r$, have negligible effect on $\mu$ and will only change $\psi$ significantly if $\bar r\leq 10^{-80}R$. The possibility of ‘gravitational collapse’ to such an extremely small relative size, which corresponds to $\sim
10^{-20}l_{\scriptsize\textrm{Planck}}$ for $R\sim
10^{28}\textrm{cm}$, will not be considered here. Instead, let us consider the more immediately realistic case $\bar r/R\sim
10^{-40}$, so that $\bar r$ is of nuclear order $10^{-12}\textrm{cm}$. Then, in contrast to the ‘plasma’ estimate (\[PlasmaEst\]), we obtain the ‘post-nucleosynthesis’ estimate $$P_{\scriptsize\textrm{pns}}={\psi\over\mu}-
{\phi\over\mu}\approx
{N\bar m\over R^{2}}+{10^{40}\over R\bar r},
\label{PNS}$$ so now the two contributions will have equal orders of magnitude.
Making the assumption that the particle model can give valid order-of-magnitude estimates, let us now try to estimate the size and mass of the actual universe. In accordance with the above estimates, let us assume that the combined potential $U$ of the gravitational and electrostatic potentials is $\approx M^2/R$. Then $$G={T\over U}={R\sum m_i\sum_i{\textrm d\textbf{x}_i\over\textrm d t}\cdot{\textrm d\textbf{x}_i\over\textrm d t}
\over M^2}.$$
To extract useful information from this, it is necessary to identify the characteristic values of the velocities that appear in the numerator. Since special relativity prohibits velocities above the speed of light $c$, while relativistic speeds are realized in nuclei, let us assume for simplicity that all velocities are $\approx c$. Then $$G\approx {RMc^2\over M^2}\approx {c^2\over\rho R^2},$$ where $\rho$ is the matter density, assumed constant. Since $G$ and $c$ are known accurately, and $\rho$ is known approximately, we can estimate the ‘radius of the universe’: $$R\approx\sqrt{c^2\over G\rho}.$$
Taking $\rho\approx 10^{-30}\textrm{g}.\textrm{cm}^{-3}$, and substituting the known values of $c$ and $G$, we find $$R\approx 10^{29}\textrm{cm},\hspace{.5cm}M\approx 10^{57}\textrm g.
\label{Size}$$
Such relationships are obtained in modern cosmology, in which they reflect the fact that the universe has a near-critical density and is therefore spatially flat. The estimates for the radius and mass of the universe refer to the current epoch and increase with time. Inflation is invoked to explain the flatness, which must otherwise be regarded as a cosmic coincidence. In scale-invariant gravity, the interpretation is different. An analogue of the flatness coincidence is a prediction that follows from the basic dynamical structure of the theory. Moreover, the radius $R$ of the universe cannot change significantly because $\mu\approx MR$ is constant. Of course, $R$ has no absolute meaning. The significance of the above estimates is that they relate human dimensions to the cosmic dimensions. The issue of whether scale-invariant gravity has any chance of replacing the standard hot big-bang model will be discussed in the next section. In this section, it only remains to estimate the relative strengths of gravity and the cosmological force.
Since $\textbf F$ increases linearly with the distance $r$ while gravity decreases as $1/r^2$, the relative strengths of the two forces will depend on $r$. The simple calculation shows, first, that the cosmological force must be attractive if the only Newtonian forces are gravity and electrostatics[^11] and, second, that at distance $\tilde{r}$ the ratio of $\textbf F$ to gravity is $${\textrm{cosmological force}\over{\textrm{gravitational force}}}\sim
{r^{3}\over R^{3}}.
\label{Ratio}$$
This result can be expected to survive in conformal gravity. It shows that on solar-system scales, for which $r/R\sim 10^{-14}$, the effect of $\textbf F$ will be $\sim 10^{-42}$ and negligible. However, for $r\sim R$, the two forces will have the same magnitude. The present derivation of the cosmological force as a scaling concomitant of gravity (and possibly other forces in certain epochs) differs markedly from all ad hoc introductions of a cosmological constant: $\textbf F$ must occur, its strength is fixed and extremely weak within the solar system, and it is always ‘tuned’ to maintain a universe of constant size.[^12] The relative strength of $\textbf F$ may increase in a ‘post-nucleosynthesis’ phase.
An Alternative Cosmology?
=========================
It seems barely possible that scale-invariant gravity could replace GR and give a viable cosmology. The standard hot Big Bang scenario looks very secure. However, scale invariance is an attractive principle, and the manner in which GR *just* fails to be scale invariant (spelled out in [@ABFO]) is mysterious. Moreover, it is precisely this ‘failure’ that makes an expanding universe possible. For this reason, the best-matching scheme, with its definite and striking predictions (\[Size\]) and (\[Ratio\]), can, at the least, act as a useful foil to the current paradigm.
In this spirit, I shall merely make a few comments. Serious conclusions must await detailed calculations using conformal gravity, the structure of which is now settled [@ABFO]. However, it looks as if the first obvious test will be passed easily: as [@ABFO] shows, conformal gravity at human, solar-system and binary-pulsar mass and distance scales appears to agree with GR as well as the scale-invariant (\[ModelAction\]) agrees with Newtonian gravity. The real problems are all in cosmology, in which the major difference between the two theories is this. In the standard model the universe is simultaneously expanding and becoming steadily more inhomogeneous. In the scale-invariant case, the universe cannot expand, since that has no meaning. It can only become more inhomogeneous – it can only ‘change its shape’.
Now it is a fact that, while remaining isotropic and homogeneous on large scales, the presently observed universe is obviously hugely more structured than it was in the past. It has changed its shape massively. The estimates of section 7 show how readily the scale-invariant potential energy can increase if the universe becomes more clumpy. Scale-invariant gravity must, in the first place, yield a cause of the Hubble red shift. The only plausible candidate that I can see is this change in the ‘potential’ of the universe induced by such clumping. It is suitably great and, according to the standard model, has been happening since the end of inflation. Therefore, the conjecture has to be that somehow the change in potential causes the Hubble red shift. This is not inherently impossible. We know that differences in the gravitational potential give rise to a gravitational red shift. Moreover, in scale-invariant gravity the increases in all forms of potential energy increase the cosmological force, which, being universal and acting on mass, is gravitational in nature. Therefore, it seems to me not entirely impossible that clumping could give rise to some kind of gravitational red shift. In the realm of classical physics, the integrated Sachs–Wolfe [@SW] effect would certainly be relatively more important than it is in Einsteinian cosmology. However, it seems to me that it must still fall far short of what is needed. Possibly more relevant is the fact that, in scale-invariant gravity, the cosmological force mimics a negative cosmological constant whose strength may have increased by orders of magnitude over cosmological epochs. Its impact on galactic dynamics is obviously a matter of some interest, as is the possibility that it might somehow have a bearing on the red shift. For the moment, I do not think that anything more can be usefully said about this crucial issue, except perhaps that scale-invariant gravity does force one to look for totally new explanations for fundamental effects. That cannot be bad. In the next section, I shall mention possible quantum effects induced by clumping.
At this stage, it is premature to try to attack detailed issues such as primordial nucleosynthesis and the origin of the microwave background. However, the isotropy and homogeneity of the observed universe on large scales and the problem of singularities seem to me to much more promising topics. I shall discuss them in the context of particle dynamics.
In considering these issues, the difference between the configuration spaces Q and Q$_0$ for theories with and without scale invariance is significant. Consider the Newtonian gravitational $N$-body problem for $N\geq 3$. It is dominated by the uniquely distinguished point in Q at which all particles coincide. In [@EOT] I have called this point (and its analogues in other configuration spaces) Alpha. The singularity of both the topology and the potential at Alpha causes the unregularizable central collisions mentioned earlier. But Alpha is not a ‘shape’ and does not belong to shape space Q$_0$ [@Kendall]. However, Q$_0$ has a rather different uniquely distinguished configuration Alpha$_0$. It is the most uniform state, at which the (negative) scale-invariant potential has an absolute maximum. In the 3- and 4-body problems, this is at the equilateral triangle and quadrilateral, respectively, whatever the masses of the particles [@Marchal]. The topology, potential and scale-invariant dynamics are very well behaved at Alpha$_0$. There cannot be anything like the Big-Bang singularity.
There is another difference. The dynamical curve of the Newtonian $N$-body problem can enter (or leave) Alpha with arbitrary energy $E$ and dilatational momentum $D$ (but only with vanishing angular momentum). If emergence from Alpha is seen as a model Big Bang, the statistics of initial states must include a spectrum of $E$ and $D$ values. In contrast, $E=D=0$ in scale-invariant dynamics. Intuitively, the fulfilment of the constraint requiring the dilatational momentum to vanish, especially in conformal gravity, where it becomes a constraint at each space point, must favour isotropic momenta distributions. At the very least, the undoubted existence of the smooth absolute maximum of the scale-invariant potential at the most uniform state must be significant. The critical points of the potential always exert a decisive influence on the classical [@Rick; @Chenciner] and quantum [@Tanner] physical processes that unfold in a configuration space. I believe scale invariance opens up new possibilities for explaining (without fine tuning and the need to postulate as yet unknown scalar fields) those features of the universe for which inflation is currently invoked.
As a final example, let me mention the flatness problem. The current theory is unable to explain why the universe seems to be expanding at almost exactly its escape velocity. In Newtonian terms, this corresponds to expansion with total energy $E$ exactly equal to zero. In the standard theory, there is no fundamental principle that can enforce $E=0$. However, we have seen that this is exactly what scale invariance does do. Of course, at the same time, it rules out expansion, predicting, in particle dynamics, a universe like a globular cluster. Nevertheless, there is a principle that enforces $E=0$.
To conclude this tentative discussion of classical cosmology, scale invariance encourages radical questioning of accepted ideas and opens up some interesting possibilities.
Possible Quantum Implications
=============================
The dimensional analysis discussed in section 6 indicates that if the universe is described by a scale-invariant law it will be necessary to revise our ideas about the constants of nature. Section 6 shows that it is in principle possible to calculate Newton’s constant G in a scale-invariant framework and suggests that the quantum of action does not have dimensions $ml^2t^{-1}$ but is a pure number. These major changes are a consequence of the elimination of time from kinematics, the transition to potentials that are homogeneous of degree -2, and the inclusion of $\mu$, the square root of the moment of inertia of the universe, in the fundamental action principle.
There is another very direct way of seeing that potentials homogeneous of degree -2 may well have far-reaching consequences. The need for such potentials is a direct consequence of having a kinetic energy quadratic in the velocities. Each velocity has the dimension of length (divided by time), and this is why we need the degree -2 in the potential. Now the quadratic kinetic energy of classical physics leads to the Laplacian in quantum physics. If one solves the time-independent Schr$\ddot{\textrm o}$dinger equation for the $1/r$ electrostatic potential, there is a mismatch between the dimensions of the potential and the Laplacian. A constant with the dimensions of length is needed. It seems to me that this is the origin of the Planck length. A scale-invariant time-independent Schr$\ddot{\textrm o}$dinger equation would need no such constant – the dimensions match without one.
The canonical approach to quantum gravity leads to the Wheeler–DeWitt equation [@DeWitt], which is very like the time-independent Schr$\ddot{\textrm o}$dinger equation [@CQG94; @EOT]. In the case of conformal gravity, it will not contain any dimensional constant. In the naivest approach, one will expect a solution of the equation, if it exists, to give (static) probabilities for different possible shapes of the configuration of the universe. In the case of the scale-invariant 3-body problem, one should then obtain probabilities for shapes of triangles. Since the scale-invariant potential for the 3-body problem has infinitely deep wells at the two-body coincidences and wave functions tend to collect in potential wells, it could well be that shapes with one short side of the triangle and two long ones will get the highest probabilities. These correspond, of course, to configurations close to a two-particle coincidence. If so, I would see this as an embryonic theoretical derivation of the fundamental ratio of scale-invariant quantum cosmology, namely the ratio of the shortest to the longest length realized in the most probable configurations. There cannot be a Planck length and a Hubble radius, but only a shortest-to-longest ratio. It is therefore possible that radical scale invariance as proposed in this paper and in [@ABFO] has the potential to derive the Planck length from first principles.
This may also have a bearing on the Hubble red shift. Since it is revealed through quantum transitions, a complete explanation of it must be based on a quantum theory of the universe. This theory is likely to be strongly affected by the key role that the total potential (both gravitational and non-gravitational) plays in the classical theory – in the ‘formula for time’ (\[clem\]) it determines how much a given amount of change ‘advances time’ and in (\[p\]) it is a factor in the canonical momenta. The potential is therefore likely to have a strong effect on the relative sizes and structures of quantum objects such as atoms and molecules, but these sizes themselves determine the potential. In the classical scale-invariant theory, the nominal initial size of the universe is conserved by the classical equations in the distinguished representation. However, without a quantum form of the theory, we cannot say what we shall measure using physical rods and clocks. Given the central role of the potential in all the basic equations of the theory, it is at least possible that it would affect rods, making them relatively shorter in states of the universe with high absolute magnitude of the potential. This would then mimic an apparent expansion of a universe that becomes more clumpy.
Despite the tentative and speculative nature of these two final sections, I hope they will serve to emphasize the potential importance of scale invariance. Weyl’s Cartesian dream represents unfinished business.
**Acknowledgements.** I have benefitted greatly from discussions with the $N$-body specialists Alain Albouy, Alain Chenciner, J$\ddot{\textrm o}$rg Elsner, Douglas Heggie, Piet Hut, Christian Marchal, Richard Montgomery, and Carles Simó. Richard suggested the use of the moment of inertia to achieve the necessary homogeneity [@RM] and the Alains and Carles showed me what could be done with the Lagrange–Jacobi relation. Ted Jacobson also helped in the early stages of this work. I thank John Briggs for discussion of quantum aspects. I have also had much help from my collaborators Edward Anderson, Brendan Foster, and Niall $\acute{\textrm O}$ Murchadha. The notion of the corrected coordinate is due in large part to Brendan. I also thank a referee and a member of the editorial board for helpful suggestions for improvement of the first draft.
Appendix: Hamiltonian Formulation
=================================
As Lanczos explains [@Lanczos], p. 169, the Hamiltonian formalism in variational mechanics is really a special case of the Lagrangian method for which the kinetic energy has a particularly simple (canonical) form, namely $T=\sum p^i\dot q_i$. Thus, the (first-order) Lagrangian of the Hamiltonian method is $$A=\int\textrm{d}t[\sum p^i\dot q_i-H(q_1,\dots,q_n;p^1,\dots,p^n)],
\label{Lag}$$ where $H$ is the normal Hamiltonian and is treated for variational purposes as if it were a potential.
From the Lagrangian (\[Lag\]), in which the $p^i$’s are coordinates on an equal footing with the $q_i$’s except that they appear as Lagrange multipliers (their velocities do not occur in the Lagrangian, unlike the $\dot q_i$’s), one calculates the Euler–Lagrange equations in the standard manner. They turn out to be the Hamiltonian equations.
To obtain a Hamiltonian treatment appropriate for scale-invariant gravity, two modifications to the standard treatment are needed. First, the Lagrange function of any Jacobi action is homogeneous of degree one in the velocities. This leads to the identity (\[QuadCon\]), which in turn has the consequence that when one attempts to calculate the Hamiltonian by the rule $H=p^i\dot q_i-{\cal L}$ an expression that vanishes identically is obtained. This problem is overcome in Dirac’s generalized Hamiltonian theory [@Dirac] by converting (\[QuadCon\]) into the quadratic constraint $${\cal H}=\sum_i{{{\textbf{p}}_{i}}\cdot
{{\textbf{p}}_{i}}\over 2m_{i}}-(E-U)=0$$ and taking $H=N\cal H$, where $N$ is a Lagrange multiplier, as the Hamiltonian. In the scale-invariant case, the constraint is of the same basic form but, first, $E=0$ and, second, $\bar{\cal L}$ is a function of the corrected coordinates $\bar q_i$ and their velocities. The quadratic constraint (in the one-dimensional case) is $$\bar{\cal H}=\sum_i{p^ip^i\over 2m_i}+(a^{\scriptsize\textrm{d}})^2U(\bar q^i)=0,
\hspace{.5cm}\bar q^i=a^{\scriptsize\textrm{d}}(q^i+a^{\scriptsize\textrm{t}}),$$ where $a^{\scriptsize\textrm{t}}$ and $a^{\scriptsize\textrm{d}}$ are the gauge variables corresponding to translations and dilatations, and the $(a^{\scriptsize\textrm{d}})^2$ appears in the constraint because $p^i=a^{\scriptsize\textrm{d}}{\partial\bar{\cal L}/\partial v^i}$. Any other constraints obtained in the Lagrangian treatment must be added with further Lagrange multipliers. In our case, we shall need to add, with appropriate multipliers, the primary constraints (\[Niall\]) corresponding to the various types of best matching. This is relatively standard, through there is, as we shall shortly see, a slight departure from Dirac’s procedure. The second modification concerns the treatment of the gauge auxiliary variables. We need to find a Hamiltonian treatment that corresponds to the free-end-point variation in the Lagrangian approach. Both modifications are achieved if we use the first-order principle $$A=\int\textrm{d}\lambda\left(\sum_ip^i\dot q^i
+p^{\scriptsize\textrm{t}}\dot a^{\scriptsize\textrm{t}}+p^{\scriptsize\textrm{d}}\dot a^{\scriptsize\textrm{d}}-
N\bar{\cal H}-N^{\scriptsize\textrm{t}}\left(p^{\scriptsize\textrm{t}}-\sum_i p^i\right)-
N^{\scriptsize\textrm{d}}\left(p^{\scriptsize\textrm{d}}a^{\scriptsize\textrm{d}}-\sum_i p^iq^i\right)\right),
\label{ConstrainedHam}$$ where $p^{\scriptsize\textrm{t}}, p^{\scriptsize\textrm{d}}$ are the canonical momenta corresponding to auxiliary gauge variables, and $N^{\scriptsize\textrm{t}},N^{\scriptsize\textrm{d}}$ are multipliers associated with the translational and dilatational constraints.
If we now vary (\[ConstrainedHam\]) with respect to the multipliers $N, N^{\scriptsize\textrm{t}},
N^{\scriptsize\textrm{d}}$, we obtain $$\bar{\cal H}=\sum_i{p^ip^i\over 2m_i}+(a^{\scriptsize\textrm{d}})^2U(\bar q^i)=0,
\hspace{.5cm}p^{\scriptsize\textrm{t}}-\sum_i p^i=0,
\hspace{.5cm}p^{\scriptsize\textrm{d}}a^{\scriptsize\textrm{d}}-\sum_i p^iq^i=0.
\label{PrimaryConstraints}$$
We now employ the free-end-point method and require the variation of the action w.r.t both the coordinates $a^{\scriptsize\textrm{t}}$, $a^{\scriptsize\textrm{d}}$ and the velocities $\dot a^{\scriptsize\textrm{t}}$, $\dot a^{\scriptsize\textrm{d}}$ of the auxiliary gauge variables to vanish. The variation w.r.t the velocities tells us $$p^{\scriptsize\textrm{t}}=0,\hspace{.5cm}p^{\scriptsize\textrm{d}}=0,$$ and in conjunction with (\[PrimaryConstraints\]) these imply $$\sum_i p^i=0,\hspace{.5cm}\sum_i p^iq^i=0.
\label{SecCons}$$
Variation with respect to the auxiliary variables $a^{\scriptsize\textrm{t}}$ and $a^{\scriptsize\textrm{d}}$ gives $${\partial U\over\partial a^{\scriptsize\textrm{t}}}=-\sum_i{\partial U\over\partial q^i}=0,
\hspace{.5cm}p^{\scriptsize\textrm{d}}+2a^{\scriptsize\textrm{d}}U+(a^{\scriptsize\textrm{d}})^2
{\partial U\over\partial a^{\scriptsize\textrm{d}}}=0.
\label{Conds}$$
The first of these conditions is the one that ensures Newton’s third law, and, since we already have $p^{\scriptsize\textrm{d}}=0$, the second gives the homogeneity condition on the potential.
Since the momenta are regarded as independent variables, we also vary w.r.t the momenta of the auxiliary variables, which gives $$\dot a^{\scriptsize\textrm{t}}=N^{\scriptsize\textrm{t}},\hspace{.5cm}
\dot a^{\scriptsize\textrm{d}}=N^{\scriptsize\textrm{d}}.
\label{Velocities}$$
Finally, Hamilton’s equations are $$\dot q^i={\partial H\over\partial p^i}=N{p^i\over m_i}+N^{\scriptsize\textrm{t}}+N^{\scriptsize\textrm{d}}q^i,
\hspace{.5cm}\dot p^i=-N(a^{\scriptsize\textrm{d}})^2{\partial U\over\partial q^i}+N^{\scriptsize\textrm{d}}p^i.
\label{HamEqs}$$
These reduce to Newton’s equations if we go over to the distinguished representation by setting $N=1$ and $N^{\scriptsize\textrm{t}}=N^{\scriptsize\textrm{d}}=0$. However, we can perfectly well specify the three multipliers freely. Then $N^{\scriptsize\textrm{t}}$ and $N^{\scriptsize\textrm{d}}$ are velocities in the unphysical gauge directions generated by translations and dilatations as in Dirac’s treatment [@Dirac], and $N$ is the unphysical ‘label speed’ at which we move along the dynamical orbit in shape space.
However, it will be noted that we have generalized Dirac’s treatment by the inclusion of auxiliary gauge variables in addition to the conventional physical variables and multipliers. The simpler and more familiar form given by Dirac corresponds to the replacement of (\[ConstrainedHam\]) by $$A^*=\int\textrm{d}\lambda\left(\sum_ip^i\dot q^i-
N\left(\sum_i{p^ip^i\over 2m_i}+(a^{\scriptsize\textrm{d}})^2U(\bar q^i)\right)-N^{\scriptsize\textrm{t}}\sum_i p^i-
N^{\scriptsize\textrm{d}}\sum_i p^iq^i\right).
\label{DiracHam}$$ Variation w.r.t the multipliers $N, N^{\scriptsize\textrm{t}},N^{\scriptsize\textrm{d}}$ gives the first member of (\[PrimaryConstraints\]) and (\[SecCons\]). However, the first member of (\[Conds\]) no longer arises from the variational principle but must be adjoined if the constraint $\sum_i p^i=0$ is to propagate. In addition, the status of $a^{\scriptsize\textrm{d}}$ is strange. Variation w.r.t it does give the second member of (\[Conds\]), but (\[Velocities\]) has no variational derivation and its second member must be adjoined in order to propagate the constraint $\bar{\cal H}=0$. We see that in the presence of scale invariance the Dirac procedure works but lacks a transparent variational basis.
More could be said about these matters, above all about a new unifying perspective on gauge theory that best matching provides. But since this is already a long paper, that must be the subject of another paper.
[99]{}
Anderson E, Barbour J, Foster B Z and Ó Murchadha N 2002 “Scale-Invariant Gravity: Geometrodynamics", *Class. Quantum Grav.* 2003 **20** 1571, arXiv:gr-qc/0211022 Barbour J, Foster B Z and Ó Murchadha N 2002 *Class. Quantum Grav.* **19** 3217; “Relativity without relativity”, arXiv:gr-qc/0012089 Anderson E and Barbour J 2002 *Class. Quantum Grav.* **19** 3249; “Interacting vector fields in relativity without relativity”, arXiv:gr-qc/0201092 Weyl H 1918 *Sitzungsber. d. Preuss. Akad. d. Wissensch.* 465 (There are English translations and commentaries of Weyl’s papers togther with much interesting material on gauge theory in: O’ Raifeartaigh L 1997 *The Dawning of Gauge Theory* (Princeton: Princeton University Press) Einstein’s objection was reported by Weyl at the end of [@Weyl1] Dirac P A M 1973 *Proc. R. Soc.* A **333** 403 Brans C and Dicke R 1962 *Phys. Rev.* **124** 935 Wetterich C 1988 *Nucl. Phys.* **B302** 645 668; Bachm$\ddot{\textrm u}$ller W and Dragon N 1989 *Nucl. Phys.* **B321** 207 Weyl H 1919 *Ann. d. Physik* **59** 101 Weyl H 1949 *Philosophy of Mathematics and Natural Science* (Princeton, Princeton University Press), p. 83 Poincar$\acute{\textrm e}$ H 1905 *Science and Hypothesis* (London, translated from the French edition of 1902) Mach E 1883 *Die Mechanik in ihrer Entwicklung historisch-kritsch dargestellt* (Leipzig: Barth); 1893 *The Science of Mechanics* (Chicago: Open Court) Barbour J 2001 “On general covariance and best matching”, in *Physics Meets Philosophy at the Planck Length*, eds. Callender C and Huggett N (Cambridge: Cambridge University Press) Einstein A 1955 *The Meaning of Relativity* (London: Methuen), Appendix II Schr$\ddot{\textrm o}$dinger E 1925 *Annalen der Physik* **77** 325 (English translation in: *Mach’s Principle: From Newton’s Bucket to Quantum Gravity*, eds. Barbour J and Pfister H (Boston: Birkh$\ddot{\textrm a}$user) Hughes V, Robinson H, and Beltran-Lopez V 1960 *Phys. Rev. Lett.* **4** 342; Drever R 1961 *Phil Mag.* **6** 683 Barbour J and Bertotti B 1982 *Proc. R. Soc. A* **382** 295 Hertz H 1956 *The Principles of Mechanics* (New York: Dover) (German original 1895) Lanczos C 1949 *The Variational Principles of Mechanics* (University of Toronto Press, Toronto)(also available from Dover, New York) Barrow J D 2002 *The Constants of Nature* (London: Jonathan Cape) Fulton T, Rohrlich F, and Witten L 1962 *Reviews of Modern Physics* **34** 442 Chenciner A and Sim$\acute{\textrm o}$ C 2001 Private communication Dziobek O 1888 *Die mathematische Theorien der Planeten-Bewegungen* (Leipzig: J A Barth); English translation: *Mathematical Theories of Planetary Motions* (New York: Dover). Barbour J 1994 *Class. Quantum Grav.* **11** 2875 Barbour J 1999 *The End of Time* (London: Weidenfeld and Nicolson; New York: Oxford University Press) Dirac P A M 1964 *Lectures on Quantum Mechanics* (New York: Yeshiva University) Poincaré H 1898 *Rev. Métaphys. Morale* **6** 1 (English translation 1913: “The Measure of Time” in *The Value of Science* (New York: Science Press) Clemence G 1957 *Rev. Mod. Phys.* **29** 2 Littlejohn R and Reinsch M 1997 *Rev. Mod. Phys.* **69** 213 Marchal C 1990 *The Three-Body Problem* (Amsterdam: Elsevier) Moeckel R 1988 *Contemporary Mathematics* **81** 1 Einstein A 1917 *Sitzungsber. d. Preuss. Akad. d. Wissensch.* **1** 142 Chenciner A 1998 *Reg. Chaotic Dynamics* **3** 93 Barbour J and Bertotti B 1977 *Nuovo Cimento* **38B** 1 Planck M 1899 *Sitzungsber. d. Preuss. Akad. d. Wissensch.* **5** 440; *Ann. d. Physik* **11** 69 Rees M J and Sciama D W 1968 *Nature **217 511; Padmanabhan T 1993 “Structure Formation in the Universe" (Cambridge: Cambridge University Press). Kendall D 1984 *Bull. London Math. Soc.* **16** 81 Tanner G, Richter K and Rost J-M 2000 *Rev. Mod. Phys.* **72** 497 DeWitt B 1967 *Phys. Rev.* **160** 1113 Richard Montgomery 1999 Private communication.***
[^1]: In Appendix II of [@Einstein1955] Einstein develops a notion of the ‘strength’ of dynamical theories that bears some relation to Poincaré’s criterion. However, it is formulated in spacetime rather than configuration-space terms.
[^2]: The situation in Newtonian mechanics is somewhat enigmatic. No absolute length itself can be observed (the overall scale is nominal), but the effect of *change of length* is observable. This matches what happens with regard to rotations. Newtonian mechanics is invariant under a time-independent rotation but not under changing rotations.
[^3]: I understand from $N$-body specialists [@LJR] that the relation (\[Iddh\]) of this section is generally known as the Lagrange–Jacobi relation. For a useful account of Lagrange’s work, see [@Dziobek]
[^4]: It will in the particle model, since positive-definite quantities are being minimized; the situation remains open in scale-invariant geometrodynamics [@ABFO], which however has a well-behaved Hamiltonian formulation.
[^5]: The auxiliary space $\Gamma$ is a fibre over the base Q, but the formalism of fibre bundles is not needed for translations and dilatations. It is for rotations, which are only mentioned in this paper. For a beautiful gauge fibre-bundle treatment of rotations in the $N$-body problem, see [@LR].
[^6]: We shall use the words *coordinate, velocity, and momentum* in the sense of Hamiltonian canonical theory.
[^7]: A full treatment of this point goes beyond the scope of this paper. I hope to discuss it in a further paper devoted to the unifying perspective on gauge theory that best matching provides.
[^8]: The scale-invariant potential in fact determines all the qualitative properties of the Newtonian $N$-body problem and is a more important concept than the ordinary potential. For example, all asymptotic limiting motions of the 3-body problem terminate at one of the critical or singular points of $U_0$. They ‘govern’ the motion [@Marchal; @Rick].
[^9]: Although $C(t)$ is epoch dependent, this does not mean that the theory contains any fundamental coupling constants with such a dependence. The epoch dependence is an artefact of the decomposition of the forces into Newtonian-type forces and a residue, which is the cosmological force.
[^10]: Physicists wonder how it will ever be possible to explain the huge difference between the strength of gravity and the remaining forces of nature. It seems to me not impossible (see the first footnote on p. 23 of [@BB1]) that, in a self-contained universe, it is somehow a consequence of the very different ways in which the potentials for gravity and electrostatics are calculated. This might have far-reaching consequences in quantum cosmology.
[^11]: The expressions (\[ModEq\]) and (\[DefC\]) simplify if there is only one value of $k$, and we can use the equation (\[ScaleGravEq\]).
[^12]: One can see qualitatively how this happens. Suppose a particle in a ‘cloud’ that forms an island universe manages to escape a significant distance from the cloud. It will then feel a restoring force that grows linearly with distance and will be rapidly pulled back into the cloud as if it were on a piece of elastic. Similarly, if the cloud becomes egg shaped, the particles at the ends of the long axis will feel an enhanced restoring force.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Two-phase superconductor tapes were produced by blending high purity magnesium diboride powder with a liquid ethylcellulose-based polymeric binder. This procedure produced a material which is easily formable with a high superconducting transition temperature (38K). We show that the bulk superconducting properties are not affected by the presence of the binder, nor is there any evidence of a chemical reaction between the superconducting particles and the binder. However, the transport properties of the material are strongly affected by the presence of the binder, which leads to a seven order of magnitude increase of the normal state resistance along with a seven order of magnitude decrease of the transport critical current density. This new material is shown to be equivalent to a system of coupled Josephson junctions.'
author:
-
-
title: 'Cellulose-Bound Magnesium Diboride Superconductivity'
---
Introduction
============
Magnesium diboride was found to be superconducting in 2001 by Nagamatsu et al. [@1]. In crystalline or polycrystalline form MgB$_2$ has a transition temperature of 39 K [@1], the highest for any simple intermetallic compound in addition of being one of the lightest. It boasts a high critical current density, surpassing 85,000 A/cm2 [@2], a high critical magnetic field, typically 8 T at 21 K [@3] and grain boundary transparency [@4]. This leads to a huge potential for applications [@magnet] in view of replacing other very successful conventional superconductors based on niobium alloys, which all have lower critical temperatures but have been used as workhorses for superconducting applications like magnets. It turns out that MgB$_2$ is a simple compound, which was already discovered in 1953 [@5]. The wide availability of this compound eliminates many steps in its fabrication, including sintering and pressing, synthesis at elevated temperatures and under reducing conditions, as is the case for other high temperature cuprate superconductors. The pairing mechanism is meditated by phonons, which makes it a conventional BCS superconductor [@BCS] with the exception of it’s two double gap structure [@double]. Thanks to its conventional pairing mechanism it would be an ideal candidate for the next generation of quantum interferences devices based on isotropic tunneling in Josephson junction.
A major roadblock in the development of magnesium boride in its quest to supplant other conventional compounds is the relative difficulty in producing reliable superconducting wires. While some techniques such as the powder in tube method [@PIT] have proven to work, they remain very brittle, similar to high temperature cuprate superconductors. The idea in this work is to explore new avenues in order to fabricate a more robust and malleable material, with a good potential for quantum interference devices and which can be easily molded in any shape. Therefore it is important to have an isotropic material, where the properties do not depend on particular orientations. Hence, we explored the possibility of obtaining such a material by using readily available polycrystalline MgB$_2$ powder as starting material and bind it using a cellulose based polymer. Cellulose tends to react only very weekly and is very robust, hence constitutes a good candidates for forming novel compounds.
Fabrication
===========
The MgB$_2$ powder was incorporated into a flexible polymeric matrix, ethylcellulose which can be naturally derived and shaped into any geometry. The binder was prepared from a mixture of Standard 45 Ethocel and a 50/50 solvent blend of butyl Cellosolve and terpineol. The MgB$_2$ powder was commercially obtained and of 99.8% purity, of particles below 44 microns and was incorporated with the prepared wet binder in ratios varying from 20% to 90%. In order to produce tapes, the MgB$_2$ paste was spread onto an Al foil substrate and then dried before depositing contacts. Some samples were also deposited directly on a plastic sample holder, which did not affect the quality of the samples. Shaping the tapes required a simple cutting tool and peeling from the flexible aluminium substrate. The surface of the tapes was imaged at high magnification using a Hitachi 4700-S field emission gun scanning electron microscope (FEGSEM) shown in figure \[SEM1\]. The surface is rough due to the polymer component as well as the jaggedness of the MgB$_2$ particles. This method is quite different from the powder in tube method since no inter-diffusion is occurring.
![SEM image of the MgB$_2$ polymer tape showing surface topography and MgB2 particles[]{data-label="SEM1"}](SEM1.eps){width="3in"}
The elemental and compound identifications were determined using energy dispersive spectroscopy (EDS) and x-ray diffraction (XRD). XRD was performed using a Philips 1050/65 x-ray diffractometer and a Philips PW 1710 diffractometer. The American Instrument Inc. x-ray generator was the source of CuK$\alpha$ radiation. The spectrum is shown in figure \[XRD\]. Except for some additional traces of magnesium oxide, no other elements other than MgB$_2$ are seen. More importantly, there is no change in the spectrum when comparing X-ray diffraction of the powder directly with the diffraction of the MgB$_2$ polymer tape, indicating that no structural change is occurring in the polymer mix.
![X-ray diffraction spectrum of a MgB$_2$ polymer tape. []{data-label="XRD"}](Xray.eps){width="3in"}
The elemental analysis, using EDS showed expected peaks for B, C, O and Mg. The carbon and oxygen peaks are attributed to the polymer component of the sample, whose constituent compounds are ethylcellulose, \[C$_6$H$_7$O$_2$(OC$_2$H$_5$)$_3]$n, and butyl acetic Cellosolve C$_4$H$_9$OCH$_2$CH$_2$OC(O)CH$_3$.
In order to investigate the possibility of a chemical change in the ethylcellulose polymer during fabrication, Fourier transform infrared spectroscopy (FT-IR) analysis was conducted shown in figure \[FTIR\]. The dried binder, the MgB2 powder and the MgB2 polymer tapes were analyzed via FT-IR separately and the strongest peaks were located and compared to the MgB$_2$ polymer tapes. The FT-IR instrument was a Bio-Rad FTS 6000 spectrometer, used at 2.5 kHz with a resolution of 8 cm$^{-1}$ for 128 scans in the range of 4000 to 450 cm$^{-1}$. Carbon black was used to determine the background spectrum. The signal was amplified using a photoacoustic (PA) accessory. In addition, the MgB$_2$ polymer tapes were crushed for greater signal amplification before placing in the PA cell. We found peaks at (1056, 1383, 2844, 2930, 3482) cm$^{-1}$, when measuring only the polymer, which is consistent with the results by Suthar et al. [@6] for ethylcellulose. The same peaks were found again in the FT-IR spectrum of the MgB$_2$-polymer tapes, showing that there is no chemical change of the ethylcellulose when mixed to the MgB$_2$ powder. The only change was a substantial weakening of the peak at 3482cm$^{-1}$, which corresponds to the O-H stretching as indicated by the circle on figure \[FTIR\]. The MgB$_2$-polymer tapes showed additional peaks at (506, 671, 1459, 1650, 671) cm$^{-1}$, corresponding to the MgB$_2$ system. We found these same peaks when measuring only the MgB$_2$ powder, which demonstrates that the ethylcellulose did not affect the MgB$_2$ system either. The positions of the peaks were all consistent with the ones measured by other groups in MgB$_2$ [@7; @8; @9]. Surprisingly, no new peak was found between 450 and 4000 cm$^{-1}$, which might correspond to a new bond created between the polymer and MgB$_2$. Overall, the FT-IR measurements showed that mixing MgB$_2$ powder with ethylcellulose does not alter its components significantly.
![FT-IR intensity for the MgB$_2$-polymer tape (red) and the sum of the intensities for the MgB$_2$ powder only plus the polymer only.[]{data-label="FTIR"}](FTIR.eps){width="3in"}
Bulk superconductivity
======================
Bulk superconductivity was tested by performing standard magnetization measurements as a function of temperature and field. The temperature dependence shows a sharp drop at $T_C=38\pm0.5$K as shown in figure \[M\] and indicates high quality superconducting MgB$_2$ particles. The dependence is very similar to observations of single polycrystals of MgB$_2$ [@Shigeta]. This means that the cellulose based binder does not alter the superconducting properties of the MgB$_2$ particles.
![The Magnetization at low fields for a MgB$_2$-polymer tape as a function of temperature.[]{data-label="M"}](MvsT.eps){width="3in"}
For magnetic fields beyond the first critical field $Bc_1$, it is possible to extract the bulk critical current density from the hysteric magnetization curve using Bean’s model and assuming that the polycrystals in the binder are well separated. For cylindrical-like grains the critical current density is then given by $J_C=3\Delta M/a$ [@Bean], where the factor 3 is a geometrical factor assuming a cylindrical shape. This geometrical factor varies by less than a factor of two for any standard geometry with a similar aspect ratio. Hence, the order of magnitude of the extracted current is insensitive to the details of our assumptions. $\Delta M$ is the difference in magnetization in international units \[A/m\] of the up-sweep and the down-sweep. $a$ (in meter) is the average diameter of the particle in the direction perpendicular to the applied field. $J_C$ shown in figure \[Jc\], therefore represents the magnetization critical current density of a single superconducting particle. It is also possible to extract the susceptibility of the entire sample, by evaluating the slope of the initial low field magnetization curve, which we found to be $\chi=-0.45\pm0.05$. In order to obtain this value we used an effective demagnetizing factor corresponding to a vertical cylinder of aspect ratio 1 [@Brandt]. This implies that only about half (by volume) of the sample is superconducting, i.e., the binder does not contribute to superconductivity. Indeed, the density of the MgB$_2$ polymer was (1.7$\pm$0.2)g/cm$^3$ as opposed to 2.6g/cm$^3$ for the density of a MgB$_2$ crystal, which amounts to 60% of MgB$_2$ particles and is comparable to the measured susceptibility.
![The hysteresic magnetization dependence as a function of applied magnetic field. The lower curve corresponds to the up-sweep, whereas the top curve shows the down sweep.[]{data-label="Jc"}](CriticalCurrent.eps){width="3in"}
Assuming a typical particle diameter of 40$\mu m$ we obtain a maximum critical current density of $J_c\simeq 10^6$A/cm$^2$, which is very similar to the results for a single polycrystal [@Shigeta] of a similar size. This shows that the superconductivity within a particle is not affected by the presence of the cellulose based binder. This is however in stark contrast, as we will see below, to the transport properties, where the coupling between superconducting particles becomes important.
Transport superconductivity
===========================
In order to measure the transport properties we made ohmic contacts using two different methods. The simplest consists in placing silver wires in the wet MgB$_2$ powder-binder mix and letting it dry and solidify. Such a sample is shown in figure \[SamplePic\], where it is mounted on a 14pin header for transport measurements. Alternatively, silver epoxy was used as well.
![Image of the MgB$_2$ polymer with contacts on a probe header.[]{data-label="SamplePic"}](SamplePic2.eps){width="3in"}
Quite strikingly, the temperature dependence of the resistance is very weak above the critical temperature, before dropping very slowly into the superconducting regime as shown in figure \[RvsT\]. While the onset is at $T_C=38.4\pm 0.5$K and very close to the value obtained from magnetization measurements, the transition width as measured from the 10% and 90% value of the normal resistance is $\Delta T=15\pm 3$K. The errors are based on the statistics of a dozen tested samples all showing very wide transitions. This has to be contrasted to the case of polycrystals or single crystals, where the temperature drop is very sharp.
![Resistance normalized to the resistance just above $T_C$ as a function of temperature. The main figure shows the dependence of the high temperature part, whereas the inset shows the temperature dependence of the normalized resistance on a logarithmic scale close to the superconducting transition. []{data-label="RvsT"}](RvsT.eps){width="3in"}
Similarly striking is the extremely high value of resistivity above $T_C$. Indeed, our samples show a range in resistivities of $\rho=3-20\Omega$ cm, which is about 7 orders of magnitude higher than typical values for crystals and polycrystals ($\rho\simeq 2.5\mu \Omega$cm) [@Xu]. A typical resistance for our sample geometry is 100$\Omega$. Moreover, the temperature dependence of the normal resistance is very weak just above $T_C$. It varies by less than 20% up to 200K. This again is very different to single crystals and polycrystals, where the residual resistance ratio (RRR) factor is about 5 and 20, respectively. Above 200K, some of our samples show almost no temperature dependence (less than 20%), whereas other samples show an increase up to a factor 10. We believe that is probably due to the mechanical rearranging of some particles.
When comparing the transport critical current density, as determined from the current voltage characteristics, to the critical current density extracted from the magnetization data we obtain a transport critical current density in the range $J_C$=0.05-0.2 A/cm$^2$, which is 7 orders of magnitude smaller than the magnetization critical current density.
Hence, this leaves us to draw the following picture of what is happening in this superconducting compound: The MgB$_2$ superconducting particles are largely unaffected by the mixing with the cellulose based binder, which holds the superconducting particles together. Macroscopic superconductivity is obtained thanks to Josephson coupling between these particles. This coupling is strong enough to allow for macroscopic superconductivity to occur. The presence of the Josephson junctions (JJ) in the system explains the wide transition observed around the critical temperature.
Further evidence that this system is dominated by a network of JJ is the nature of the magnetic field induced transition. As shown in figure \[RvsB\], the magnetic field induced transition from normal to superconducting is extremely wide (about 6T) as measured from the 10% to 90% value of the normal resistance. This is very different from single crystals or polycrystals, which exhibit a much narrower transition range.
![The normalized resistance as a function of the magnetic field at 4.2K[]{data-label="RvsB"}](RvsB.eps){width="3in"}
This picture of a network of JJ also explains the very low values of critical transport currents observed. Indeed, the Ambegaokar Baratoff relation [@JJ] for JJ implies an inverse dependence between the critical current and the normal resistance ($J_C\sim R_N^{-1}$). Hence, the 7 order of magnitude increase of the normal resistance as compared to the single crystal can be directly related to the observed 7 order magnitude decrease in critical currents, providing further evidence that our sample structure is dominated by JJ. Interestingly, varying the ratio of the cellulose binder to MgB$_2$ does not significantly alter the transport properties, indicating that the coupling between superconducting particles is not affected by the ratio of cellulose to MgB$_2$. Therefore, the cellulose binder simply “holds” the superconducting particles together without modifying any of the intrinsic properties. Within this picture it is then possible to greatly enhance the critical current density by simply reducing the size of the superconducting particles, which would lead to an increase proportional to $J_C\sim a^{-2}$ [@JJN], where $a$ is the typical diameter of the superconducting particles. Consequently, the normal resistivity would decrease as $\rho_N\sim a$. For reasonable values of particle sizes (60nm) this would lead to a critical current density of the order of $J_C\simeq 10^5$ A/cm$^2$ and therefore represent a very interesting system for various applications.
Conclusion
==========
In summary, we discovered an efficient production path to making MgB$_2$ tapes through an initial formable step involving a cellulose based binder. The material exhibits a high transition temperature close to that of crystalline MgB$_2$. We showed that the binder does not alter any of the chemical properties and constitutes an optimal way to form a new compound, which is mechanically strong and formable. The main drawback is the very low critical current density. However, the current density could be increased substantially by using much smaller particle sizes, hence opening the door for many potential applications for high temperature superconductivity and possibly new quantum interference devices based on the intrinsic Josephson junctions.
Acknowledgment {#acknowledgment .unnumbered}
==============
Samir Elouatik is gratefully acknowledged for his help in analyzing the results as well as support from NSERC, FQRNT, RQMP and INTRIQ.
[1]{}
J. Nagamatsu, N. Nakagawa, T. Muranaka, Y. Zenitani, J. Akimitsu, Nature 410 (2001) 63-64. S. Jin, H. Mavoori, C. Bower, R. B. van Dover, Nature 411 (2001) 563-565. V. P. S. Awana, R. Rawat, A. Gupta, M. Isobe, K. P. Singh, A. Vajpayee, H. Kishan, E. Takayama-Muromachi, A. V. Narlikar, Solid State Communications 139 (2006) 306-309. Y. Bugoslavsky, G. K. Perkins, X. Qi, L. F. Cohen, A. D. Caplin, Nature 410 (2001) 563-565. K. Vinod, R.G. Abhilash Kumar and U. Syamaprasad, Supercond. Sci. Technol. 20 (2007) R1–R13. M. E. Jones, R. E. Marsh, J. Am. Chem. Soc. 76 (1954) 1434-1436. S. Jin, H. Mavoori, C. Bower and R.B. van Dover, Nature 411 (2001), 563-565. J. Bardeen, L. N. Cooper, and J. R. Schrieffer, Phys. Rev. 106, (1947) 162-164. E.J. Choi, D. Roundy, H. Sun, M.L. Cohen, S.G. Loule, Nature 418, (2002) 758-760. V. Suthar, A. Pratap, H. Raval, Bulletin of Materials Science 23 (2000) 215-219.
C. S. Sundar, A. Bharathi, M. Premila, T. N. Sairam, S. Kalavathi, G. L. N. Reddy, V. S. Sastry, Y. Hariharan, T. S. Radhakrishnan, Condensed Matter (2001). K. M. Subhedar, R. S. Hyam, V. Ganesan, S. H. Pawar, Physica C: Superconductivity 449 (2006) 73-79. J. Kortus, I. I. Mazin, K. D. Belashchenko, V. P. Antropov, L. L. Boyer, Physical Review Letters 86 (2001) 4656. I. Shigeta, T. Abiru, K. Abe, A. Nishida, Y. Matsumoto, Physica C: Superconductivity 392–396 (2003) 359–363. C.P. Bean, Rev. Mod. Phys. 36 (1964) 31. E.H. Brandt, Physica B 284-288 (2000) 743-744. M. Xu, H. Kitazawa, Y. Takano, J. Ye, K. Nishida, H. Abe, A. Matsushita, N. Tsujii, and G. Kido, Appl. Phys. Lett., 79 (2001) 2779-2781. V. Ambegaokar and A. Baratoff, Phys. Rev. Lett. 10, (1963) 486-489. J. R. Clem, B. Bumble, S. I. Raider, W. J. Gallagher, and Y. C. Shih, Phys. Rev. B 35, (1987) 6637-6642.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'V.Ferenczi'
title: OPERATORS ON SUBSPACES OF HEREDITARILY INDECOMPOSABLE BANACH SPACES
---
=msbm10 =msbm7 =msbm5 ===
Introduction
============
By [*space (resp. subspace)*]{}, we shall always mean infinite dimensional space (resp. subspace). A space $X$ is [*hereditarily indecomposable*]{} if no two subspaces of $X$ are in a direct sum. In the whole article, $X$ stands for a hereditarily indecomposable complex Banach space. It was shown in [@GM] that every operator from $X$ to $X$ is of the form ${\mbox{$\lambda$}}I+S$, where $I$ is the identity map and $S$ is strictly singular. We generalize this result by showing that for every subspace $Y$ of $X$, every operator from $Y$ to $X$ is of the form ${\mbox{$\lambda$}}I+S$, where $I$ is the inclusion map and $S$ is strictly singular (this was proved in [@GM] in a particular case).
In fact, it is a consequence of the following lemma that it is enough to prove this for every subspace $Y$ with a basis with constant $2$.
Lemma
=====
#### Lemma 1
[*Let $Y$ be H.I. and $S$ an operator from $Y$ to some Banach space. Let $Z \subset Y$. Then $S$ is strictly singular if and only if $S_{/Z}$ is strictly singular.*]{}
The direct implication is clear. Now suppose that $S_{/Z}$ is strictly singular while $S$ is not. Then there is a subspace ${\mbox{$Y'$}}$ of $Y$, and $c > 0$ such that $\forall y \in {\mbox{$Y'$}}, {\mbox{$\|S(y)\|$}} \geq c {\mbox{$\|y\|$}}$; furthermore, given ${\mbox{$\epsilon$}}< c/(1+{\mbox{$\|S\|$}})$, there is a subspace ${\mbox{$Z'$}}$ of $Z$ with ${\mbox{$\|S_{/{\mbox{$Z'$}}}\|$}} \leq {\mbox{$\epsilon$}}$. As $Y$ is H.I., there exist two unit vectors $y \in {\mbox{$Y'$}}$, $z \in {\mbox{$Z'$}}$ with ${\mbox{$\|y-z\|$}} \leq {\mbox{$\epsilon$}}$. Then $ c-{\mbox{$\epsilon$}}\leq {\mbox{$\|S(y-z)\|$}} \leq {\mbox{$\|S\|$}}{\mbox{$\epsilon$}}$, a contradiction.
#### Consequence
In particular, this lemma can be applied to an operator from $Y \subset X$ to $X$. Now suppose that the claim of the article is true for every subspace with a basis with constant $2$. Let $Y \subset X$, and $T:Y \rightarrow X$. We know that $Y$ contains a normalized basic sequence with constant $2$; let $Z$ be the subspace associated to it. By hypothesis, there exists ${\mbox{$\lambda$}}$ such that $T_{/Z}-{\mbox{$\lambda$}}I_{/Z}$ is strictly singular. By the lemma, $T-{\mbox{$\lambda$}}I_{/Y}$ is still strictly singular.
A filter on block-subspaces
===========================
A normalized basic sequence $(y_i)_{i \in \N}$ in $X$ will be denoted by $y$. We denote by $[y]$ the closed subspace generated by $y$, by $I_y$ the inclusion map from $[y]$ to $X$, by $y^{k}$ the normalized basic sequence $(y_{i+k})_{i \in \N}$. For an operator $T$ such that the restriction of $T$ to $[y]$ is defined, $T_{/y}$ stands for this restriction. If $(y_i)_{i \in \N}$ is a basic sequence such that for every $i$, $y_i \neq 0$, then $|y|$ denotes the normalized basic sequence $(y_i/{\mbox{$\|y_i\|$}})_{i \in \N}$. By ${\mbox{$y'$}}\subset y$, we mean that ${\mbox{$y'$}}$ is a normalized block basic sequence of $y$. Let $e$ be a normalized basic sequence in $X$ with constant $2$. Let $J=\{y / y \subset e\}$.
#### Definition 1
Given $y$ and $z$ in $J$, we say that they are [*similar*]{}, and write [*$y \sim z$*]{} if $\Sigma {\mbox{$\|z_i-y_i\|$}} < +\infty$.
#### Properties
Similarity is an equivalence relation. Furthermore, for every $y \sim z$ and every ${\mbox{$y'$}}\subset y$, there exists ${\mbox{$z'$}}\subset z$ such that ${\mbox{$y'$}}\sim {\mbox{$z'$}}$. Indeed, for $k \in \N$, let ${\mbox{$z'$}}_k$ have the same coordinates on $(z_i)$ as ${\mbox{$y'$}}_k$ on $(y_i)$: the normalization $|{\mbox{$z'$}}|$ of $({\mbox{$z'$}}_k)_{k \in \N}$ is similar to ${\mbox{$y'$}}$.
#### Definition 2
[*For $y$ and $z$ in $J$, we say that $y \leq z$ if $\exists {\mbox{$z'$}}\subset z / y \sim {\mbox{$z'$}}$.*]{}
It is a consequence of the properties of similarity that $\leq$ is a preordering.
It is also a filter. Indeed let $y$ and $z$ in $J$; using the fact that $X$ is H.I., it is possible to find ${\mbox{$y'$}}\subset y$ and ${\mbox{$z'$}}\subset z$ with ${\mbox{$y'$}}\sim {\mbox{$z'$}}$. We have that ${\mbox{$y'$}}\leq y$ and ${\mbox{$y'$}}\leq z$.
#### Definition 3
We define a filter on the set of block subspaces of $[e]$ by letting $Y \leq Z$ if there exists $y \leq z$ with $Y=[y]$, $Z=[z]$.
#### Definition 4
Let $\cal U$ be an ultrafilter on $J$. For $y \in J$, let $B_{y}$ be the Banach space of bounded operators from $[y]$ to $X$, and let ${\mbox{$\|.\|$}}_y$ be the seminorm $\lim_{k \rightarrow +\infty}{\mbox{$\|._{/y^k}\|$}}$.
Let $\cal B$ be the quotient space of $l_{\infty}((B_{y})_{y \in J})$ by the kernel of the seminorm $\lim_{\cal U}{\mbox{$\|.\|$}}_{y}$.
#### Lemma 2
[*Let $y \in J$ and $z \in J$ be similar . The operator $p_{yz}$ from $[y]$ to $[z]$ defined by $\forall i \in \N, p_{yz}(y_i)=z_i$ is bounded and ${\mbox{$\|p_{yz}-I_y\|$}}_y=0 $.*]{}
A morphism from ${\cal L}(Y,X)$ to $\cal B$
===========================================
#### Definition 5
Let $T:Y \rightarrow X$ with $Y=[y]$. For $z \leq y$, let ${\mbox{$y'$}}\subset y$ such that $z \sim {\mbox{$y'$}}$, let $T_z$ be the element $T p_{z{\mbox{${\scriptstyle y}'$}}}$ of $B_z$. Let ${\mbox{$\widetilde{T}$}}$ be the element of $\cal B$ associated to $(T_z)_{z \leq y}$ (the value of $T_z$ for the other values of $z$ has no effect on the value of ${\mbox{$\widetilde{T}$}}$, take for example $T_z=0$).
This definition does not depend on the choice of ${\mbox{$y'$}}$. Indeed, let ${\mbox{$y'$}}_1~\sim~{\mbox{$y'$}}_2$ be two choices of ${\mbox{$y'$}}$; let $T_{iz}$ be associated to ${\mbox{$y'$}}_i$; then $T_{1z}-T_{2z}=T(p_{z{{\scriptstyle}y}'_1}-p_{z{{\scriptstyle}y}'_2})$ so that by Lemma 2, ${\mbox{$\|T_{1z}-T_{2z}\|$}}_z=0$.
It is easy to check that $z \rightarrow {\mbox{$\|T_z\|$}}_{z}$ is increasing, so that ${\mbox{$\|{\mbox{$\widetilde{T}$}}\|$}}$ is a simple limit.
Let $B$ be the set of elements of $\cal B$ of the form ${\mbox{$\widetilde{T}$}}$.
#### Remark 1
If $T:Y \rightarrow X$ and $U:Y \rightarrow X$, then ${\mbox{$\widetilde{T+U}$}}={\mbox{$\widetilde{T}$}}+{\mbox{$\widetilde{U}$}}$.
By Lemma 2, for all $y \sim z$, ${\mbox{$\tilde{p_{yz}}$}}={\mbox{$\tilde{I_y}$}}$.
Lemmas
======
#### Lemma 3
[*Let $T:Y \rightarrow X$ with $Y=[y]$, and let $z \leq y$. Then ${\mbox{$\tilde{T_z}$}}={\mbox{$\widetilde{T}$}}$.*]{}
Let ${\mbox{$y'$}}\subset y$ such that $z \sim {\mbox{$y'$}}$. Let ${\mbox{$z'$}}\subset z$. For $k$ in $\N$, let $t_k=p_{z{\mbox{${\scriptstyle y}'$}}}({\mbox{$z'$}}_k)$. It has the same coordinates on $({\mbox{$y'$}}_i)$ as ${\mbox{$z'$}}_k$ on $(z_i)$, so ${\mbox{$z'$}}\sim |t|$. We have
$$\sum_k |{\mbox{$\|t_k\|$}}-1| \leq \sum_k {\mbox{$\|t_k-{\mbox{$z'$}}_k\|$}}
\leq 4 \sum_i {\mbox{$\|{\mbox{$y'$}}_i-z_i\|$}},$$
so $\sum_k |{\mbox{$\|t_k\|$}}-1| $ converges.
Furthermore, defining $T_{{\mbox{${\scriptstyle z}'$}}}$ as $Tp_{{\mbox{${\scriptstyle z}'$}}|t|}$ and $T_z$ as $Tp_{z{\mbox{${\scriptstyle y}'$}}}$, we get for every $k$:
$${\mbox{$\|(T_{{\mbox{${\scriptstyle z}'$}}}-(T_z)_{{\mbox{${\scriptstyle z}'$}}})({\mbox{$z'$}}_k)\|$}} =
{\mbox{$\|T(|t|_k-t_k)\|$}} \leq
{\mbox{$\|T\|$}} | {\mbox{$\|t_k\|$}}-1 | .$$
These two points imply that ${\mbox{$\|T_{{\mbox{${\scriptstyle z}'$}}}-(T_z)_{{\mbox{${\scriptstyle z}'$}}}\|$}}_{{\mbox{${\scriptstyle z}'$}}}=0$; so ${\mbox{$\tilde{T_z}$}}={\mbox{$\widetilde{T}$}}$.
#### Consequences
If ${\mbox{$y'$}}\subset y$ then ${\mbox{$\widetilde{T_{/{\mbox{${\scriptstyle y}'$}}}}$}}={\mbox{$\widetilde{T}$}}$.
Let ${\mbox{$\widetilde{T}$}}$ and ${\mbox{$\widetilde{U}$}}$ be in $B$, with $T:Y \rightarrow X$ and $U:Z \rightarrow X$; let $t \leq y$ and $t \leq z$; then ${\mbox{$\widetilde{T}$}}+{\mbox{$\widetilde{U}$}}={\mbox{$\widetilde{T}$}}_t+{\mbox{$\widetilde{U}$}}_t={\mbox{$\widetilde{T_t+U_t}$}}$. This proves that $B$ is a linear space.
>From now on, $Y$ (resp. ${\mbox{$Y'$}}$, $Z$, ${\mbox{$Z'$}}$) stands for a block subspace $[y]$ (resp. $[{\mbox{$y'$}}]$, $[z]$, $[{\mbox{$z'$}}]$). Recall that $Y \leq Z$ means $y \leq z$.
#### Proposition 1
*Let $S:Y \rightarrow X$. Then the following properties are equivalent:*
\(a) ${\mbox{$\widetilde{S}$}}={\mbox{$\tilde{0}$}}$.
\(b) $S$ is strictly singular.
\(a) means $ \forall {\mbox{$\epsilon$}}> 0, \exists z \leq y
\ s.t.\ {\mbox{$\|S_{z}\|$}}_{z} \leq {\mbox{$\epsilon$}}$ or equivalently, the assertion: $$(c)\ \forall {\mbox{$\epsilon$}}> 0, \exists {\mbox{$y'$}}\subset y,
\exists k\ s.t.\ {\mbox{$\|S_{{\mbox{${\scriptstyle y}'$}}^k}\|$}} \leq 2{\mbox{$\epsilon$}}.$$ This last assertion implies (b) and (b) implies (c) is shown to be true in [@LT].
#### Lemma 4
[*Let $T:Y \rightarrow X$. Let $Z$ be a block subspace of $X$. There exists ${\mbox{$Y'$}}\subset Y$, ${\mbox{$T'$}}: {\mbox{$Y'$}}\rightarrow Z$ with ${\mbox{$\widetilde{T}$}}={\mbox{$\widetilde{{\mbox{$T'$}}}$}}$.*]{}
First notice that if $S:Y \rightarrow X$ is such that $\Sigma {\mbox{$\|S(y_i)\|$}}$ converges, then ${\mbox{$\widetilde{S}$}}={\mbox{$\tilde{0}$}}$.
Now let $T:Y \rightarrow X$. If ${\mbox{$\widetilde{T}$}}={\mbox{$\tilde{0}$}}$ then one can choose ${\mbox{$T'$}}=0$, ${\mbox{$Y'$}}=Y$. Suppose now ${\mbox{$\widetilde{T}$}}\neq {\mbox{$\tilde{0}$}}$. By the first consequence of Lemma 3 and Proposition 1, passing to a subspace, we may assume that $T$ is an isomorphism.
By projections on big enough intervals, selecting block vectors, and using repeatedly the fact that $X$ is H.I., it is possible to build a normalized block basic sequence ${\mbox{$y'$}}\subset y$ and a (non-normalized) block basic sequence $(z_i)_{i \in \N}$ in $Z$ such that the sequence $\Sigma {\mbox{$\|T({\mbox{$y'$}}_i)-z_i\|$}}$ converges.
The operator ${\mbox{$T'$}}$ defined on $[{\mbox{$y'$}}]$ by ${\mbox{$T'$}}({\mbox{$y'$}}_i)=z_i$ satisfies ${\mbox{$\tilde{{\mbox{$T'$}}}$}}={\mbox{$\widetilde{T}$}}$ by the first remark, and takes its values in $Z$.
A product on $B$
================
#### Definition 6
Let ${\mbox{$\widetilde{T}$}}\in B$ with $T:Z \rightarrow X$. Let ${\mbox{$\widetilde{U}$}}\in B$. Let $V \leq Z$. By Lemma 4, we can assume that ${\mbox{$\widetilde{U}$}}$ is associated to $U:Y \rightarrow V$. We want to define ${\mbox{$\widetilde{T}$}}{\mbox{$\widetilde{U}$}}$ as ${\mbox{$\widetilde{T_v U}$}}$. We need to show that ${\mbox{$\widetilde{T_v U}$}}$ does not depend on the choice of $T$ and $U$.
First, for $i=1,2$, let $V_i \leq Z$, and $U_i:Y_i \rightarrow V_i$ be such that ${\mbox{$\tilde{U_i}$}}={\mbox{$\widetilde{U}$}}$. Let ${\mbox{$z'$}}_i \subset z$ with $v_i \sim {\mbox{$z'$}}_i$. Let ${\mbox{$y'$}}\leq y_1$ and ${\mbox{$y'$}}\leq y_2$. Then $${\mbox{$\widetilde{T_{v_2}U_2}$}}-{\mbox{$\widetilde{T_{v_1}U_1}$}}=
{\mbox{$\widetilde{(T_{v_2}U_2)_{{\mbox{${\scriptstyle y}'$}}}}$}}-{\mbox{$\widetilde{(T_{v_1}U_1)_{{\mbox{${\scriptstyle y}'$}}}}$}}=
{\mbox{$\widetilde{T_{v_2}U_{2{\mbox{${\scriptstyle y}'$}}}-T_{v_1}U_{1{\mbox{${\scriptstyle y}'$}}}}$}}.$$
Now $$T_{v_2}U_{2{\mbox{${\scriptstyle y}'$}}}-T_{v_1}U_{1{\mbox{${\scriptstyle y}'$}}}=
T(p_{v_2 {{{\scriptstyle}z}'}_2} U_{2{\mbox{${\scriptstyle y}'$}}}-p_{v_1 {{{\scriptstyle}z}'}_1} U_{1{\mbox{${\scriptstyle y}'$}}})$$
$$T_{v_2}U_{2{\mbox{${\scriptstyle y}'$}}}-T_{v_1}U_{1{\mbox{${\scriptstyle y}'$}}} =
T[(p_{v_2 {{{\scriptstyle}z}'}_2}-I_{v_2})U_{2{\mbox{${\scriptstyle y}'$}}}
+(U_{2{\mbox{${\scriptstyle y}'$}}}-U_{1{\mbox{${\scriptstyle y}'$}}})+(I_{v_1}-p_{v_1 {{{\scriptstyle}z}'}_1}) U_{1{\mbox{${\scriptstyle y}'$}}}].$$
Using Remark 1, and the fact that the space of strictly singular operators is a two-sided ideal, we get that $${\mbox{$\widetilde{T_{v_2}U_2}$}}-{\mbox{$\widetilde{T_{v_1}U_1}$}}={\mbox{$\tilde{0}$}}.$$
We now prove that ${\mbox{$\widetilde{T_v U}$}}$ does not depend on the choice of $T$. Let for $i=1,2$ $T_i:Z_i \rightarrow X$ be such that ${\mbox{$\tilde{T_i}$}}={\mbox{$\widetilde{T}$}}$. Let $V \leq Z_1$ and $V \leq Z_2$ and $U:Y \rightarrow V$ be a representative for ${\mbox{$\widetilde{U}$}}$. Then ${\mbox{$\widetilde{(T_2)_v U}$}}-{\mbox{$\widetilde{(T_1)_v U}$}}={\mbox{$\widetilde{((T_2)_v-(T_1)_v)U}$}}={\mbox{$\tilde{0}$}}$.
So we can define ${\mbox{$\widetilde{T}$}}{\mbox{$\widetilde{U}$}}$ as ${\mbox{$\widetilde{T_v U}$}}$ without ambiguity.
#### Remark 2
If $T:Z \rightarrow X$ and $U:Y \rightarrow Z$, then ${\mbox{$\widetilde{T}$}}{\mbox{$\widetilde{U}$}}= {\mbox{$\widetilde{TU}$}}$. Using appropriate representatives one can then show that $B$ is an algebra; in particular ${\mbox{$\tilde{1}$}}$ the common value associated to all $I_y$ for $y \in J$ is neutral for the multiplication.
#### Proposition 2
[*$B$ is a field.*]{}
Let ${\mbox{$\widetilde{T}$}}\neq {\mbox{$\tilde{0}$}}$, and let us choose a representative $T:Y \rightarrow Z$ such that $Z$ is a block subspace. By Proposition 1 and the first consequence of Lemma 3, we may assume that $T:Y \rightarrow Z$ is an isomorphism onto. Then ${\mbox{$\widetilde{T}$}}{\mbox{$\widetilde{T^{-1}}$}} = {\mbox{$\widetilde{T T^{-1}}$}}= {\mbox{$\widetilde{I_{/Z}}$}}={\mbox{$\tilde{1}$}}$; in the same way ${\mbox{$\widetilde{T^{-1}}$}} {\mbox{$\widetilde{T}$}}= {\mbox{$\tilde{1}$}}$; so ${\mbox{$\widetilde{T}$}}$ is invertible. As $B$ is not a singleton (${\mbox{$\tilde{1}$}} \neq {\mbox{$\tilde{0}$}}$), $B$ is a field.
$B$ is a Banach algebra.
========================
#### Lemma 5
[*For all ${\mbox{$\lambda$}}\in B$,*]{} $${\mbox{$\|{\mbox{$\lambda$}}\|$}}=
\inf\{{\mbox{$\|T\|$}}: y \in J,\ T \in {\cal L}(Y,X),\ {\mbox{$\widetilde{T}$}}={\mbox{$\lambda$}}\}.$$
If ${\mbox{$\lambda$}}= {\mbox{$\tilde{0}$}}$, the result is clear. Now assume ${\mbox{$\lambda$}}\neq {\mbox{$\tilde{0}$}}$. For every $Y$, if $T$ belongs to ${\cal L}(Y,X)$ then ${\mbox{$\|{\mbox{$\widetilde{T}$}}\|$}} \leq {\mbox{$\|T_y\|$}}_y \leq {\mbox{$\|T\|$}}$, so that ${\mbox{$\|{\mbox{$\lambda$}}\|$}} \leq \inf_{{\mbox{$\widetilde{T}$}}={\mbox{$\lambda$}}}{\mbox{$\|T\|$}}$.
Furthermore, let $T:Z \rightarrow X$ satisfy ${\mbox{$\lambda$}}={\mbox{$\widetilde{T}$}}$. For every ${\mbox{$\epsilon$}}> 0$, there exists ${\mbox{$z'$}}\subset z$ so that ${\mbox{$\|T_{{\mbox{${\scriptstyle z}'$}}}\|$}}_{{\mbox{${\scriptstyle z}'$}}} \leq (1+{\mbox{$\epsilon$}}){\mbox{$\|{\mbox{$\lambda$}}\|$}}$. So there exists $k$ such that ${\mbox{$\|T_{{\mbox{${\scriptstyle z}'$}}^k}\|$}} \leq (1+{\mbox{$\epsilon$}})^2 {\mbox{$\|{\mbox{$\lambda$}}\|$}}$. As ${\mbox{$\widetilde{T_{{\mbox{${\scriptstyle z}'$}}^k}}$}}={\mbox{$\widetilde{T}$}}$, we have $ \inf_{{\mbox{$\widetilde{T}$}}={\mbox{$\lambda$}}} {\mbox{$\|T\|$}} \leq {\mbox{$\|{\mbox{$\lambda$}}\|$}}$.
We say that a representative $T$ for ${\mbox{$\widetilde{T}$}}$ is [*${\mbox{$\epsilon$}}$-minimal*]{} if ${\mbox{$\|T\|$}} \leq (1+{\mbox{$\epsilon$}}){\mbox{$\|{\mbox{$\widetilde{T}$}}\|$}}$. We have proved that if $T$ belongs to ${\cal L}(Y,X)$, then for any $Z \leq Y$ and any ${\mbox{$\epsilon$}}> 0$, we can find an ${\mbox{$\epsilon$}}$-minimal representative for ${\mbox{$\widetilde{T}$}}$ defined on a block subspace of $Z$.
#### Proposition 3
[*The norm on $B$ is an algebra norm.*]{}
Let ${\mbox{$\widetilde{T}$}}$ and ${\mbox{$\widetilde{U}$}}$ be in $B$. Let ${\mbox{$\epsilon$}}> 0$. Let $T:Y \rightarrow X$ be an ${\mbox{$\epsilon$}}$-minimal representative for ${\mbox{$\widetilde{T}$}}$. Choose an ${\mbox{$\epsilon$}}$-minimal representative for ${\mbox{$\widetilde{U}$}}$ from some $Z$ into $Y$ (it is possible by the previous remark and Lemma 4).
Then ${\mbox{$\|{\mbox{$\widetilde{T}$}}{\mbox{$\widetilde{U}$}}\|$}}={\mbox{$\|{\mbox{$\widetilde{TU}$}}\|$}}
\leq {\mbox{$\|TU\|$}} \leq {\mbox{$\|T\|$}}{\mbox{$\|U\|$}}
\leq (1+{\mbox{$\epsilon$}})^2 {\mbox{$\|{\mbox{$\widetilde{T}$}}\|$}}{\mbox{$\|{\mbox{$\widetilde{U}$}}\|$}}$.
As ${\mbox{$\epsilon$}}$ is arbitrary, the norm on $B$ is an algebra norm.
#### Proposition 4
[*$B$ is a Banach space.*]{}
We show that any normally converging series converges in $B$.
Let $({\mbox{$\lambda$}}_n)_{n \in \N}$ be a normally converging series. Let $T_0:Y_0 \rightarrow X$ be a $1$-minimal representative for ${\mbox{$\lambda$}}_0$. Given $T_{n-1}:Y_{n-1} \rightarrow X$, by the remark at the end of Lemma 5, we can find a $1$-minimal representative $T_n:Y_n \rightarrow X$ for ${\mbox{$\lambda$}}_n$ with $Y_n$ a block subspace of $Y_{n-1}$.
Let $y \in J$ such that $\forall n, y_n \in Y_n$. We define an operator $H$ on $Y$ by $H(y_n)=\sum_{k=0}^n T_k(y_n)$. Let us evaluate ${\mbox{$\|(H-\sum_{k=0}^n T_k)_{/y^n}\|$}}$.
Let $u=\sum_{i=n}^{+\infty}a_i y_i$ be a vector in $[y^n]$. Then
$$\left(H-\sum_{k=0}^n T_k \right)(u)=
\sum_{i=n}^{+\infty}
a_i \left(\sum_{k=0}^{i}T_k(y_i)-\sum_{k=0}^{n} T_k(y_i) \right).$$
$$\left(H-\sum_{k=0}^n T_k\right)(u)=
\sum_{i=n}^{+\infty}
a_i \left(\sum_{k=n+1}^{i}T_k(y_i) \right).$$
$$\left(H-\sum_{k=0}^n T_k \right)(u)=
\sum_{k=n+1}^{+\infty} T_k
\left(\sum_{i=k}^{+\infty}a_i y_i \right).$$
So $${\mbox{$\|(H-\sum_{k=0}^n T_k)(u)\|$}}
\leq 3 \left(\sum_{k=n+1}^{+\infty} {\mbox{$\|T_k\|$}}\right) {\mbox{$\|u\|$}}.$$
In particular, $H$ is a bounded operator. Furthermore, $$\lim_{n \rightarrow +\infty}
{\mbox{$\|(H-\sum_{k=0}^n T_k)_{/y^n}\|$}}=0.$$ By Lemma 5, $\lim_{n \rightarrow +\infty}
{\mbox{$\|{\mbox{$\tilde{H}$}}-\sum_{k=0}^{n}{\mbox{$\lambda$}}_k\|$}}=0$. This proves that $\sum_{k=0}^{+\infty} {\mbox{$\lambda$}}_k$ converges in $B$ to ${\mbox{$\tilde{H}$}}$.
Conclusion
==========
The algebra $B$ is a Banach algebra and also a field. By Gel’fand-Mazur theorem, we can identify it to the field of complex numbers.
Now let $T$ belong to ${\cal L}(Y,X)$. Then ${\mbox{$\widetilde{T-{\mbox{$\widetilde{T}$}}I_{/Y}}$}}={\mbox{$\tilde{0}$}}$, that is $T-{\mbox{$\widetilde{T}$}}I_{/Y}$ is strictly singular.
So for any block subspace $Y$ of $[e]$, all operators on $Y$ are of the required form. As $e$ is arbitrary, this is true for any subspace with a basis with constant $2$, and thus for any subspace.
[AA]{}
W.T. GOWERS and B. MAUREY, [*The unconditional basic sequence problem*]{}, J.Amer.Math.Soc. [**6**]{} (1993), 851-874.
J. LINDENSTRAUSS and L. TZAFRIRI, [*Classical Banach spaces I*]{}, Springer-Verlag, New York (1977).
Equipe d’Analyse et de Mathématiques Appliquées,
Université de Marne-la-Vallée,
2, rue de la Butte Verte,
93166 Noisy le Grand Cedex, France.
[email protected].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The finite-temperature magnetic properties of Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys have been investigated. It is shown that the temperature-dependent magnetic behaviour of alloys, composed of originally magnetic and non-magnetic elements, cannot be described properly unless the coupling between magnetic moments at magnetic atoms (Fe,Co) mediated through the interactions with induced magnetic moments of non-magnetic atoms (Pd,Pt) is included. A scheme for the calculation of the Curie temperature ($T_C$) for this type of systems is presented which is based on the extended Heisenberg Hamiltonian with the appropriate exchange parameters $J_{ij}$ obtained from [*ab-initio*]{} electronic structure calculations. Within the present study the KKR Green’s function method has been used to calculate the $J_{ij}$ parameters. A comparison of the obtained Curie temperatures for Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys with experimental data shows rather good agreement.'
author:
- 'S. Polesya$^a$, S. Mankovsky$^a$, O. Sipr$^b$, W. Meindl$^c$, C. Strunk$^c$, H. Ebert$^a$'
title: 'Finite-temperature magnetism of Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys'
---
Introduction
============
Whether a magnetic material is technologically useful or not depends on its properties at finite temperatures. However, the [ *ab-initio*]{} treatment of finite-temperature magnetism remains a challenge despite ongoing progress in this field during the last decades.
In this context, itinerant-electron 3d-transition metals and their alloys receive particular interest. For these systems, finite-temperature magnetic properties cannot be described successfully neither within the collective-electron Stoner model nor within the local-moment model based on the Heisenberg Hamiltonian (for an overview see Refs. [@Kueb00; @Mohn03]). The Stoner model, which treats the transition to the paramagnetic state as vanishing of the local magnetic moments, accounting thereby only for the longitudinal spin fluctuations, grossly overestimates the Curie temperature $T_C$. A bigger success has been achieved using the Heisenberg model, which accounts for temperature-induced transverse spin fluctuations and characterizes the paramagnetic state by orientational disorder within the system of localized magnetic moments. However, a magnitude of the moments is assumed to be unchanged upon fluctuations. In some recent studies the Heisenberg model approach has been combined with [*ab-initio*]{} band-structure calculations, that allow to evaluate the exchange coupling parameters from first principles [@LKAG87]. In this way, trends of the Curie or Néel temperature with composition can be quantitatively described for many systems, in particular, for transition metal monoxides [@FDE09], dilute magnetic semiconductors [@FSK04] or transition metals [@TKDB03].
Despite the rather satisfying results obtained within this combined approach for several systems, the description of finite-temperature magnetism of transition metals and alloys still suffers from many problems owing to the restrictions of the Heisenberg model. For some itinerant-electron systems, e.g. Ni, the thermally-induced longitudinal spin fluctuations play a crucial role in describing properly the temperature-dependent magnetisation and obtaining the correct value for the critical temperature. A phenomenological theory of finite temperature magnetism, which accounts for both types of fluctuations on the same footing, was developed in the past [@Hub79; @Mor85; @Has79; @KMP77]. This theory was used in combination with [*ab-initio*]{} electronic structure calculations to describe temperature dependent magnetic properties of Fe, Co and Ni [@UK96; @RJ97; @RKMJ07]. In this way a much better agreement with experiment, as compared to calculations based on the Heisenberg model, was obtained. In particular, proper accounting for longitudinal fluctuations results in the vanishing of local magnetic moments on Ni atoms above $T_C$ [@RKMJ07], in agreement with experiment.
Other interesting itinerant-electron systems in this context are alloys or compounds composed of originally magnetic and non-magnetic elements. Such systems exhibit the so called covalent magnetism [@WZM+81; @MS93], where magnetisation of the ’non-magnetic’ atoms is governed by the spontaneously magnetised atoms via the strong spin-dependent hybridisation of their electronic states. The Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys considered in the present work belong to this type of systems. To describe the temperature-dependent magnetism of such systems on the basis of Heisenberg model, one obviously has to account properly for the behavior of the Pd/Pt sub-lattices. Only a few [*ab-initio*]{} studies of finite-temperature magnetism of systems of this kind have been done so far. Similarly to the work mentioned above [@UK96; @RJ97; @RKMJ07] these studies were based on a generalisation of the classical Heisenberg Hamiltonian in one or another way to account for the different character of magnetism on different types of atoms. Mryasov et al. [@MNG05] have investigated the compound FePt and showed that the anomalous temperature-dependence of its magneto-crystalline anisotropy energy (MAE) is due to the induced Pt magnetic moments. In another study of these authors [@Mry05], a crucial role of the magnetic moment induced on Rh was demonstrated for the stabilisation of the ferromagnetic state of FeRh and for the control of the antiferromagnet-ferromagnet phase transition. Lezaic et al. [@LME+06] emphasized the need to account for longitudinal fluctuations of magnetic moments induced on Ni atoms for a proper description of the temperature-dependence of the spin-polarisation at the Fermi energy $E_{F}$ in the half-metallic ferromagnet NiMnSb. Sandratskii et al. [@SSS07] investigated several ways for accounting for the induced magnetic moments within the spin-spiral approach used for the calculation of the exchange coupling parameters in NiMnSb and MnAs. Using these results Sandratskii et al. [@SSS07] have studied finite-temperature magnetic properties of NiMnSb; their findings are consistent with the findings of Lezaic et al [@LME+06].
In this work we introduce an [ *ab-initio*]{} method to describe finite-temperature magnetism of systems with spontaneous and induced magnetic moments. The method is based on an extension of the Heisenberg Hamiltonian by adding a term which describes the induced magnetic moments within the linear response formalism. Our approach relies on a combination of [*ab-initio*]{} band-structure calculations with Monte Carlo (MC) simulations based on the extended Heisenberg model. To test this approach, we investigate finite-temperature magnetic properties of Pd-rich Fe$_x$Pd$_{1-x}$ alloys, with Fe concentrations up to 20 at.%, as well as of ordered and disordered Co$_{3}$Pt, CoPt and CoPt$_{3}$ alloys that are interesting both for fundamental reasons and for possible use in industrial applications because of their high magnetic anisotropy [@CHZ+00; @MLR+08]. Theoretical investigations of finite-temperature magnetism of these systems failed so far to reproduce experimental results with satisfactory accuracy [@KGS99]. We demonstrate in the present work that a combination of [*ab-initio*]{} band-structure calculations with the Monte Carlo (MC) simulations based on the extended Heisenberg model gives satisfying agreement between theoretical and experimental values of critical temperatures. We found that despite their small magnitudes, the moments induced on non-magnetic atoms (Pd, Pt) have an important influence on finite-temperature magnetic order.
Experimental detailes
=====================
The Fe$_x$Pd$_{1-x}$ films were thermally evaporated onto oxidized silicon substrates from separate effusion cells for Pd and Fe in an ultra high vacuum system (base pressure $5 \dot 10^{-11}$ mbar). The film thicknesses were between 15 and 20 nm. The deposition rate of the two components could be controlled independently, resulting in an accuracy of the Fe concentration of $1\%$. Auger spectroscopy on a sample with nominally $7\%$ Fe content provided an independent value of $7.8\%$ Fe for this film. After deposition, the films were patterned into a six-terminal Hall-bar geometry. Measurements of the anomalous Hall effect at temperatures between 2K and 300 K provided the magnetization $M(T)$, from which the Curie temperatures of the films were deduced.
Theoretical approach
====================
Ground-state calculations
-------------------------
Within the present work, spin-polarised electronic structure calculations for the ground-state have been performed using the multiple scattering KKR (Korringa-Kohn-Rostoker) Green’s function method [@Ebe00] in the scalar-relativistic approximation. The local spin density approximation (LSDA) for density functional theory was used with the parametrisation for the exchange-correlation potential due to Vosko, Wilk, and Nusair [@VWN80]. The potential was treated within the atomic sphere approximation (ASA) with the radii of the spheres around Fe/Co and Pd/Pt sites chosen by requiring the ratios of the corresponding volumes to be the same as for the pure elements. For the angular momentum expansion of the Green’s function a cutoff of $l_{max} = 3$ was applied. For substitutionally disordered alloys, the self-consistent coherent potential approximation (CPA) method was employed. A geometry optimisation was performed, i.e., the lattice constants of the alloys have been obtained by minimisation of the total energy.
Extended Heisenberg Hamiltonian
-------------------------------
The finite temperature properties of the investigated systems were studied by Monte Carlo simulations based on the Heisenberg model with the underlying Hamiltonian given by
$$\begin{aligned}
\label{Heisenberg}
H_{ex} &=& -\sum_{ij}\tilde{J}^{M-M}_{ij}\vec{M}_i\vec{M}_j
-\sum_{ij}\tilde{J}^{M-m}_{ij}\vec{M}_i\vec{m}_j \nonumber \\
&& -\sum_{ij}\tilde{J}^{m-m}_{ij}\vec{m}_i\vec{m}_j\;.\end{aligned}$$
Here the classical Hamiltonian was generalised to allow an application to itinerant-electron systems consisting of magnetic and non-magnetic atoms having magnetic moments $M_i$ and $m_i$, respectively, connected with corresponding exchange coupling parameters. The dependence of the induced magnetic moments $m_i$ on a specific magnetic configuration is treated via linear response formalism (for details see the Appendix).
We suppose that the induced magnetic moments on Pd or Pt atoms are governed only by the magnetic moments of the Fe or Co atoms in Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys, respectively, arranged in the first neighbor shell around the non-magnetic atom, so that
$$\begin{aligned}
\label{A3}
\vec{m}_i &=& \sum_{j\in M} X^{m-M}_{ij}\vec{M}_j = X_i^{m-M} \sum_{j\in M} \vec{M}_j \;.\end{aligned}$$
The notation $\sum_{j\in M}$ means that the sum includes only such terms where $j$ is a site with an inducing magnetic moment.
In this case (see Appendix) the susceptibility $X^{m-M}_{ij}$ can be approximated by the value found for the ground state of a system with well defined collinear spontaneous (Fe, Co) and induced (Pd, Pt) magnetic moments:
$$\begin{aligned}
\label{A4}
X_i^{m-M} &=& \frac{m_i}{\sum_{j\in M}{M}_j}\;.\end{aligned}$$
The exchange coupling parameters $\tilde{J}_{ij}$ between atoms $i$ and $j$ in the Eq. (\[Heisenberg\]) were obtained via the formula of Lichtenstein et al. [@LKAG87]:
$$\label{eq_jij}
J_{ij} \: = \: -\frac{1}{4\pi} \, \mbox{Im} \int^{E_F} \mathrm{d} E\; \mathrm{Tr} \:
(t^{-1}_{i\uparrow} -t^{-1}_{i\downarrow}) \,
\tau^{ij}_{\uparrow} \, (t^{-1}_{j\uparrow} -t^{-1}_{j\downarrow})
\, \tau^{ji}_{\downarrow} \; .$$
with the relation
$$\label{eq_jij1}
\tilde{J}_{ij} = \frac{J_{ij}}{|\vec{m}_i||\vec{m}_j|} \; .$$
In Eq. (\[eq\_jij\]), $t_{i\sigma}$ and $\tau^{ij}_{\sigma}$ are the spin ($\sigma$) and site ($i, j$) dependent single-site and scattering path operator matrices occurring within the KKR formalism [@Wei90].
Evaluation of $T_{C}$ {#meth-Ac}
---------------------
The Curie temperature $T_C$ was evaluated with the Monte Carlo (MC) method [@Bin97] using the standard Metropolis importance sampling algorithm [@LB00], on the basis of the model Hamiltonian in Eq. (\[Heisenberg\]). The number of atoms in the MC unit cell for different concentrations was taken between 1728 and 4000 and $T_C$ was determined from the peak position of the temperature-dependent susceptibility. The Fe/Co magnetic moments were treated during MC simulation as localised and changed only their orientation. On the other hand, the magnetic moments on Pd/Pt atoms could change their absolute value as well as the orientation in accordance with the changed magnetic configuration around these atoms. Eq. (\[A3\]) implies that the magnetic moments on Pd/Pt are proportional to the vector sum of magnetic moments at neighboring Fe/Co atoms with only nearest neighbors taken into account. This means that each MC step consists of 1) change of the orientation of a magnetic moment on the Fe/Co atoms, 2) search for all nearest neighbour Pd/Pt atoms and calculation of the orientation and absolute value of their moments using the susceptibilities $X^{m-M}$ via Eq. (\[A4\]). The change of the energy of the total system is due to both effects. For disordered alloys the resulting $T_C$ values in addition were averaged over up to 20 different configurations.
Obviously, the approach described above, accounts for the contribution of spin polarised ’non-magnetic’ atoms to the exchange interactions between the magnetic atoms. As this contribution is temperature-dependent, it allows a corresponding description of the temperature dependent magnetisation. In particular, it accounts for longitudinal spin fluctuations occurring on the ’non-magnetic’ sub-lattice. Below it will be shown that this rather simple scheme gives rather good agreement with experimental data for the systems under consideration.
Results for Fe$_x$Pd$_{1-x}$ alloys
===================================
[*Ab-initio*]{} calculations
----------------------------
The scheme for calculation of temperature dependent magnetic properties, described above, was used to investigate disordered Fe$_x$Pd$_{1-x}$ alloys with Fe concentration up to 20 at.%. The exchange coupling parameters $J^{Fe-Fe}_{ij}$ and $J^{Fe-Pd}_{ij}$, shown in Fig. \[Fig1\], have a similar dependency on the distance $R_{ij}$ for all investgated alloys. In the Pd reach limit (Fe concentration $x < 0.2$) the exchange coupling parameters $J^{Fe-Fe}_{ij}$ corresponding to the average distance between magnetic atoms is rather small and does not allow to create long-range magnetic order in the system, as was demonstrated by corresponding restricted MC simulations.
![\[Fig1\] Exchange coupling parameters $J^{Fe-Fe}$ (a) and $J^{Fe-Pd}$ (b) for Fe$_{x}$Pd$_{1-x}$ alloys at different concentrations. The dashed lines represent results for ordered FePd$_3$ (note that in this case for the Fe atoms the nearest neighbour interaction $J^{Fe-Fe}$ is absent because there are only Pd nearest neighbours). ](Jij_fefe_diso.eps "fig:"){width="7.5cm"}a ![\[Fig1\] Exchange coupling parameters $J^{Fe-Fe}$ (a) and $J^{Fe-Pd}$ (b) for Fe$_{x}$Pd$_{1-x}$ alloys at different concentrations. The dashed lines represent results for ordered FePd$_3$ (note that in this case for the Fe atoms the nearest neighbour interaction $J^{Fe-Fe}$ is absent because there are only Pd nearest neighbours). ](Jij_fepd_diso.eps "fig:"){width="7.5cm"}b
![\[Fig2\] Results of spin-spiral calculations for ordered FePd$_3$ and disordered Fe$_{0.25}$Pd$_{0.75}$: (a) the energy of spin spirals as a function of the wave vector $q = \frac{\pi}{a}(0,0,q_z)$, obtained for the disordered alloy (full circles), for the ordered compound with non-zero Pd magnetic moments (full squares) and for the ordered compound with Pd magnetic moments equal to 0 (open squares); (b) - magnetic moment of inequivalent Pd atoms in ordered FePd$_3$ as a function of the wave vector $q_z$. Fe and Pd2 occupy the sites $(0,0,0)$ and $(\frac{1}{2},\frac{1}{2},0)$, respectively, Pd1 atoms occupy the sites $(0,\frac{1}{2},\frac{1}{2})$ and $(\frac{1}{2},0,\frac{1}{2})$. ](Fe_0.25Pd_0.75_Spin_Spiral_CMP_ordered_vs_disordered.eps "fig:"){width="7.5cm"} ![\[Fig2\] Results of spin-spiral calculations for ordered FePd$_3$ and disordered Fe$_{0.25}$Pd$_{0.75}$: (a) the energy of spin spirals as a function of the wave vector $q = \frac{\pi}{a}(0,0,q_z)$, obtained for the disordered alloy (full circles), for the ordered compound with non-zero Pd magnetic moments (full squares) and for the ordered compound with Pd magnetic moments equal to 0 (open squares); (b) - magnetic moment of inequivalent Pd atoms in ordered FePd$_3$ as a function of the wave vector $q_z$. Fe and Pd2 occupy the sites $(0,0,0)$ and $(\frac{1}{2},\frac{1}{2},0)$, respectively, Pd1 atoms occupy the sites $(0,\frac{1}{2},\frac{1}{2})$ and $(\frac{1}{2},0,\frac{1}{2})$. ](FePd3_Pd_mmom_vs_q_spinspiral.eps "fig:"){width="7.5cm"}
According to the experimental findings [@CS65], the alloys exhibit ferromagnetic order for Fe concentrations above 0.1%. In addition, previous experimental [@Ger58; @CWK65; @CMS65; @Lon68; @CT71; @Cra60; @BCTS75; @AWF+88] and theoretical [@Kim66; @Ber81; @MP82; @MS93; @JHZQ93; @OZD86] investigations on the magnetic properties of Fe$_x$Pd$_{1-x}$ alloys have shown a strong host polarisation by magnetic Fe impurities leading to a giant magnetic moment per impurity atom, up to $12.9 \mu_B$. A wide-spread regime of magnetised Pd atoms leads to ferromagnetic order in diluted Fe$_x$Pd$_{1-x}$ alloys despite a large distance between the magnetic Fe atoms.
The crucial role of the induced magnetic moment on Pd for the Fe-Fe exchange interactions can be demonstrated by an analysis of the energy of spin spirals as a function of a wave vector, shown in Fig. \[Fig2\] for the disordered alloy Fe$_{0.25}$Pd$_{0.75}$ in comparison with the results for the ordered compound FePd$_{3}$. The calculations have been performed for spin spirals along the $z$ direction with the Fe magnetic moments tilted by $90^o$ with respect to the $z$-axis. For the ordered system an increase of the wave vector of the spin spirals is accompanied first by an increase in energy reflecting the stability of the ferromagnetic order in the system. A further increase of the wave vector above $q_z > \pi/2a$, leads to a decrease of the energy of the spin spiral (Fig. \[Fig2\]a). This behavior is governed by a decrease of the Pd magnetic moments at these wave vectors (see Fig. \[Fig2\]b), that diminishes their role in the Fe-Fe exchange.
The role of Pd becomes clearly visible for the spin spirals in FePd$_{3}$ with the Fe-Pd exchange interactions being suppressed. This can be achieved by forcing the Pd induced magnetic moments to be perpendicular to the Fe magnetic moments, and therefore to be equal to 0 (see Appendix). In this case the minimum of the spin-spiral energy corresponds to an AFM state, i.e. at $q = \pi/a$, that originates from Fe-Fe exchange interaction (see Fig. \[Fig1\]). In the case of the disordered alloy, the dependence of the spin-spiral energy on the wave vector is different because the random distribution of the Fe atoms allows Fe atoms to be nearest neighbours with a strong FM interaction. Due to this, the system retains the FM order at all values of wave vector $\vec{q}$.
![\[Fig\_DOS\] a) Pd DOS in Fe$_x$Pd$_{1-x}$ for $x = 0.003$ and $x =
0.15$ for spin-up (upper panel) and spin-down (lower panel) states. Ground state characteristics of Fe$_x$Pd$_{1-x}$ alloys vs. Fe concentration: b) density of states of Pd at the Fermi level; c) Pd spin magnetic moments; d) Fe spin magnetic moments. ](CMP_DOS.eps "fig:"){width="6.cm"} ![\[Fig\_DOS\] a) Pd DOS in Fe$_x$Pd$_{1-x}$ for $x = 0.003$ and $x =
0.15$ for spin-up (upper panel) and spin-down (lower panel) states. Ground state characteristics of Fe$_x$Pd$_{1-x}$ alloys vs. Fe concentration: b) density of states of Pd at the Fermi level; c) Pd spin magnetic moments; d) Fe spin magnetic moments. ](FePd_ground_state.eps "fig:"){width="8.5cm"}
While the ground-state magnetic properties of Fe$_x$Pd$_{1-x}$ alloys can essentially be understood on the basis of [*ab-initio*]{} electronic structure calculations, a theoretical description of finite-temperature properties faces many difficulties due to the itinerant-electron nature of magnetism. Theoretical investigations based on the results of [*ab-initio*]{} electronic structure calculations have been performed for example by Mohn and Schwarz [@MS93]. They used the model approach formulated by Bloch et al. [@BESV75] to describe the magnetic behavior of a system characterised by the coexistent local- and itinerant-electron magnetism. Within this approach the system is characterised by two interacting subsystems: (i) one having local magnetic moments showing a Curie-Weiss-like behaviour and (ii) an itinerant electron subsystem magnetically polarised by the effective Weiss field, with the corresponding parameters found by [*ab-initio*]{} electronic structure calculations.
The spin moment on every Pd atom is induced by the magnetic moment of the Fe atoms and all surrounding induced Pd magnetic moments (see Appendix). The rather big absolute value and large region of Pd induced magnetic moments around an Fe atom is a result of the high magnetic spin susceptibility of pure Pd and Fe$_x$Pd$_{1-x}$ alloys with small Fe concentration, that is determined by a large Pd density of states (DOS) at the Fermi level, $n(E_F)$ (see Fig. \[Fig\_DOS\]a). In turn, $n(E_F)$ decreases with the increase of Fe content in the alloy (Fig. \[Fig\_DOS\]b), resulting in a decrease of the partial magnetic susceptibility of the Pd atoms. Thus, at very small Fe concentrations the induced Pd spin moment can extend to big distances - inducing shell-by-shell a spin moment in the Pd subsystem. This polarisation mechanism decays with the distance from the magnetic impurity. When the Fe concentration increases, the regions with the induced moments overlap and, as was pointed in M. Shimizu et al. [@SK68; @TS65], when the Fe concentration is larger than 0.1% the Pd magnetic properties can be described well by band calculations, as done in our present calculations using the CPA alloy theory.
Figure \[Fig\_DOS\], c and d, represent the spin magnetic moments of Pd and Fe versus the Fe content in Fe$_x$Pd$_{1-x}$. As is seen, the Fe magnetic moments change only slowly with the increase of Fe concentration, while the variation of the Pd spin magnetic moments is rather pronounced. This can be directly connected to the decrease of the Pd DOS at the Fermi level.
![\[Clu\_siscept\] Magnetic moment distribution in Pd around a single Fe impurity, as a function of the distance from the magnetic atom. Full squares represent the Pd magnetic moments self-consistently obtained within a cluster with 3 atomic shells of Pd around an Fe impurity embedded into a Pd host, full circles give the results obtained for the same system but with the exchange potential switched off. The open circles represent the induced magnetic moments in Pd calculated within linear response formalism using Eq. (\[Delta\_m\]) in the Appendix. ](mmom_Fe_clu_in_Pd_mod.eps){width="8.5cm"}
For a more detailed analysis, we investigated the properties of the induced Pd magnetic moment using the [*ab-initio*]{} calculations. In particular, we studied the distribution of the Pd magnetic moment in a Pd host in the limit of very small Fe concentrations, i.e. around a single Fe impurity. To see that the induced magnetic moment at every Pd atom is determined not only by the Fe magnetic moment but also by the surrounding Pd magnetic moments, one can compare the un-enhanced induced Pd spin moments created by only one Fe atom with the total induced spin moments in Pd. The total induced magnetic moment distribution in Pd can be found by solving the system of equations (\[Minduced2\]) within a selected region around an Fe atom (see Appendix). In addition, in the present work the moment distribution has been obtained by self-consistent electronic structure calculations instead of using linear response formalism. Fig. \[Clu\_siscept\] shows the slow decay of the induced magnetic moment with the distance (full squares). The corresponding un-enhanced induced moment in these calculations has been obtained by suppressing the effective exchange B-field for the Pd atoms during the SCF-cycle. These un-enhanced magnetic moments compare very well with those obtained from linear response formalism (Eq. (\[Delta\_m\])) (open circles). These are shown in Fig. \[Clu\_siscept\] also for larger distances. One can see that the decrease with the distance of the Pd un-enhanced spin moment is very fast compared to the enhanced one. For the nearest Pd neighbors of Fe atom these values differ approximately by a factor of 2, while the difference for the next nearest neighbors is already an order of magnitude. The local exchange enhancement, well approximated within the linear approach at small values of the induced magnetic moments, should keep the ratio of these two values approximately constant. The obtained results give evidence for a more complicated picture of the creation of the induced magnetic moment in accordance to the description given in the Appendix.
The effect of temperature-induced magnetic disorder within the Fe-subsystem was analysed within [*ab-initio*]{} calculations, describing magnetic disorder within the uncompensated Disordered Local Moment (DLM) approximation. Using this approximation an effective alloy of two types of Fe atoms with opposite spin directions and having different concentrations is treated using the CPA alloy theory. In this way one can study the dependence of the induced magnetic moment of individual Pd atoms on the average magnetic moment in the system. Fig. \[M\_Pd\_vs\_M\_Fe\] shows that the magnetic disorder in the Fe subsystem (assumed to be temperature induced) is accompanied by a decease of the total magnetic moment in the system and results in a decease of the induced magnetic moment in the Pd subsystem. One can see a rather good linear dependence of the induced Pd magnetic moment as a function of the magnetic moment of the Fe subsystem, for nearly all Fe concentrations. Only in the limiting case of low Fe concentration (1%), a noteworthy deviation from linear behavior is observed. This deviation will influence the final results in a Curie temperature evaluation correspondingly.
![\[M\_Pd\_vs\_M\_Fe\] Results of [*ab-initio*]{} calculations for induced Pd moment in three Fe-Pd alloys, using the uncompensated DLM approximation as a function of the Fe average magnetic moment. ](Pd_mmom_vs_Fe_mmom.eps){width="8.5cm"}
Finite-temperature magnetism of Fe$_x$Pd$_{1-x}$ alloys
-------------------------------------------------------
The temperature dependent magnetic properties of Fe$_x$Pd$_{1-x}$ alloys were investigated by performing Monte Carlo simulations. They show an absence of an ordered FM state for the alloys with Fe concentration up to $x \approx
0.2$ if the Fe-Fe exchange interactions mediated by Fe-Pd interactions are neglected. This clearly demonstrates the importance of these interactions. The Curie temperature for disordered Fe$_{0.2}$Pd$_{0.8}$ alloy in this case is around 60K, much lower than observed experimentally (around 400 K). To illustrate the role of Fe-Pd interactions in the formation of magnetic order, Fig. \[T\_C2\] shows spin configurations obtained within MC simulations for ($T = 0.1 K$) for the Fe$_x$Pd$_{1-x}$ alloy with $x = 0.02$. Fig. \[T\_C2\]a represents a result for the case that the Fe-Pd exchange interactions are neglected and therefore it shows only the magnetic moments of Fe. Fig. \[T\_C2\]b represents the spin configuration when the Fe-Pd first-neighbour interactions are taken into account and shows only those Fe and Pd magnetic moments which give a contribution to the total energy of a system given by Eq. (\[Heisenberg\]).
A comparison of $T_C$ obtained within the MC simulation based on the Hamiltonian in Eq. (\[Heisenberg\]) with the experimental data for Fe$_x$Pd$_{1-x}$ alloys is shown in Fig. \[T\_C\]. Obviously, a rather good agreement is obtained for the whole concentration range. It should be emphasized once more that all parameters for the model Hamiltonian (Eq. (\[Heisenberg\])) are obtained within [*ab-initio*]{} electronic structure calculations using the scheme described in the Appendix. Of course, going beyond the various approximations the final results can be improved to get better agreement with the experimental results. Fig. \[T\_C\] shows that the theoretical result obtained by Mohn and Schwarz [@MS93] for Fe$_x$Pd$_{1-x}$ alloys at small Fe concentrations is also in a good agreement with experiment. However, it should be emphasized that this work is based on a semi-phenomenological approach.
![\[T\_C2\] The distribution of spin moments in unit cell used within the MC simulations without (a) and with (b) the induced Pd moments taken into account. Large arrows correspond to the Fe magnetic moments, small arrows to Pd magnetic moments. ](Fe_2_T0_conf.eps "fig:"){width="3.5cm"}a ![\[T\_C2\] The distribution of spin moments in unit cell used within the MC simulations without (a) and with (b) the induced Pd moments taken into account. Large arrows correspond to the Fe magnetic moments, small arrows to Pd magnetic moments. ](FePd_2_T0_conf.eps "fig:"){width="3.5cm"}b
![\[T\_C\] (a) The Curie temperature for different Fe concentrations in Fe$_x$Pd$_{1-x}$ alloys: present results vs theoretical results of Mohn and Schwarz [@MS93] and experimental data: Expt. 1 – present work, Expt. 2 [@CWK65], Expt. 3 [@Cra60], Expt. 4 [@YCTF75]. ](FePd_Tc.eps){width="8.5cm"}
Results for Co$_x$Pt$_{1-x}$ alloys
===================================
Magnetic moments and exchange coupling constants
------------------------------------------------
---------------------- ------------ ------ ----------------------------- -----------------------------
system $a$ $ \mu_{\mathrm{spin}}$ (Co) $ \mu_{\mathrm{spin}}$ (Pt)
a.u. $\mu_{B}$ $\mu_{B}$
Co$_{3}$Pt ordered 6.98 1.82 0.36
disordered 7.06 1.88 0.25
\[0.5ex\] CoPt ordered 7.23 1.93 0.37
disordered 7.23 2.03 0.27
\[0.5ex\] CoPt$_{3}$ ordered 7.31 1.76 0.26
disordered 7.37 2.19 0.25
---------------------- ------------ ------ ----------------------------- -----------------------------
: Equilibrium lattice constants and magnetic moments at Co and Pt atoms for ordered and disordered Co$_x$Pt$_{1-x}$ alloys.[]{data-label="tab-mom"}
The calculated equilibrium lattice constants for ordered and disordered Co$_x$Pt$_{1-x}$ alloys are shown in Tab. \[tab-mom\]. For ordered CoPt, we used a simplified $L1_{0}$ geometry, assuming $c$=$a$ instead of $c$=0.98$a$ found in experiment. The magnetic moments of the Co and Pt atoms for the equilibrium lattice constants obtained in the scalar-relativistic mode are presented in Table\[tab-mom\]. More details about ground-state properties of Co-Pt can be found in our earlier study [@SMME08].
The exchange coupling constants $J_{ij}$ for the investigated systems were evaluated via Eq. (\[eq\_jij\]). The dependence of $J_{ij}$ on the distance between the atoms $i$ and $j$ is displayed in Fig. \[fig-jxc\] with the left panels showing the situation when both $i$ and $j$ are Co atoms and the right panels showing the situation when $i$ is a Co atom and $j$ is a Pt atom. If experimental lattice constants were used instead of equilibrium lattice constants, the $J_{ij}$ constants would change slightly but both the trends and the values would remain similar as in Fig. \[fig-jxc\].
One can see from Fig. \[fig-jxc\] that for Co$_{3}$Pt and CoPt the coupling between the moments on the Co atoms do not differ very much from the results for their disordered counter parts, Co$_{0.75}$Pt$_{0.25}$ and Co$_{0.50}$Pt$_{0.50}$ respectively. However, the situation changes dramatically for CoPt$_{3}$ (lower left panel in Fig. \[fig-jxc\]). The pronounced difference between the data for the ordered and the disordered system stems mainly from the fact that there are no Co atoms present for some coordination spheres around a central Co atom in ordered CoPt$_{3}$. Concerning the coupling between moments on Co and Pt atoms, the degree of long-range order has a larger influence than for the Co–Co coupling. For ordered alloys, the $J_{ij}$ constants significantly vary also with composition. For disordered alloys, on the other hand, the $J_{ij}$ constants do not vary very much with composition. For ordered CoPt$_{3}$, there is a surprisingly strong Co-Co coupling between atoms which are 2.83 $a$ apart. We verified that for larger distances no comparable strong coupling occurs.
Curie temperatures
------------------
---------------------- --------------------- --------------- ---------------
system model $T_{C}$ \[K\] $T_{C}$ \[K\]
ordered disordered
fcc Co MC 1100
experiment 1388
Mohn-Wohlfarth 3523
\[0.5ex\] Co$_{3}$Pt MC, Co-Co only 800 750
MC, Co-Co and Co-Pt 900 880
experiment – 1100
Mohn-Wohlfarth 1803 1120
\[0.5ex\] CoPt MC, Co-Co only 360 620
MC, Co-Co and Co-Pt 620 760
experiment 727 830
Mohn-Wohlfarth 1964 850
\[0.5ex\] CoPt$_{3}$ MC, Co-Co only $-$180 370
MC, Co-Co and Co-Pt 150 520
experiment 288 468
Mohn-Wohlfarth 241 510
---------------------- --------------------- --------------- ---------------
: Curie temperatures $T_{C}$ for fcc Co and for ordered and disordered Co$_x$Pt$_{1-x}$ alloys. Experimental results [@Dah85; @CDK86] are shown together with results of our Monte-Carlo calculations with either both Co-Co and Co-Pt coupling or with only Co-Co coupling included. For comparison, results obtained by relying on the Mohn-Wohlfarth theory are also shown [@QSC94; @KGS99; @GCS01]. Negative $T_{C}$ implies antiferromagnetic ordering.[]{data-label="tab-tc"}
The Curie temperatures $T_{C}$ of the investigated Co-Pt systems evaluated by means of the Monte-Carlo technique are shown in Tab. \[tab-tc\]. In addition, results for fcc Co are given in this table. The two theoretical Curie temperatures correspond to two different Hamiltonians used to describe the magnetic coupling. The first $T_{C}$ (denoted as “Co-Co only” in Tab. \[tab-tc\]) corresponds to the standard Heisenberg Hamiltonian for magnetic moments only on the Co atoms. The second $T_{C}$ (denoted as “Co-Co and Co-Pt”) corresponds to the extended Heisenberg Hamiltonian (Eq. (\[Heisenberg\])), accounting for the coupling between moments on Co atoms $\tilde{J}^{Co-Co}_{ij}$ as well as for the coupling between moments on Co and Pt atoms $\tilde{J}^{Co-Pt}_{ij}$, with the moments on Pt atoms determined via Eqs. (\[A3\]) and (\[A4\]). The values of $T_{C}$ obtained earlier by relying on the semi-empirical Mohn-Wohlfarth theory [@MW87] were taken from Qi et al. [@QSC94] for fcc Co, from Kashyap et al. [@KGS99] for ordered Co-Pt compounds and from Ghosh et al. [@GCS01] for disordered Co$_x$Pt$_{1-x}$ alloys. Experimental values [@Dah85; @CDK86] are shown for comparison. Note that an experimental value for $T_{C}$ for ordered Co$_{3}$Pt is not available because the ordered phase is not stable for this composition.
As is seen, the Mohn-Wohlfarth theory [@MW87] gives reasonable agreement with experiment at small Co concentration. However, increasing the Co content leads to large discrepancies between the theory and experiment. This results from the limitations of the Mohn-Wohlfarth theory: it was developed for homogeneous itinerant-electron systems, which is not the case for Co$_x$Pt$_{1-x}$ alloys.
Our [*ab-initio*]{} scheme based on an extended Heisenberg Hamiltonian (both Co-Co and Co-Pt coupling included) accounts quantitatively for the trends of $T_{C}$ with the composition and with the degree of long-range order. If the coupling mediated via moments at Pt atoms is not included, the results are unrealistic. This is especially true for ordered CoPt$_{3}$, where an [*antiferromagnetic*]{} order is established at finite temperatures (reflected by a negative value for $T_{C}$) if the coupling between moments on Co and Pt is ignored.
Conclusions
===========
As was shown in the present work by the examples of Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys, the finite-temperature magnetism of alloys composed of magnetic and non-magnetic elements requires to account for the exchange interactions between magnetic atoms, mediated by the exchange interaction with non-magnetic atoms. This inplies in particular that one has to account properly for the induced magnetic moment within the Monte Carlo simulations which are based in the present work on a corresponding extension of the standard Heisenberg Hamiltonian. The approach presented suggests to describe the induced magnetic moment on non-magnetic atoms within linear response formalism, being proportional to the vector sum of magnetic moments of neighboring magnetic atoms. This ansatz allows for substantial technical simplifications and leads to substantial improvement of the results when compared to simpler schemes.
The finite-temperature calculations for Fe$_x$Pd$_{1-x}$ and Co$_x$Pt$_{1-x}$ alloys performed within this approach give the dependence of $T_{C}$ on the composition as well as on the degree of long-range order in good agreement with experimental data. The case of ordered CoPt$_{3}$ also demonstrates that even if the coupling between nearest inducing moments is antiferromagnetic, the magnetic order can still be ferromagnetic due to the effect of coupling between inducing and induced moments. A mere inspection of the $J_{ij}$ constants thus cannot serve as a reliable indicator of ferromagnetic or antiferromagnetic order at $T \neq 0K$.
Acknowledgement
===============
This work was supported by the DFG within the SFB 689 “Spinphänomene in reduzierten Dimensionen” as well as the project Eb 154/20 “Spin polarisation in Heusler alloy based spintronics systems probed by SPINAXPES”, and by GA CR within the project 202/08/0106.
Induced magnetic moments {#sec:App}
========================
The magnetic moment induced on site $i$ (Pd or Pt) by the exchange field $\vec{B}^{xc}_j$ due to magnetic moments at site $j$ can be calculated within the linear response formalism using the expression [@DFVE01]:
$$\begin{aligned}
\label{Delta_m}
\hspace{-1cm}
\vec{m}^0_i(\vec{r}) & = & - \frac{1}{\pi} Im \int^{E_F} dE Tr
\vec{\sigma} \nonumber \\
& &\times \int_{\Omega_j} d^3r'
G(\vec{r},\vec{r}\,',E) H_j^{xc}
(\vec{r}\,')G(\vec{r}\,',\vec{r},E) \nonumber \\
& = & \int_{\Omega_j} \chi_{ij}(\vec{r},\vec{r}\,') \vec{B}^{xc}_j(r')d^3r'\end{aligned}$$
Here $H^{xc}_j(\vec{r}\,') = \vec{\sigma}\vec{B}^{xc}_j(\vec{r}\,') =
\vec{\sigma}\vec{e}_M B^{xc}_j(\vec{r}\,')$ with $\vec{\sigma}$ - the matrix of Pauli matrices [@Ros61], $\vec{e}_M$ the unit vector in the direction of spontaneous magnetic moment of atom $j$, $B^{xc}_j(\vec{r})$ the local exchange field at the site $j$. Note that, neglecting relativistic effects, the magnetic moment $\vec{m}^0_i$ induced by a neighboring magnetic atom is parallel to the direction $\vec{e}_M$ of the magnetic moment $\vec{M}_j$ of this atom.
The total induced magnetic moments on a Pd or Pt atom is represented as a response to the exchange field of all surrounding atoms by the following expression:
$$\begin{aligned}
\label{Minduced}
\vec{m}_i(\vec{r}) & = & \sum_{j\in M} \int_{\Omega_j^M} \chi^{m-M}_{ij}(\vec{r},\vec{r}\,')
\vec{B}^{xc,M}_j(\vec{r}\,')d^3r' \\
&& +\sum_{j\in m} \int_{\Omega_j^m} \chi^{m-m}_{ij}(\vec{r},\vec{r}\,')
\vec{B}^{xc,m}_j(\vec{r}\,')d^3r' \nonumber \\
&& + \chi^0_i(\vec{r}) \vec{B}^{xc,m}_i(\vec{r}) \;. \nonumber \end{aligned}$$
Analogous to $\sum_{j\in M}$ defined at Eq. (\[A3\]), the sum $\sum_{j\in m}$ means summation over sites with induced magnetic moments.
This can be reformulated in terms of local magnetic moments $M_i$ of Fe (Co) and $m_i$ of Pd (Pt)
$$\begin{aligned}
\label{Minduced2}
\hspace{-1cm}
m_i & = &
\sum_{j\in M} \tilde{\chi}^{m-M}_{ij} \vec{M}_j
+ \sum_{j \in m} \tilde{\chi}^{m-m}_{ij} \vec{m}_j
+ \tilde{\chi}_{ii} \vec{m}_i\end{aligned}$$
with
$$\begin{aligned}
\label{BXC_ind_suppl}
\vec{M}_i =
\int_{\Omega_{WS}} d^3r \vec{M}_i (\vec{r}); \;\;\;\; \vec{m}_i =
\int_{\Omega_{WS}} d^3r \vec{m}_i (\vec{r})\;.\end{aligned}$$
Here we use the following reformulation for the first term in Eq. (\[Minduced\]), that is more convenient for the model implementation:
$$\begin{aligned}
\label{chi_m}
\hspace{-1cm}
&& \int_{\Omega_j^M} \chi^{m-M}_{ij}(r,r') \vec{B}^{xc,M}_j(r')d^3r' \\
&& = \vec{M}_j \int_{\Omega_j^M} \chi^{m-M}_{ij}(r,r')
\frac{\vec{B}^{xc,M}_j(r')}{M_j}d^3r' \nonumber \\
&& = \tilde{\chi}^{m-M}_{ij}
\vec{M}_j \nonumber\end{aligned}$$
For the second term, using a linearised expression for the exchange potential in the case of a small induced magnetic moment on Pd and Pt sites [@DFVE01; @ME06], one can write analogously:
$$\begin{aligned}
\label{chi_mind}
\hspace{-1cm}
&& \int_{\Omega_j^m} \chi^{m-m}_{ij}(r,r') \vec{B}^{xc,m}_j(r')d^3r'
\\
& = & \vec{m}_j \int_{\Omega_j^m} \chi^{m-m}_{ij}(r,r') \frac{\delta
V_j^{xc}[n,m]}{\delta m(\vec{r})}|_{m_{spin} = 0}
\frac{m_j(\vec{r})}{m_j}d^3r' \nonumber \\
&=& \tilde{\chi}^{m-m}_{ij} \vec{m}_j\,. \nonumber\end{aligned}$$
Solving the system of equations (\[Minduced2\]) for a restricted region around a magnetic impurity atom gives the distribution of the induced magnetic moments on the non-magnetic atoms. This can be done for the ground state ($T = 0$K). Alternatively, without any approximations, one can get these values within [*ab-initio*]{} calculations for embedded magnetic atoms by using the CPA alloy theory, assuming an uniform distribution of the induced magnetic moment.
Strictly spoken, one can go beyond the linear approximation in the expansion of the exchange potential. However, the linear approximation makes the use of this scheme in subsequent Monte Carlo simulations much easier.
By making an additional simplification one can restrict to one response function $X^{m-M}_{ij}$ within the Monte-Carlo simulations. This quantity is defined to give the induced magnetic moment as a response to the exchange fields of only the surrounding nearest neighbour magnetic atoms:
$$\begin{aligned}
\label{Minduced3}
\hspace{-1cm}
\vec{m}_i & = &
\sum_{j \in M} X^{m-M}_{ij} \vec{M}_j.\end{aligned}$$
[54]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, , ), p. .
, ** (, , ), p. .
, , , , ****, ().
, , , , , , , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, ****, ().
, ** (, , ).
, ****, ().
, , , ****, ().
, ****, (), <http://link.aps.org/doi/10.1103/PhysRevLett.77.334>.
, ****, ().
, , , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , , , ****, ().
, , , , , , , , , , , ****, ().
, , , , , ****, ().
, in **, edited by (, , ), vol. of **, p. .
, , , ****, ().
, ** (, , ).
, ****, ().
, ** (, , ).
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , , , , , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, , , , ****, (), <http://link.aps.org/doi/10.1103/PhysRevB.78.144403>.
, Ph.D. thesis, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , , ****, ().
, ** (, , ).
, ****, ().
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We consider the ordinary and noncommutative Dirac-Born-Infeld theories within the open string $\sigma$-model. First, we propose a renormalization scheme, hybrid point splitting regularization, that leads directly to the Seiberg-Witten description including the two-form $\Phi$. We also show how such a form appears within the standard renormalization scheme just by some freedom in changing variables. Second, we propose a Wilson factor which has the noncommutative gauge invariance on the classical level and then compute the $\sigma$-model partition function within one of the known renormalization scheme that preserves the noncommutative gauge invariance. As a result, we find the noncommutative Yang-Mills action.\
PACS : 11.25.-w, 11.15.-q\
Keywords: $\sigma$-models, strings, Yang-Mills theory
author:
- |
Oleg Andreev[^1] [^2] and Harald Dorn[^3]\
\
Humboldt–Universität zu Berlin, Institut für Physik\
Invalidenstraße 110, D-10115 Berlin, Germany
title: 'On Open String $\sigma$-Model and Noncommutative Gauge Fields'
---
hep-th/9912070\
HU Berlin-EP-99/64
Introduction
=============
The open string $\sigma$-model was a hot topic in the eighties as it provided a basic tool to derive the low energy effective action (Dirac-Born-Infeld action) that is nonperturbative in $\ap$ [@FT; @C; @P] [^4]. Later it was also realized that it is useful for D-brane physics [@P1]. Recently, it has been pointed out by Seiberg and Witten that a special renormalization scheme, a point splitting regularization, results in a rather peculiar situation where the space-time (brane) coordinates do not commute (see [@SW] and refs. therein). The purpose of this paper is to further develop ideas about the appearance of noncommutative geometry by the open string $\sigma$-models.
Our conventions and some features of the quantization of open strings ending on D-branes that are relevant to our discussion are the following:
1\. The $\sigma$-model action is given by $$\label{ac}
S=\frac{1}{4\ai }\int_{\text{D}}d^2z\,\,
g_{ij}\pd_aX^i\pd^aX^j-i
\int_{\pd{\text D}}d\tau \,\Bigl(
{\frac{1}{2}}B_{ji}\dot X^iX^j+A_i(X)\dot X^i\Bigr)\,\,\,\,+\varphi
\quad,$$ where D means the string world-sheet (disk or half plane) whose boundary is $\pd\text D$. $g_{ij},\,\,B_{ij},\,\,\varphi$ are the constant metric, antisymmetric tensor and dilaton fields, respectively. We also included in the Abelian background gauge field $A_i(X)$. $X^i$ map the world-sheet to the brane and $i,\,j=1,\dots , p+1$. The world-sheet indices are denoted by $a,b$.
At this point, let us make a couple of remarks. First, we keep the explicit dependence on the constant dilaton field because it plays an important role in what follows. Second, for a constant $B$ field $B_{ij}\epsilon^{ab}\pd_aX^i\pd_bX^j$ is a total derivative, so we write this term as a boundary interaction.
2\. A natural object to compute within the path integral is the $\sa$-model partition function $$\label{pf}
Z[\varphi,g,B,A,\ap ]=\int {\cal D}X\,{\text e}^{-S}
\quad,$$ that results in the open string low energy effective action (Born-Infeld action) as well as the D-brane action (Dirac-Born-Infeld action) (for a review and refs., see, [@T; @P1]).
To compute the partition function one first integrates over the internal points of the disc to reduce the integral to the boundary and next splits the integration variable $X^i$ in the constant and the non-constant part as $X^i(\tau)=x^i+\xi^i(\tau)$. As a result, one gets the infinite set of the vertices $F,\,\pd F,\pd^2F$ etc [^5]. Next the perturbation theory in $\ap$ is used to compute the leading terms of $Z$. The derivative-independent part of the partition function proves to be the Dirac-Born-Infeld action namely, $$\label{dbi}
Z[\varphi,g,B,A,\ap ] =S_{\text{DBI}}+O(\pd F)
={\text e}^{-\varphi}\int [dx] \,\sqrt{\det (g+2\ai (F+B))}+O(\pd F)
\quad,$$ where $F_{ij}=\pd_iA_j-\pd_jA_i$ and $[dx]=\frac{d^{p+1}x}{(2\ai )^{p+1}}$. The above result is based on the use of one of the standard renormalization schemes like Pauli-Villars or $\zeta$-function which was originally used by Fradkin and Tseytlin [@FT].
It is natural to ask, whether different renormalization schemes used to compute the path integral lead to the same answer. It was understood in the eighties that the ambiguity in the structure of $Z$ related to the choice of a renormalization scheme should be a particular case of the field redefinition ambiguity present in the effective action reconstructed from the S-matrix. In other words, only structures in $Z$ that are invariant under the field redefinition are relevant.
3\. Seiberg and Witten pointed out that in the framework of the point splitting regularization scheme the leading terms of the perturbation theory are summed into the noncommutative version of the Dirac-Born-Infeld action namely, $$\label{ndbi}
\hat Z[\hat\varphi,G,\theta,\hat A,\ap ]
=\hat S_{\text{DBI}}+\dots
={\text e}^{-\hat\varphi}\int [dx]
\sqrt{\det (G+2\ai\hat F)}+\dots
\quad,$$ where $\hat F_{ij}=\pd_i\hat A_j-\pd_j\hat A_i-i\hat A_i\ast\hat A_j+
i\hat A_j\ast\hat A_i$. Here the $\ast$-product is defined by $$\label{star}
f(x)\ast g(x)={\text e}^{\frac{i}{2}\theta^{ij}
\frac{\pd}{\pd y^i}\frac{\pd}{\pd z^j}}f(x+y)g(x+z)\vert_{y=z=0}
\quad.$$ It is well-known that such a product is noncommutative but associative.
As we mentioned above, the different renormalization schemes are equivalent, so using a change of variables one has to get $$\label{eq}
\hat S_{\text{DBI}}=S_{\text{DBI}}+\text{total derivative terms}+O (\pd F)
\quad.$$ In the case of interest the change of variables found by Seiberg and Witten is given by [@SW] $$\label{nv}
\begin{split}
G&=(g-2\ai B)g^{-1}(g+2\ai B)
\quad\quad\,\,,\, \quad
\theta=-(2\ai )^2(g+2\ai B)^{-1}B(g-2\ai B)^{-1}
\quad, \\
\hat\varphi&=\varphi+{\frac{1}{2}}\ln\det \Bigl(G(g+2\ai B)^{-1}\Bigr)
\quad,\quad
\hat F=F+F\theta F-(A\theta\pd)F+O(\theta^2)
\quad.
\end{split}$$ Note that the last relation is simplified in the case of a constant field $\hat F$. Explicitly, $$\label{F}
\hat F=(1+F\theta)^{-1}F
\quad.$$
Moreover, it was also conjectured in [@SW] that the use of a suitable regularization that interpolates between the both mentioned above results in [^6] $$\label{fdbi}
\mathbf{Z}[\boldsymbol{\varphi},\mathbf{G},\boldsymbol{\theta},\mathbf{A},\ap ]
=\mathbf{S}_{\text{DBI}}+\dots
={\text e}^{-\boldsymbol{\varphi}}\int [dx]
\sqrt{\det (\mathbf{G}+2\ai(\mathbf{F}+\boldsymbol{\Phi}))}+\dots
\quad,$$ where $\boldsymbol{\Phi}$ is some two-form. In this case the relations are modified to $$\label{fnv}
(\mathbf{G}+2\ai\boldsymbol{\Phi})^{-1}=
-\frac{1}{2\ai }\boldsymbol{\theta}+
(g+2\ai B)^{-1}
\quad,\quad
\boldsymbol{\varphi}=\varphi+{\frac{1}{2}}\ln\det\Bigl(\frac{\mathbf{G}+2\ai\boldsymbol{\Phi}}
{g+2\ai B}\Bigr)
\quad.$$
Seiberg-Witten conjecture via $\sigma$-model
=============================================
The aim of this section is to show how the Seiberg-Witten conjecture may be simply derived within the $\sigma$-model approach. First, we start from the point splitting renormalization scheme. Then we propose how to modify it to get the desired result. Second, we start with the $\zeta$-function renormalization scheme and use the change of variables ($\sigma$-model couplings) to get . In the both cases $\Phi$ naturally appears due to freedom in choosing new variables.
Hybrid point splitting regularization
-------------------------------------
Following the ideas sketched in the introduction, we split the integration variable $X^i$ and moreover include the $B\dot\xi\xi$ term into the kinetic term that is usual for the problem at hand [@C; @SW]. So we have for the partition function $$\label{pf-ps}
\begin{split}
&Z[\varphi,g,B, A,\ap ]
={\text e}^{-\varphi}\int [dx]\,\,
\langle\,{\text e}^{-S_{int}}\,\rangle
\quad,\quad
\langle\,\dots\,\rangle=\int{\cal D\xi}\,{\text e}^{-S_0}\dots
\quad, \\
&S_0={\frac{1}{2}}\int_{\pd{\text D}}d\tau d\tau^{\prime}\,
\xi^{\text{\tiny T}}\hat N^{-1}\xi
\quad,\quad
S_{int}=-i\int_{\pd{\text D}}d\tau \, A(x+\xi)\dot \xi
\quad,
\end{split}$$ where $\hat N=-\ap G^{-1}\ln(\tau -\tau^{\prime})^2+\frac{i}{2}\theta
\epsilon (\tau-\tau^{\prime})$. The matrices $G,\,\theta$ are given by Eq. .
Expanding the interaction part in $A$ and doing the Fourier transform, the partition function is $$\label{pf-ps1}
Z[\varphi,g,B, A,\ap ]=
{\text e}^{-\varphi}\int [dx]
\sum_{n=0}^{\infty}\frac{i^n}{n!}\idotsint d\tau_n \idotsint d^{p+1}k_n \,\,
{\text e}^{i\kappa x}
A(k_1)\dots A(k_n)\,\,\langle\,\prod_{j=1}^n\dot\xi
{\text e}^{ik_j\xi (\tau_j)}\,
\rangle_{G,\,\theta}
\,\,,$$ where $\kappa=k_1+\dots +k_n$.
From now let us stick to the point splitting regularization. We are aware of the problem to define what is meant by the point splitted path integral for the partition function. Instead of giving a general answer, we define for our purpose a hybrid point splitting regularization. As used in [@SW] the point splitting regularization is not defined per se, but only in connection with a partial summation of perturbation theory that includes $B\dot\xi\xi$ term into the kinetic term. This results in an additive part $\frac{i}{2}\theta
\epsilon(\tau-\tau^{\prime})$ in the corresponding propagator. Everywhere the propagator ends on vertices which contain a $\tau$-derivative, part of the regularization is to drop the $\pd\epsilon$ term. We use this rule throughout, also for possible operator expressions inside functional determinants. But we insist that the determinants for the remaining expressions are treated within the standard renormalization scheme like Pauli-Villars or $\zeta$-function. As a consequence we get a generalization of the corresponding formula in [@SW] $$\label{path}
\langle\,\prod_{j=1}^n\dot\xi
{\text e}^{ik_j\xi (\tau_j)}\,\rangle_{G,\,\theta}=J\,
{\text e}^{-\frac{i}{2}\sum _{i>j}k_i\boldsymbol{\theta}k_j
\epsilon (\tau_i-\tau_j)}
\,\langle\,\prod_{j=1}^n\dot\xi
{\text e}^{ik_j\xi (\tau_j)}\,
\rangle_{G,\,\theta_0}
\quad,$$ where $\boldsymbol{\theta}=\theta-\theta_0$. As to $J$, it is the quotient of functional determinants. We have to care of it because it is relevant for what follows. It is evident that there exists freedom in choosing the new parameter $\theta_0$. It turns out that exactly such freedom is responsible for the two-form $\Phi$. Let us go on to see that this is indeed the case. Undoing the Fourier transform we arrive at [^7] $$\label{pf-ps2}
\mathbf{Z}[\varphi,\mathbf{G},\boldsymbol{\theta},\mathbf{A},\ap ]= J {\text e}^{-\varphi}\int [dx]
\sum_{n=0}^{\infty}
\idotsint d\tau_n\,
{\text e}^{\frac{i}{2}\sum_{i>j}\pd^i\boldsymbol{\theta}\pd^j
\epsilon (\tau_i-\tau_j)}
\langle\, \prod_{j=1}^n
\mathbf{A}(x_j+\xi_j)\dot \xi (\tau_j)\vert_{x_1=\dots x_n=x}
\,\rangle_{G,\,\theta_0}
\,\,,$$ where $\pd^i=\frac{\pd}{\pd x_i}$.
An important point we should stress here is that the above redefinition of $\theta$ automatically leads to a proper redefinition of the kinetic term. This time we have $$\label{k}
\mathbf{S}_0={\frac{1}{2}}\int_{\pd{\text D}}d\tau d\tau^{\prime}\,
\xi^{\text{\tiny T}} \mathbf{G}N^{-1}\xi
-\frac{i}{2}\int_{\pd{\text D}}d\tau \,\xi^{\text{\tiny T}}\boldsymbol{\Phi}\dot\xi
\quad,\quad
N=-\ap\ln(\tau-\tau^{\prime})^2
\quad,$$ with the new matrices for the metric and antisymmetric field, given by the corresponding inversion of the relations $$\label{met-b}
\mathbf{G}=\Bigl(G^{-1}-\frac{1}{(2\ai )^2}\theta_0G\theta_0\Bigr)^{-1}
\quad,\quad
\boldsymbol{\Phi}=-\frac{1}{(2\ai )^2}\mathbf{G}\theta_0G
\quad.$$ A simple algebra shows that the such defined $\boldsymbol{\theta},\,\mathbf{G},\,\boldsymbol{\Phi}$ obey the Seiberg-Witten relation .
Since our regularization prescription allows us to treat the functional determinants along the lines of [@FT] we find for $J$ $$J=
\Biggl(\frac{\text{Det}\hat N_{G,\theta}}{\text{Det}\hat N_{G,\theta_0}}
\Biggr)^{{\frac{1}{2}}}
=
\Biggl(\frac{\det (g+2\ai B)}{\det (\mathbf{G}+2\ai\boldsymbol{\Phi})}\Biggr)^{{\frac{1}{2}}}
\quad.$$ The latter may be absorbed by the corresponding redefinition of the dilatonic field (string coupling constant). Thus we get the last relation in .
Let us now define the Wilson factor as $$\label{Wl}
P\ast \exp \Bigl(i\int\,\mathbf{A}(x+\xi)\dot \xi\Bigr)=\sum_{n=0}^{\infty}\frac{i^n}{n!}
\idotsint d\tau_n\,
{\text e}^{\frac{i}{2}\sum_{i>j}\pd^i\boldsymbol{\theta}\pd^j
\epsilon (\tau_n-\tau_m)}
\prod_{j=1}^n
\mathbf{A}(x_j+\xi_j)\dot \xi (\tau_j)\vert_{x_1=\dots x_n=x}
\quad.$$ It is clear that this factor is equal to the usual path ordering $P$ applied to the exponential in the sense of the $\ast$-product.
Finally, the partition function becomes $$\label{pf-ps3}
\mathbf{Z}[\boldsymbol{\varphi},\mathbf{G},\boldsymbol{\theta},\mathbf{A},\ap ]
=\,{\text e}^{-\boldsymbol{\varphi}}
\int [dx]\int{\cal D}\xi\,
{\text e}^{-\mathbf{S}_0}\,P\ast \exp \Bigl(i\int_{\pd{\text D}}d\tau\,\mathbf{A}(x+\xi)\dot\xi\Bigr)
\quad.$$
Let us now specialise the gauge field to $$\label{cf}
\mathbf{A}_i(X)=\frac{1}{2}\mathbf{f}_{ji}X^j
\quad,$$ with a constant matrix $\mathbf{f}$. Then the related noncommutative field strength is simply given by $\mathbf {F}=\mathbf{f}-\mathbf{f}\boldsymbol{\theta}\mathbf{f}$.
For the gauge field the Wilson factor turns out to be $$\label{cfW}
P\ast \exp \Bigl(\frac{i}{2}\int_{\pd{\text D}}d\tau\,(x+\xi)^{\text{\tiny T}}\mathbf{f}\dot\xi\Bigr)=
\exp \Bigl(\frac{i}{2}\int_{\pd{\text D}}d\tau\,(x+\xi)^{\text{\tiny T}}\mathbf{F}\dot\xi\Bigr)
\quad.$$ Now we just have the functional integral treated in [@FT] and find $$\label{fin}
\mathbf{Z}[\boldsymbol{\varphi},\mathbf{G},\boldsymbol{\theta},\mathbf{A},\ap ]=
\mathbf{S}_{\text{DBI}}=
\,{\text e}^{-\boldsymbol{\varphi}}
\int [dx]
\sqrt{\det (\mathbf{G}+2\ai(\mathbf{F}+\boldsymbol{\Phi}))}
\quad.$$ With this formula we have finished an explicit calculation proving that the partition function for a constant field strength within the hybrid point splitting renormalization scheme is given by the Dirac-Born-Infeld action referring to the noncommutative gauge field. This has been argued for in [@SW] by referring to the analogy of the involved structures compared to the commutative case.
$\zeta$-function regularization
-------------------------------
As we have mentioned in the introduction, the different renormalization schemes used to compute the path integral lead to the same result after a proper change of the variables ($\sigma$-model couplings) is done. In fact, such a change of the couplings is nothing but the resummation of the perturbation theory in $\ap$. After this is understood, it immediately comes to mind to realize a resummation by changing the path integral variables. A possible way to do this is to take a new variable as a function of $\ap$. Let us now show how it works. Specialising to the $\zeta$-function renormalization scheme [@FT], we define the path integral measure in the same way as it was done in [@AMT]. So we have $$\label{pt}
\begin{split}
&Z[\varphi,g,B,A,\ap ]={\text e}^{-\varphi}\int [dx]\sqrt{\det g}\,\,
\langle\,{\text e}^{-S_{int}}\,\rangle
\quad,\quad
\langle\,\dots\,\rangle=\int{\cal D\xi}\,{\text e}^{-S_0}\dots
\quad, \\
&S_0={\frac{1}{2}}\int_{\pd{\text D}}d\tau d\tau^{\prime}\,
\xi^{\text{\tiny T}}gN^{-1}\xi
\quad,\quad
S_{int}=-i\ai\int_{\pd{\text D}}d\tau \, \xi^{\text{\tiny T}}(B+F)\dot\xi+O(\pd F)
\quad,
\end{split}$$ Above we have rescaled $X^i$ as $X^i\rightarrow\sqrt{2\ai }X^i$. $N$ is the boundary value of the Neumann function.
Now let us change the variable $\xi^i$ as $\xi^i=(g^{-1}\Lambda\boldsymbol{\xi})^i\,$ [^8]. In fact, the measure is defined in such a way that the effect is only due to the action $S$. The latter becomes $$\label{na}
S_0={\frac{1}{2}}\int_{\pd{\text D}}d\tau d\tau^{\prime}\,
\boldsymbol{\xi}^{\text{\tiny T}}\mathbf{G}N^{-1}\boldsymbol{\xi}
\quad,\quad
S_{int}=-i\ai\int_{\pd{\text D}}d\tau \, \boldsymbol{\xi}^{\text{\tiny T}}\bigl(
-\frac{1}{(2\ai )^2}\mathbf{G}\boldsymbol{\theta}\mathbf{G}+\Lambda^{\text{\tiny T}}g^{-1}F
g^{-1}\Lambda\bigr)\boldsymbol{\xi}+O(\pd F)
\,\,,$$ where $\mathbf{G}=\Lambda^{\text{\tiny T}}g^{-1}\Lambda,\,
\boldsymbol{\theta}=-(2\ai )^2\Lambda^{-1}B\Lambda^{\text{\tiny T}-1}$.
It is evident that as far as $\Lambda$ depends on $\ap$ we automatically get a resummation of the perturbation theory because the vertices ($\sigma$-model couplings) are redefined [^9]. It is also clear which role the dilaton field $\varphi$ has to play. It is responsible for the cancellation of the poles in $\ap$ within the perturbation theory i.e., the $\ap$-dependence of $Z$ always looks like $Z\sim (\ap )^{-\frac{p+1}{2}}\Bigl(1+O(\ap )\Bigr)$. Note that $\Lambda$ is an arbitrary function in $g,\,B,\,\ap $. Moreover $B$ plays a key role as it is responsible for $\ap$-dependence of $\Lambda$ because $g$ is dimensionless. One can also consider $\Lambda$ as a function in $\mathbf{G},\,\theta,\,\ap $. In this case the change of the variables is given by $\xi^i=(\Lambda^{\text{\tiny T}-1}\mathbf{G}\boldsymbol{\xi})^i$.
In general, it is not clear how to exactly compute the partition function. The best what we can do is to find the leading terms within the perturbation expansion as it was done in [@FT; @AT; @T2]. For our purposes the one-loop approximation is also sufficient. So the problem is reduced to finding the corresponding determinant. It is straightforward to write down a solution of the problem [@FT] $$\label{det}
S_{\text{DBI}}={\text e}^{-\varphi}\int [dx]\sqrt{\det g}\,\,
\sqrt{\det (1-\frac{1}{2\ai }\boldsymbol{\theta} \mathbf{G}+
2\ai\Lambda^{-1}F\Lambda^{\text{\tiny T}-1}\mathbf{G}\,)}
\quad.$$ In doing transformations with the determinant the important thing one should keep in mind is that the resummation of the perturbation theory assumes that the partition function depends on $\ap$ as $Z\sim (\ap )^{-\frac{p+1}{2}}\Bigl(1+O(\ap )\Bigr)$. As to the determinant, we postulate that it is given by $\sqrt{\det (A+2\ai B)}$ with some matrices $A$ and $B$. So it is clear that we should get rid of the $\frac{1}{\ap}$-term. The latter argument assumes $$\label{det1}
S_{\text{DBI}}={\text e}^{-\varphi}\sqrt{\det (g+2\ai B)\,}
\int [dx]\,\,
\sqrt{\det (1+(1-\frac{1}{2\ai }\boldsymbol{\theta}\mathbf{G})^{-1}
2\ai\Lambda^{-1}F\Lambda^{\text{\tiny T}-1}\mathbf{G}\,)}
\quad.$$ Above we have used that $\det(g-\frac{1}{2\ai }g\boldsymbol{\theta}\mathbf{ G})=\det (g+2\ai B)$.
Moreover we bring $\Lambda^{-1}$ to a form $$\label{del}
\Lambda^{-1}=(\mathbf{G}^{-1}-\frac{1}{2\ai }\boldsymbol{\theta} )\Delta
\quad,$$ where $\Delta$ is a new matrix that depends on $\mathbf{G},\,\boldsymbol{\theta},\,\ap $. This results in $$\label{det2}
S_{\text{DBI}}={\text e}^{-\varphi}\sqrt{\det (g+2\ai B)\,}
\int [dx]\,\,
\sqrt{\det (1+2\ai F\Delta^{\text{\tiny T}}(\mathbf{G}^{-1}+\frac{1}{2\ai }\boldsymbol{\theta})
\Delta)\,}
\quad.$$ Our postulate for the determinant yields the following $\ap$-dependence of $\Delta^{\text{\tiny T}}(\mathbf{G}^{-1}+\frac{1}{2\ai }\boldsymbol{\theta})\Delta$ $$\label{del1}
\Delta^{\text{\tiny T}}(\mathbf{G}^{-1}+\frac{1}{2\ai }\boldsymbol{\theta})\Delta
=\frac{1}{2\ai }\Gamma+\Sigma(\ap )
\quad,$$ where $\Sigma(\ap )$ is regular in the limit $\ap\rightarrow 0$.
Substituting into we find $$\label{det3}
S_{\text{DBI}}={\text e}^{-\boldsymbol{\varphi}}
\int [dx]\,\,
\sqrt{\det(1+F\Gamma )\,}
\sqrt{\det (\Sigma^{-1}+2\ai (1+F\Gamma)^{-1}F\,)}
\quad.$$ Above we have also defined a new dilatonic field as $$\label{dilaton}
\boldsymbol{\varphi}=\varphi+{\frac{1}{2}}\ln\frac{\det \Sigma^{-1}}{\det(g+2\ai B)}
\quad.$$
Furthermore due to our postulate for the determinant we can represent $\Sigma^{-1}(\ap )$ as $\Sigma^{-1}(\ap )=\Pi+2\ai\Omega$ with some $\ap$-independent matrices $\Pi$ and $\Omega$. So the Eq. is rewritten in the following form $$\label{det4}
S_{\text{DBI}}={\text e}^{-\boldsymbol{\varphi}}
\int [dx]\,\,
\sqrt{\det(1+F\Gamma)\,}
\sqrt{\det (\Pi+2\ai (\Omega+(1+F\Gamma)^{-1}F)\,)}
\quad.$$
Let us now see what the matrices $\Sigma,\Pi,\Omega$ are. To do so, it is worth to remark that $\Lambda_{ij}=g_{ij}$ is a trivial transform in a sense that it does not lead to any resummation of the perturbation theory. So we can consider it as a unity in a space of all transforms. It is reasonable to normalize transforms with respect of this unity by requiring that $\Delta$ has the following asymptotic behaviour $\Delta_i^j=\delta_i^j
+O(\ap )$ as $\ap\rightarrow 0$. On the other hand, it also seems reasonable because $\delta_i^j$ is the only $\ap$-independent tensor with such an index structure we have. Then it immediately follows from that $\Gamma=\boldsymbol{\theta}$. This allows to define a new field strength $\mathbf{F}$ as $\mathbf{F}=(1+F\boldsymbol{\theta})^{-1}F$ that of course coincides with the Seiberg-Witten definition . However, the point is that we have not used the fact that two theories should be in the same class of equivalence as gauge theories. As a result, the Dirac-Born-Infeld action is rewritten as $$\label{det5}
S_{\text{DBI}}={\text e}^{-\boldsymbol{\varphi}}
\int [dx]\,\,
\sqrt{\det(1+F\boldsymbol{\theta})\,}
\sqrt{\det (\Pi+2\ai (\Omega+\mathbf{F})\,)}
\quad.$$ Furthermore it is clear from the general covariance of the partition function that $\Pi$ must be treated as a metric $\mathbf{G}$. As to $\Omega$, it coincides with $\Phi$ from Eq.. Of course, , , are equivalent to the relations .
Finally, we get $$\label{det6}
S_{\text{DBI}}={\text e}^{-\boldsymbol{\varphi}}
\int [dx]\,\,
\sqrt{\det(1+F\boldsymbol{\theta})\,}
\sqrt{\det (\mathbf{G}+2\ai (\Phi+\mathbf{F})\,)}
\quad.$$
What we actually need is only the last factor of the integrand to get the equivalence . To fix this problem, one subtle point we should remind is that different formulations coincide up to derivative terms. It also assumes that we can integrate by parts that of course makes no sense for constant field strengths. In fact, we have to consider slowly varying field strengths. As a consequence, the relation between $F$ and $\mathbf{F}$ becomes more involved [@SW]. In the leading order of $\boldsymbol{\theta}$ it is given by $$\label{dF}
\mathbf{F}=F-F\boldsymbol{\theta}F-(A\boldsymbol{\theta}\pd)F+O(\boldsymbol{\theta}^2)
=\mathbf{F}_0-(A\boldsymbol{\theta}\pd)F+O(\boldsymbol{\theta}^2)
\quad.$$ It is easy to see that in this order the actions and coincide namely, $$\int [dx]\,\biggl(\sqrt{\det(1+F\boldsymbol{\theta})\,}
\sqrt{\det (G+2\ai (\Phi+\mathbf{F}_0 )\,)}
-\sqrt{\det (G+2\ai (\Phi+\mathbf{F}_0-(A\boldsymbol{\theta}\pd)F)\,)}\biggr)
=0
\quad.$$ Unfortunately calculations become more and more involved as far as we consider higher order terms.
To conclude, let us comment on the background independence of the action $\mathbf{S}_{\text{DBI}}$. Since the action $S_{\text{DBI}}$ is explicitly invariant the expression we find in above is also invariant. Next what we drop is total derivative terms, so the action $\mathbf{S}_{\text{DBI}}$ is invariant at least in the leading order in $\boldsymbol{\theta}$ modulo total derivatives.
Alternative way
================
In fact, there are two ways in getting the noncommutative Yang-Mills theory with the $\ast$-product structure within the $\sigma$-model. The first one [*a la*]{} Seiberg-Witten is to start from the Wilson factor that has the ordinary gauge invariance on the classical level and then get the noncommutative gauge invariance on the quantum level just by using a proper renormalization scheme. The second way we propose, based on our experience with the hybrid renormalization scheme, is to start from the Wilson factor that has the noncommutative gauge invariance on the classical level and use a regularization that maintains it on the quantum level [^10]. So our aim in this section is to show how to realize this proposal.
Let us define the Wilson factor as $$\label{nW}
W[C]=P\ast \exp \Bigl(i\int_Cd\tau\,A\dot X\Bigr)
=\sum_{n=0}^{\infty}i^n
\idotsint d\tau_n\, H(\tau_{12})\dots H(\tau_{n-1n})A\dot X(\tau_1)\ast\dots\ast
\,A\dot X(\tau_n)
\quad,$$ where the $\ast$-product is defined with respect to a translational mode of $X$. Such a factor coincides with the one defined in . This is clear just by substituting the expansion of unity $1=H(\tau_{12})\dots H(\tau_{n-1n})+(\text{all perms.})$ [^11] and changing the variables $\tau$ in such a way to get the ordering $\tau_1>\tau_2>\dots>\tau_n$. Note that the definition is nothing but a slightly modified version of the non-Abelian Wilson factor. Of course, it is simply to fit the non-Abelian case into the above definition just by allowing the gauge field to be a $N\times N$ hermitian matrix and taking a trace. Explicitly, $$\label{naW}
W[C]=TrP\ast \exp \Bigl(i\int_Cd\tau\,A\dot X\Bigr)
\quad.$$
An important point we should stress here is that the Wilson factor as it is defined in or is [*almost*]{} invariant [^12] under the gauge transformation of the noncommutative Yang-Mills theory namely, $\delta_{\lambda}A_i=\pd_i\lambda-iA_i\ast\lambda +i\lambda\ast A_i$. The latter is clear from an analogy with the non-Abelian Yang-Mills theory where the gauge invariant expression is given by a trace of $P\exp\int A{\dot X}$. Of course, this is easy to see by directly doing the gauge transforms. Fortunately for us, what saves the day is that we are interested in the partition function. Indeed, as we saw splitting the integration variable $X^i$ automatically provides the integral over the zero mode (translational mode) $x^i$. This is exactly what we need because in the noncommutative case the integral does the same job as the trace in the non-Abelian case, i.e. it provides the cyclic property $\int d^{p+1}x\,f(x)\ast g(x)=\int d^{p+1}x\,g(x)\ast f(x)$ that is crucial for the gauge invariance. Thus the partition function is gauge invariant.
To be more precise, the partition function is given by $$\label{z}
Z[A]=\int{\cal D}X\,{\text e}^{-S_0}\,P\ast \exp \Bigl(i\int_Cd\tau\,A_i\dot X^i\Bigr)
\quad,\quad
S_0=\frac{1}{4\ai }\int_{\text{D}}d^2z\,\,G_{ij}\pd_aX^i\pd^aX^j
\quad.$$
To see that this partition function indeed leads to the noncommutative Yang-Mills theory it is instructive to compute its asymptotic behaviour as $\ap\rightarrow 0$. The main point is that we have to be careful to preserve the noncommutative gauge invariance. It turns out that the renormalization scheme based on the $\zeta$-function or some of its modification does this job [@FT; @AT]. The computations are analogous to the ordinary non-Abelian case [@T2]. As a result, we find $$\label{nYM}
Z[A]=\int [dx]\sqrt{\det G\,}
\Bigl(1+\frac{1}{4}(2\ai )^2F_{ij}\ast F_{kl}G^{ik}G^{jl}+
O(\ap{}^{\frac{5}{2}}) \Bigr)
\quad,$$ with $F_{ij}=\pd_iA_j-\pd_jA_i-iA_i\ast A_j +iA_j\ast A_i$.
It is straightforward to generalize these computations for the non-Abelian case. The only new thing that appears is $Tr$.
Finally, let us briefly show how to incorporate SUSY within the above formalism. For simplicity, let us consider the NSR formalism. In other words, we add a set of the fermionic fields $\psi^i$ whose metric also is $G$. Following [@AT], it is simply to suggest what the Wilson factor should be. Moreover, the formalism developed in this paper allows to use the super-space notations. Thus, the Wilson factor is given by $$\label{snW}
\mathbf{W}[C]=\mathbf{P}\ast \exp \Bigl(i\int_Cd\boldsymbol{\tau}\,A D\mathbf{X}\Bigr)
=\sum_{n=0}^{\infty}i^n
\int .\,.\,.\int
d\boldsymbol{\tau}_n\mathbf{H}(\boldsymbol{\tau}_{12})\,.\,.\,.\,
\mathbf{H}(\boldsymbol{\tau}_{n-1n})\,
AD\mathbf{X}(\boldsymbol{\tau}_1)
\ast\dots\ast
AD\mathbf{X}(\boldsymbol{\tau}_n)
\,,$$ where we use the super-space notations namely, $d\boldsymbol{\tau}=d\tau d\vartheta,\,
\mathbf{H}(\boldsymbol{\tau}_{nm})=H(\tau_{nm})+\vartheta_n\vartheta_m\delta(\tau_{nm}),\,
\mathbf{X}^i=X^i+\vartheta\psi^i,\,D=\vartheta\pd_{\tau}-\pd_{\vartheta}$. As in the bosonic case the noncommutative multiplication law is defined in terms of the translational modes of $X^i$.
It is a simple matter to check that the expression is [*almost*]{} gauge invariant as well as to compute the partition function in the leading order in $\ap$. The result is again given by the noncommutative Yang-Mills action (see Eq.). It is also straightforward to fit the non-Abelian case into the above definition just by doing in the same way way as we did in the bosonic case.
Concluding Comments
====================
Motivated by our experience with the $\sigma$-model, we would like to propose an exactly gauge invariant version of the Wilson factor within the noncommutative Yang-Mills theory, $$\label{nW-loop}
{\cal W}[C]=\frac{1}{V}\int d^{p+1}x\sqrt{\det G\, }\,\, TrP\ast \exp \Bigl(i\int_Cd\tau\,A\dot X
\Bigr)
\quad,\quad \text{ with}\quad V=\int d^{p+1}x\sqrt{\det G\,}
\quad.$$
Some discussions of the Wilson factors and Dirac-Born-Infeld action that have some interference with what we described in above are due to [@refs].
We would like to thank A. Tseytlin for comments and reading the manuscript. The work of O.A. is supported in part by the Alexander von Humboldt Foundation, and by Russian Basic Research Foundation under Grant No. 9901-01169. The work of H.D. is supported in part by DFG.
[99]{} E. S. Fradkin, and A. A. Tseytlin, 123. A. Abouelsaood, C. G. Callan, C. R. Nappi, and S. A. Yost, 599;\
C. G. Callan, C. Lovelace, C. N. Nappi, and S. A. Yost, 525. J. Dai, R. G. Leigh, and J. Polchinski, 2073. A. A. Tseytlin, “Born-Infeld action, supersymmetry and string theory”, Report No. Imperial/TP/98-99/67, hep-th/9908105, to appear in the Yuri Golfand memorial volume, ed. M. Shifman, World Scientific (2000). J. Polchinski, “TASI lectures on D-branes”, Report No. NSF-ITP-96-145, hep-th/9611050. N. Seiberg, and E. Witten, 032. B. Pioline, and A. Schwarz, 021. O. Andreev, R. Metsaev, and A. Tseytlin, Yad. Fiz. 51 (1990) 564; Sov. J. Nucl. Phys. 51 (1990) 359. O. Andreev, and A. Tseytlin, 205. A. A. Tseytlin, 81;\
H. Dorn, and H.-J. Otto, Zeitschr. f. Phys. C32 (1986) 599. C. Chu and P. Ho, 151.\
A. Fayyazuddin, and M. Zabine, “A note on bosonic strings in constant $B$ field”, Report No. USITP-99-8, hep-th/9911018. L. Cornalba, and R. Schiappa, “Matrix Theory Star Products from the Born-Infeld Action”, Report No. CTP$\sharp$2887, hep-th/9907211; L. Cornalba, “D-brane Physics and Noncommutative Yang-Mills Theory”, hep-th/9909081;\
N. Ishibashi, S. Iso, H. Kawai, Y. Kitazawa, “Wilson loops in Noncommutative Yang Mills”, Report No. KEK-TH-649, hep-th/9910004;\
K. Okuyama, “A Path Integral Representation of the Map between Commutative and Noncommutative Gauge Fields”, Report No. KEK-TH-655, hep-th/9910138;\
P. Watts, “Noncommutative String Theory, the R-Matrix, and Hopf Algebras”, Report No. DIAS-STP-99-13, hep-th/9911026;\
J. Ambjorn, Y.M. Makeenko, J. Nishimura, R.J. Szabo, “Finite N Matrix Models of Noncommutative Gauge Theory”, Report No. NBI-HE-99-44, hep-th/9911041;\
T. Lee, “Noncommutative Dirac-Born-Infeld Action for D-brane”, Report No. KIAS-P99109 , hep-th/9912038.
[^1]: e-mail: [email protected]
[^2]: Permanent address: Landau Institute, Moscow, Russia
[^3]: e-mail: [email protected]
[^4]: For a recent review of this issue see [@T] and references therein.
[^5]: For the sake of simplicity, we use the matrix notations here and below.
[^6]: It should be noted that though analogous formulae first appeared in the context of NCSYM [@Pio] their appearance within the $\sigma$-model approach is due to Seiberg and Witten.
[^7]: We change our notations due to specializing the renormalization scheme. So, $\mathbf{A}$ denotes the gauge variable in this scheme etc.
[^8]: In fact, it is nothing but $GL(p+1)$ transforms whose matrices depend on $\ap$.
[^9]: Note that $\Lambda=g+2\ai B$ recovers the Seiberg-Witten relations .
[^10]: Some similar motivations are provided by the quantization of open strings in a constant $B$ field background where the deformation parameter $\theta$ is explicitly related with zero modes (see, e.g., [@Fa] and refs. therein).
[^11]: $H$ means the Heaviside step function.
[^12]: It becomes gauge invariant after the integration over the translational mode of $X$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- Carlos García Argos
bibliography:
- '../bibliography/refs\_cosmic.bib'
- '../bibliography/refs\_silicon.bib'
- '../bibliography/refs\_powering.bib'
- '../bibliography/refs\_tracking.bib'
- '../bibliography/refs\_misc.bib'
- '../bibliography/refs\_physics.bib'
- '../bibliography/refs\_upgrade.bib'
- '../bibliography/refs\_vcsels.bib'
- '../bibliography/refs\_dedx.bib'
title: |
PhD Thesis\
A Silicon Strip Detector for the Phase II High Luminosity Upgrade of the ATLAS Detector at the Large Hadron Collider
---
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract:
- |
[Let $\{R_t, \; 0\leq t\leq 1\}$ be a symmetric $\alpha$-stable Riemann-Liouville process with Hurst parameter $H > 0$. Consider a translation invariant, $\beta$-self-similar, and $p$-pseudo-additive functional semi-norm $||.||$. We show that if $H>\beta +1/p$ and $\gamma = (H-\beta-1/p)^{-1}$, then $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; =\; - \k\;\in\; [-\infty, 0),$$ with $\k$ finite in the Gaussian case $\alpha = 2$. If $\alpha < 2$, we prove that $\k$ is finite when $R$ is continuous and $H>\beta +1/p + 1/\alpha$. We also show that under the above assumptions, $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||X||\leq \ee \rcr
\; =\; - \k\;\in\; (-\infty, 0),$$ where $X$ is the linear $\alpha$-stable fractional motion with Hurst parameter $H\in(0,1)$ (if $\alpha = 2$, then $X$ is the classical fractional Brownian motion). These general results cover many cases previously studied in the literature, and also prove the existence of new small deviation constants, both in Gaussian and Non-Gaussian frameworks.]{}
- |
[Soit $\{R_t, \; 0\leq t\leq 1\}$ un processus de Riemann-Liouville $\alpha$-stable symétrique avec paramètre de Hurst $H > 0$. Considérons une semi-norme fonctionnelle $||.||$ invariante par translation, $\beta$-autosimilaire et $p$-pseudo-additive. Nous montrons que si $H>\beta +1/p$ et $\gamma = (H-\beta-1/p)^{-1}$ alors $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; =\; - \k\;\in\; [-\infty, 0),$$ avec $\k$ finie dans le cas gaussien $\alpha = 2$. Lorsque $\alpha < 2$, nous montrons que $\k$ est finie quand $R$ est continu et $H>\beta +1/p + 1/\alpha$. Nous montrons aussi que sous ces hypothèses $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||X||\leq \ee \rcr
\; =\; - \k\;\in\; (-\infty, 0),$$ où $X$ est le mouvement fractionnaire linéaire $\alpha$-stable avec paramètre de Hurst $H\in(0,1)$ (lorsque $\alpha = 2$, $X$ est le mouvement brownien fractionnaire usuel). Ces résultats généraux recouvrent de nombreux cas précédemment étudiés dans la littérature et prouvent l’existence de nouvelles constantes de petites déviations, aussi bien dans le cadre gaussien que non gaussien.]{}
---
[**Keywords:**]{} Fractional Brownian motion - Gaussian process - Linear fractional stable motion - Riemann-Liouville process - Small ball constants - Small ball probabilities - Small deviations - Stable process - Wavelets.
[**MSC 2000:**]{} 60E07, 60G15, 60G17, 60G18
Introduction
============
Let $X$ be a random process whose sample paths belong to some functional normed space $\lpa \FF, \|\cdot\|\rpa$. The investigation of the small deviations (or small ball probabilities) of $X$ deals with the asymptotics of $$\pb\lcr ||X||\leq \ee \rcr$$ when $\ee\downarrow 0$, and has proved to be a difficult problem with increasing number of applications in Probability, Analysis, Complexity ... etc. We refer to the recent surveys [@LS] [@Lf] for a detailed account on this subject.
In the literature, this problem is usually studied for a particular class of processes and under a particular norm. It remains a great challenge, a kind of “mission impossible”, to find some principle describing small deviations for general classes of processes and norms, rather than tackle the problem case by case.
The unique successful attempt in this direction was made by W. Stolz [@St1] [@St2], who obtained estimates $$\label{estim}
-\infty \; < \; \liminf_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr
||X||\leq \ee \rcr \; \leq \; \limsup_{\ee\downarrow 0}
\ee^{\gamma}\log\pb\lcr ||X||\leq \ee \rcr \; < \; 0,$$ where $\{X_t, \; 0\leq t\leq 1\}$ is a Brownian motion (or more generally a continuous Gaussian process with covariance function similar to that of fractional Brownian motion), and $\gamma$ a finite positive parameter depending on the behavior of the (semi-)norm $\|\cdot\|$ on linear combinations of Schauder functions. Many classical semi-norms fell into Stolz’s scope: $L_p$-norms, Hölder and Sobolev semi-norms, Besov norms... etc, and his estimates resumed from a general point of view a lot of situations which were previously studied (see the references quoted in [@St1] [@St2]).
The next important issue is the [*existence of the limit*]{} in (\[estim\]), and this is the matter of the present paper. Our main result says that if $||.||$ is a translation invariant, $\beta$-self-similar and $p$-superadditive functional semi-norm - see Definition 1 for more precisions about these notions, and $\{R_t, \; 0\leq t\leq 1\}$ is a symmetric $\alpha$-stable Riemann-Liouville process with Hurst parameter $H > \beta + 1/p$ ($R$ can be viewed as a fractionally integrated symmetric $\alpha$-stable Lévy process, see Section 2 for a precise definition), then $$\label{limit1}
\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; =\; - \k\;\in\; [-\infty, 0)$$ with the rate $\gamma = (H-\beta-1/p)^{-1}$. From the technical point of view, the main ingredient of the proof is a stochastic superadditive inequality which is based upon the extrapolation-homogeneity of $R$, and is then easy to combine with an exponential Tauberian theorem and the standard subadditivity arguments.
Our framework has two secondary, but non-negligible advantages with respect to [@St2]. [*Non-Gaussian*]{} stable processes are included, as well as smooth Gaussian ones like integrated Brownian motion of arbitrary parameter.
A special effort is needed to prove that the constant $\k$ in (\[limit1\]) is [*finite*]{}. In order to get the suitable lower bound on the small deviation probabilities, we use roughly the same method as in [@St1] [@St2], except that we decompose $R$ along Daubechies’ wavelet bases, since Schauder system not smooth enough when $H$ becomes too large.
In the last part of this paper, we extend the above results to a class of self-similar processes with long range dependence, the so-called unilateral linear fractional stable motions. These processes can be viewed as a possible generalization of fractional Brownian motion with an underlying stable noise. Thanks to a nice argument essentially due to of Li and Linde [@LL1], the problem is reduced to a study of the Schauder decomposition of the associated long memory process.
We conclude this article with a brief survey of concrete results. It seems that our theorems synthesize everything that is known about the existence of finite small deviation constants for continuous fractional processes under translation-invariant semi-norms. Last but not least, we can harvest a bunch of new constants, both in Gaussian and Non-Gaussian situations.
Still, a major drawback of the above wavelet methods is that they exclude discontinuous processes. As a rule, proving lower bound probabilities for processes with jumps requires completely different discretization techniques [@Mo] [@Ta], and we have no idea how much time it will take to enclose them into a global strategy.
Preliminaries {#norms}
=============
Parametrization: $(\beta, p)$-semi-norms
----------------------------------------
Let $\II$ be the set of all closed bounded intervals of $\R$. Consider $\FF$ a linear space of functions from $\R$ to $\R$ and, for each $I\in\II$, let $\FF_I$ be a linear space of functions from $I$ to $\R$ such that $f_I\in\FF_I$ for every $f\in\FF$, where $f_I$ stands for the restriction of $f$ to $I$.
We define a semi-norm $||\cdot||$ on $\FF$ as a family $\lacc ||\cdot||^{}_I, \; I\in\II\racc$ of functionals mapping $\FF_I$ to $\R^+$ such that $||\lambda f||^{}_I = |\lambda| ||f||^{}_I$ and $||f + g||^{}_I\leq ||f||^{}_I + ||g||^{}_I$ for every $\lambda\in\R, f, g\in\FF_I$. We will use the notation $||f||^{}_I = ||f_I||^{}_I$ for every $f\in\FF$, $I\in\II$. In the remainder of this paper we will assume that $||\cdot||$ satisfies the following assumptions, which are verified by all the classical semi-norms:
\(A) $||\cdot||^{}_I\leq ||\cdot||^{}_J$ for every $I, J\in\II$ such that $I\subset J$. (Contractivity)
\(B) $||f||^{}_{I-c} = ||f(\cdot-c)||^{}_I$ for every $f\in\FF$, $I\in\II$ and $c\in\R$. (Translation-invariance)
Let $\beta\in\R$, $p\in (0, +\infty]$ and $||\cdot||$ be a contractive and translation-invariant semi-norm on $\FF$. We say that $||\cdot||$ is an [*upper $(\beta, p)$-semi-norm*]{} if it satisfies the following properties:
[*(C)*]{} $||f(c\,\cdot)||^{}_{I/c} = c^\beta \ ||f||^{}_I$ for every $f\in\FF$, $I\in\II$ and $c > 0.\;\;\;\;\;\;\;\;\;$ [*($\beta$-self-similarity)*]{}
[*(D)*]{} For every $a_0<\ldots<a_n\in\R$ and $f\in\FF$
$$\lacc
\begin{array}{llr}
||f||^{}_{[a_0,a_n]} \;\geq\; {\lpa ||f||_{[a_0, a_1]}^p
+ \cdots + ||f||_{[a_{n-1},a_n]}^p\rpa}^{1/p}
& \mbox{{\em if} $p < +\infty$,} & \\
& & \;\mbox{\em ($p$-superadditivity)} \\
||f||^{}_{[a_0,a_n]} \;\geq\; \sup\lpa ||f||^{}_{[a_0, a_1]},
\ldots, ||f||^{}_{[a_{n-1},a_n]}\rpa
& \mbox{{\em if} $p = +\infty$.} &
\end{array}
\right.$$
In the following we will denote by $\UU(\beta,p)$ the set of upper $(\beta, p)$-semi-norms, and set $\UU$ for the union of all $\UU(\beta,p)$’s.
*(a) Of course it suffices to take $n=2$ in the definition of $p$-superadditivity. We wrote the property in this form in order to make it symmetric with the $p$-subadditivity and the corresponding lower $(\beta, p)$-semi-norms, which will appear just below.*
\(b) From the inequality ${(a+b)}^q \geq a^q + b^q$ for every $a, b \geq 0$ and $q\geq 1$, it follows that $\UU(\beta,p)\subset\UU(\beta,p')$ whenever $p'\geq p$.
\(c) In the definition of upper $(\beta,p)$-semi-norm, one can ask for the possible values of the parameters $\beta$ and $p$. By contractivity, homogeneity and translation-invariance, it is easy to see that if $ 0 < ||\Un||_{[0,1]} < \infty$ (where $\Un$ stands for the unit function), then necessarily $\beta + 1/p \leq 0$. However, the examples (c)-(f) below show that this inequality is no more true whenever $||\Un||_{[0,1]} = 0$.
We stress that most of the usual semi-norms belong to $\UU$:
*(a) The supremum semi-norm, which is given by $$||f||^{}_I\; =\; \sup_{t\in I}|f(t)|$$ for every $I\in\II$, belongs to $\UU(0,+\infty)$.*
\(b) The $L_p$-semi-norm, $p\geq 1$, which is given by $$||f||^{}_I\; =\; {\lpa\int_I {\lva f(s)\rva}^p\, ds \rpa}^{1/p}$$ for every $I\in\II$, belongs to $\UU(-1/p,p)$.
\(c) The $\aa$-Hölder semi-norm, $0\leq \aa < 1$, which is given by $$||f||^{}_I\; =\; \sup_{s<t\in I}\frac{|f(t)-f(s)|}{{|t-s|}^{\aa}}$$ for every $I\in\II$, belongs to $\UU(\aa,+\infty)$. In particular, the oscillation semi-norm ($\aa = 0$) belongs to $\UU(0,+\infty)$. Similarly, the Calderón-Zygmund semi-norm $$||f||^{}_I\; =\; \sup_{s<t\in I}\frac{|2f((t+s)/2)-f(s) - f(t)|}{(t-s)}$$ belongs to $\UU(1,+\infty)$, and the $\aa$-Lipschitz semi-norm, $\aa > 1$, which is given by $$||f||^{}_I = \;
\sup_{s<t\in I}\frac{|f^{(n)}(t)-f^{(n)}(s)|}{{|t-s|}^{\aa-n}}$$ where $n < \aa < n+1$ (and by its Calderón-Zygmund analogue for $\aa = n$), belongs to $\UU(\aa,+\infty)$.
\(d) The (strong) $p$-variation semi-norm, $p\geq 1$, which is given by $$||f||^{}_I\; =\; {\lpa \sup_{t_0 < \ldots < t_n \in I }
\sum_{i = 1}^n {|f(t_i)-f(t_{i-1})|}^p\rpa}^{1/p}$$ for every $I\in\II$, belongs to $\UU(0,p)$.
\(e) The $(\aa, p)$-Sobolev semi-norm, $p\geq 1$ and $0\leq\aa +1/p < 1$, which is given by $$||f||^{}_I\; =\; {\lpa\int_{I}\int_{I}
{\lpa\frac{|f(t)-f(s)|}
{{|t-s|}^{\aa +1/p}}\rpa}^p ds\, dt\rpa}^{1/p}$$ for every $I\in\II$, belongs to $\UU(\aa -1/p, p)$.
\(f) The $(\aa, p, q)$-Besov semi-norm, $\aa > 0$ and $p, q \geq 1$, which is given by $$||f||^{}_I\; =\;
{\lpa\int_0^{|I|}{\lpa\frac{\omega_{p,I}(t,f)}
{t^{\aa}}\rpa}^q \frac{dt}{t}\rpa}^{1/q}$$ for every $I\in\II$, where $$\omega_{p,I}(t,f) = \sup_{|h|\leq t}{\lpa\int_{I_h}
{|f(x-h)-f(x)|}^p dx\rpa}^{1/p},$$ $I_h = \lacc x\in I\,|\, x-h\in I\racc$ and $|I|$ stands for the Lebesgue measure of $I$, belongs to $\UU(\aa-1/p, +\infty)$. Notice that the usual $(\aa, p, q)$-Besov semi-norm, where the $L_p$-semi-norm is added in the definition (see e.g. [@CKR]), does [*not*]{} belong to $\UU$ since $L_p$-semi-norms have a different self-similarity index.
It follows from Remark (b) above that the concept of upper $(\beta, p)$-semi-norm is not sharp enough for our further purposes. We need some kind of inverse property to $(D)$, in order to work with a family of [*disjoint*]{} sets of semi-norms. Let $\CC^l_K$ (resp. $\CC^0_K$) denote the class of all $l$-times continuously differentiable (resp. continuous) functions with compact support. For technical reasons, we may have to make the following assumptions on $||\cdot||$ which, again, are verified by all the classical semi-norms:
\(E) $\exists\;l\in\inte$ such that $\CC^l_K\subset \FF$. (Smooth-finiteness)
\(F) For every $I\in\II$ and $f, f_n \in\FF_I$, $$f_n\rightarrow f\; \mbox{uniformly on $I$}\;\Longrightarrow\;||f||_I\leq\limsup ||f_n||_I.\;\;\;\;\;\;\;\mbox{(Lower semi-continuity)}$$
[*Clearly, $l$-smooth-finiteness yields $||f||_I < +\infty$ for every $f$ of class $\CC^l$ and every $I\in\II$.*]{}
Let $\beta\in\R$, $p\in (0, +\infty]$ and $||\cdot||$ be a contractive, translation-invariant, smooth-finite, lower semi-continuous and $\beta$-self-similar semi-norm on $\FF$. We say that $||\cdot||$ is a [*lower $(\beta, p)$-semi-norm*]{} if it satisfies the following property:
[*(G)*]{} There exists $C_p > 0$ such that for every $a_0<\ldots<a_n\in\R$ and $f\in\CC^l_K$ verifying $f(a_0) = \cdots = f(a_n) = 0$,
$$\lacc\begin{array}{llr}
||f||^{}_{[a_0,a_n]} \;\leq\;
C_p {\lpa ||f||_{[a_0, a_1]}^p + \cdots +
||f||_{[a_{n-1},a_n]}^p\rpa}^{1/p} & \mbox{{\em if}
$p < +\infty$,} & \\
& & \mbox{\em ($p$-subadditivity)} \\
||f||^{}_{[a_0,a_n]} \;\leq\; C_{\infty}
\sup\lpa ||f||^{}_{[a_0, a_1]}, \ldots,
||f||^{}_{[a_{n-1},a_n]}\rpa & \mbox{{\em if} $p = +\infty$.}
& \end{array}
\right.$$
In the following we will denote by $\LL(\beta,p)$ the set of lower $(\beta, p)$-semi-norms, and set $\LL$ for the union of the $\LL(\beta,p)$’s. We will say that $||\cdot||$ is a [*$(\beta, p)$-semi-norm*]{} if it belongs to $\NN(\beta,p) = \UU(\beta,p)\cap\LL(\beta,p)$ and analogously we will set $\NN$ for the union of $\NN(\beta,p)$’s.
*(a) Since the constant $C_p$ may be larger than 1, it is not enough to take $n = 2$ in the definition of $p$-subadditivity, contrary to $p$-superadditivity.*
\(b) From the inequality ${(a_0 + \cdots + a_n)}^q \leq a_0^q + \cdots + a_n^q$ for every $a_0,\ldots, a_n\geq 0$ and $0 \leq q\leq 1$, it follows that $\LL(\beta,p)\subset\LL(\beta,p')$ whenever $p'\leq p$.
\(c) Again, one can ask for the possible values of the parameters $\beta$ and $p$ in the definition of a lower $(p,\beta)$-semi-norm. By contractivity, homogeneity and translation-invariance, it is easy to see that when the inequalities in (G) hold for the function $f=\Un$ and for all partitions, then necessarily $\beta + 1/p
\geq 0$. This inequality is actually true for all available examples, but we got stuck in proving that in full generality.
\(d) It is not difficult to see that $\NN(\beta, p)\cap\NN(\beta', p') =
\emptyset$ as soon as $(\beta, p)\neq(\beta', p')$.
Notice that each of the above examples (a)-(d) belongs to $\LL$, and hence to $\NN$. Actually in each case we had chosen the smallest possible parameter $p$, i.e. we could have written $||\cdot||\in\LL(\beta, p)$ as well as $||\cdot||\in\UU(\beta, p)$. This fact is trivial for examples (a)-(c) where we can take $C_p = 1$, a bit more involved in example (d) where we have to take $C_p = 2^{1- 1/p}$ as soon as $p \geq 1$ (notice that here the condition $f(a_0) = \cdots = f(a_n) = 0$ is essential).
However, $(\aa, p, q)$-Besov and $(\aa, p)$-Sobolev semi-norms do not belong to $\LL$, because in these cases the constant $C_p$ may depend on the subdivision $a_0 < \ldots < a_n$. These two examples are important in certain contexts, and this is why we would like to introduce a weaker definition, which will be given in terms of the evaluation of $||\cdot||$ along specific families of functions.
If $l\in\inte$ and $\psi\in\CC^l_K$, we introduce $\Psi=\lacc\psi_{jn}, \; n \in \z, \; j\ge 0 \racc$, the two-parametric subset of $\CC^l_K$ defined by $$\psi_{jn}(t)= \psi(2^j(t-n))$$ for every $n\in\z, j\ge 0, t\in\R$.
Let $\beta\in\R$, $p\in(0,+\infty]$ and $||\cdot||$ be a lower semi-continuous, $l$-smooth-finite semi-norm on $\FF$. We say that $||\cdot||$ is a [*lower $(\beta, p)$-semi-norm in the wide sense*]{} with respect to $\Psi$, if it satisfies the following property:
[*($\rm \tilde G$)*]{} There exists $C_p > 0$ such that for every $j\ge 0$, arbitrary $x_1,\ldots,x_m\in\R$, and arbitrary $n_1,\ldots, n_m\in\inte$ such that the supports of the functions $\{\psi_{j,n_i}, \; 1\le i\le m\}$ have disjoint interiors, $$\lacc
\begin{array}{ll}
{\lva\lva {{\displaystyle}\sum_{i=1}^{m} x_i \psi_{jn_i} }\rva\rva}_{[0,1]} \;
\leq\; C_p\, 2^{\beta j} \
{\lpa {{\displaystyle}\sum_{i=1}^{m} |x_i|^p }\rpa}^{1/p}
& \mbox{{\em if} $p < +\infty$,} \\
& \\
{\lva\lva {{\displaystyle}\sum_{i=1}^{m} x_i \psi_{jn_i} }\rva\rva}_{[0,1]} \;
\leq\; C_\infty\, 2^{\beta j} \
{{\displaystyle}\sup_{1\le i\le m} |x_i| }
& \mbox{{\em if} $p = +\infty$.}
\end{array}
\right.$$
Similarly we will denote by $\Lde(\beta,p)$ the set of lower $(\beta, p)$-semi-norms in the wide sense, and set $\Lde$ for the union of the $\Lde(\beta,p)$’s. Analogously, we define $\Nde(\beta,p) = \UU(\beta,p)\cap\Lde(\beta,p)$ and set $\Nde$ for the union of the $\Nde(\beta,p)$’s.
*(a) In the sequel, $\Psi$ will be either a family of sufficiently smooth wavelet functions, or some Schauder system on $[0,1]$.*
\(b) Condition ($\rm \tilde G$) just means that (G) holds for linear combinations of specific functions with the same “frequency”, and it is a well-known condition to obtain lower bounds for small deviation probabilities in a Gaussian framework [@St1] [@St2]. Actually we will use it for the same purposes, but sometimes in a more general context, see Section \[lower\].
\(c) At first sight, neither translation-invariance nor self-similarity are required in the definition of $\Lde(\beta,p)$. Actually, working with a specific family of function allows us to combine these necessary properties in the single inequality given by ($\rm \tilde G$).
\(d) Again we can prove that $\LL(\beta,p)\subset\LL(\beta,p')$ (resp. $\Nde(\beta, p)\cap\Nde(\beta', p') = \emptyset$) whenever $p'\leq p$ (resp. $(\beta, p)\neq(\beta', p')$).
\(e) It is a bit tedious but not difficult to see that $(\aa, p, q)$-Besov semi-norm belongs to $\Nde(\aa,+\infty)$ and that $(\aa, p)$-Sobolev semi-norm belongs to $\Nde(\aa -1/p,p)$, both with respect to the Schauder system.
[*In the following we will mainly consider functions restricted to $[0,1]$. By minor abuse of notations, and for the sake of brevity, we will set $||f|| = ||f||_{[0,1]}$ for every $f\in\FF$, although such statements like $"||\cdot||\in\UU(\beta, p)"$ or $"||\cdot||\in\LL(\beta, p)"$ will always refer to the family $||\cdot|| = \lacc ||\cdot||^{}_I, \; I\in\II\racc$. By $C$ we will always mean a positive finite constant independent of the involved parameters, and whose value may change from line to line.*]{}
Riemann-Liouville processes and their associated linear fractional stable motions
---------------------------------------------------------------------------------
Let $\lacc Z_t, \; t\geq 0\racc$ be a symmetric $\alpha$-stable process with index $\alpha\in(0,2]$, i.e. $Z$ is a Lévy process whose characteristic function is given by $$\esp\lcr e^{\rm{i} \lambda Z_t}\rcr\; =\; e^{-t|\lambda|^{\alpha}}$$ for every $t\geq 0$ and $\lambda\in\R$. It is well-known and easy to see that for every $H > 0$ the following stochastic integral: $$R^H_t\; =\;\int_0^t (t-s)^{H-1/\alpha} dZ_s,$$ is well-defined for every $t > 0$. We set $R^H_0 = 0$ and call $\lacc R^H_t, \; t\geq 0\racc$ the [*Riemann-Liouville process*]{} (in abridged form: RLP) with Hurst parameter $H$. This latter terminology is motivated by the following $H$-self-similarity property of $R^H$: for every $c > 0$ $$\lacc R^H_{ct}, \; t\geq 0\racc\;\elaw\;\lacc c^H R^H_t, \;
t\geq 0\racc.$$ When no confusion is possible, we will drop the subscript $H$ and write $R=R^H$ for the sake of brevity. Notice that $R$ has no stationary increments, unless $H = 1/\alpha$. However $R$ shares some kind of extrapolation-homogeneity, which will be important in the sequel: namely, if we set $$R_{a,t}\; =\;\int_a^t (t-s)^{H-1/\alpha} dZ_s$$ for every $t\geq a\geq 0$, then the following equality in law holds: $$\lacc R_{a,a+t}, \; t\geq 0\racc\;\elaw\;\lacc R_t, \; t\geq 0\racc.$$ Besides, $R$ shares an equally important independence property: if $\Rde_{a,t} = R_t - R_{a,t}$, then for every $a \geq 0$ the processes $\lacc R_{a,a+t}, \; t\geq 0\racc$ and $\{\Rde_{a,a+t}, \; t\geq 0\}$ are independent. These three properties are easy to see as respective direct consequences (in reverse order) of the independence, stationarity, and stability of the increments of $Z$.
Of course, $Z$ itself is an RLP with Hurst parameter $H = 1/\alpha$. Notice also that up to normalization constants, the family of Riemann-Liouville processes is closed with respect to time-integration. In particular, the $m$-times integrated Brownian motion is an RLP with parameters $\alpha=2$ and $H=m+ 1/2$.
The Riemann-Liouville process is closely related to $\lacc X^H_t,\; t\geq 0\racc$, the so-called [*linear stable fractional motion*]{} (in abridged form: LFSM) with Hurst parameter $H$. $X^H$ can be defined through the following (independent) decomposition: $X^H = R^H + \W^H$ where $\W^H_0 = 0$ and $$\W^H_t=\int_0^{+\infty} \lpa (t+s)^{H-1/\alpha} - s^{H-1/\alpha}
\rpa\; d\Zde_s$$ for every $t> 0$, $\Zde$ being an independent copy of $Z$. We call $\W^H$ the [*long memory process*]{} (in abridged form: LMP) associated to $X^H$. Notice that the stochastic integral defining $\W^H$ diverges a.s. at $+\infty$ as soon as $H\geq 1$. Hence, $X^H$ can be defined only for $H\in(0,1)$. Again, when no confusion is possible, we will drop the subscript $H$ and write $X = X^H$ (resp. $\W = \W^H$) for the sake of brevity.
$X$ is also $H$-self-similar but, contrary to $R$, does not share the above extrapolation-homogeneity property. Instead, its increments are truly stationary: $$\lacc X_{s+t} - X_s, \; t\geq 0\racc\;\elaw\;\lacc X_t, \;
t\geq 0\racc$$ for every $s \geq 0$, and $X$ is a so-called [*$H$-sssi process*]{}. We refer to the monographs [@ST] [@EM] for an extensive account on these latter processes. Notice that in the Gaussian case $\alpha = 2$, $X$ coincides with the canonical fractional Brownian motion, up to some numerical factor (see Proposition 7.2.6 in [@ST]). Moreover, an alternative definition of $X$ can then be given through the following (“well-balanced”) decomposition: $$X_t=\int_{-\infty}^{+\infty} \lpa |t+s|^{H-1/2} -
|s|^{H-1/2} \rpa\; \dot{B}(ds),$$ where $\dot{B}$ is the usual white noise (see Exercise 7.2 in [@ST]). When $\alpha < 2$, the corresponding integral $$\label{balanced}
\tilde X_t=\int_{-\infty}^{+\infty} \lpa |t+s|^{H-1/\alpha} -
|s|^{H-1/\alpha} \rpa\;\dot{Z}(ds)$$ also makes sense, where $\dot{Z}$ is the so-called symmetric $\alpha$-stable noise. However, $\tilde X$ is no more equivalent to $X$ (see Theorem 7.4.5 in [@ST]). Actually, many other “bilateral” $H$-sssi $\alpha$-stable processes can be constructed, which are all non-equivalent except in the Gaussian case (see Definition 7.4.1 and Theorem 7.4.5 in [@ST]). Hence, our unilateral LFSM process is just [*one*]{} possible stable generalization of fractional Brownian motion. However, we will restrict our study to this unilateral process, even though our results can probably be adapted to some other situations.
Notice that the process $M$ is [*smooth*]{} on $(0, +\infty)$, so that $X$ and $R$ exhibit similar local properties. In particular it is well-known (see Chapters 9-12 in [@ST]) that $R$ (resp. $X$) admits a continuous version if and only if $$\alpha = 2\;\;\;\;\;\mbox{or}\;\;\;\;\; H > 1/\alpha.$$ In Section 6 we will give some other local properties of $R$ and $X$, related to the examples of semi-norms (a)-(f) listed above. We finally refer to [@MaR] for a more thorough comparison between $R$ and $X$, in the context of econometric applications.
Existence of the small deviation constant for RLP {#main}
=================================================
We can now state the main result of this paper.
\[t31\] Let $||\cdot||\in\UU(\beta, p)$ and $R$ be an RLP with Hurst parameter $H$. Assume that $H>\beta +1/p$ and set $\gamma = (H-\beta-1/p)^{-1}$. Then there exists $\k\;\in\; (0, +\infty]$ such that $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; = \; - \k.$$
*(a) It is interesting to note that the stability index $\alpha$ does not directly show up in the expression of the small deviation rate.*
\(b) This result says nothing about the finiteness of the constant $\k$, which is of course a very important feature. To this aim, one needs to show a lower bound for small deviation probabilities with appropriate order - see Section \[lower\]. While [*a priori*]{} nothing indicates that the order $\gamma$ of small deviation probabilities we propose in Theorem \[t31\] is the right one, it is indeed correct for [*all*]{} available examples, as soon as $p$ is chosen as small as possible - see Section \[survey\].
The next two paragraphs are devoted to the proof of Theorem \[t31\] and we fix once and for all an upper $(\beta, p)$-semi-norm $||\cdot||$, and an RLP $R$ with Hurst parameter $H>\beta +1/p$. We begin in proving a kind of stochastic superadditivity property, which will be our crucial argument.
Stochastic superadditivity
--------------------------
Let $X$ and $Y$ be two real-valued random variables. We say that $X$ is [*stochastically larger*]{} than $Y$ if and only if $$\pb\lacc X \le r\racc\; \leq \; \pb \lacc Y \leq r\racc$$ for every $r\in\R$, and we write in this case $X\; \ss\; Y$. If $X$ and $Y$ are positive, then clearly $X\; \ss\; Y$ entails that $$\esp\lcr\exp -\lambda X\rcr\; \leq \;\esp\lcr\exp -\lambda Y\rcr$$ for every $\lambda \geq 0$.
\[in1\] Suppose that $p < \infty$. Let $R_1$, $R_2$ be two independent copies of $R$ and set $q = p(H-\beta) > 0$. Then for every $a, b\geq 0$ $$(a+b)^q \ ||R||^p\; \ss\; a^q\ ||R_1||^p + b^q \ ||R_2||^p.$$
[*Proof.*]{} Fix $a, b > 0$ and set $c = a + b$. We first appeal to the $p$-superadditivity of $||\cdot||$ and get $$||R||_{[0,c]}^p\; \ge\; ||R||_{[0,a]}^p + ||R||_{[a,c]}^p.$$ By self-similarity for both $||\cdot||$ and $R$, we also have $$\begin{aligned}
|| R ||_{[0,c]}^p & \elaw & ||c^H R(c^{-1}\cdot)||_{[0,c]}^p \;
=\; c^{p(H -\beta)}|| R ||_{[0,1]}^p \; =\; c^q || R ||^p\end{aligned}$$ and, similarly, $$||R||_{[0,a]}^p \;\elaw\; a^q || R_1 ||^p\;\;\; \mbox{and}\;\;\;
||R||_{[0,b]}^p \;\elaw\; b^q || R_2 ||^p.$$ Putting everything together, we now see that it is enough to show $$\label{bp1}
||R||_{[0,a]}^p + ||R||_{[a,c]}^p\; \ss\;
||R||_{[0,a]}^p + ||R_a||_{[a,c]}^p,$$ because of the translation-invariance of $||\cdot||$ and the extrapolation-homogeneity of $R$. Here and throughout this section, $R_a$ denotes the process $\{R_{a,t},\, t\ge a\}$ For every $a\geq 0$, set $\F_a$ for the filtration generated by $\lacc Z_s,\; 0 \leq s \leq a\racc$ and $\pb_a$ for the conditional probability with respect to $\F_a$. We clearly have $$\begin{aligned}
\pb\lcr\ ||R||_{[0,a]}^p + ||R||_{[a,c]}^p \leq r \rcr & = &
\esp\lcr\pb_a \lcr\ ||R||_{[a,c]}^p \leq r - ||R||_{[0,a]}^p \rcr\rcr \\
& = & \esp \lcr \pb_a \lcr ||\tilde R_a + R_a||_{[a,c]}^p \leq r -
||R||_{[0,a]}^p \rcr\rcr\end{aligned}$$ for every $r\in\R$. Since $R_a$ is a conditionally Gaussian process under $\pb_a$ (see [@ST] p.153-154), we can apply Anderson’s inequality under $\pb_a$ and get (recall that $\tilde R_a$ is $\F_a$-measurable) $$\pb_a\lcr ||\tilde R_a + R_a ||_{[a,c]}^p \le r -
||R||_{[0,a]}^p \rcr\; \leq\;\pb_a\lcr||R_a||_{[a,c]}^p \leq r -
||R||_{[0,a]}^p \rcr.$$ Now, since $R_a$ is independent of $\F_a$, we can average backwards and obtain $$\pb\lcr ||R||_{[0,a]}^p + ||R||_{[a,c]}^p \leq r\rcr\; \leq\;\pb\lcr
||R||_{[0,a]}^p + ||R_a||_{[a,c]}^p \leq r \rcr$$ for every $r\in\R$, as desired for (\[bp1\]).
The next proposition covers the simpler situation when $p=\infty$, and follows mainly the outline of Theorem 2.1 in [@LL1].
\[in2\] Suppose that $ p = +\infty$ and set $q= H - \beta > 0$. For every $a,
b\geq 0$ and $r\in\R$ $$\pb\lcr (a+b)^q ||R||\,\leq\, r \rcr\; \leq\; \pb\lcr a^q ||R||\,\leq\,r
\rcr\;\pb\lcr b^q ||R||\,\leq\, r \rcr.$$
[*Proof.*]{} Fix $a, b > 0$ and set $c = a + b$. Again we appeal to the $\infty$-superadditivity of $||\cdot||$ and get $$||R||_{[0,c]} \geq \sup\lacc ||R||_{[0,a]}, ||R||_{[a,c]} \racc.$$ Therefore, $$\begin{aligned}
\pb\lcr ||R||_{[0,c]}\leq r \racc & \leq & \pb\lcr ||R||_{[0,a]}\leq r,
||R||_{[a,c]}\leq r\racc \; =\; \esp\lcr {\bf 1}_{\lacc ||R||_{[0,a]
}\leq r \racc}
\pb_a \lcr ||R||_{[a,c]}\le r\rcr\rcr\end{aligned}$$ for every $r\in\R$, where $\pb_a$ is defined as above. By Anderson’s inequality, independence of $R_a$ with respect to $\F_a$, extrapolation-homogeneity of $R$ and translation-invariance of $||\cdot||$, we obtain $$\begin{aligned}
\pb_a \lcr ||R||_{[a,c]}\le r\racc & = & \pb_a \lcr
||\tilde R_a+R_a||_{[a,c]}\le r\rcr\;\leq\; \pb_a \lcr
||R_a||_{[a,c]}\le r\rcr \; =\;\pb \lcr||R||_{[0,b]}\le r\rcr.\end{aligned}$$ Averaging backwards, we get $$\pb\lcr ||R||_{[0,c]}\le r \rcr\; \leq\;
\pb\lcr ||R||_{[0,a]}\le r\rcr\pb\lcr||R||_{[0,b]}\le r\rcr$$ for every $r\in\R$. Now this yields the desired inequality $$\pb\lcr c^q ||R||\,\leq\, r \rcr\; \leq\; \pb\lcr a^q ||R||\,
\leq\,r \rcr\;\pb\lcr b^q ||R||\,\leq\, r \rcr,$$ since by self-similarity of $R$ and $||\cdot||$ $$\pb\lcr ||R||_{[0,t]} \leq r \rcr\; =\; \pb\lcr t^q ||R|| \leq r \rcr$$ for every $t \geq 0$ and $r\in\R$.
We are now ready to proceed to the proof of Theorem \[t31\].
Proof of Theorem \[t31\]
------------------------
Again we first consider the case $p<\infty$. Proposition \[in1\] yields the following decisive inequality for Laplace transforms: $$\esp\lcr \exp - (a+b)^q ||R||^p \rcr \;\le\; \esp\lcr \exp -a^q
||R||^p \rcr \ \esp\lcr \exp - b^q ||R||^p \rcr$$ for every $a, b \geq 0$ and $q = p(H-\beta)$ as above. This entails that the log-Laplace transformation of $||R||$ defined by $$\Psi(h) = \log \esp \lcr \exp -h^{q} ||R||^p \rcr$$ for every $h \geq 0$, is a continuous negative function which satisfies $\Psi(a+b) \leq \Psi(a) + \Psi(b)$ for every $a, b \geq 0$. By the standard subadditivity argument, we obtain $$\lim_{h\to\infty} \frac{\Psi(h)}h \; = \; \inf_{h \geq 0}
\frac{\Psi(h)}h\; = \; -C \;\in \; [-\infty, 0)$$ and, returning to the Laplace transform, $$\lim_{\lambda\to +\infty}\lambda^{1/q}\; \esp\lcr \exp -\lambda
||R||^p \rcr\; =\; -C.$$ Notice that $q>1$ by assumption. Hence we can appeal to de Bruijn’s exponential Tauberian theorem (see Theorem 4.12.9 in [@BGT], or Theorem 3.5 in [@LS] for a more comfortable formulation), which yields $$\lim_{\ee\to 0}\ee^{1/(q-1)}\;\pb\lcr ||R||^p \le \ee \rcr\;
=\; -\k \; =\; -(q-1){\lpa C/ q \rpa}^{q/(q-1)},$$ and finally $$\lim_{\ee\to 0}\ee^{\gamma}\;\pb\lcr ||R||
\le \ee \rcr\; =\; - \k \;\in \; [-\infty, 0)$$ with $\gamma = (H-\beta-1/p)^{-1}$. This completes the proof of Theorem \[t31\] when $p < +\infty$.
In the case $p= +\infty$ we do not even need Laplace transformation, since it follows directly from Proposition \[in2\] that the function $$\Psi(h) = -\log \pb\lcr ||R|| \le h^{\beta-H} \rcr$$ is subadditive. Again this entails $$\lim_{h\to\infty} \frac{\Psi(h)}h \; = \; \inf_{h \geq 0}
\frac{\Psi(h)}h\; = \; -\k \;\in \; [-\infty, 0),$$ and we obtain $$\lim_{\ee\to 0}\ee^{\gamma}\;\pb\lcr ||R|| \le \ee \rcr\; =\; -
\k \;\in \; [-\infty, 0)$$ with $\gamma = (H-\beta)^{-1}$, as desired when $p = +\infty$.
Lower bounds: finiteness of the constant for continuous RLP {#lower}
===========================================================
In this section we obtain a suitable lower bound for small deviation probabilities which will allow us to prove, under certain conditions, that the constant $\k$ from Theorem \[t31\] is actually finite whenever $||\cdot||\in \LL(\beta, p)$ as well. Unfortunately, our method is only efficient in the [*continuous*]{} case, i.e. when $$\alpha = 2\;\;\;\;\;\mbox{or}\;\;\;\;\; H > 1/\alpha.$$ An explanation for this understandable, but important restriction on $R$ will be given later. Our result reads as follows:
\[t41\] Let $||\cdot||\in\LL(\beta, p)$ and $R$ be a continuous $\alpha$-stable RLP with Hurst parameter $H$. Suppose that $H > \beta + 1/p$ if $\alpha = 2$ and $H > \beta+ 1/p+1/\alpha$ if $\alpha < 2$. Then $$\label{linf}
\liminf_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr
||R||\leq \ee \rcr
\; > \; -\infty,$$ with $\gamma = (H-\beta-1/p)^{-1}$.
Combining Theorem \[t31\] and Theorem \[t41\] yields readily the following fairly general small deviation theorem for continuous Riemann-Liouville processes:
\[c41\] Let $||\cdot||\in\NN(\beta, p)$ and $R$ be a continuous $\alpha$-stable RLP with Hurst parameter $H$. Suppose that $H > \beta + 1/p$ if $\alpha = 2$ and $H > \beta+ 1/p+1/\alpha$ if $\alpha < 2$. Then there exists $\k\in (0,\infty)$ such that $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; = \; -\k,$$ with $\gamma = (H-\beta-1/p)^{-1}$.
We stress that if the parameter $H$ is not too large, more precisely if $H<2$, then it is possible to obtain the lower bound for small deviation probabilities under a weaker assumption than $||\cdot||\in\LL(\beta, p)$. Here we just need that $||\cdot||\in\Lde(\beta, p)$ with respect to $$\Psi\; =\; \lacc\psi_{jn}=\psi(2^jt-n+1),\; 1\le
n \le 2^j,\; j\ge 0\racc,$$ the Schauder system generated by the triangular function $\psi(t)=\Un_{[0,1]} (t)(1-|2t-1|)$.
Notice that on each level $j$ the supports of the functions $\psi_{jn}$ have disjoint interiors, so that here $||\cdot||\in\Lde(\beta, p)$ w.r.t. $\Psi$ simply means that $${\lva\lva
{{\displaystyle}\sum_{n=1}^{2^j} x_n \psi_{jn} }
\rva\rva} \;
\leq\; C_p\, 2^{\beta j} {\lpa {{\displaystyle}\sum_{n=1}^{2^j}
|x_n|^p }\rpa}^{1/p}$$ for $p < +\infty$, and with an obvious modification for $p=\infty$. We have an analogous result to Theorem \[t41\]:
\[t42\] Let $||\cdot||\in\Lde(\beta, p)$ w.r.t. Schauder system $R$ be a continuous $\alpha$-stable RLP with Hurst parameter $H < 2$. Suppose that $H > \beta + 1/p$ if $\alpha = 2$ and $H >
\beta+ 1/p+1/\alpha$ if $\alpha < 2$. Then $$\begin{aligned}
\liminf_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr
||R||\leq \ee \rcr
& > & -\infty,\end{aligned}$$ with $\gamma = (H-\beta-1/p)^{-1}$.
And, of course, we get the corresponding corollary:
\[c42\] Let $||\cdot||\in\Nde(\beta, p)$ w.r.t. Schauder system and $R$ be a continuous $\alpha$-stable RLP with Hurst parameter $H < 2$. Suppose that $H > \beta + 1/p$ if $\alpha = 2$ and $H > \beta+ 1/p+1/\alpha$ if $\alpha < 2$. Then there exists $\k\in (0,\infty)$ such that $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; = \; -\k,$$ with $\gamma = (H-\beta-1/p)^{-1}$.
Decorrelating stable arrays
---------------------------
In this paragraph we prove a crucial lower bound, which is a generalization of Lemma 2.1 in [@St2] to arrays of symmetric $\alpha$-stable random variables, and which will be useful for both Theorems \[t41\] and \[t42\]. It relies on a version of Šidák’s inequality for stable variables recently obtained in this framework by G. Samorodnitsky [@Sa].
\[samo\] Let $M, h > 0$ and $\lacc y_{jn},\; 1\leq n \leq M 2^{h j}, \;
j\geq 0\racc$ be an array of identically distributed symmetric $\alpha$-stable random variables. Let $z, \delta > 0$ be such that $\delta< z$ if $\alpha = 2$ and $ \delta < z - h/\alpha$ if $\alpha < 2$. Let $m>0$ be an integer and set $$d_j\; =\; d_j(m)\; =\;
2^{z(j-m) - \delta|j-m|}.$$ Then there exists a constant $C$ depending only on $M,h, z,
\delta$ such that $$\pb\lcr |y_{jn}| \leq d_j, \; 1\leq n \leq M 2^{h j}, \; j\geq 0\rcr \;
\geq \; \exp - C 2^{h m}.$$
[*Proof*]{}. Up to normalization, the case $\alpha = 2$ is just the statement of Lemma 2.1 in [@St2]. Hence we can concentrate on the case $\alpha < 2$, and first notice that Lemma 2.1 in [@Sa] entails the following decorrelation inequality: $$\pb\lcr |y_{jn}| \leq d_j, \; 1\leq n \leq M 2^{h j}, \; j\geq 0\rcr \;
\geq \; \prod_{j\geq 0}
{\pb\lcr |y| \leq d_j\rcr}^{M 2^{h j}},$$ where $y$ is some symmetric $\alpha$-stable random variable. We can decompose the right-hand side into $$\prod_{j\geq 0} {\pb\lcr |y| \leq d_j\rcr}^{M 2^{h j}}\; =\;
\lpa\prod_{j\geq m} {\pb\lcr |y| \leq d_j\rcr}^{M 2^{h j}}\rpa
\lpa \prod_{j= 0}^{m-1} {\pb\lcr |y| \leq d_j\rcr}^{M 2^{h j}}\rpa.$$ To estimate the infinite product, we use the following well-known tail behavior of $y$ (see e.g. Property 1.2.15 in [@ST]): $$\label{tail}
\lim_{r\uparrow +\infty} r^{\alpha} \pb\lcr |y| > r\rcr\; =\; K_1\,
\in\, (0, +\infty).$$ This yields $$\begin{aligned}
\log \prod_{j\geq m} {\pb\lcr |y| \leq d_j\rcr}^{M 2^{h j}} & =
& M\;\sum_{j\geq m} 2^{h j} \log\lpa 1 - \pb\lcr |y| >
d_j\rcr\rpa\\
& \geq & -C \sum_{j\geq m} 2^{h j}\, \pb\lcr |y| >
2^{(z - \delta)(j-m)}\rcr\\
& \geq & -C \sum_{j\geq m} 2^{h j} 2^{-\alpha(z
- \delta)(j-m)}\\
& \geq & - C\, 2^{h m},\end{aligned}$$ where in the last inequality we used $h -\alpha(z -\delta) < 0$. The estimate of the finite product is even simpler. Since $y$ has a positive density in the neighbourhood of the origin, we have $$\lim_{\ee\downarrow 0} \ee^{-1} \pb\lcr |y| \leq \ee\rcr\; =\;
K_2\, \in\, (0, +\infty).$$ This entails $$\begin{aligned}
\log \prod_{j = 0}^{m-1} {\pb\lcr |y| \leq d_j\rcr}^{M 2^{h j}} &
= & M\;\sum_{j = 0}^{m-1} 2^{h j} \log\pb\lcr |y| \leq d_j\rcr\\
& \geq & -C \sum_{j = 0}^{m-1} 2^{h j}\lpa 1 \, +\, \log\lpa
2^{(z + \delta)(m-j)}\rpa\rpa\\
& \geq & -C 2^{h m}\sum_{j = 0}^{m-1} (m-j)\, 2^{h (j-m)}\\
& \geq & - C\, 2^{h m},\end{aligned}$$ where in the last inequality we used $h > 0$. Putting everything together now yields $$\pb\lcr |y_{jn}| \leq d_j, \; 1\leq n \leq M 2^{h j}, \; j\geq 0\rcr \; \geq \; \exp - C 2^{h m}$$ for a constant $C$ not depending on $m$.
[*In the Non-Gaussian case, it is easy to see that the condition $\delta< z- h/\alpha$ is also necessary, because of the heavy tails of $\alpha$-stable random variables: indeed, if the $y_{jn}$’s are mutually independent and if $z = \delta +
h/\alpha$, then it follows from (\[tail\]) that $$\pb\lcr |y_{jn}| \leq d_j, \; 1\leq n \leq M 2^{h j}, \; j\geq 0\rcr \; = \; 0.$$* ]{}
Some elements of wavelet theory {#wavelets}
-------------------------------
The proof of Theorem \[t41\] relies mainly on a suitable wavelet decomposition of $R$, which we recall here for the sake of completeness. In this paragraph we fix once and for all a semi-norm $\|.\|\in\LL(\beta, p)$ which is $\ell$-smooth-finite, and $R$ a continuous RLP satisfying the assumptions of Theorem \[t41\].
There exist two functions $\varphi, \psi\in\CC^{\ell}_K$ (“wavelet parents”) such that $\psi$ has vanishing moments up to order $\ell$: $$\label{mom}
\int_{-\infty}^{\infty} t^k \psi(t)\, dt\; =0\; \qquad 0\le k \le \ell,$$ and such that the wavelet functions $\lacc \psi_{jn},\; n\in\z,\; j>0\racc$ and $\lacc \varphi_{n}, \; n\in\z\racc$, respectively defined by $$\psi_{jn}(t)\;= \; 2^{j/2}\psi(2^jt-n)\;\;\;\;\;
\mbox{and}\;\;\;\;\;\varphi_{n}(t)\; =\;\varphi(t-n),$$ form an orthogonal base of $L_2(\R)$. We refer to Daubechies’ construction [@Dau], Section 6.4 for the definition and Section 7.1 for smoothness properties of these compactly supported wavelets. A useful book is also [@M].
Fix $[-D,D]$ an interval containing the support of $\psi$. Consider $I,$ the integration operator on compactly supported functions: $$If(t)\;= \;\int_{-\infty}^t f(s)\, ds$$ for every such $f$ and $t \in \R$. We set $I^0$ for the identity operator and $I^k$ for the $k$-th iteration of $I$, $k\geq 1$. Since for every $k\geq 0$ $$I^{k+1}f(t)\;= \;\frac{1}{k!}\int_{-\infty}^t f(s) (t-s)^k\, ds,$$ the moment condition (\[mom\]) on $\psi$ entails that the functions $I^k\psi$ are also supported by the interval $[-D,D]$ for $0\le k\le \ell$. In particular, $$\label{border}
I^k\psi(\pm D)\; = \;0, \qquad
0\le k\le \ell.$$ It is well-known that under our assumptions on $H$, the process $\Un_{[0,D+1]} R$ belongs a.s. to $L_2(\R)$ (see [@ST] Chapter 11). Hence, we can write its wavelet decomposition: $$\label{ltwo}
\Un_{[0,D+1]} R\; =\; \sum_{j,n} r_{jn} \psi_{jn} + \sum_{n} r_n
\varphi_{n},$$ where $$r_{jn}\;=\; \int_0^{D+1} R(s) \psi_{jn}(s)\, ds \;\;\;\;\;\mbox{and}
\;\;\;\;\; r_n\;=\; \int_0^{D+1} R(s) \varphi_{n}(s) ds.$$ Actually more can be said about the convergence of the series on the right-hand side in (\[ltwo\]). Namely, again from our assumption on $H$, it is well-known that $R$ is locally $\eta$-Hölder for some $\eta > 0$ [@KM] [@Tak], and in particular the function $R$ coincides on $[-D,2D+1]$ with an $L_2$-function which is [*globally*]{} $\eta$-Hölder. We know then (see e.g. Theorem 7 in [@M], Chapter 6) that the series (\[ltwo\]) converges to $R$ [*uniformly*]{} on $[0,1]$.
Besides, in (\[ltwo\]) we can delete each $\psi_{jn}$ and $\varphi_n$ whose support does not overlap with $[0,1]$, and the remaining series still converges to $R$ uniformly on $[0,1]$. Since $\varphi$ and $\psi$ have compact support, this leads to a decomposition of $R$ into a functional array of exponential size: there exists a constant $M$ depending only on $D$ such that $$R\; =\; \sum_{j\geq 1}\lpa\sum_{|n|\leq M2^j} r_{jn}
\psi_{jn}\rpa\; + \;\sum_{|n|\leq M} r_n \varphi_{n}.$$
Proof of Theorem \[t41\]
------------------------
We may (and will) suppose that $\ell > H$. For every $n\in\z$ we will set $\psi_{0n} = \phi_n$ and $r_{0n} = r_n$ for the sake of concision. By lower semi-continuity and triangle’s inequality, we clearly have $$\begin{aligned}
||R|| & \leq & \sum_{j\geq 0}\lva\lva \sum_{|n|\leq M2^j} r_{jn}
\psi_{jn}\rva\rva.\end{aligned}$$ Besides, since the parent functions have support in $[-D,D]$, for each $j\geq 0$ we can split the family $$\lacc \psi_{jn}, \; |n|\leq M2^j \racc$$ into $2D$ subfamilies such that in each subfamily the functions have supports with disjoint interiors. Suppose first that $p < +\infty$. Condition (G) yields $$\label{ineqr}
||R|| \; \leq \; C \,\sum_{j\geq 0}{\lpa\sum_{|n|\leq M2^j}
|r_{jn}|^p ||\psi_{jn}||^p\rpa}^{1/p}.$$ On the one hand, by homogeneity, $\beta$-self-similarity, contractivity and smooth-finiteness of $||\cdot||$, it is easy to see that for every $j,n$, $$\label{smoopsi}
||\psi_{jn}||\; \leq \; C\, 2^{(1/2 + \beta)j}$$ for some constant $C$ independent of $j, n$. On the other hand, if $j\geq 1$, $$\begin{aligned}
\label{rjn}
r_{jn} &=& 2^{j/2}\int_0^{+\infty}\!\!\!\! R(s)\psi(2^j s - n)\, ds
\nonumber\\
&=& 2^{-j/2} \int_0^{+\infty}\!\!\!\! R(2^{-j} s)\psi(s - n)\,
ds\nonumber\\
&\elaw& 2^{-(H + 1/2)j} \int_0^{+\infty}\!\!\!\! R(s)\psi(s - n)\,
ds\nonumber\\
& = & 2^{-(H + 1/2)j} \int_{\R} R(s + n)\psi(s)\, ds,\end{aligned}$$ where in the penultimate inequality we appealed to the $H$-self-similarity of $R$. Plugging (\[smoopsi\]) and (\[rjn\]) into (\[ineqr\]) yields $$\label{ineqr2}
||R|| \; \leq \; C \,\sum_{j\geq 0} 2^{-(H-\beta-1/p)j}\lpa
\sup_{|n|\leq M2^j} |r_{jn}'|\rpa,$$ where $r_{jn}'$ stand for the [*renormalized*]{} wavelet coefficients of $R$, viz. $$\label{reno}
r_{0n}'\; =\; r_{0n} \;\;\;\;\mbox{and}\;\;\;\;
r_{jn}'\;\elaw\;\int_{\R} R(s + n)\psi(s)\, ds\;\;\mbox{if $j\geq 1.$}$$ Notice that one can write the wavelet coefficients from (\[reno\]) in the following integral form: if $j\ge 1$, then $r_{jn}' = \tau_{n} y_{jn}$, where $$\begin{aligned}
\tau_{n}^{\alpha} & = & \int_0^{+\infty} {\lpa \int_{\R} {\lpa u + n - s
\rpa}_+^{H'}\psi(u) \; du\rpa}^{\alpha}\; ds\end{aligned}$$ $H'=H-1/\alpha$, and $\lacc y_{jn}, \; |n| \leq M 2^j, \; j\geq 0\racc$ is an array of identically distributed symmetric $\alpha$-stable random variables. We first aim to prove $$\label{sup}
\sup_{n} \tau_{n}\; < \; +\infty.$$ To get this uniform bound, we first recall that $\psi$ has its support in $[-D, D]$, so that the integral defining $\tau_{n}^{\alpha}$ can be rewritten as follows: $$\begin{aligned}
\tau_{n}^{\alpha} & = & \int_0^{n + D} {\lpa \int_{-D}^D {\lpa u + n - s
\rpa}_+^{H'}\psi(u) \; du\rpa}^{\alpha}\; ds.\end{aligned}$$ We cut the domain of integration over $s$ into $[0, n-2D]$ and $[n-2D, n +D]$. The first integral is given by $$I^1_{n}\; =\; \int_0^{n - 2D} {\lpa \int_{-D}^D {\lpa u + n - s
\rpa}^{H'}\psi(u) \; du\rpa}^{\alpha}\; ds.$$ We first transform $$\int_{-D}^D {\lpa u + n - s \rpa}^{H'}\psi (u)\, du$$ through $\ell$ successive integrations by parts. Recalling (\[border\]), we see that each time the border terms $I^m\psi(\pm D)$, $0\leq m\leq \ell$, vanish. In the end, we obtain $$\begin{aligned}
\label{ijn1}
I^1_{n}& = & C\,\int_0^{n - 2D} {\lpa \int_{-D}^D {\lpa u + n - s \rpa}^{H'
- \ell}I^{\ell}\psi(u) \; du\rpa}^{\alpha}\; ds\nonumber\\
& \leq & C\,\int_0^{n - 2D} {\lpa D + n - s \rpa}^{\alpha(H' - \ell)}\;
ds\nonumber\\
& \leq & C\,\int_D^{+\infty} {s}^{\alpha(H - \ell) -1}\; ds,\end{aligned}$$ where in the first inequality we used $\ell>H> H'$. The integral on the right-hand side is clearly finite and independent of $n$.
The integral over the second domain is given, after a change of variable, by $$\begin{aligned}
\label{ijn2}
I^2_{n}= \int_0^{3D} {\lpa \int_{-D}^D {\lpa u + 2D - s\rpa}_+^{H'}
\psi(u) \; du\rpa}^{\alpha} ds &
\leq & C\, \int_0^{3D} {\lpa \int_0^{3D}u^{H'}\, du\rpa}^{\alpha} ds.\end{aligned}$$ Since $H' > -1/2$, the integral on the right-hand side is again finite and independent of $n$. Putting (\[ijn1\]) and (\[ijn2\]) together yields (\[sup\]) as desired.
The following upper bound on $||R||$ is a direct consequence of (\[ineqr2\]) and (\[sup\]): $$\label{iner}
||R||\; \leq\; C\sum_{j\geq 0} 2^{-(H-\beta - 1/p)j}\lpa\sup_{1\leq
|n|\leq M 2^j} |y_{jn}|\rpa,$$ where $\lacc y_{jn}, \; |n| \leq M 2^j, \; j\geq 0\racc$ is an array of identically distributed symmetric $\alpha$-stable random variables on $\R$.
The end of the proof is now standard and follows [@St2], Theorem 3.1. Choose $\delta > 0$ such that $$\lacc\begin{array}{ll}
\delta <H-\beta-1/p & \mbox{if $\alpha=2$,}\\
\delta <H-\beta-1/p - 1/\alpha & \mbox{if $\alpha < 2$.}
\end{array} \right.$$ Let $m$ be a positive integer. Set $$\label{dj}
d_j\; =\; 2^{(H-\beta-1/p) (j-m) - \delta |j - m|}$$ for every $j\geq 0$. On the one hand, it is clear from (\[iner\]) and (\[dj\]) that $$\lacc |y_{jn}| \leq d_j, \; 1\leq |n|\leq M 2^j,\; j\geq 0\racc\;
\subset\; \lacc ||R||\, \le \,C(\delta)\, 2^{-(H-\beta-1/p)m} \racc.$$ On the other hand, it follows from Lemma \[samo\] (with $h = 1$) that $$\pb\lcr |y_{jn}| \leq d_j, \; 1\leq |n|\leq M 2^j,\; j\geq 0\rcr\;
\geq\;\exp - C 2^m .$$ Since $H-\beta-1/p=1/\gamma$, we obtain $$\liminf_{m\to \infty} 2^{-m}\log\pb\lcr ||R||\leq C(\delta)
2^{-m/\gamma} \rcr \; > \; -\infty$$ which is equivalent to $$\liminf_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; > \; -\infty,$$ as desired. The proof is complete in the case $p < \infty$. The case $p = \infty$ can be handled exactly in the same way, replacing (\[ineqr\]) by $$\begin{aligned}
||R|| & \leq & C \,\sum_{j\geq 0} \lpa \sup_{|n|\leq M2^j}|r_{jn}| ||\psi_{jn}||\rpa.\end{aligned}$$
*(a) It is clear from the above proof that the assumption $\|.\|\in\LL(\beta, p)$ is not necessary to get the lower bound. We just need $\|.\|\in\Lde(\beta, p)$ w.r.t. $\Psi$, a wavelet family generated by parents $\{\varphi, \psi\}$ that are smooth enough. But the required smoothness depends on the parameter $H$, and this would lead to much heavier notations. For this reason we prefer stating Theorem \[t41\] in this form, save for the loss of generality.*
\(b) There is at least an example where $\Lde(\beta, p)$ is really more relevant than $\LL(\beta, p)$. It is well-known (see [@M] Chapter 6) that $(\aa, p, q)$-Besov semi-norms, $\aa > 0$ and $p, q
\geq 1$, are equivalent to sequential norms on the wavelet coefficients. More precisely, $$\left\| \sum_n x_n \phi_{n}+
\sum_{j,n} x_{jn} \psi_{jn}\right\| \; \sim\;
\left( \sum_n |x_n|^p \right)^{1/p}
\; + \; \left\| \lacc 2^{(\aa-1/p)j} \left(
\sum_{n} |x_{jn}|^p
\right)^{1/p}\racc_{j\ge 1}
\right\|_{\ell_q}.$$ Hence, with our notations, it is clear that the $(\aa, p, q)$-Besov semi-norm belongs to $\Lde(\aa -1/p, p)$ w.r.t. any wavelet basis $\Psi$. Hence, when $p = +\infty$, Theorem \[t31\] yields the existence of the constant for $R$ with $\gamma \;=\; (H-\aa)^{-1}$. Note that this rate is in accordance with the results of [@St2], which covered the range of parameters $\alpha=2$ and $0<H<1$ (and with no condition on $p$).
\(c) In the Gaussian case, the smooth-finiteness of the semi-norm is obtained for free when $\|R\|$ is a.s. finite. Indeed, if $\H$ denotes the reproducing kernel Hilbert space associated with $R$, then it is well-known that $\H$ contains functions with $(H+1/2)$-th derivative in $L_2$. On the other hand, from a 0-1 law for Gaussian measures on linear spaces, the finiteness of $\|R\|$ with positive probability yields that $\|f\|<\infty$ for every $f\in \H$. Hence, $\|.\|$ is $\ell$-smooth-finite as soon as $\ell>H+1/2$. In the non-Gaussian case, the challenging question whether $\|R\|<\infty$ a.s. implies the smooth-finiteness of the semi-norm remains open.
\(d) For the first efficient use of wavelet methods in similar problems to small deviation probability, we refer to [@AT] where the optimal finite-dimensional approximation of fractional Brownian motion is considered.
Proof of Theorem \[t42\]
------------------------
The outline of the proof is the same as that for Theorem \[t41\], except that we use Schauder system and provide new estimates for the corresponding coefficients. Recall that from our assumptions, $0 < H < 2 $ in the Gaussian case and $1/\alpha<H<2$ in the non-Gaussian case. Again we set $H' = H - 1/\alpha$. We exclude the case $H' = 0$ where $R$ is Brownian motion: the result follows then directly from Theorem 1 in [@St1].
Since $R$ is continuous, we can decompose it on $[0,1]$ along the Schauder system: $$R_t\; =\; \sum_{j\geq 0} \RC_j(t) + R_1 t,$$ for every $t\in [0,1]$. Here, for every $j\geq 0$, $$\RC_j(t)\; =\; \sum_{n=1}^{2^j} r_{jn}\,\psi_{jn}(t)$$ where we set $\tk = (n-1)2^{-j}$, $\tkk = (n-1/2)2^{-j}$, $\tkkk =n2^{-j}$ and $$r_{jn} \; =\; 2 R_{\tkk} - R_{\tk} - R_{\tkkk}.$$ We first suppose that $p<\infty$. Condition ($\rm \tilde G$) entails $$\label{inr}
||\RC_j||\; \leq\; C \ 2^{j\beta}
{\lpa \sum_{n=1}^{2^j}{|r_{jn}|}^p\rpa}^{1/p}.$$ Notice that the coefficients $r_{jn}$ can be rewritten as $r_{jn}= \sigma_{jn} y_{jn}$, where $$\begin{aligned}
\label{sjnr}
\sigma_{jn}^{\alpha} & = & \int_0^{+\infty} {\lva 2
{\lpa \tkk - s\rpa}_+^{H'} - {\lpa\tk - s\rpa}_+^{H'}
- {\lpa\tkkk - s \rpa}_+^{H'} \rva}^{\alpha}\; ds\end{aligned}$$ (we use the notation $u_+ = u\wedge 0$ for every $u\in\R$), and $\lacc y_{jn},\; 1\leq n \leq 2^j, \;
j\geq 0\racc$ is an array of identically distributed symmetric $\alpha$-stable random variables. For each $
1\leq n \leq 2^j, \; j \geq 0$, we will now give an upper bound on $\sigma_{jn}$ depending only on $j$.
If $n = 1$, then $$\begin{aligned}
\sigma_{j1}^{\alpha} & = & \int_0^{2^{-j}} {\lva 2 {\lpa 2^{-(j+1)}
- s\rpa}_+^{H'} - {\lpa 2^{-j} - s \rpa}^{H'} \rva}^{\alpha}\; ds\;
\leq \; C\, \int_0^{2^{-j}} s^{\alpha H'}\; ds\; = \; C 2^{-\alpha H j}.\end{aligned}$$ If $n > 1$, then we set $\tkl = (n-2)2^{-j}\geq 0$ and cut the domain of integration in (\[sjnr\]) into $[0,\tkl]$ and $[\tkl, +\infty)$. Reasoning as above, we see that the second integral $$\begin{aligned}
I^2_{jn} & = & \int_{\tkl}^{\tkkk} {\lva 2 {\lpa \tkk - s\rpa}^{H'}_+
- {\lpa\tk - s\rpa}^{H'}_+ - {\lpa\tkkk - s \rpa}^{H'} \rva}^{\alpha}\; ds,\end{aligned}$$ is bounded from above by $C\, 2^{-\alpha H j}$. We estimate the first integral as follows: $$\begin{aligned}
I^1_{jn} & = & \int_0^{\tkl} {\lva 2 {\lpa \tkk - s\rpa}^{H'} -
{\lpa\tk - s\rpa}^{H'} - {\lpa\tkkk - s \rpa}^{H'} \rva}^{\alpha}\; ds
\\
&=& 2^{-j} \int_2^{n} {\lva 2 {\lpa \tkk - \tkkk+2^{-j}u\rpa}^{H'} -
{\lpa\tk - \tkkk+2^{-j}u \rpa}^{H'} -
\lpa 2^{-j}u \rpa^{H'} \rva}^{\alpha}\; du
\\
&=& 2^{-\alpha Hj} \int_2^{n} {\lva 2 {\lpa u-1/2 \rpa}^{H'} -
{\lpa u-1 \rpa}^{H'} -
{u}^{H'} \rva}^{\alpha}\; du
\\
&\le& C\ 2^{-\alpha Hj} \int_2^{\infty} u^{(H'-2)\alpha}\; du
\\
&=& C\ 2^{-\alpha Hj},\end{aligned}$$ where the last equality comes from $H<2$. Putting everything together yields now $$\sigma_{jn}\; \leq \; C\, 2^{-H j}$$ and, recalling (\[inr\]), $$\label{inr2}
||\RC_j||\; \leq\; C 2^{(\beta + 1/p - H)j}\lpa
\sup_{1\leq n\leq 2^j} |y_{jn}|\rpa.$$ By lower semi-continuity and triangle’s inequality, this entails $$\begin{aligned}
||R||\; \leq\; C \lpa |R_1|\; +\; \sum_{j\geq 0}
2^{(\beta + 1/p - H)j}\lpa\sup_{1\leq |n|\leq M 2^j} |y_{jn}|\rpa\rpa\end{aligned}$$ and since $R_1$ has a symmetric $\alpha$-stable law, we can finish the proof as in Theorem \[t41\].
*(a) In the latter proof, the estimate on $I^1_{jn}$ becomes too crude when $H\geq 2$. This is the principal reason why we need to introduce smoother wavelets in Theorem \[t41\].*
\(b) As we said before, the proofs of Theorems \[t41\] and \[t42\] work only for continuous processes. The main reason for this comes from Lemma \[samo\]. On the one hand, the exclusion of the boundary value $z = \delta + h/\alpha$ in the non-Gaussian case cancels important case $H = 1/\alpha$. On the other hand, Šidák’s inequality makes us handle the $\|\RC_j\|$’s as if they were [*independent*]{}. But if $R$ is discontinuous, then it is possible that its jumps have a significant influence on every level $\RC_j$, so that our estimate of the series $\sum_j \|\RC_j\|$ may not be realistic anymore.
\(c) In the Non-Gaussian case and when $H < \beta + 1/p + 1/\alpha$, it is easy to see that the statements of Theorems \[t41\] and \[t42\] are false, even if $H > \beta + 1/p$. Suppose for example that $H < 1/\alpha$. It is well-known (see [@ST] Chapter 10) that a.s. $$\sup_{0\leq t\leq 1} |R_t|\; = \; +\infty,$$ so that Theorems \[t41\] and \[t42\] cannot hold with respect to the supremum norm (i.e. $\k = +\infty$ in Theorem \[t31\]), although here $\beta = 0$ and $p = +\infty$. Similarly, if $H = 1/\alpha$, then clearly $$\sup_{0\leq s <t\leq 1}\lpa \frac{|R_t - R_s|}{{|t-s|}^{\eta}}\rpa\;
= \; +\infty\;\;\;\;\mbox{a.s.}$$ for every $\eta > 0$, so that Theorems \[t41\] and \[t42\] cannot hold with respect to any $\eta$-Hölder semi-norm ($\beta =\eta,
p = +\infty$). It is natural to conjecture that these two theorems remain true in full generality when $H = \beta + 1/p + 1/\alpha$, but in view of the above Remark (b), we probably need different methods.
Lower bounds for LMP: Finiteness of the constant for continuous LFSM {#motion}
====================================================================
Our aim in this section is to extend Theorem \[t42\] to $X$, the continuous LFSM which we defined in Section 2. Recall that $X$ admits an independent decomposition $X = R + M$ where $R$ (resp. $M$) is an RLP (resp. an LMP) with the same parameters. The following theorem can be viewed as a generalization of Lemma 3.2 in [@LL1].
\[t51\] Let $||\cdot||\in\Lde(\beta, p)$ w.r.t. the Schauder system and $M$ be an LMP with parameters $\alpha\in (1, 2]$, $H\in (1/\alpha,1)$. Suppose that $H > \beta + 1/p$ if $\alpha = 2$ and $H > \beta+ 1/p+1/\alpha$ if $\alpha\neq 2$. Then $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||M||\leq \ee \rcr
\; = \; 0,$$ where $\gamma=(H-\beta-1/p)^{-1}$.
Using Theorems \[t31\] and \[t42\] along with the elementary independence argument developed in [@LL1] p. 1334, Theorem \[t51\] yields readily the desired small deviation theorem for $X$:
\[t52\] Let $||\cdot||\in\Nde(\beta, p)$ w.r.t. the Schauder system and $X$ be a continuous LFSM with parameters $\alpha\in (1, 2]$, $H\in (1/\alpha,1)$. Suppose that $H > \beta + 1/p$ if $\alpha = 2$ and $H > \beta+ 1/p+1/\alpha$ if $\alpha\neq 2$. Then $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||X||\leq
\ee \rcr
\; = \; \lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||R||\leq \ee \rcr
\; = \; -\k$$ where $\k\in (-\infty, 0),
\gamma = (H-\beta-1/p)^{-1}$ and $R$ is the RLP associated to $X$.
We now proceed to the proof of Theorem \[t51\], which is quite analogous to the proof of Theorem \[t42\], except that our estimates on the Schauder coefficients will not be uniform in time-argument since $M$ has a singularity at 0.
Proof of Theorem \[t51\]
------------------------
Again we set $H' = H- 1/\alpha\in (-1,1)$ and exclude the trivial situation when $H' = 0$. Since $M$ is continuous, we can decompose it on $[0,1]$ along the Schauder system: for every $t\in [0,1]$ $$M_t\; =\; \sum_{j\geq 0} \MC_j(t) + M_1 t,$$ where $$\MC_j(t) = \sum_{n=1}^{2^j} m_{jn}\,\psi_{jn}(t)$$ and $m_{jn} \; =\; 2 M_{\tkk} - M_{\tk} - M_{\tkkk}$ as in the proof of Theorem \[t42\]. We first consider the case $p < +\infty$. Condition ($\rm \tilde G$) entails $$\label{inm}
||\MC_j||\; \leq\; C 2^{\beta j}{\lpa
\sum_{n=1}^{2^j}{|m_{jn}|}^p\rpa}^{1/p}$$ and we just need to evaluate the coefficients $m_{jn} = \sigma_{jn} y_{jn}$, where $$\begin{aligned}
\sigma_{jn}^{\alpha} & = & \int_0^{+\infty} {\lva 2 {\lpa s +
\tkk\rpa}^{H'} - {\lpa s + \tk\rpa}^{H'} -
{\lpa s + \tkkk\rpa}^{H'} \rva}^{\alpha}\; ds
\\
&=& 2^{-H\alpha j} \int_0^{+\infty} {\lva 2 {\lpa u + n- 1/2 \rpa}^{H'}
- {\lpa u+n-1\rpa}^{H'} - {\lpa u + n\rpa}^{H'} \rva}^{\alpha}\; du,\end{aligned}$$ and $\lacc y_{jn},\; 1\leq n \leq 2^j, \; j\geq 0\racc$ is an array of identically distributed symmetric $\alpha$-stable random variables. Notice first that from the so-called Cooper’s formula $$2 {\lpa u + n- 1/2 \rpa}^{H'} - {\lpa u+n-1\rpa}^{H'} -
{\lpa u + n\rpa}^{H'} =
- H'(H'-1)
\int_0^{1/2} \int_{u+n-1+\theta}^{u+n-1/2+\theta} v^{H'-2} dv d\theta$$ Clearly, since $H' < 2$, this entails $${\lva 2 {\lpa u + n- 1/2 \rpa}^{H'} - {\lpa u+n-1\rpa}^{H'} -
{\lpa u + n\rpa}^{H'} \rva} \le
C \min \left\{ (u+n-1)^{H'-2} ,1 \right\}.$$ We now fix some $h \in (0,1)$ and first estimate $\sigma_{jn}$ when $2^{h j} < n \leq 2^j$: $$\begin{aligned}
\sigma_{jn}^{\alpha} & \le & C 2^{-H\alpha j}
\int_0^{+\infty} (u+n-1)^{\alpha(H'-2)} \; du
\nonumber \\
&\le& C 2^{-H\alpha j} (n-1)^{\alpha(H'-2)+1}
\nonumber \\
&=& C 2^{-H\alpha j} 2^{-h\alpha(2-H)j}\ .
\label{sig1}\end{aligned}$$ The estimate of $\sigma_{jn}$ when $1 \leq n\leq 2^{h j}$ is even simpler: $$\begin{aligned}
\sigma_{jn}^{\alpha} & \le & C 2^{-H\alpha j}
\lpa
\int_1^{+\infty} u^{\alpha(H'-2)} \; du
+
\int_0^{1} C \; du \rpa\;\le\; C 2^{-H\alpha j} .
\label{sig2}\end{aligned}$$ Recalling (\[inm\]), we get $$\begin{aligned}
||\MC_j|| &\leq& C \, 2^{\beta j} {\lpa
\sum_{n=1}^{2^{h j}} \sigma_{jn}^p +
\sum_{n = 2^{h j}+1}^{2^j} \sigma_{jn}^p
\rpa}^{1/p} \sup_{1\le n\le 2^j} |y_{jn}|
\\
&\le& C\, 2^{-(H-\beta) j} {\lpa
2^{h j} + 2^{j} 2^{-h(2-H)pj}
\rpa}^{1/p} \sup_{1\le n \le 2^j} |y_{jn}|.\end{aligned}$$ The balance is attained at $h=(1+(2-H)p)^{-1}\in (0,1)$, whence $$\label{inm2}
||\MC_j|| \leq
C \, 2^{-(H-\beta-h/p) j} \sup_{1\le n \le 2^j} |y_{jn}|.$$ Using (\[inm2\]) instead of (\[inr2\]), we can proceed as in the proof of Theorem \[t42\] and obtain $$\liminf_{\ee\downarrow 0}\ee^{\gamma'}\log\pb\lcr ||M||\leq \ee \rcr
\; > \; -\infty,$$ with $\gamma'=(H-\beta-h/p)^{-1}< \gamma$. In particular $$\liminf_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||M||\leq \ee \rcr
\; = \; 0,$$ which completes the proof of Theorem \[t51\] for $p<\infty$.
The case $p = +\infty$ requires more careful estimates. Condition ($\rm \tilde G$) entails $$\label{inm5}
||\MC_j||\; \leq\; C 2^{\beta j}\,\sup_{1\leq n\leq 2^j} |m_{jn}|.$$ Again we fix some $h\in(0,1)$. Using (\[inm5\]), (\[sig1\]), and (\[sig2\]) we can write, $$\begin{aligned}
||\MC_j||\; &\leq&\; C\, 2^{\beta j}
\sup_{1\le n \le 2^j} \sigma_{jn} |y_{jn}|
\\
&\leq& C\, 2^{-(H-\beta) j} \lpa
\sup_{1\le n \le 2^{hj}} |y_{jn}| +
2^{-h_1 j}\sup_{2^{hj} < n \le 2^j} |y_{jn}| \rpa,\end{aligned}$$ where we set $h_1=(2-H)h$. For every integer $m>0$ we focus on the event $$\Omega_m\; = \;\lacc |y_{jn}| \leq d_{jn}(m), \; 1\leq n \leq
2^j, \; j \geq 0\racc$$ where we set $m_1=[m/h]$, $\delta\in (0, H-\beta)$ and $$d_{jn}(m)\; =\; \lacc\begin{array}{ll}
2^{(H-\beta)(j-m_1)-\delta|j-m_1|}& \mbox{if $1\leq n \leq 2^{hj}$}\\
2^{(H-\beta+h_1)(j-m)-\delta|j-m|}& \mbox{if $2^{hj} < n \leq 2^{j}.$}
\end{array}\right.$$ Take now $h$ small enough such that $$H-\beta < H- \beta + h_1 < (H -\beta)/h.$$ On the one hand, it is clear that $$\label{inm3}
\Omega_m\; \subset\; \lacc \sum_{j\ge 0} ||\MC_j||\; \leq\;
C 2^{-(H-\beta+h_1)m} \racc.$$ On the other hand, it follows from decorrelation argument of Lemma \[samo\] that $$\label{inm4}
\log\pb\lcr \Omega_m\rcr\; \geq
- \, C\, \lpa 2^{h m_1} + 2^{m} \rpa\; \geq\; -\, C \, 2^m.$$ Using (\[inm3\]) and (\[inm4\]), we can now finish the proof exactly as in the case $p < +\infty$.
*(a) Contrary to [@LL1], the above proof does not require any entropy argument and relies only on an elementary estimate of the Schauder coefficients.*
\(b) It would be quite interesting to calculate the optimal rate $$\gamma_0\; =\; \inf\lacc\gamma > 0\;\lva\; \lim_{\ee\downarrow
0}\ee^{\gamma}\log\pb\lcr ||M|| \leq \ee \rcr\; = \; 0\racc.\right.$$ We did not study this question in details but it seems plausible that $\gamma_0 = 0$: indeed, $M$ is a $\CC^{\infty}$ process on $(0, +\infty)$, and one may expect subexponential rate for its small deviation probabilities.
Examples: particular semi-norms {#survey}
===============================
In this section, we place all our results in the context of the previous literature, and show what is new. We try to be as exhaustive as possible as far as RLP’s and LFSM’s are concerned. We refer to the surveys [@LS] [@Lf] for more information about other processes and for further references.
Everywhere $R$ (resp. $X$) will be an RLP (resp. an LFSM) with Hurst parameter $H > 0$ and stability index $\alpha\in (0,2]$, whereas $\|.\|$ will be a semi-norm in $\NN(\beta, p)$ or $\Nde(\beta, p)$ for some $\beta\in\R$ and $p\in (0, +\infty]$. We always suppose that $H > \beta + 1/p$ and set $$\gamma = (H-\beta-1/p)^{-1}$$ for our small deviation rate. $\k$ will stand for the small deviation constant appearing on the right-hand side of Theorem \[t31\]. Sometimes we will call them just “the rate” and “the constant”. Let us begin with the most classical case, which deserves of course a particular mentioning.
Brownian motion
---------------
Brownian motion $B$ is an RLP (or an LFSM) with parameters $H=1/2$ and $\alpha=2$, hence it clearly satisfies the assumptions of Corollary \[c42\] (or Theorem \[t52\]). Notice that here our Theorem \[t42\] just amounts to Stolz’s lower bound criterion (Theorem 1 in [@St1]), so that in the present paper the novelty comes only from Theorem \[t31\]. Corollary \[c42\] entails the existence of Brownian small deviation constants under almost all the classical semi-norms (with the regrettable exception of certain Besov semi-norms for which we were unable to prove the appropriate superadditivity index). In other words, we get $$\lim_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr ||B||\leq \ee \rcr
\; =\; - \k\;\in\; (-\infty, 0),$$ where the dependence of $\gamma$ on the semi-norm $||\cdot||$ is given by the following table:
$||\cdot||$ $\gamma$
-------------------------- ---------------
Supremum 2
$L_p$ 2
$\aa$-Hölder $2/(1- 2\aa)$
$p$-variation $2p/(p- 2)$
$(\aa, p)$-Sobolev $2/(1- 2\aa)$
$(\aa, \infty, q)$-Besov $2/(1- 2\aa)$
The existence of a finite small deviation constant for Brownian motion under $L_p$ and supremum norms is a very classical result, and in this situation the constant $\k$ can even be more or less explicitly computed (see [@LS] and the references therein). The existence of the constant under Hölder semi-norms was first proved by Baldi and Roynette [@BR], and it is actually a direct consequence of Stolz’s upper bound criterion (Theorem 2 in [@St1]) and of the standard subadditivity argument involving the Markov property of Brownian motion. In the three last examples, Stolz’s results do provide the right rate, viz. $$-\infty \; < \; \liminf_{\ee\downarrow 0}\ee^{\gamma}\log\pb\lcr
||B||\leq \ee \rcr \; \leq \; \limsup_{\ee\downarrow 0}
\ee^{\gamma}\log\pb\lcr ||B||\leq \ee \rcr \; < \; 0$$ with the above corresponding $\gamma$, but here the existence of the limit does not follow directly. In these three cases, our result seems to be new.
It is of course quite tantalizing to [*compute*]{} these new Brownian constants. As a rule, this kind of (hard) problems requires analytical techniques depending heavily on the choice of the semi-norm, and it seems difficult to tackle them through a global study.
We stress that the above table matches quite accurately the sample path properties of Brownian motion. It is well-known that $B$ is $\aa$-Hölder if and only if $\aa < 1/2$ and has finite $p$-variation if and only if $p > 2$. Besides, it has finite $(\aa, p)$-Sobolev semi-norm if and only if $\aa < 1/2$ [@CHS].
Besov semi-norms deserve a special remark since no necessary and sufficient conditions seem to be available. Roynette [@Ro] had proved that Brownian motion has finite $(\aa, p, q)$-Besov semi-norm whenever one of the three following situations occurs:
$\aa$ $p$ $q$
-------- ---------- --------------
$<1/2$ $>1/\aa$ $\ge 1$
$<1/2$ $=1/\aa$ $=1$
$=1/2$ $>2$ $ = +\infty$
Notice that the latter situation is rather specific, since here small deviation probabilities vanish when $\varepsilon$ is small enough, thus the investigation is pointless.
We are able to prove the existence of a finite small deviation constant in the following situation:
$\aa$ $p$ $q$
-------- ------------- ----------
$<1/2$ $= +\infty$ $\geq 1$
Gaussian fractional processes
-----------------------------
In this paragraph $\alpha = 2$ and $H > 0$, the case $H =1/2$ being implicitly excluded. Up to normalization constant our process $X$ is the usual fractional Brownian motion, while $R$ can be viewed as the fractionally integrated Brownian motion. Before stating our results, we will give a list of local properties of $R$, most of which are classical. As we mentioned in Section 2, all these local properties remain true for $X$ as soon as $H < 1$.
$R$ is $\aa$-Hölder (resp. $\aa$-Lipschitz) if and only if $H > \aa$. In particular, $R$ has finite $p$-variation if $H> 1/p$ and only if $H\geq 1/p$ (the boundary value being excluded when $H\geq 1/2$ - see Theorem 5.4 in [@DN], but the situation $H = 1/p < 1/2$ seems to have escaped investigation). Finally, $R$ has finite $(\aa,p)$-Sobolev semi-norm if and only if $H >\aa$ [@CHS].
Less is known about Besov semi-norms. In [@CKR], Proposition 4.1 and Theorem 4.3, it is proved that if $H < 1$ and $0 < 1/p < \aa$, then $X$ has finite $(\aa,p,\infty)$-Besov semi-norm if and only if $H\geq\eta$ (when $H = \eta$, the small deviation probabilities vanish for small $\ee$).
With the same notations as in the preceding paragraph, the following table is a direct consequence of Corollary \[c41\] and of Remark (b) at the end of paragraph 4.3. Recall that it is available for $R$ and for $X$ with the same finite constant $\k$, as soon as $H > \beta + 1/p$.
$||\cdot||$ $\gamma$
-------------------------- -------------
Supremum $1/H$
$L_p$ $1/H$
$\aa$-Hölder $1/(H-\aa)$
$p$-variation $p/(Hp- 1)$
$(\aa, \infty, q)$-Besov $1/(H-\aa)$
If in addition $H < 2$, then we can add another line:
$||\cdot||$ $\gamma$
-------------------- -------------
$(\aa, p)$-Sobolev $1/(H-\aa)$
The existence of $\k$ for $X$ under the supremum norm was established in [@LL1] and [@S2]. In [@LL1], it is also proved for $R$ when $H < 1$. When $H\geq 1$, the existence of $\k$ is mentioned in [@LL2]. Under $L_p$-norms, the right rate for $R$ can be found in [@BL]. Specialists seem to be aware of the existence of the constant [@LfL2], but up to our knowledge this is not explicitly proved in literature. Notice that the case $p=2$ is rather special because of its Hilbert space background. Quite recently, the [*exact*]{} value of $\k$ under the $L_2$-norm was found for integrated Wiener process [@CL] [@KS], and for fractional Brownian motion [@Br].
In all other situations, our results seem to be new.
Non-Gaussian stable Lévy processes
----------------------------------
Non-Gaussian symmetric $\alpha$-stable Lévy processes are RLP’s with parameters $\alpha < 2$ and $H = 1/\alpha$. They have a.s. discontinuous sample paths, hence in the present paper they are only concerned by Theorem \[t31\].
Recall that they have finite $p$-variation if and only if $p > \alpha$ [@BG]. In the case $1<\alpha<2$, it is shown in [@CKR] Theorem 6.1 that they have finite $(1/\alpha, p, \infty)$-Besov norm if and only if $1\leq p <\alpha$.
Our Theorem \[t31\] yields the following table which can be read as in the preceding paragraph, except that here we know nothing [*a priori*]{} about the finiteness of $\k$.
$||\cdot||$ $\gamma$
--------------- ------------------------
Supremum $\alpha$
$L_p$ $\alpha$
$p$-variation $\alpha p/(p- \alpha)$
Under the supremum norm, the finiteness of $\k$ dates back to Taylor [@Ta] and Mogulskii [@Mo]. Notice that in this case the existence of $\k$ follows readily from the Markov property, so that our Theorem \[t31\] is a bit useless here. Under $L_p$-norms, the existence and the finiteness of $\k$ seems to belong to some ancient folklore, but we could not find any precise reference in the literature. We remark that this follows readily from Theorem \[t31\] and Taylor-Mogulskii’s lower bound. Z. Shi also indicated us that in this case the problem of $L_p$-small deviations can be reduced to some specific large deviation probabilities for occupation measures which were previously studied by Donsker and Varadhan. Finally, it is shown in [@Shi] that the case $p=2$ is again rather specific since it can be derived from the $L_p$-small deviations of Wiener process, thanks to an identity in law due to Donati-Martin, Song, and Yor [@DSS].
Nothing is known about the finiteness of $\k$ for the $p$-variation norm, but there is no reason to believe that the rate $\gamma$ indicated above is not the right one. This will be the matter of further study.
Non-Gaussian fractional processes
---------------------------------
In this paragraph $\alpha < 2$ and $H > 0$, the case $H = 1/\alpha$ being implicitly excluded. As in the Gaussian framework, we begin with listing local properties of $R$, most of which can be found in the monograph [@ST]. Again, all these local properties remain true for $X$ when it is well-defined.
$R$ is sample-bounded iff $H \geq 1/\alpha$ and belongs locally to $L_p$ if and only if $H > 1/\alpha -1/p$. $R$ is $\aa$-Hölder (resp. $\aa$-Lipschitz) if and only if $H\geq \aa +1/\alpha$. This latter fact is less classical and was proved for $X$ by Takashima [@Tak] (see also [@KM]). Notice that the ”if" part is quite straightforward from $$|R_t - R_s |\; \le\; C\,
||Z||_{\infty}\, |t - s|^{\min\{1,H-1/\alpha\}}$$ for all $s, t\in [0,1]$. It is interesting to note that contrary to the Gaussian case, the boundary value $\eta=H-1/\alpha$ is included here. In particular, this entails that $R$ has finite $p$-variation if and only if $H\geq 1/p +1/\alpha$ [@ChG].
Nothing seems to be known about Sobolev or Besov semi-norms. From Hölder continuity, we see that $R$ has finite $(\aa, p)$-Sobolev semi-norm if $H > \aa + 1/\alpha$, yet this does not seem to be a sharp estimate.
When $R$ is continuous and when $H > \beta + 1/p + 1/\alpha$, we get exactly the same tables as in paragraph 6.2.
The only result which we are aware of in this direction is due to G. Samorodnitsky [@Sa], who had given just a bit less precise bounds in the small deviation problem for LFSM’s ($1/\alpha<H<1$) under the supremum norm. In the present paper we find the right rate, prove that the constant exists, and our condition on $H$ is the best possible since $R$ is not bounded anymore when $H<1/\alpha$. In general, our results match quite accurately the local sample-path properties of $R$. Namely, except for Sobolev and Besov norms, the only situations which are not covered are the critical values of $H$ w.r.t. Hölder and $p$-variation semi-norms.
When $R$ is not continuous, i.e. when $H < 1/\alpha$, the only relevant classical semi-norms are the $L_p$-norms, as soon as $H>1/\alpha-1/p$. Our Theorem \[t31\] yields $\gamma=1/H$ but nothing is known about the finiteness of the constant $\k$. We believe that $1/H$ is still the right rate in this ultimate situation.
Concluding remarks
==================
\(a) In general, it is not true that $H$ is the only parameter of a self-similar stable process which is involved in the exponential behavior of its small deviation probabilities. Indeed, [@Sa] delivers an example of an $\alpha$-stable, $H$-sssi process whose small deviation rate depends both on $H$ and $\alpha$. Let us also recall the general results of [@Ry] about stable measures, which entail that $$\gamma\;\le\; \frac \alpha{1-\alpha}$$ for arbitrary $\alpha$-stable processes with $0<\alpha<1$.
\(b) The [*symmetry*]{} assumption on the stable processes is a crucial one, since our proofs are heavily based on two Gaussian results: Anderson’s and Šidák’s inequalities. Nevertheless, it seems likely that everything remains true as soon as $Z$ is not a subordinator.
\(c) Theorem \[t31\] can be easily extended to multi-dimensional RLP’s having some symmetry, since they also share the crucial extrapolation-homogeneity property. However, adaptating the proof of Theorem \[t41\] in a multi-dimensional context raises some technical difficulties related to Šidák’s inequality.
\(d) As we indicated before, there are many different generalizations of fractional Brownian motion in non-Gaussian case. If $\tilde X$ is the well-balanced linear fractional stable motion defined in (\[balanced\]), then the proof of Theorem \[t31\] fails in the absence of extrapolation-homogeneity. However, we believe that the rates should be the same.
\(e) Although our results are designed for nice translation-invariant semi-norms, it is worthwhile to mention that there exist elaborated techniques concerning norms which do not share any translation-invariance. See especially [@LfL1], [@LfL2] and [@Shi], where weighted $L_p$-norms are handled respectively for Brownian motion, fractional Brownian motion, and symmetric $\alpha$-stable processes.
[*Acknowledgements:*]{} We are much indebted to Charles Suquet for valuable advice concerning wavelet techniques, and to Werner Linde and Zhan Shi for helpful discussions. The first-named author was partially supported by grants RFBR 02-01-00265 and NSh-2258.2003.1. This work was carried out during a visit at the Technische Universität Berlin of the second-named author, who would like to thank Michael Scheutzow for his kind hospitality.
[10]{}
and [M. Taqqu.]{} Rate optimality of wavelet series approximations of fractional Brownian Motion. To appear in [*Journal of Fourier Analysis and Applications.*]{}
and [B. Roynette.]{} Some exact equivalents for the Brownian motion in Hölder semi-norm. [*Prob. Th. Rel. Fields*]{} [**93**]{} (4), pp. 457-484, 1992.
and [W. Linde.]{} Small ball probabilities of fractional Brownian sheets via fractional integration operators. [*J. Theoret. Probab.*]{} [**15**]{} (3), pp. 589-612, 2002.
and [Z. Shi.]{} Small ball estimates for Brownian motion under a weighted sup-norm. [*Studia Sci. Math. Hung.*]{} [**36**]{} (1-2), pp. 275-289, 2001.
and [J. L. Teugels.]{} [*Regular Variation.*]{} Cambridge University Press, Cambridge, 1987.
and [R. K. Getoor.]{} Some theorems on stable processes. [*Trans. Am. Math. Soc.*]{} [**95**]{}, pp. 263-273, 1960.
and [A. A. Mogulskii.]{} On probabilities of small deviations for stochastic processes. [*Sib. Adv. Math.*]{} [**1**]{} (1), pp. 39-63, 1991.
Small ball constants and tight eigenvalue asymptotics for fractional Brownian motions. [*J. Theoret. Probab.*]{} [**16**]{} (1), pp. 87-100, 2003.
and [W. V. Li.]{} A functional LIL for symmetric stable processes. [*Ann. Probab.*]{} [**28**]{} (1), pp. 258-277, 2000.
Quadratic functionals and small ball probabilities for the $m$-fold integrated Brownian motion. [*Ann. Probab.*]{} [**31**]{}(2), pp. 1052-1077, 2003.
and [O. E. Galkin]{}. On maps of bounded $p$-variation with $p > 1$. [*Positivity*]{} [**2**]{} (1), pp. 19-45, 1998.
and [B. Roynette.]{} Quelques espaces fonctionnels associés à des processus gaussiens. [*Stud. Math.*]{} [**107**]{} (2), pp. 171-204, 1993.
and [Q.-M. Shao.]{} Convergence of integrals of uniform empirical and quantile processes. [*Stoch. Proc. Appl.*]{} [**45**]{} (2), pp. 283-294, 1993.
. SIAM, 1992.
and [M. Yor.]{} Symmetric stable processes, Fubini’s theorem, and some extensions of the Ciesielski-Taylor identities in law. [*Stoch. Stoch. Rep.*]{} [**50**]{} (1-2), pp. 1-33, 1994.
and [R. Norvaiša]{}. An introduction to $p$-variation and Young integrals. With emphasis on sample functions of stochastic processes. [*MaPhySto Lect. Notes*]{} [**1**]{}, Univ. of Aarhus, 1998.
and [M. Maejima.]{} [*Self-similar processes.*]{} Princeton University Press, Princeton, 2002.
and [Z. Shi.]{} Chung’s law for integrated Brownian motion. [*Trans. Amer. Math. Soc.*]{} [**350**]{} (10), pp. 4253-4264, 1998.
and [M. Maejima.]{} Hölder continuity of sample paths of some self-similar stable processes. [*Tokyo J. Math.*]{} [**14**]{} (1), pp. 93-100, 1991.
and [W. V. Li.]{} Small ball problems for Brownian motion and for the Brownian sheet. [*J. Theoret. Probab.*]{} [**6**]{} (3), pp. 547-577, 1993.
and [Q.-M. Shao.]{} Small ball probabilities for Gaussian processes with stationary increments under Hölder norms. [*J.Theoret.Probab*]{} [**8**]{} (2), pp. 361-386, 1995.
and [W. Linde.]{} Existence of small ball constants for fractional Brownian motions. [*C. R. Acad. Sci. Paris*]{} [**326**]{} (11), pp. 1329–1334, 1998.
and [W. Linde.]{} Approximation, metric entropy and small ball estimates for Gaussian measures. [*Ann. Probab.*]{} [**27**]{} (3), pp. 1556-1578, 1999.
and [Q.-M. Shao.]{} Small ball estimates for Gaussian processes under Sobolev type norms. [*J. Theoret. Probab.*]{} [**12**]{} (3), pp. 699-720, 1999.
and [Q.-M. Shao.]{} Gaussian Processes: Inequalities, Small Ball Probabilities and Applications. In: [*Stochastic processes: Theory and methods, Handbook of Statistics*]{} [**19**]{}, pp. 533-597, 2001.
Asymptotic behavior of small ball probabilities. In: [*Probab. Theory and Math. Statist. Proc. VII International Vilnius Conference (1998)*]{}. Vilnius, VSP/TEV, pp. 453-468, 1999.
and [W. Linde.]{} Approximation and entropy numbers of Volterra operators with application to Brownian motion. [*Mem. Amer. Math. Soc.*]{} [**745**]{}, 2002.
and [W. Linde.]{} Small deviations of weighted fractional processes and average non-linear approximation. Submitted to [*Stochastic Processes and their Applications*]{}, 2002.
and [P. M. Robinson.]{} Alternative forms of fractional Brownian motion. [*J. Stat. Plann. Inference*]{} [**80**]{} (1-2), pp. 111-122, 1999.
Cambridge studies in advanced mathematics, Cambridge, 1992.
Small deviations in a space of trajectories. [*Theor. Probab. Appl.*]{} [**19**]{}, pp. 726-736, 1974.
Mouvement brownien et espaces de Besov. [*Stoch. Stoch. Rep.*]{} [**43**]{} (3-4), pp. 221-260, 1993.
Asymptotic behavior of stable seminorms near the origin. [*Ann. Probab.*]{} [**14**]{} (1), pp. 287-298, 1986.
Lower tails of self–similar stable processes. [*Bernoulli*]{} [**4**]{} (1), pp. 127-142, 1998.
and [M. S. Taqqu.]{} [*Stable non-Gaussian random processes.*]{} Chapman & Hall, New York, 1994.
A note on small ball probability of a Gaussian process with stationary increments. [*J. Theoret. Probab.*]{} [**6**]{} (3), pp. 595-602, 1993.
A Gaussian correlation inequality and its application to the existence of small ball constant. [*Preprint*]{}, 1999.
Lower tails of quadratic functionals of symmetric stable processes. [*Prépublication de l’Université Paris 6*]{}, 1999.
Une méthode élémentaire pour l’évaluation des petites boules browniennes. [*C. R. Acad. Sci. Paris*]{} [**316**]{} (11), pp. 1217-1220, 1993.
Small ball probabilities for Gaussian processes under non-uniform norms. [*J. Theoret. Probab.*]{} [**9**]{} (3), pp. 613-630, 1996.
Sample path properties of ergodic self-similar processes. [*Osaka J. Math.*]{} [**26**]{} (1), pp. 159-189, 1989.
Sample path properties of a transient stable process. [*J. Math. Mech.*]{} [**16**]{}, pp. 1229-1246, 1967.
-- -------------------------------------- --------------------------------------
Mikhail A. Lifshits Thomas Simon
Faculty of Mathematics and Mechanics Équipe d’Analyse et Probabilités
St-Petersburg State University Université d’Évry-Val d’Essonne
198504, Stary Peterhof Boulevard François Mitterrand
Bibliotechnaya pl., 2 91025 Évry Cedex
Russia France
E-mail: [[email protected]]{} E-mail: [[email protected]]{}
-- -------------------------------------- --------------------------------------
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We give a partial answer to the question for the precise value of the nuclear dimension of UCT-Kirchberg algebras raised by W. Winter and J. Zacharias in [@WZ10]. It is shown that every Kirchberg algebra in the UCT-class with torsion free $K_1$-group has nuclear dimension 1.'
address: 'Dominic Enders Department of Mathematical Sciences, University of Copenhagen Universitetsparken 5, DK-2100 Copenhagen Ø, Denmark'
author:
- Dominic Enders
date: 'May 20, 2014'
title: 'On the nuclear dimension of certain UCT-Kirchberg algebras'
---
Introduction
============
In 2010, Winter and Zacharias introduced the notion of nuclear dimension for $C^*$-algebras as a non-commutative analogue of topological covering dimension ([@WZ10]). Since then it has clearly become one of the most important concepts in the Elliott classification program, i.e. the classification of simple, nuclear $C^*$-algebras by $K$-theoretic data. In particular, finite-dimensionality with respect to this notion of dimension constitutes one of the fundamental regularity properties for the $C^*$-algebras in question. It is part of the Toms-Winter conjecture that in this situation finite nuclear dimension is presumably equivalent to $\mathcal{Z}$-stability and therefore a necessary assumption in order to obtain classification by the Elliott invariant. This conjecture received a great deal of attention and the equivalence mentioned above has been partially verified ([@SWW14], [@Win12]). On the other hand, large classes of simple, nuclear $C^*$-algebras with finite nuclear dimension have been successfully classified by their Elliott invariant (e.g. [@Lin11], [@Win12]).
However, once a $C^*$-algebra having finite nuclear dimension belongs to a class which is classified by some invariant, it is a natural question to ask for the precise value of its dimension and how it can be read off from its invariant. In this paper we study this question for the class of UCT-Kirchberg algebras, i.e. purely infinite, simple, separable, nuclear $C^*$-algebras satisfying the UCT which are, due to Kirchberg and Phillips, completely classified by their $K$-groups. In [@WZ10], Winter and Zacharias showed that all these algebras have nuclear dimension at most 5 and asked whether the precise value of their dimension is determined by algebraic properties of their $K$-groups, such as torsion ([@WZ10 Problem 9.2]). Here, we give a partial answer to their question and give the optimal dimension estimate in the absence of $K_1$-torsion. More precisely, we show that the nuclear dimension of UCT-Kirchberg algebras with torsion free $K_1$-groups equals 1 (Theorem \[main\]).
Our methods of proof used in the computation of the nuclear dimension differ from the arguments in the original estimate by Winter and Zacharias in two essential ways. First, we do not make use of Cuntz-Pimsner algebras, as done in [@WZ10], but employ crossed product models for UCT-Kirchberg algebras which are provided by the work of R[ø]{}rdam ([@Ror95]). Second, we do not construct completely positive approximations for the identity map on a given Kirchberg algebra $A$ directly. Instead, we study a certain family of homomorphisms $\iota_n\colon A\rightarrow M_n(A)$ and show that the $\iota_n$ can be approximated in a decomposable manner. While the error we have to make in the approximation of each $\iota_n$ is bounded from below, the lower bound will be small for large $n$. In a second step, we show how to combine the $\iota_n$ in order to construct a homomorphism $A\rightarrow M_k(A)$ which admits decomposable approximations of the same quality as the $\iota_n$, but in addition induces an isomorphism on $K$-theory. By Kirchberg-Phillips classification, this map can be perturbed to the identity map on $A$ and by that provides suitable approximations for estimating the nuclear dimension of $A$.
This paper is organized as follows: First, right now, we define certain matrix embeddings for crossed products which are the main objects of study in this paper.
\[iota\] Let $A$ be a $C^*$-algebra, $\alpha\in{\operatorname{Aut}}(A)$ an automorphism and $A\rtimes_\alpha\mathbb{Z}$ the corresponding full crossed product. For $n\in\mathbb{N}$, $n\geq 2$, we let $$\iota_n\colon A\rtimes_\alpha\mathbb{Z} \rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$$ denote the embedding given by $$\begin{tabular}{ccc}$a \mapsto\begin{pmatrix}
\alpha^{-1}(a) & 0 & \cdots& 0 \\ 0 & \alpha^{-2}(a) & \ddots& \vdots\\ \vdots & \ddots& \ddots& 0 \\ 0 & \cdots& 0 & \alpha^{-n}(a)
\end{pmatrix}$
& and
& $U \mapsto\begin{pmatrix}
0 & \cdots & \cdots & 0 & U^n \\
1 & \ddots &&& 0 \\
0 & \ddots & \ddots & & \vdots \\
\vdots & \ddots & \ddots & \ddots & \vdots \\
0 & \cdots & 0 & 1 &0
\end{pmatrix}$\end{tabular}$$ for $a\in A$ resp. for the unitary $U\in\mathcal{M}(A\rtimes_\alpha\mathbb{Z})$ implementing the action $\alpha$.
In section \[section approximation\], we show that these maps $\iota_n$ admit completely positive approximations in an order zero fashion, at least up to an error which is small for large $n$. The $K$-theory of these maps can be computed in many cases of interest, this is done in section \[section k-theory\]. A combination of these results yields the dimension estimate for Kirchberg algebras in section \[section main\]. We finish with some remarks about possible generalizations of the technique developed in this paper.
Decomposable approximations for $\iota_n$ {#section approximation}
=========================================
We want to study certain maps which admit completely positive approximations of the same type as in the original definition of nuclear dimension in [@WZ10]. In our situation, however, we won’t be able to find approximations of arbitrary small error (which would rather lead to a notion of nuclear dimension for maps). Therefore we need to keep track of the precision of approximation that we can get for a given map. This will be done using the following definition.
\[def decomposition\] Let $\alpha\colon A\rightarrow B$ be a map between $C^*$-algebras $A$ and $B$. Given $d\in\mathbb{N}_0$, $\epsilon>0$ and a finite subset $\mathcal{G}$ of $A$, we say that $\alpha$ admits a piecewise contractive, completely positive, $d$-decomposable $(\mathcal{G},\epsilon)$-approximation if there exists $(F,\psi,\varphi)$ such that $F$ is a finite-dimensional $C^*$-algebra, and such that $\psi\colon A\rightarrow F$ and $\varphi\colon F\rightarrow B$ are completely positive maps satisfying
1. $\|(\varphi\circ\psi)(a)-\alpha(a)\|<\epsilon$ for all $a\in\mathcal{G}$;
2. $\psi$ is contractive;
3. $F$ decomposes as $F=F^{(0)}\oplus\cdots\oplus F^{(d)}$ such that $\varphi_{|F^{(i)}}$ is a c.p.c. order zero map for each $i=0,...,d$.
In this terminology, the notion of nuclear dimension for a $C^*$-algebra $A$, as defined by Winter and Zacharias in [@WZ10], can be formulated as follows:
$$\begin{tabular}{rc}
\multirow{2}{*} {${\operatorname{dim}_{{\operatorname{nuc}}}}(A)\leq d\quad\Leftrightarrow$} &
${\operatorname{id}}_A$ admits piecewise contractive, c.p., $d$-decomposable, \\ & $(\mathcal{G},\epsilon)$-approximations for all finite $\mathcal{G}\subset A$ and all $\epsilon>0$.
\end{tabular}$$
We first note that the approximations of \[def decomposition\] for $^*$-homomorphisms are well-behaved with respect to composition, orthogonal sum and approximate unitary equivalence. We collect some of these elementary permanence results in the following lemma.
\[permanence\] Let $*$-homomorphisms $\xymatrix{A \ar[r]^\alpha & B \ar[r]^{\beta_1,\beta_2} & C\ar[r]^\gamma & D}$ between $C^*$-algebras $A,B,C,D$, a finite subset $\mathcal{G}\subset B$, $\epsilon>0$ and $d\in\mathbb{N}_0$ be given. If both $\beta_1$ and $\beta_2$ admit piecewise contractive, completely positive, $d$-decomposable $(\mathcal{G},\epsilon)$-approximations $(F_i,\psi_i,\varphi_i)$, $i\in\{1,2\}$, then the following holds:
1. The composition $\gamma\circ\beta_1$ admits a piecewise contractive, completely positive, $d$-decomposable $(\mathcal{G},\epsilon)$-approximation.
2. For any $\alpha$-preimage $\mathcal{G}'$ of $\mathcal{G}$, the composition $\beta_1\circ\alpha$ admits a piecewise contractive, completely positive, $d$-decomposable $(\mathcal{G'},\epsilon)$-approximation.
3. The orthogonal sum $\begin{pmatrix}\beta_1 & 0 \\ 0 & \beta_2\end{pmatrix}\colon B\rightarrow M_2(C)$ admits a piecewise contractive, completely positive, $d$-decomposable $(\mathcal{G},\epsilon)$-approximation.
4. If $\beta_1\sim_{a.u.}\delta$ for some $\delta\colon B\rightarrow C$, then $\delta$ admits a piecewise contractive, completely positive, $d$-decomposable $(\mathcal{G},\epsilon)$-approximation.
We only name suitable approximating systems and leave the details to the reader. For (1) one chooses $(F_1,\psi_1,\gamma\circ\varphi_1)$, while for (2) $(F_1,\psi_1\circ\alpha,\varphi_1)$ works. For (3) consider $\left(F_1\oplus F_2,\psi_1\oplus\psi_2,\begin{pmatrix}\varphi_1 & 0 \\ 0 & \varphi_2\end{pmatrix}\right)$ with respect to the decomposition $(F_1\oplus F_2)^{(i)}=F_1^{(i)}\oplus F_2^{(i)}$. Given $\delta=\lim\limits_{n\rightarrow\infty}{\operatorname{Ad}}(U_n)\circ\beta_1$ in (4), one checks that the triple $(F,\psi_1,{\operatorname{Ad}}(U_n)\circ\varphi_1)$ works for $n$ sufficiently large.
Next, we apply a ’cut and paste’-technique similar to the one used in [@WZ10 section 7] to the maps $\iota_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ of Definition \[iota\]. This essentially reduces a decomposable approximation of $\iota_n$ to a decomposable approximation of the coefficient algebra $A$.
\[approximation1\] Let a nuclear $C^*$-algebra $A$, an automorphism $\alpha\in{\operatorname{Aut}}(A)$, a finite subset $\mathcal{G}\subset A\rtimes_\alpha\mathbb{Z}$ and $\epsilon>0$ be given. Then the following holds for the maps $\iota_n$ as defined in \[iota\]: For all sufficiently large $n$ there exists a c.c.p. map $\psi_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A)$ and two $\ast$-homomorphisms $\Lambda_n^0,\Lambda_n^1\colon M_n(A)\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ such that the diagram $$\begin{xy}
\xymatrix{A\rtimes_\alpha\mathbb{Z} \ar[rr]^{\iota_n} \ar[dr]_{\psi_n} && M_n(A\rtimes_\alpha\mathbb{Z}) \\ & M_n(A) \ar[ur]_{\Lambda_n^0+\Lambda_n^1}}
\end{xy}$$ commutes up to $\epsilon$ on $\mathcal{G}$, i.e. $\|\iota_n(x)-((\Lambda_n^0+\Lambda_n^1)\circ\psi_n)(x)\|<\epsilon$ holds for all $x\in\mathcal{G}$.
Assume $A\subseteq\mathcal{B}(H)$ for some Hilbert space $H$. By nuclearity of $A$ we have $A\rtimes_\alpha\mathbb{Z}=A\rtimes_{\alpha,r}\mathbb{Z}$ ([@BO08 Theorem 4.2.4]) and may therefore identify the crossed product with the $C^*$-subalgebra of $\mathcal{B}(\ell^2(\mathbb{Z})\otimes H)$ generated by the operators
$$\begin{array}{rclr}
a(\delta_i\otimes\xi)&=&\delta_i\otimes(\alpha^{-i}(a))(\xi),&a\in A \\
U(\delta_i\otimes\xi)&=&\delta_{i+1}\otimes\xi
\end{array}$$ respectively by the ideal therein generated by $A$ in the non-unital case.
Denote by $P_n$ the projection onto the subspace $\ell^2(\{1,...,n\})\otimes H$. Compression by $P_n$ gives a c.p.c. map from $A\rtimes_\alpha\mathbb{Z}$ to $M_n(A)$. We need to put suitable weights on the entries of the elements of $\{P_nxP_n\}_{x\in\mathcal{G}}$. As in [@WZ10], this will be done by Schur multiplication (i.e. entrywise multiplication) with a suitable positive contraction $\kappa_n\in M_n(\mathbb{C})$. We choose
$$\kappa_n=\frac{1}{m+1}\begin{pmatrix}
1 & 1 & \cdots & 1 & 1 & 1 & 1 & \cdots & 1 & 1 \\
1 & 2 & \cdots & 2 & 2 & 2 & 2 & \cdots & 2 & 1 \\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \vdots & \iddots & \vdots & \vdots \\
1 & 2 & \cdots & m-1 & m-1 & m-1 & m-1 & \cdots & 2 & 1 \\
1 & 2 & \cdots & m-1 & m & m & m-1 & \cdots & 2 & 1 \\
1 & 2 & \cdots & m-1 & m & m & m-1 & \cdots & 2 & 1 \\
1 & 2 & \cdots & m-1 & m-1 & m-1 & m-1 & \cdots & 2 & 1 \\
\vdots & \vdots & \iddots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
1 & 2 & \cdots & 2 & 2 & 2 & 2 & \cdots & 2 & 1 \\
1 & 1 & \cdots & 1 & 1 & 1 & 1 & \cdots & 1 & 1
\end{pmatrix}$$ if $n=2m$ is even, respectively
$$\kappa_n=\frac{1}{m+1}\begin{pmatrix}
1 & 1 & \cdots & 1 & 1 & 1 & \cdots & 1 & 1 \\
1 & 2 & \cdots & 2 & 2 & 2 & \cdots & 2 & 1 \\
\vdots & \vdots & \ddots & \vdots & \vdots & \vdots & \iddots & \vdots & \vdots \\
1 & 2 & \cdots & m & m & m & \cdots & 2 & 1 \\
1 & 2 & \cdots & m & m+1 & m & \cdots & 2 & 1 \\
1 & 2 & \cdots & m & m & m & \cdots & 2 & 1 \\
\vdots & \vdots & \iddots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots \\
1 & 2 & \cdots & 2 & 2 & 2 & \cdots & 2 & 1 \\
1 & 1 & \cdots & 1 & 1 & 1 & \cdots & 1 & 1
\end{pmatrix}$$ if $n=2m+1$ is odd. It is not hard to see that the map on $M_n(A)$ given by $y\mapsto\kappa_n\ast y$, where $\ast$ denotes Schur multiplication, is then in fact completely positive and contractive. Therefore, $\psi_n$ defined as $$\begin{array}{cc}\psi_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A),&x\mapsto\kappa_n\ast(P_nxP_n)\end{array}$$ is also a c.p.c. map.
Now let $j\colon M_n(A)\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ be the canonical embedding. Using the unitary element
$$V:=\begin{pmatrix}
0 & \cdots & \cdots & 0 & U^{n-1} \\
U^{-1} & \ddots &&& 0 \\
0 & \ddots & \ddots & & \vdots \\
\vdots & \ddots & \ddots & \ddots & \vdots \\
0 & \cdots & 0 & U^{-1} &0
\end{pmatrix}\in M_n(\mathcal{M}(A\rtimes_\alpha\mathbb{Z})),$$ we consider the $^*$-homomorphisms $\Lambda_n^0,\Lambda_n^1\colon M_n(A)\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ given by $\Lambda_n^0=j$ and $\Lambda_n^1=Ad\left(V^m\right)\circ j$.\
We may assume that the elements of $\mathcal{G}$ are of the form $x=a_xU^{k_x}$ for suitable contractions $a_x\in A$ and $k_x\in\mathbb{Z}$, let $K:=\max_{x\in\mathcal{G}}|k_x|$. We claim that, for $n>\frac{K}{\epsilon}$, the maps $\Lambda_n^0,\Lambda_n^1$ and $\psi_n$ defined above provide an approximation of $\iota_n$ as claimed. For convenience we only consider the case of even $n$, the calculations for the odd case are essentially identical. Given $x=aU^{k}\in\mathcal{G}$ with $k\geq 0$, one checks that $\iota_n(x)-((\Lambda_n^0+\Lambda_n^1)\circ\psi_n)(x)$ equals $$\frac{1}{(\frac{n}{2}+1)}\begin{pmatrix}
0&0&0&k\cdot I_k-\mu_k \\ k\cdot I_{(\frac{n}{2}-k)} &0&0&0\\ 0& k\cdot I_k - \mu_k &0&0 \\0&0&k\cdot I_{(\frac{n}{2}-k)}&0
\end{pmatrix}\ast\iota_n(x)$$ where $$\mu_k=\begin{pmatrix}
0 & 0 & \cdots &&& \cdots & 0 \\
0 & 1 & 0 &&&& \vdots \\
\vdots & 0 & 2 & \ddots \\
&&\ddots & \ddots & \ddots \\
&&&\ddots & 2 & 0 & \vdots \\
\vdots &&&&0&1&0\\
0&\cdots&&&\cdots&0&0\end{pmatrix}\in M_k(\mathbb{C}),$$ $I_d$ is the $d\times d$ unit matrix and $\ast$ denotes again Schur multiplication. Due to the special off-diagonal form of these elements, an entrywise norm estimate immediately shows $$\|\iota_n(x)-((\Lambda_n^0+\Lambda_n^1)\circ\psi_n)(x)\|\leq\frac{k}{(\frac{n}{2}+1)}\|x\|\leq\frac{K}{n}<\epsilon.$$
Next, we show that the maps $\iota_n$ of \[iota\] admit decomposable approximations in the sense of Definition \[def decomposition\] provided that the coefficient algebra has finite nuclear dimension.
\[approximation2\] Given a $C^*$-algebra $A$ with ${\operatorname{dim}_{{\operatorname{nuc}}}}(A)=d<\infty$, a finite subset $\mathcal{G}\subset A$ and $\epsilon>0$, the maps $\iota_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ defined in \[iota\] admit a piecewise contractive, completely positive, $(2d+1)$-decomposable $(\mathcal{G},\epsilon)$-approximation for all sufficiently large $n$.
Choose $n$ large enough such that there exist maps $\psi_n$ and $\Lambda_n^0,\Lambda_n^1$ as in Lemma \[approximation1\] satisfying $\|\iota_n(x)-\left((\Lambda_n^0+\Lambda_n^1)\circ\psi_n\right)(x)\|<\frac{\epsilon}{3}$ for all $x\in\mathcal{G}$. Since ${\operatorname{dim}_{{\operatorname{nuc}}}}(M_n(A))={\operatorname{dim}_{{\operatorname{nuc}}}}(A)=d$ by [@WZ10 Corollary 2.8], there exists a piecewise contractive, completely positive, $d$-decomposable $(\psi_n(\mathcal{G}),\frac{\epsilon}{3})$-approximation of ${\operatorname{id}}_{M_n(A)}$, i.e. a finite-dimensional $C^*$-algebra $F=F^{(0)}\oplus...\oplus F^{(d)}$, a c.p.c. map $\psi'\colon M_n(A)\rightarrow F$ and c.p.c. order zero maps $\varphi'^{(i)}\colon F^{(i)}\rightarrow M_n(A)$, $i=0,...,d$, such that $$\left\|\left(\sum_{i=0}^d\varphi'^{(i)}\circ\psi'\right)(\psi_n(x))-\psi_n(x)\right\|<\frac{\epsilon}{3}$$ holds for all $x\in\mathcal{G}$. Putting these two approximations together, i.e. setting $G:=F\oplus F$ with decomposition $G=\bigoplus_{j=0,1}\bigoplus_{i=0}^d G^{(i,j)}$, where $G^{(i,j)}:=F^{(i)}$, and considering $$\begin{xy}\xymatrix{
A\rtimes_\alpha\mathbb{Z} \ar[rr]^{\iota_n} \ar[dr]_(0.4){(\psi'\circ\psi_n)\oplus(\psi'\circ\psi_n)\;\;\;\;\;} && M_n(A\rtimes_\alpha\mathbb{Z}) \\
& G\oplus G=\bigoplus\limits_{j=0,1}\;\bigoplus\limits_{i=0}^d G^{(i,j)} \ar[ur]_(0.7){\sum\limits_{j=0,1}\;\sum\limits_{i=0}^d\left(\Lambda_n^j\circ\varphi'^{(i)}\right)}
}\end{xy},$$ one finds each $\Lambda_n^j\circ\varphi'^{(i)}$ to be an order zero map and further $$\begin{array}{rl}&\left\|\sum\limits_{j=0}^1\sum\limits_{i=0}^d\left(\Lambda_n^j\circ\varphi'^{(i)}\right)((\psi'\circ\psi_n)(x))-\iota_n(x)\right\| \\ \\
\leq & \left\|\sum\limits_{j=0}^1\Lambda_n^j(\psi_n(x))-\iota_n(x)\right\| +2\left\|\left(\sum\limits_{i=0}^d\varphi'^{(i)}\circ\psi'\right)(\psi_n(x))-\psi_n(x)\right\| \\ \\ < & \epsilon
\end{array}$$ for all $x\in\mathcal{G}$. In other words, $\left(G,(\psi'\circ\psi_n)^{\oplus 2},\sum_{i,j}\left(\Lambda_n^j\circ\varphi'^{(i)}\right)\right)$ provides a piecewise contractive, completely positive, $(2d+1)$-decomposable $(\mathcal{G},\epsilon)$-approximation for $\iota_n$.
The $K$-theory of $\iota_n$ {#section k-theory}
===========================
In this section we compute the map on $K$-theory induced by $\iota_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ from \[iota\]. Using a rotation argument we first express $K_*(\iota_n)$ in terms of $K_*(\alpha)$. Under suitable assumptions, this allows us to read off $K_*(\iota_n)$ from the Pimsner-Voiculescu sequence associated to $(A,\alpha)$.
\[homotopy\] The homomorphism $\iota_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ defined in \[iota\] is homotopic to the diagonal embedding $j_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ given by $$\begin{tabular}{ccc}$a \mapsto\begin{pmatrix}
\alpha^{-1}(a) & 0 & \cdots& 0 \\ 0 & \alpha^{-2}(a) & \ddots& \vdots\\ \vdots & \ddots& \ddots& 0 \\ 0 & \cdots& 0 & \alpha^{-n}(a)
\end{pmatrix}$
& and
& $U \mapsto\begin{pmatrix}
U \\ & U \\ &&\ddots \\&&&U
\end{pmatrix}$\end{tabular}$$ for $a\in A$ resp. for the unitary $U\in\mathcal{M}(A\rtimes_\alpha\mathbb{Z})$ implementing the action $\alpha$.
Let $(V_t)_{t\in[0,1]}$ be a path of unitaries in $M_n(\mathbb{C})$ connecting the shift-unitary $V_0=\sum_{i=1}^{n-1} e_{i,i+1}+e_{n,1}$ to identity matrix $V_1=1_n$ and consider the path $(\kappa_t)_{t\in[0,1]}$ of homomorphisms $\kappa_t\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ given by
$$\begin{tabular}{ccc}$a \mapsto\begin{pmatrix}
\alpha^{-1}(a) & 0 & \cdots& 0 \\ 0 & \alpha^{-2}(a) & \ddots& \vdots\\ \vdots & \ddots& \ddots& 0 \\ 0 & \cdots& 0 & \alpha^{-n}(a)
\end{pmatrix}$
& and
& $U \mapsto V_t\ast\begin{pmatrix}
U & U^2 & \cdots & U^n \\
1 & U & \ddots & \vdots \\
\vdots & \ddots & \ddots & U^2 \\
U^{-n+2} & \cdots & 1 & U
\end{pmatrix}$\end{tabular}$$ where $\ast$ denotes Schur multiplication. It is straightforward to check that $\kappa_t(U)$ is a continuous unitary path in $M_n(\mathcal{M}(A\rtimes_\alpha\mathbb{Z}))$, hence we only have to make sure that each $\kappa_t$ is well-defined, i.e. compatible with the action $\alpha$: $$\begin{array}{rl}
&\left(\kappa_t(U)\kappa_t(a)\kappa_t(U)^*\right)_{ij} \\
= & \sum_{k=1}^n \kappa_t(U)_{ik}\alpha^{-k}(a)\kappa_t(U^*)_{kj} \\
= & \sum_{k=1}^n (V_t)_{ik}\overline{(V_t)_{jk}}U^{(k-i+1)}\alpha^{-k}(a)U^{-(k-j+1)} \\
= & \left(\sum_{k=1}^n (V_t)_{ik}\overline{(V_t)_{jk}}\right) U^{-i+1}aU^{j-1} \\
= & \delta_{ij}\cdot U^{-i+1}aU^{j-1} \\
= & \delta_{ij}\cdot \alpha^{-i+1}(a) \\
= & \left(\kappa_t(\alpha(a))\right)_{ij}
\end{array}$$ This shows that $\kappa_t$ provides a homotopy between $\kappa_0=\iota_n$ and $\kappa_1=j_n$.
\[k-theory\] Given $n\in\mathbb{N}$, let $\iota_n\colon A\rtimes_\alpha\mathbb{Z}\rightarrow M_n(A\rtimes_\alpha\mathbb{Z})$ be the homomorphism defined in \[iota\]. If both boundary maps in the Pimsner-Voiculescu sequence associated to $(A,\alpha)$ vanish, we find $K_*(\iota_n)=n\cdot {\operatorname{id}}_{K_*(A\rtimes_\alpha\mathbb{Z})}$. The same conclusion holds if instead either $K_0(A)=0$ or $K_1(A)=0$.
We know by Lemma \[homotopy\] that $(\iota_n)_*=\sum_{i=1}^n (\widehat{\alpha^i})_*$ where $\hat{\beta}$ denotes the canonical extension of an automorphism $\beta\in{\operatorname{Aut}}(A)$ satisfying $\alpha\circ\beta=\beta\circ\alpha$ to an automorphism of $A\rtimes_\alpha\mathbb{Z}$ via $\widehat{\beta}(U)=U$. Using $(\widehat{\alpha^i})_*=(\widehat{\alpha}_*)^i$, it suffices to show that $\widehat{\alpha}$ induces the identity map on $K$-theory. By the Pimsner-Voiculescu sequence ([@Bla98 Theorem 10.2.1]) and the naturality thereof, we have the following commutative diagram with exact rows $$\begin{xy}\xymatrix{
\ar[r] & K_*(A) \ar[r]^{1-\alpha_*} \ar[d]^{\alpha_*} & K_*(A) \ar[r]^(.4){i_*} \ar[d]^{\alpha_*} & K_*(A\rtimes_\alpha\mathbb{Z}) \ar[r]^{\partial_*} \ar[d]^{\widehat{\alpha}_*} & K_{*+1}(A) \ar[r]^{1-\alpha_{*+1}} \ar[d]^{\alpha_{*+1}} & K_{*+1}(A) \ar[r] \ar[d]^{\alpha_{*+1}} & \\
\ar[r] & K_*(A) \ar[r]^{1-\alpha_*} & K_*(A) \ar[r]^(.4){i_*} & K_*(A\rtimes_\alpha\mathbb{Z}) \ar[r]^{\partial_*} & K_{*+1}(A) \ar[r]^{1-\alpha_{*+1}} & K_{*+1}(A) \ar[r] & }\end{xy}$$ where $i$ denotes the canonical inclusion of $A$ into the crossed product.
In the first case, i.e. $\partial_*=0$, we have an isomorphism $K_*(A)/(1-\alpha_*)K_*(A)\cong K_*(A\rtimes_\alpha\mathbb{Z})$ induced by $i_*$. Hence $\widehat{\alpha}_*\circ i_*=i_*\circ\alpha_*=i_*$ and the claim follows by surjectivity of $i_*$.
The second case is treated similarly. If $K_i(A)=0$, then $K_{i+1}(A\rtimes_\alpha\mathbb{Z})\cong K_{i+1}(A)/(1-\alpha_{i+1})K_{i+1}(A)$ and $\widehat{\alpha}_{i+1}={\operatorname{id}}$ follows as in the first case. On the other hand, the boundary map $\partial_i$ identifies $K_i(A\rtimes_\alpha\mathbb{Z})$ with $ker(1-\alpha_{i+1})\subseteq K_{i+1}(A)$. Hence $\partial_i\circ\widehat{\alpha}_i=\alpha_{i+1}\circ\partial_i=\partial_i$, which shows that also $\widehat{\alpha}_i={\operatorname{id}}$.
Nuclear dimension of Kirchberg algebras {#section main}
=======================================
Combining the results of section \[section approximation\] and \[section k-theory\] with the $K$-theoretical classification results by Kirchberg and Phillips, we obtain our main result, the computation of the nuclear dimension for UCT-Kirchberg algebras in the absence of torsion in $K_1$.
\[main\] Let $B$ be a Kirchberg algebra in the UCT-class. If $K_1(B)$ is torsion free, then ${\operatorname{dim}_{{\operatorname{nuc}}}}(B)=1$.
Since ${\operatorname{dim}_{{\operatorname{nuc}}}}(B)={\operatorname{dim}_{{\operatorname{nuc}}}}(B\otimes\mathbb{K})$ by [@WZ10 Corollary 2.8], we may assume that $B$ is stable. In this case, $B$ can be realized as a crossed product $B=A\rtimes_\alpha\mathbb{Z}$ with $A$ an AF-algebra by [@Ror02 Corollary 8.4.11]. This implies that, given any finite subset $\mathcal{G}$ of $B=A\rtimes_\alpha\mathbb{Z}$ and $\epsilon>0$, we can use Proposition \[approximation2\] to find a natural number $n$ such that both $\iota_n\colon B\rightarrow M_n(B)$ and $\iota_{n+1}\colon B\rightarrow M_{n+1}(B)$ (as defined in \[iota\]) admit piecewise contractive, completely positive, 1-decomposable $(\mathcal{G},\epsilon)$-approximations $(F_i,\psi_i,\varphi_i)$, $i\in\{n,n+1\}$. Using the existence part of Kirchberg-Phillips classification ([@Ror02 Theorem 8.4.1]), we can choose an automorphism $\omega$ on $M_n(B)$ which induces $\omega_*=-{\operatorname{id}}$ on $K$-theory. Now consider the embedding $\iota$ given by $$\iota=\begin{pmatrix}\iota_{n+1} & 0 \\ 0 & \omega\circ\iota_n\end{pmatrix}\colon B\rightarrow M_{2n+1}(B).$$ By Lemma \[permanence\], $\iota$ also admits a piecewise contractive, completely positive, 1-decomposable $(\mathcal{G},\epsilon)$-approximation. By Proposition \[k-theory\] and the choice of $\omega$ we further have $$K_*(\iota)=K_*(\iota_{n+1})+K_*(\omega)\circ K_*(\iota_n)=(n+1)\cdot{\operatorname{id}}-n\cdot{\operatorname{id}}={\operatorname{id}}.$$ Since $B$ satisfies the UCT, this shows that the class of $\iota$ is invertible in $KK(B,M_{2n+1}(B))$ and its inverse is, using [@Ror02 Theorem 8.4.1] again, induced by a $*$-isomomorphism $\varrho\colon M_{2n+1}(B)\rightarrow B$. The uniqueness part of Kirchberg-Phillips classification ([@Ror02 Theorem 8.4.1]) implies that $\varrho\circ\iota$ is approximately unitarily equivalent to ${\operatorname{id}}_B$. Hence, by Lemma \[permanence\], ${\operatorname{id}}_B$ also admits a piecewise contractive, completely positive, 1-decomposable $(\mathcal{G},\epsilon)$-approximation.
Since $\mathcal{G}$ and $\epsilon$ were arbitrary, this shows ${\operatorname{dim}_{{\operatorname{nuc}}}}(B)\leq 1$. Because $B$ is not an AF-algebra, the nuclear dimension of $B$ must in fact be equal to 1.
For the general case, i.e. for Kirchberg algebras without any $K$-theory constraints, we immediately get the following. Note that the same estimate has by now also be obtained in [@MS13 Theorem 7.1] and [@BEMSW14 Corollary 3.4] by methods completely different from the one developed here. These results are more general though, they do not require the UCT.
A Kirchberg algebra in the UCT-class has nuclear dimension at most 3.
This is proven exactly as in [@WZ10 Theorem 7.5] except for using the improved estimate ${\operatorname{dim}_{{\operatorname{nuc}}}}(\mathcal{O}_\infty)=1$ which is given by Theorem \[main\].
The strategy of this paper can be used to obtain dimension estimates in more general situations (cf. Remark \[Efren\] below). In fact, there are only two ingredients needed: First, one needs sufficient classfication results, i.e. both existence and uniqueness results for homomorphisms between $C^*$-algebras in the class under consideration. Second, one requires the existence of a family of homomorphisms which admit good approximations in a decomposable manner. These homomorphisms should further induce maps on the level of invariants which can be used to build up an invertible element. In that case, one can proceed as in Theorem \[main\] to obtain dimension estimates. However, the proof of \[main\] seems to be limited to infinite $C^*$-algebras. This is because we require an automorphism $\omega$ which gives a sign on $K$-theory, i.e. which satisfies $\omega_*=-{\operatorname{id}}$. However, the map $-{\operatorname{id}}$ will not be an automorphism as soon as the invariant involves some non-trivial order structure on the $K$-groups. Therefore one cannot find such $\omega$ in this case.
\[Efren\] After having recieved a preprint of this paper, Ruiz, Sims and Tomforde estimated the nuclear dimension of certain graph $C^*$-algebras in [@RST13], following the strategy outlined above. Their result includes some, but not all Kirchberg algebras covered by Theorem \[main\]. Moreover, using classification results involving filtered-$K$-theory they also obtain estimates in the non-simple case.
\[AT-case\] In the presence of torsion in $K_1$, one can still represent stable UCT-Kirchberg algebras $B$ as crossed products $A\rtimes_\alpha\mathbb{Z}$. We can no longer choose $A$ to be an AF-algebra, but one can settle for $A$ to be an A$\mathbb{T}$-algebra instead. Furthermore, the crossed product can be constructed in such a way that the boundary maps in the associated Pimsner-Voiculescu sequence vanish (Theorem 3.6 of [@Ror95] and its proof) and therefore Proposition \[k-theory\] applies. Following the lines of \[main\] and using nuclear one-dimensionality of A$\mathbb{T}$-algebras, this gives an alternative proof for the general estimate ${\operatorname{dim}_{{\operatorname{nuc}}}}(B)\leq 3$. However, in this case it looks like the maps $\iota_n$ from \[iota\] even admit $2$-decomposable approximations when contructed more carefully. This would give an improved dimension estimate in the torsion case.
[3]{} S. Barlak, D. Enders, H. Matui, G. Szab[ó]{} and W. Winter. The Rokhlin property vs. Rokhlin dimension 1 on unital Kirchberg algebras, arXiv preprint [math.OA/1312.6289v2](http://arxiv.org/abs/1312.6289), 2014. B. Blackadar. [*K-theory for operator algebras*]{}, volume 5 of [*Mathematical Sciences Research Institute Publications*]{}. Cambridge University Press, Cambridge, second edition, 1998. N. P. Brown and N. Ozawa. [*$C^*$-algebras and finite-dimensional approximations*]{}, volume 88 of Graduate Studies in Mathematics, American Mathematical Society, Providence, RI, 2008. H. Lin. Asymptotic unitary equivalence and classification of simple amenable $C^*$-algebras, [*Invent. Math.*]{}, 183(2):385-450, 2011. H. Matui and Y. Sato. Decomposition rank of UHF-absorbing $C^*$-algebras, arXiv preprint [math.OA/1303.4371v2](http://arxiv.org/abs/1303.4371), to appear in [*Duke Math. J.*]{}, 2013. M. R[ø]{}rdam. Classification of certain infinite simple $C^*$-algebras, [*J. Funct. Anal.*]{} 131, 415-458, 1995. M. Rørdam. [*Classification of nuclear simple $C^*$-algebras*]{}, Encyclopaedia Math. Sci., vol. 126, Springer-Verlag, Berlin 2002, pages 1-145 E. Ruiz, A. Sims and M. Tomforde. The nuclear dimension of graph $C^*$-algebras, arXiv preprint [math.OA/312.0507](http://arxiv.org/abs/1312.0507), 2013. Y. Sato, S. White and W. Winter. Nuclear dimension and $\mathcal{Z}$-stability, arXiv preprint [math.OA/1403.0747](http://arxiv.org/abs/1403.0747), 2014. W. Winter. Nuclear dimension and $\mathcal{Z}$-stability of pure $C^*$-algebras, [*Invent. Math.*]{}, 187(2):259-342, 2012. W. Winter and J. Zacharias. The nuclear dimension of $C^*$-algebras, [*Adv. Math.*]{} 224:461-498, 2010.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We evaluate the effect of the bottom-quark mass on resummed transverse momentum ($q_{T}$) distributions of supersymmetric Higgs bosons at the Tevatron and LHC. The mass of the bottom quark acts as a non-negligible momentum scale at small $q_{T}$ and affects resummation of soft and collinear radiation in this region. The improved treatment of the $b$-quark mass and kinematical effects leads to observable modifications in the resummed predictions for both colliders.'
author:
- |
Alexander Belyaev$^{1}$, Stefan Berge,$^{2}$ Pavel M. Nadolsky,$^{3}$\
Fredrick I. Olness,$^{2}$ and C.-P. Yuan$^{1}$
title: 'Heavy-flavor effects in supersymmetric Higgs boson production at hadron colliders '
---
Introduction\[sec:Introduction\]
================================
Understanding the nature of electroweak symmetry breaking is the central challenge for high-energy physics. The search for Higgs bosons, assumed to be responsible for the generation of gauge-boson and fermion masses, is the primary task for the existing and future colliders.
The Higgs sector may be represented by one complex scalar doublet, as it is economically realized in the Standard Model (SM), or by two or more doublets, as it takes place in the Minimal Supersymmetric Standard Model (MSSM) and its extensions. An important feature of MSSM is that, for large values of $\tan\beta$, the Yukawa couplings of the $b$-quarks to the neutral Higgs bosons $\mH$ (where $\mH=h$, $H$, or $A$) are strongly enhanced compared to the SM $b\bar{b}H_{SM}$ Yukawa coupling. Consequently, production of supersymmetric Higgs bosons in $b\bar{b}$ fusion can have a large cross section in supersymmetric extensions of the Standard Model [@Carena:2002es; @Spira:1997dg; @Assamagan:2004mu; @Balazs:1998nt; @Diaz-Cruz:1998qc].
The partonic processes contributing to the inclusive Higgs boson production with enhanced $b\bar{b}\mH$ coupling are represented by (a) $b\bar{b}\rightarrow\mH$; (b) $gb\rightarrow\mH b$; and (c) $gg\rightarrow b\bar{b}\mH$ scattering. The three processes (a,b,c) all give rise to the same hadronic final states, with two $B$-mesons appearing in different, but overlapping, regions of phase space. The distinction between the three processes depends very much on the factorization scheme adopted for the QCD calculation, as has been recently reviewed in Ref. [@Belyaev:2005nu]. The $\mH$${b\bar{b}}$ processes have been extensively studied recently in SM and MSSM scenarios [@Dicus:1998hs; @Campbell:2002zm; @Maltoni:2003pn; @Boos:2003yi; @Hou:2003fm; @Dittmaier:2003ej; @Harlander:2003ai; @Dawson:2003kb; @Kramer:2004ie; @Field:2004nc; @Dawson:2004sh; @Dawson:2005vi].
As shown in Refs. [@Belyaev:2002zz; @Belyaev:2005ct], the correct model for the transverse momentum distribution of the Higgs boson is crucial for unambiguous reconstruction of the Higgs boson mass in the $\mH\rightarrow\tau\tau$ decay channel. It is also important for discriminating the signal events from the backgrounds by examining the $q_{T}$ distribution of the Higgs boson in ${\mH}b\bar{b}$ associated production, followed by ${\mH}\rightarrow b\bar{b}$ decay [@Carena:2000yx]. The transverse momentum ($q_{T}$) distributions of Higgs bosons may be sensitive to the mass $m_{b}$ of the bottom quark when $q_{T}$ is comparable to $m_{b}$. In Refs. [@Berge:2005rv; @Belyaev:2005bs] , we study the effect of the initial-state multiple parton radiation and heavy-quark masses on the transverse momentum distribution in the $b\bar{b}\rightarrow\mH$ process. Here we summarize the results of those two papers.
\[sec:Formalism\]Transverse Momentum Resummation for Massive Quarks
===================================================================
The resummed differential cross section for inclusive production of Higgs bosons in scattering of initial-state hadrons $A$ and $B$ takes the form [@Collins:1985kg]$$\frac{d\sigma}{dQ^{2}dydq_{T}^{2}}=\int_{0}^{\infty}\frac{{\rm b}d{\rm b}}{2\pi}\, J_{0}(q_{T}{\rm b})\, W({\rm b},Q,x_{A},x_{B},m_{b})\,\,+\,\, Y(q_{T},Q,y,m_{b}),\label{WYDY}$$ where $y$ is the rapidity of the Higgs boson, $x_{{A,B}}\equiv Qe^{\pm y}/\sqrt{S}$ are the Born-level partonic momentum fractions, $S$ is the square of the center-of-mass energy of the collider, and $J_{0}(q_{T}{\rm b})$ is the Bessel function. The resummed form factor $W$ is given in impact parameter ($\mathrm{b}$) space and factorizes as$$W({\rm b},Q,x_{{A}},x_{{B}},m_{b})=\frac{\pi}{S}\,\sum_{j,k}\sigma_{jk}^{(0)}\, e^{-\mathcal{S}({\rm b},Q,m_{b})}\,\,{\mathcal{\overline{P}}}_{j/A}(x_{A},{\rm b},m_{b})\,\,{\mathcal{\overline{P}}}_{k/B}(x_{{B}},{\rm b},m_{b}),\label{WCSS}$$ where the summation is performed over the relevant parton flavors $j$ and $k$. Here, $\sigma_{jk}^{(0)}$ is a product of the Born-level prefactors, $e^{-{\mathcal{S}}({\rm b},Q,m_{b})}$ is an exponential of the Sudakov integral $$\begin{aligned}
{\mathcal{S}}({\rm b},Q,m_{b})\equiv\int_{b_{0}^{2}/{\rm b}^{2}}^{Q^{2}}\frac{d\bar{\mu}^{2}}{\bar{\mu}^{2}}\biggl[{\mathcal{A}}(\alpha_{s}(\bar{\mu}),m_{b})\,\mathrm{ln}\biggl(\frac{Q^{2}}{\bar{\mu}^{2}}\biggr)+{\mathcal{B}}(\alpha_{s}(\bar{\mu}),m_{b})\biggr],\label{Sudakov}\end{aligned}$$ with $b_{0}\equiv2e^{-\gamma_{E}}\approx1.123$, and ${\mathcal{\overline{P}}}_{j/A}(x,{\rm b},m_{b})$ are the ${\rm b}$-dependent parton distributions for finding a parton of type $j$ in the hadron $A$. In the perturbative region (${\rm b}^{2}\ll\nolinebreak\Lambda_{QCD}^{-2}$), the distributions ${\mathcal{\overline{P}}}_{j/A}(x,\mathrm{b},m_{b})$ factorize as$$\begin{aligned}
\left.\overline{{\mathcal{P}}}_{j/A}(x,\mathrm{b},m_{b})\right|_{\mathrm{b}^{2}\ll\Lambda_{QCD}^{-2}} & = & \sum_{a=g,u,d,...}\,\int_{x}^{1}\,\frac{d\xi}{\xi}\,{\mathcal{C}}_{j/a}(x/\xi,\mathrm{b},m_{b},\mu_{F})\, f_{a/A}(\xi,\mu_{F})\label{CxF}\end{aligned}$$ into a convolutions of the Wilson coefficient functions ${\mathcal{C}}_{j/a}(x,\mathrm{b},m_{b},\mu_{F})$ and $k_{T}$-integrated parton distributions $f_{a/A}(\xi,\mu_{F})$. The Sudakov exponential and ${\rm b}$-dependent parton densities resum contributions from soft and collinear multi-parton radiation, respectively. $Y\equiv\mbox{PERT}-\mbox{ASY}$ is the difference between the finite-order cross section (PERT) and its asymptotic expansion in the small-$q_{T}$ limit (ASY).
The Higgs cross sections depend on the mass $m_{b}$ of the bottom quark. The distributions $\overline{{\mathcal{P}}}_{j/A}(x,{\rm b},m_{b})$ for the heavy quarks ($j=c,b$) cannot be reliably evaluated at all impact parameters if a conventional factorization scheme, such as the zero-mass variable-flavor number (ZM-VFN, or massless) scheme, is used. The reason is that ${m_b}$ acts as an additional large momentum scale, which, depending on the value of ${\rm b}$, introduces large logarithms $\ln^{n}({m_b}{\rm b})$ or non-negligible terms $\propto({m_b}{\rm b})^{n}$. The situation encountered here is reminiscent of the heavy-quark contributions to the DIS structure functions $F_{i}(x,Q^{2}),$ which are not adequately described by the conventional factorization schemes at either small or large momentum transfers $Q^{2}$ (see, for instance, [@Tung:2001mv]). To work around this complication, Ref. [@Nadolsky:2002jr] proposed to formulate the CSS formalism in a general-mass variable flavor number (GM-VFN) scheme [@Collins:1998rz], which correctly evaluates the heavy-quark mass effects at all momentum scales. Among all GM-VFN factorization schemes, the S-ACOT scheme [@Collins:1998rz; @Kramer:2000hn] was found to be well-suited for the efficient calculation of the CSS resummed cross sections. In particular, in this heavy-quark CSS (CSS-HQ) formalism [@Nadolsky:2002jr] the dependence on $m_{b}$ is dropped in all ${\cal O}(\alpha_{s})$ terms in Eq. (\[WYDY\]) except for $\overline{{\mathcal{P}}}_{b/A}(x,{\rm b},m_{b})$.
The dependence of the bottom-quark parton density $\overline{{\mathcal{P}}}_{{b}/p}(x,{\rm b},{m_b})$ on the impact parameter is shown in Fig. \[fig:PbA\]. The ZM-VFN parton density $\overline{{\mathcal{P}}}_{{b}/p}(x,{\rm b},{m_b})$ is not properly defined below the threshold $\mu_{F}={m_b}$ (or above ${\rm b}=b_{0}/{m_b}$). It was continued to large ${\rm b}$ in the previous calculations using an effective “ZM-VFN” approximation described in Ref. [@Berge:2005rv]. The S-ACOT parton density $\overline{{\mathcal{P}}}_{{b}/p}(x,{\rm b},{m_b})$ is well-defined at all ${\rm b}.$ It reduces to the ZM-VFN result at ${\rm b}\ll b_{0}/{m_b}$ and is strongly suppressed at ${\rm b}\gg b_{0}/{m_b}$. The suppression is caused by the decoupling of the heavy quarks in the parton densities at $\mu_{F}$ much smaller than ${m_b}$ (${\rm b}$ much larger than $b_{0}/{m_b}$). Consequently the impact of the non-perturbative contributions from ${\rm b}\gtrsim1\mbox{\, GeV}^{-1}$ is reduced in the heavy-quark channels compared to the light-quark channels.
(a)(b)
The massless (“ZM-VFN”) calculation therefore underestimates the true behavior at ${\rm b}>\nolinebreak0.1\mbox{ GeV}^{-1}$ and small $q_{T}$. This effect can be seen in Fig. \[fig:bbh\_lhc\], which displays $d\sigma/dq_{T}$ for $b\bar{b}\rightarrow\mH$ boson production at (a) the Tevatron and (b) LHC.[^1] At the Tevatron, the $q_{T}$ maximum shifts in the “ZM-VFN” approximation to larger $q_{T}$ by about $2$ GeV out of $11.7$ GeV (about $17\,$%). For a Higgs mass $M_{H}=200$ GeV, the maximum of $d\sigma/dq_{T}$ shifts by about 1.9 GeV out of $12.7$ GeV. At the LHC, the difference between the “ZM-VFN” and S-ACOT calculations is smaller compared to the Tevatron, because the influence of the ${\rm b}>0.1\,\mbox{GeV}^{-1}$ region is reduced at smaller momentum fractions $x$ probed at the LHC [@Qiu:2000hf]. The maximum of the $q_{T}$ distribution shifts in the “ZM-VFN” approximation by about $1.3$ GeV (9% out of $14.1$ GeV) to larger $q_{T}$. The results for other Higgs masses can be found in Ref. [@Berge:2005rv].
Numerical Comparison with PYTHIA\[sec:Numerical-results\]
=========================================================
The full $q_{T}$ dependence of the $b\bar{b}\rightarrow\mH$ process is affected by constraints on phase space available for QCD radiation (less relevant at small $q_{T}$). We illustrate the interplay of various effects by comparing the CSS-HQ resummation to the PYTHIA Monte Carlo program [@Sjostrand:2001yu]. We focus on production of the CP-odd Higgs particle $A$ for $\tan\beta=50$ (predictions for the other Higgs bosons can be obtained by rescaling the $b\bar{b}A$ coupling).
As compared to the CSS-HQ formalism, the PYTHIA calculation does not include contributions generated from the ${\cal C}$-functions and $Y$-term, and it evaluates the soft parton contributions at ${\cal O}(\alpha_{s}).$ Therefore, we start by comparing the PYTHIA $q_{T}$ distribution to the resummed $W$-term $W(1,1,0)$ in Eq. (\[WYDY\]), with the functions ${\cal A},$ ${\cal B}$, and ${\cal C}$ in Eqs. (\[Sudakov\]), (\[CxF\]) being evaluated at orders $\alpha_{s},$ $\alpha_{s}$, and $\alpha_{s}^{0}$, respectively. The orders of $\alpha_{s}$ in ${\cal A},$ ${\cal B}$, and ${\cal C}$ are shown as the arguments of $W(1,1,0)$.
It is evident from Fig. \[Fig:one\] that the shapes of $W(1,1,0)$ and PYTHIA $q_{T}$ distribution are very different, though the integrated rates (*i.e.,* the areas under the two curves) are about the same. The $q_{T}$ distribution from PYTHIA is narrower and peaks at lower $q_{T}$ than $W(1,1,0)$. The large discrepancy between the two curves is in contrast to the case of $W$ and $Z$ production via light-quark scattering, where the above two calculations predict similar, though not identical, $q_{T}$ distributions [@Balazs:1997xd].
A closer examination reveals that additional features must be implemented in the resummed cross section in order to reliably describe the $q_{T}$ distributions of Higgs bosons produced via $b\bar{b}$ fusion.
- The kinematical effects account for a large part of the disparity between $W(1,1,0)$ and PYTHIA. The bottom-quark PDF is a rapidly decreasing function of $x$ in the probed range of $x$. Consequently, approximations for the true partonic kinematics (especially those made for the light-cone momentum fractions $x$) may have a strong impact on the rate of $b\bar{b}$ scattering. This feature should be contrasted to the behavior of the light-quark PDF’s in $W$ and $Z$ production, which include a substantial valence component and vary slower with $x$. As a result, the kinematical approximations are less consequential in the $W$ and $Z$ case.\
[ ]{}\
When PYTHIA generates QCD radiation, the kinematical distributions of the final-state particles, including the quarks and gluons from the QCD showering, are modified to satisfy energy-momentum conservation at each stage of the showering. In the resummation calculation, information about the exact parton kinematics is included in the finite-order term (PERT). The resummed cross section is therefore expected to be closer to PYTHIA once the ${\cal O}(\alpha_{s})$ finite term, PERT(1)-ASY(1), is included. In the $W(1,1,0)$ calculation, the emitted gluons are assumed not to carry any momentum at all in the soft limit. To compensate for small, but nonzero energy of the soft gluon emissions, we introduce a “kinematical correction” (KC) in the W and ASY terms. This correction modifies the minimal values of partonic momentum fractions $x_{A}$ and $x_{B}$ in order to account for reduction of phase space available for collinear QCD radiation at large $q_{T}$.
- The lowest-order cross section $W(1,1,0)$ does not evaluate effects of the bottom-quark mass, which is first included in the ${\cal C}$-function of order $\alpha_{s}$. Also, additional, though not complete, $O(\alpha_{s}^{2})$ contributions arise in the Sudakov form factors inside PYTHIA when the next-to-leading order PDF’s are used. To account for both features, we evaluate the $W$ term at one higher order (2,2,1) and include the $m_{b}$ dependence using the CSS-HQ scheme.
(a)(b)
Thus, our full prediction TOT(1) is obtained by adding ${\textrm{W}}_{\textrm{KC}}^{\textrm{CSS-HQ}}$(2,2,1) (evaluated in the CSS-HQ formalism with the kinematical correction) and PERT(1), and subtracting ${\textrm{ASY}}_{\textrm{KC}}$(1). It is shown for $M_{A}=100$ GeV at the Tevatron in Fig. \[fig:three\](a) and $M_{A}=300$ GeV at the LHC in Fig. \[fig:three\](b). TOT(1) (solid line) is compared to the fixed-order prediction PERT(1) (dashed) and the PYTHIA prediction (dot-dashed). As one can see, the results for Tevatron and LHC are qualitatively similar. TOT(1) is closer to the PYTHIA prediction than $W(1,1,0)$, though the two distributions are not identical. The PYTHIA $q_{T}$ distribution peaks at lower $q_{T}$ than TOT(1). In the large $q_{T}$ region, the TOT(1) rate is larger than the PYTHIA rate.
Finally, Fig. \[fig:four\] shows the integrated cross section as a function of the minimal $q_{T}$ in the calculation for the Tevatron (left) and LHC (right). This is another way to illustrate the differences in the shapes of $q_{T}$ distributions obtained in the resummation, fixed-order, and PYTHIA calculations.
Conclusion\[sec:Conclusion\]
============================
Multiple parton radiation in $b$-quark scattering is conspicuously sensitive to effects of large bottom-quark mass $m_{b}$ and phase-space constraints on collinear emissions. Both $m_{b}$ dependence and phase-space dependence tangibly modify the shape of Higgs $q_{T}$ distributions in the $b\bar{b}\rightarrow\mH$ processes. The two types of effects were consistently implemented within the CSS resummation formalism for heavy-quark scattering [@Nadolsky:2002jr; @Berge:2005rv; @Belyaev:2005bs], realized in a massive (GM-VFN) factorization scheme. These corrections act on different $q_{T}$ regions. When the dependence on $m_{b}$ is taken into account, the position of the peak in the $d\sigma/dq_{T}$ distribution shifts to a lower $q_{T}$ value, leaving the rate at large $q_{T}$ essentially unchanged. The kinematical correction is effective in the high-$q_{T}$ region, where it largely reduces the Higgs production rate. As a result, we obtain an improved prediction for the full $q_{T}$ spectrum of Higgs bosons, an important piece of information needed for the future Higgs searches.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was supported in part by the U.S. Department of Energy under grant DE-FG03-95ER40908, contract W-31-109-ENG-38, and the Lightner-Sams Foundation. We also acknowledge the support in part by the U.S. National Science Foundation under awards PHY-0354838 and PHY-0244919.
[100]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
() (), .
, , , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, ****, (), .
(), .
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , ****, ().
() (), .
, , , ****, ().
, , (), .
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
(), ****, ().
, ****, ().
, , (), .
[^1]: Fig. \[fig:bbh\_lhc\] does not specify the overall normalization of $q_{T}$ distributions. It is valid for both Standard Model and supersymmetric Higgs bosons, since at leading order the supersymmetric result can be obtained by rescaling the Standard Model $b\bar{b}H_{SM}$ coupling: $g_{b\bar{b}\{ h,H,A\}}^{MSSM}=\{-\sin\alpha,\cos\alpha,\sin\beta\,\gamma_{5}\} g_{bbH}^{SM}/\!\cos\beta$. The net effect of $m_{b}$ on $q_{T}$ distributions will be the same for the SM and MSSM Higgs bosons, up to an overall normalization constant.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Two-dimensional (random) walks in cones are very natural both in combinatorics and probability theory: they are interesting in their own right and also in relation to other discrete structures. While walks restricted to the first quadrant have been well studied, the case of planar, non-convex cones—equivalent to the three-quarter plane after a linear transform—has been approached only recently. In this article we develop an analytic approach for the enumeration of walks in three quadrants. The advantage of this method is the uniform treatment of models corresponding to different step sets. After splitting the three quadrants into two symmetric convex cones, the method is composed of three main steps: write a system of functional equations satisfied by the counting generating function, which may be simplified into one single equation under symmetry conditions; transform the functional equation into a boundary value problem; and finally solve this problem, using a new concept of anti-Tutte’s invariant. The result is a contour-integral expression for the generating function. Such systems of functional equations also appear in queueing theory, namely, in the Join-the-Shortest-Queue model, which is still open in the non-symmetric case.[^1]'
author:
- 'K. Raschel[^2] [^3] & A. Trotignon[^4]'
bibliography:
- 'bibl.bib'
title: On walks avoiding a quadrant
---
---------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------
![Walks in the quarter plane and in three-quarter plane[]{data-label="fig:pictures_1/4_3/4"}](./figures/SimulationWalksQuarter "fig:") ![Walks in the quarter plane and in three-quarter plane[]{data-label="fig:pictures_1/4_3/4"}](./figures/SimulationWalks3Quarter.eps "fig:")
---------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------
Introduction
============
#### Context.
Two-dimensional (random) walks in cones are very natural both in combinatorics and probability theory: they are interesting in their own right and also in relation to other discrete structures, see [@BMMi-10] and references therein. Most of the attention has been devoted to the case of convex cones (equivalent to the quarter plane, after a linear transform), see Figure \[fig:pictures\_1/4\_3/4\], left. Thanks to an appealing variation of techniques, which complement and enrich each other (from combinatorics [@MiRe-09; @BMMi-10], complex analysis [@FaRa-12; @Ra-12], probability theory [@DeWa-15], computer algebra [@BoKa-10; @BoChHoKaPe-17], Galois difference equations [@DrHaRoSi-17]), one now has a very good understanding of these quadrant models, most of the time via their generating function, which counts the number of walks of length $n$, starting from a fixed point, ending at an arbitrary point $(i,j)$ and remaining in the cone (see below). Throughout the present work, all walk models will be assumed to have small steps, i.e., jumps in $\{-1,0,1\}^2$, see Figure \[fig:some\_models\] for a few examples. Let us recall a few remarkable results:
- *Exact expressions* exist for the generating function (to illustrate the variety of techniques, remark that the generating functions are infinite series in [@MiRe-09], positive part extractions of diagonals in [@BMMi-10], contour integrals on quartics in [@FaRa-12; @Ra-12], integrals of hypergeometric functions in [@BoChHoKaPe-17], etc.);
- The *algebraic nature of the trivariate generating function* is known: it is D-finite (that is, satisfies a linear differential equation with polynomial coefficients) if and only if a certain group of birational transformations is finite [@BMMi-10; @BoKa-10; @KuRa-12]. More recently, the differential algebraicity (existence of non-linear differential equations) of the generating function has also been studied [@BeBMRa-17; @DrHaRoSi-17];
- The *asymptotics of the number of excursions* (an excursion is a path joining two given points and remaining in the cone) [@BMMi-10; @FaRa-12; @DeWa-15; @BoRaSa-14] is known. Although the full picture is still incomplete, the asymptotics of the total number of walks is also obtained in several cases [@BMMi-10; @FaRa-12; @DeWa-15; @Du-14; @BoChHoKaPe-17].
Almost systematically, the starting point to solve the above questions is a functional equation that is satisfied by the generating function—it corresponds to the intuitive step-by-step construction of a walk, and will be stated later on, see and .
Given the vivid interest in combinatorics of walks confined to a quadrant, it is very natural to consider next the non-equivalent case of non-convex cones, as in particular the union of three quadrants $$\mathcal C=\{(i,j)\in\mathbb Z^2: i{\geqslant}0 \text{ or } j{\geqslant}0\},$$ see Figure \[fig:pictures\_1/4\_3/4\]. Following Bousquet-Mélou [@BM-16], we will also speak about walks avoiding a quadrant. Although walks avoiding a quarter plane have many common features with walks in a quarter plane, the former cited model is definitely much more complicated. To illustrate this fact, let us recall [@BM-16] that the simple walk (usually the simplest model, see Figure \[fig:some\_models\]) in three quadrants has the same level of complexity as the notoriously difficult Gessel’s model [@BoKa-10; @BMGessel-16] in the quadrant!
---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ----------------------------------------------------------------------
$\ { $\ { $\ { $\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E,W,N,S} \testb{NE,NW,SE,SW} \testb{E,W,SE,NW} \testb{NE,S,W} \testb{NE,W,E,SW}
}\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $ }\ $ }\ $ }\ $
Simple walk Diagonal walk Gouyou-Beauchamps Kreweras Gessel
---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ----------------------------------------------------------------------
Three-quadrant walks have been approached only recently. In [@BM-16], Bousquet-Mélou solves the simple walk and the diagonal walk (see Figure \[fig:some\_models\] for a representation of these step sets) starting at various points. She obtains an exact expression of the generating function and derives several interesting combinatorial identities, among which a new proof of Gessel’s conjecture via the reflection principle. Mustapha [@Mu-19] computes the asymptotics of the number of excursions for all small step models, following [@DeWa-15; @BoRaSa-14] (interestingly and in contrast with combinatorics, the probabilistic results [@DeWa-15; @Mu-19] on random walks in cones do not really depend on convexity). Using an original connection with planar maps, Budd [@Bu-17] obtains various enumerating formulas for planar walks, keeping track of the winding angle. These formulas can be used to enumerate simple walks in the three-quarter plane [@Bu-17; @MiSi-18]. As recalled in [@BM-16], the problem of diagonal walks on the square lattice was also raised in 2001 by David W. Wilson in entry A060898 of the OEIS [@OEIS].
In this article we develop the analytic approach of [@FaIa-79; @FaIaMa-17; @Ra-12] to walks in three quadrants, thereby answering to a question of Bousquet-Mélou in [@BM-16 Sec. 7.2].
#### Strategy.
Once a step set $\mathcal S$ is fixed, our starting point is a functional equation satisfied by the generating function $$\label{eq:generating_function_3/4}
C(x,y)=\sum_{n{\geqslant}0}\sum_{(i,j)\in\mathcal C}c_{i,j}(n)x^{i}y^{j} t^{n},$$ where $c_{i,j}(n)$ counts $n$-step $\mathcal S$-walks going from $(0,0)$ to $(i,j)$ and remaining in $\mathcal C$. Stated in , this functional equation translates the step-by-step construction of three-quadrant walks and takes into account the forbidden moves which would lead the walk into the forbidden negative quadrant. At first sight, this equation is very similar to its one-quadrant analogue (we will compare the equations to in Section \[subsec:kernel\_functional\_equations\]), the only difference is that negative powers of $x$ and $y$ arise: this can be seen in the definition of the generating function and on the functional equation as well, since the right-hand side of the latter involves some generating functions in the variables $\frac{1}{x}$ and $\frac{1}{y}$. This difference is fundamental and the methodology of [@BMMi-10; @Ra-12] (namely, performing algebraic substitutions or evaluating the functional equation at well-chosen complex points) breaks down, as the series are no longer convergent.
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0);
The idea in [@BM-16] is to see $\mathcal C$ as the union of three quarter planes, and to state for each quadrant a new equation, which is more complicated but (by construction) may be evaluated. Our strategy follows the same line: we split the three-quadrant in two convex cones (of opening angle $\frac{3\pi}{4}$, see Figure \[fig:three-quarter\_split\]) and write a system of two functional equations, one for each domain. The drawbacks of this decomposition is that it increases the complexity:
- There are two functional equations instead of one;
- The functional equations involve more unknowns (corresponding to the diagonal and close-to-diagonal terms) in their right-hand sides, see Appendix \[app:proof\_lem1\].
On the other hand:
- The fundamental advantage is that the new equations may be evaluated—and ultimately will be solved;
- Unexpectedly, this splitting of the cone allows us to relate the combinatorial model of walks avoiding a quadrant to an interesting class of space inhomogeneous walks, among which a well-known problem in queueing theory: the Join-the-Shortest-Queue (JSQ) model, see Figure \[fig:JSQ\].
#### Three-quadrant walks and space inhomogeneous walks.
Doing two simple changes of variables (one for each wedge, see in particular ), the decomposition of the three-quarter plane shown on Figure \[fig:three-quarter\_split\] is equivalent to splitting a half-plane into the union of two quadrants and a half-line, see Figure \[fig:half-plane\_split\]. We end up with a *space inhomogeneous model* in the half-plane. On the $y$-axis, the step set is composed of mixed steps from the step sets of the left and right quadrants. In particular, starting with a symmetric step set in the three-quarter plane, say the simple walk, one obtains (with the terminology of Figure \[fig:some\_models\]) Gouyou-Beauchamps’ model in the left quadrant and Gessel’s model in the right one, see Figure \[fig:half-plane\_split\_symm\] on the left. This model is equivalent to study Gessel’s step set in the quadrant, killed on the $x$-axis and reflected on the $y$-axis, see Figure \[fig:half-plane\_split\_symm\] in the middle. A related model is studied in [@BeOwRe-18; @XuBeOw-19]: in these articles, the authors work on walks in the quadrant with different weights on the boundary, see Figure \[fig:half-plane\_split\_symm\] on the right, and give some results on the nature of the generating function of such walks.
A related, simpler model (that we don’t solve in the present paper) would be to split the full plane into two half-planes and a boundary axis, to consider in each of the three regions a (different) step set, and to solve the associated walk model, see Figure \[fig:plane\_split\], right. Some other space inhomogeneous walk models have been investigated in [@BoMuSi-15; @Mu-19; @BuKa-19], but this notion of inhomogeneity does not match with ours. Indeed, a simple but typical example in [@BoMuSi-15; @Mu-19; @BuKa-19] consists in dividing $\mathbb Z^2$ into the odd and even lattices, and to assign to each point of the even (resp. odd) lattice a certain step set $\mathcal S$ fixed a priori (resp. another step set $\mathcal S'$), see Figure \[fig:plane\_split\], left.
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (-1,0) – (4.5,5.5); (0,-1) – (5.5,4.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (3,-3) – (2,-3); (3,-3) – (2,-4); (3,-3) – (3,-4); (3,-3) – (4,-2); (3,-3) – (3,-2); (-3,3) – (-4,3); (-3,3) – (-3,4); (-3,3) – (-2,4); (-3,3) – (-3,2); (-3,3) – (-4,2); (3,3) – (2,3); (3,3) – (2,2); (3,3) – (3,2); (3,3) – (4,4); (3,3) – (3,4);
(1,0) – (1,6.5) – (6.5,6.5) – (6.5,0); (-1,0) – (-6.5,0) – (-6.5,6.5) – (-1,6.5); (-6.5,0) grid (6.5,6.5); (-1,0) – (-1,6.5); (1,0) – (1,6.5); (0,-0.5) – (0,6.5); (-6.5,0) – (6.5,0); (0,0) – (0,6.5); (4,4) – (3,4); (4,4) – (4,5); (4,4) – (5,5); (4,4) – (5,4); (4,4) – (4,3); (0,4) – (1,4); (0,4) – (-1,4); (0,4) – (0,3); (0,4) – (0,5); (0,4) – (-1,5); (-4,4) – (-5,4); (-4,4) – (-4,5); (-4,4) – (-4,3); (-4,4) – (-5,5); (-4,4) – (-3,3);
(1,0) – (1,6.5) – (6.5,6.5) – (6.5,0); (-1,0) – (-6.5,0) – (-6.5,6.5) – (-1,6.5); (-6.5,0) grid (6.5,6.5); (-1,0) – (-1,6.5); (1,0) – (1,6.5); (0,-0.5) – (0,6.5); (-6.5,0) – (6.5,0); (0,0) – (0,6.5); (4,4) – (3,4); (4,4) – (5,4); (4,4) – (5,5); (4,4) – (3,3); (0,4) – (1,4); (0,4) – (-1,4); (0,4) – (1,5); (0,4) – (-1,5); (-4,4) – (-3,4); (-4,4) – (-5,4); (-4,4) – (-3,3); (-4,4) – (-5,5);
(0,0) – (0,6.5) – (6.5,6.5) – (6.5,0); (0,0) grid (6.5,6.5); (0,0) – (0,6.5); (0,0) – (6.5,0); (4,4) – (5,4); (4,4) – (5,5); (4,4) – (3,4); (4,4) – (3,3); (4,0) – (5,0); (4,0) – (5,1); (4,0) – (3,0); (0,4) – (1,4); (0,4) – (1,5); (0,0) – (1,0); (0,0) – (1,1);
(0,0) – (0,6.5) – (6.5,6.5) – (6.5,0); (0,0) grid (6.5,6.5); (0,0) – (0,6.5); (0,0) – (6.5,0); (0,0)–(1,1); (1,1)–(2,1); (2,1)–(1,0); (1,0)–(4,0); (4,0)–(5,1); (5,1)–(3,1); (3,1)–(5,3); (5,3)–(2,3); (2,3)–(1,2); (1,2)–(0,2); (0,2)–(2,4); (2,4)–(3,4); (3,4)–(4,5); (4,5)–(3,5); (3,5)–(4,6); (4,6)–(6,6); in [1,2,...,6]{} (,0) circle (0.1cm); in [1,2,...,6]{} (0,) circle (0.1cm); (0,0) circle (0.1cm);
(-5.5,-5.5) – (-5.5,5.5) – (5.5,5.5) – (5.5,-5.5); (-5.5,-5.5) grid (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); in [-4,-2,0,2,4]{} in [-5,-3,-1,1,3,5]{} (,) circle (0.1cm); in [-5,-3,-1,1,3,5]{} in [-4,-2,0,2,4]{} (,) circle (0.1cm); in [-4,-2,0,2,4]{} in [-4,-2,0,2,4]{} (,) circle (0.1cm); in [-5,-3,-1,1,3,5]{} in [-5,-3,-1,1,3,5]{} (,) circle (0.1cm);
(-5.5,1) – (-5.5,5.5) – (5.5,5.5) – (5.5,1); (-5.5,-1) – (-5.5,-5.5) – (5.5,-5.5) – (5.5,-1); (-5.5,-5.5) grid (5.5,5.5); (-5.5,0) – (5.5,0); (0,3) – (0,4); (0,3) – (0,2); (0,3) – (-1,3); (0,3) – (1,4); (0,-3) – (0,-2); (0,-3) – (0,-4); (0,-3) – (-1,-3); (0,-3) – (1,-3); (0,-3) – (1,-2); (0,-3) – (-1,-4); (0,-3) – (1,-4); (0,0) – (0,1); (0,0) – (-1,-1); (0,0) – (1,0);
Let us now present the JSQ model. This is a model with (say) two queues, in which (as its name suggests) the arriving customers choose the shortest queue; if the two queues happen to have the same length, then a queue is chosen according to an a priori fixed probability law. From a random walk viewpoint, this means splitting the quarter plane in two octants (cones of opening angle $\frac{\pi}{4}$) as on Figure \[fig:JSQ\]. In general, the service times depend on the servers, and thus the transition probabilities are different in the upper and lower octants (one speaks about spatially inhomogeneous random walks, and of the general asymmetric JSQ). On the other hand, when the probability laws are symmetric in the diagonal, the model is said symmetric. Classical references are [@Fa-79; @Ia-79; @AdWeZi-91; @FoMD-01; @KuSu-03] and [@FaIaMa-17 Chap. 10]. Surprisingly, the non-symmetric JSQ is still an open problem: a typical interesting problem in queueing theory would be to compute a closed-form expression for the stationary distribution.
[c c]{}
(6.5,1.5) circle (0.3 cm); (6.5,1.5) circle (0.08 cm); (-7.75,0) node[Arrivals]{}; (9.5,0) node[Departures]{}; (-5.25,0 ) – ( -7.25,-2 ) – (-7.25,2 ) – cycle; (0,2) rectangle (6,5); (-2,2) rectangle (6,5); in [-2,-1,...,6]{} (,2) – (,5) ; (3,-2) rectangle (6,-5); (-2,-2) rectangle (6,-5); in [-2,-1,...,6]{} (,-2) – (,-5) ; (-5,0.25) – (-2.25,3.5); (-5,-0.25) – (-2.25,-3.5); (6.25,3.5) – (9,0.25); (6.25,-3.5) – (9,-0.25);
&
(0,0) rectangle (5.5,5.5); (-1.5,-1.5) grid (5.5,5.5); (0,0) – (5.5,5.5); (0,-1.5) – (0,5.5); (-1.5,0) – (5.5,0); (3,3) – (4,3); (3,3) – (2,3); (3,3) – (3,4); (3,3) – (3,2); at (2.8,3.1) ; (3,1) – (3,2); (3,1) – (2,1); (3,1) – (3,0); at (2.8,0.1) ; (1,3) – (0,3); (1,3) – (1,2); (1,3) – (2,3); at (0.25,2.8) ;
Let us briefly notice that quadrant walks could also be treated with a JSQ approach, by decomposing the quarter plane into two octants as on Figure \[fig:JSQ\], see e.g. [@KuSu-03] for asymptotic results.
#### Main results: a contour-integral expression for the generating function.
Throughout this paper we will do the following assumption:
1. \[main\_hyp\]The step set $\mathcal S$ is symmetric (i.e., if $(i,j)\in\mathcal S$ then $(j,i)\in\mathcal S$) and does not contain the jumps $(-1,1)$ and $(1,-1)$.
An exhaustive list of which step sets obey \[main\_hyp\] is given on Figures \[fig:symmetric\_models\_finite\] and \[fig:symmetric\_models\_infinite\]. We are not able to deal with asymmetric walks (as we are unable to solve the asymmetric JSQ model, see above), because of the complexity of the functional equations. The jumps $(-1,1)$ and $(1,-1)$ are discarded for similar reasons: they would lead to additional terms in the functional equation (see Figure \[fig:diag\_bigstep\]).
[1]{}
---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- --
$\ { $\ { $\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E,W,N,S} \testb{NE,S,W} \testb{N,E,SW} \testb{NE,S,W,N,E,SW}
}\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $ }\ $ }\ $
Simple walk Kreweras Reverse Kreweras Union Kreweras
---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- --
[1]{}
---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ----------------------------------------------------------------------
$\ { $\ { $\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E,NE,N,SW} \testb{W,NE,S,SW} \testb{E,W,N,S,SW} \testb{E,W,N,S,NE}
}\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $ }\ $ }\ $
---------------------------------------------------------------------- ---------------------------------------------------------------------- ---------------------------------------------------------------------- ----------------------------------------------------------------------
Our main result is a contour-integral expression for the diagonal section $$D(x)=\sum_{n{\geqslant}0,i{\geqslant}0}c_{i,i}(n)x^it^n.$$ We shall see later that knowing $D(x)$ actually suffices to give a complete solution to the problem (i.e., to find an expression for $C(x,y)$ in ). Let us postpone to Theorem \[thm:first\_main\] the very precise statement, and instead let us give now the main idea and the shape of the solution. We will show that $$\label{eq:presentation_main_theorem}
D(x)=w'(x)f(w(x))\int g(u,w(z)) \frac{w'(z)}{w(x)-w(z)}{\,\mathrm{d}z},$$ where $f$ and $g$ are algebraic functions. The integral in is taken over a quartic curve, constructed from the step set of the model. The function $w$ is interpreted as a conformal mapping for the domain bounded by the quartic, and its algebraic nature heavily depends on the model under consideration: it can be algebraic (finite group case) or non-D-finite (otherwise).
#### Five consequences of our main results.
Our first contribution is about methodology: we show that under the symmetry condition \[main\_hyp\], three-quadrant walk models are exactly solvable, in the sense that their generating function admits an explicit (contour-integral) expression .
The second point is that our techniques allow to compare walks in a quadrant and walks in three quadrants. More precisely, it is proved in [@Ra-12] that the generating function counting quadrant walks ending on the horizontal axis can typically be expressed as $$\label{eq:presentation_1/4}
\widetilde f(x)\int \widetilde g(z)\frac{w'(z)}{w(x)-w(z)}{\,\mathrm{d}z},$$ with the same function $w$ as in but different functions $\widetilde f$ (rational) and $\widetilde g$ (algebraic). Though simpler, Equation is quite similar to . This similarity opens the way to prove combinatorial formulas relating the two models.
Our third corollary is a partial answer to two questions raised by Bousquet-Mélou in [@BM-16], that we briefly recall: first, could it be that for any step set associated with a finite group, the generating function $C(x, y)$ is D-finite? Second, could it be that for the four step sets \[Kreweras, reverse Kreweras, union Kreweras (see Figure \[fig:symmetric\_models\_finite\]) and Gessel (Figure \[fig:some\_models\])\], for which \[the quadrant generating function\] is known to be algebraic, $C(x, y)$ is also algebraic?
The expression rather easily entails that if $w$ is algebraic (which will correspond to the finite group case, see Section \[sec:preliminaries\]), the generating function $D(x)$ is D-finite, being the Cauchy integral of an algebraic function. On the other hand, when the group is infinite the function $w$ is non-D-finite by [@Ra-12 Thm. 2], and the expression uses non-D-finite functions (note, this does not a priori imply that $D(x)$ itself is non-D-finite, but does provide some evidence).
Next, although we do not solve them, the expression provides a way to attack the following questions:
- Starting from the integral , various asymptotic questions concerning quadrant models are solved in [@FaRa-12] (asymptotics of the excursions, of the number of walks returning to one axis, etc.). Similar arguments should lead to the asymptotics of walks in three quadrants. Remember, however, that the asymptotics of the excursion sequence is already found in [@Mu-19].
- A further natural question (still unsolved in the quadrant case) is to find, in the finite group case, a concrete differential equation (or minimal polynomial in case of algebraicity) for the generating function, starting from the contour integrals or . It seems that the technique of creative telescoping could be applied to the contour integral expressions.
- Several interesting (and sometimes surprising) combinatorial identities relating quadrant walks to three-quadrant walks are proved in [@BM-16] (in particular, a proof of the former Gessel’s conjecture by means of simple walks in $\mathcal C$ and the reflection principle). Moreover, Bousquet-Mélou asks in [@BM-16] whether $C(x,y)$ could differ from (a simple D-finite series related to) the quadrant generating function by an algebraic series? Taking advantage of the similarity between and provides a starting point to answer this question.
Finally, along the way of proving our results, we develop a noteworthy concept of anti-Tutte’s invariant, namely a function $g$ such that ($\overline{y}$ denoting the complex conjugate number of $y\in\mathbb C$) $$\label{eq:anti_Tutte}
g(y)=\frac{1}{g(\overline{y})}$$ when $y$ lies on the contour of . The terminology comes from [@BeBMRa-17], where a function $g$ satisfying to $g(y)=g(\overline{y})$ is interpreted as a Tutte invariant and is strongly used in solving the models. Originally, Tutte introduced the notion of invariant to solve a functional equation counting colored planar triangulations, see [@Tu-95]. Tutte’s equation is rather close to functional equations arising in two-dimensional counting problems. Interestingly, a function $g$ as in appears in the book [@CoBo-83], which proposes an analytic approach to quadrant walk problems (the latter is more general than [@FaIaMa-17] in the sense that it works for arbitrarily large positive jumps, i.e., not only small steps). In [@CoBo-83] it is further assumed that $g(\overline{y})=\overline{g(y)}$, so that with one has $\vert g(y)\vert=1$, and $g$ may be interpreted as a conformal mapping from the domain bounded by contour of onto the unit disc.
#### Equations with (too) many unknowns.
What about non-symmetric models? From a functional equation viewpoint, the latter are close to random walks with big jumps [@FaRa-15; @BoBMMe-18] or random walks with catastrophes [@BaWa-17], in the sense that the functional equation has more than two unknowns in its right-hand side. One idea to get rid of these extra terms is to transform the initial functional equation, as in [@BM-16], where Bousquet-Mélou solves the simple and diagonal models, starting from non-symmetric points ($(-1,0)$, for instance). Another idea, present in [@BoBMMe-18], is to extend the kernel method by computing weighted sums of several functional equations, each of them being an algebraic substitution of the initial equation. However, finding such combinations is very difficult in general.
From the complex analysis counterpart [@FaIaMa-17; @Ra-12; @FaRa-15], equations with many unknowns become systems of boundary value problems, which seem not to have a solution in the literature. It is also shown in [@FaIaMa-17 Chap. 10] that the asymmetric JSQ is equivalent to solving an integral Fredholm equation for the generating function, but again, no closed-form expression seems to exist.
#### A conjecture.
Although it is not directly inspired by our work, let us state the following. Consider an arbitrary finite group step set $\mathcal S$ (not necessarily satisfying to \[main\_hyp\] but with small steps). We conjecture that the generating function for walks in three quadrants $C(x,y)$ is algebraic as soon as the starting point $(i_0,j_0)\in\mathcal C$ is such that $i_0=-1$ or $j_0=-1$.
This conjecture is motivated by an analogy with the quarter plane, in which the following result holds: a finite group model (having group $G$) with starting point at $(i,j)$ is algebraic if and only if the orbit-sum $$\sum_{g\in G}(-1)^gg(x^{i+1}y^{j+1})$$ is identically zero, see [@BMMi-10; @BoKa-10; @KuRa-12]. Taking $i=-1$ in the sum above (which obviously is not possible in the quadrant case!) yields a zero orbit-sum—more generally, the orbit-sum of any function depending on only one of the two variables $x,y$ is zero.
#### Structure of the paper.
$ $
$\bullet$ Section \[sec:preliminaries\]: statement of various functional equations satisfied by the generating functions (in particular Lemma \[lem:functional\_equation\_sym\]), definition of the group of the model, study of the zero-set of the kernel
$\bullet$ Section \[sec:expression\_GF\]: statement of a boundary value problem (BVP) satisfied by the diagonal generating function (Lemma \[lem:BVP\]), resolution of the BVP (Theorems \[thm:first\_main\] and \[thm:second\_main\])
$\bullet$ Appendix \[app:expression\_Gluing\]: list and properties of conformal mappings used in Theorems \[thm:first\_main\] and \[thm:second\_main\]
$\bullet$ Appendix \[app:RHBVP\]: important statements from the theory of BVP
$\bullet$ Appendix \[app:proof\_lem1\]: proof of the main functional equation stated in Lemma \[lem:functional\_equation\_sym\]
#### Acknowledgments.
We are most grateful to Marni Mishna for her constant support and many enlightening discussions. We would like to also thank Alin Bostan, Irina Ignatiouk-Robert, Sami Mustapha and Michael Wallner for various discussions. Finally, we thank an anonymous referee for his/her numerous suggestions, which in particular led us to obtain series expansions of the contour integrals given in our main theorem.
Preliminaries {#sec:preliminaries}
=============
Kernel functional equations {#subsec:kernel_functional_equations}
---------------------------
The starting point is to write a functional equation satisfied by the generating function , which, as explained in the introduction, translates the step-by-step construction of a walk. Before dealing with this functional equation, let us define some important objects.
First of all, a step set $\mathcal S\subset \{-1,0,1\}^2$ is characterized by its inventory (or jump polynomial) $\sum_{(i,j)\in\mathcal{S}}x^i y^j$ as well as by the associated kernel $$\label{eq:kernel}
K(x,y)=xy \Bigg(t\sum_{(i,j)\in\mathcal{S}}x^i y^j -1\Bigg).$$ The kernel is a polynomial of degree two in $x$ and $y$, which we can write as $$\label{eq:kernel_expanded}
K(x,y)=\widetilde{a}(y)x^2+\widetilde{b}(y)x+\widetilde{c}(y) = a(x)y^2+b(x)y+c(x),$$ where $$\label{eq:coeff_kernel}
\left\{
\begin{array}{l l l}
a(x)=tx\sum_{(i,1)\in\mathcal{S}}x^i, & b(x)=tx\sum_{(i,0)\in\mathcal{S}}x^i -x, & c(x)=tx\sum_{(i,-1)\in\mathcal{S}}x^i,\\
\widetilde{a}(y)=ty\sum_{(1,j)\in\mathcal{S}}y^j, & \widetilde{b}(y)=ty\sum_{(0,j)\in\mathcal{S}}y^j -y, & \widetilde{c}(y)=ty\sum_{(-1,j)\in\mathcal{S}}y^j.
\end{array}
\right.$$ Define further $\delta_{-1,-1}=1$ if $(-1,-1)\in\mathcal S$ and $\delta_{-1,-1}=0$ otherwise. In the three-quarter plane, we can generalize Equation (12) in [@BM-16 Sec. 2.1] and deduce the following equation satisfied by $C(x,y)$ defined in : $$\label{eq:functional_equation_3/4}
K(x,y)C(x,y)=c(x)C_{-0}(x^{-1})+\widetilde{c}(y)C_{0-}(y^{-1})-t\delta_{-1,-1} C_{0,0}-xy,$$ where $$ C_{-0}(x^{-1})= \sum_{n{\geqslant}0,i{\leqslant}0} c_{i,0}(n)x^{i}t^{n},\quad
C_{0-}(y^{-1}) = \sum_{n{\geqslant}0,j{\leqslant}0} c_{0,j}(n)y^{j}t^{n}\quad\text{and}\quad
C_{0,0}=\sum_{n{\geqslant}0}c_{0,0}(n)t^{n}.$$ In comparison, let us recall the standard functional equation in the case of the quarter plane $$\mathcal Q=\{(i,j)\in\mathbb Z^2: i{\geqslant}0 \text{ and } j{\geqslant}0\}.$$ By [@BMMi-10 Lem. 4] and using similar notation as above, the generating function $$\label{eq:generating_function_1/4}
Q(x,y)=\sum_{n{\geqslant}0}\sum_{(i,j)\in\mathcal Q}q_{i,j}(n)x^{i}y^{j} t^{n}$$ satisfies the equation $$\label{eq:functional_equation_1/4}
K(x,y)Q(x,y)=c(x)Q_{-0}(x)+\widetilde{c}(y)Q_{0-}(y)-t\delta_{-1,-1} Q_{0,0}-xy,$$ where $$\begin{gathered}
\label{eq:sections_1/4}
Q_{-0}(x)= \sum_{n{\geqslant}0,i{\geqslant}0} q_{i,0}(n)x^{i}t^{n},\quad
Q_{0-}(y)=\sum_{n{\geqslant}0,j{\geqslant}0} q_{0,j}(n)y^{j}t^{n}\quad\text{and}\quad
Q_{0,0}=\sum_{n{\geqslant}0}q_{0,0}(n)t^{n}.\end{gathered}$$ At first sight, the two functional equations and are very similar. However, due to the presence of infinitely many terms with positive and negative valuations in $x$ or $y$, the first one is much more complicated, and almost all the methodology of [@BMMi-10; @Ra-12] (namely, performing algebraic substitutions or evaluating the functional equation at well-chosen complex points) breaks down, as the series are no longer convergent.
The idea in [@BM-16] is to see $\mathcal C$ as the union of three quarter planes, and to state for each quadrant a new equation, which is more complicated but (by construction) may be evaluated. Our strategy follows the same line: we split the three-quadrant cone in two domains (two cones of opening angle $\frac{3\pi}{4}$, see Figure \[fig:three-quarter\_split\]) and write two functional equations, one for each domain.
Functional equations for the $\frac{3\pi}{4}$-cone walks
--------------------------------------------------------
In this section and in the remainder of our paper, we shall use two different step sets, $\widehat{\mathcal S}$ and $\mathcal S$. The first one, $\widehat{\mathcal S}$, will refer to the main step set, corresponding to the walks in the three-quarter plane we are counting. Accordingly, we will rename all quantities associated to the main step set with a hat, for instance the kernel will be denoted by $\widehat{K}(x,y)$. The second step set, $\mathcal S$, is associated to $\widehat{\mathcal S}$ after the change of variable . Quantities with no hat will be associated to the step set $\mathcal S$, for instance the kernel $K(x,y)$. In order not to make the notation to heavy and because in this case there is no possible ambiguity, the only exception to this rule will be the coefficients $c_{i,j}(n)$ (with no hat), which will always correspond to $\widehat{\mathcal S}$.
Having said that, we start by splitting the domain of possible ends of the walks into three parts: the diagonal, the lower part $\{i{\geqslant}0, j{\leqslant}i-1\}$ and the upper part $\{j{\geqslant}0, i{\leqslant}j-1\}$, see Figure \[fig:three-quarter\_split\]. We may write $$\label{eq:equation_cut3parts}
C(x,y)=\widehat{L}(x,y)+\widehat{D}(x,y)+\widehat{U}(x,y),$$ where $$\widehat{L}(x,y)=\sum\limits_{\substack{i {\geqslant}0 \\ j{\leqslant}i-1 \\n{\geqslant}0}}c_{i,j}(n)x^i y^j t^n, \quad
\widehat{D}(x,y)=\sum\limits_{\substack{i{\geqslant}0 \\ n{\geqslant}0}}c_{i,i}(n)x^i y^i t^n \quad\text{and}\quad
\widehat{U}(x,y)=\sum\limits_{\substack{j {\geqslant}0 \\ i{\leqslant}j-1 \\n{\geqslant}0}}c_{i,j}(n)x^i y^j t^n.$$ Let $\delta_{i,j}=1$ if $(i,j)\in\mathcal S$ and $0$ otherwise.
=\[rectangle,draw,fill=white\] (0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (-1,0) – (4.5,5.5); (0,-1) – (5.5,4.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); at (3,-3) [$\widehat{L}(x,y)$]{}; at (-3,3) [$\widehat{U}(x,y)$]{}; at (7.4,6.5) [$\widehat{D}(x,y)$]{}; at (2.8,6.5) [$\widehat{D}^u(x,y)$]{}; at (7.6,4) [$\widehat{D}^\ell(x,y)$]{};
\[lem:functional\_equation\_sym\] For any step set which satisfies \[main\_hyp\] and starts at $(0,0)$, one has $$\begin{gathered}
\label{eq:functional_equation_sym}
\widehat{K}(x,y)\widehat{L}(x,y)=-\frac{1}{2}xy+txy\left(\delta_{-1,-1}x^{-1}y^{-1}+ \delta_{-1,0}x^{-1} \right)\widehat{L}_{0-}(y^{-1}) + \frac{1}{2}t\delta_{-1,-1}\widehat{D}(0,0)
\\
-xy\left(-\frac{1}{2}+t\left(\frac{1}{2}(\delta_{1,1}xy+\delta_{-1,-1}x^{-1}y^{-1})+ \delta_{0,-1}y^{-1} + \delta_{1,0}x \right) \right)\widehat{D}(x,y),\end{gathered}$$ with $\widehat{L}_{-0}(y^{-1})= \sum_{n{\geqslant}0,j<0} c_{0,j}(n)y^{j}t^{n}$.
The proof of Lemma \[lem:functional\_equation\_sym\] is postponed to Appendix \[app:proof\_lem1\], as it is elementary but a bit technical. The functional equation for non-symmetric models (as well as for symmetric models with non-diagonal starting points) is commented in Appendix \[app:proof\_lem1\]. Here we will only consider symmetric models starting at $(0,0)$, but notice that our study can be easily generalized to arbitrary diagonal starting points.
In order to simplify the functional equation , we perform the change of variable $$\label{eq:change_var}
\varphi(x,y)=(xy, x^{-1}).$$ Then becomes $$\label{eq:functional_equation_octant}
K(x,y)L(x,y)=c(x)L_{-0}(x)-x\left(x\widetilde{a}(y)+\frac{\widetilde{b}(y)}{2}\right)D(y)+\frac{1}{2}txD(0)-\frac{1}{2}xy,$$ where $K(x,y) =xy(t\sum_{(i,j)\in \mathcal{S}}x^{i-j}y^{i}-1)=x\widehat{K}(\varphi(x,y))$, $L_{-0}(x)=\sum_{n{\geqslant}0,j{\geqslant}1} c_{0,-j}x^j t^n$ and similarly $$\label{eq:def_diagonal}
L(x,y) = \widehat{L}(\varphi(x,y)) = \sum\limits_{\substack{i {\geqslant}1 \\ j{\geqslant}0 \\n{\geqslant}0}}c_{j,j-i}(n)x^i y^j t^n
\quad \text{and}\quad
D(y) = \widehat{D}(\varphi(x,y)) = \sum\limits_{\substack{i{\geqslant}0 \\ n{\geqslant}0}} c_{i,i}(n)y^i t^n.$$
The change of coordinates $\varphi$ simplifies the resolution of the problem, as the functional equation is closer to a (solvable) quadrant equation; compare with . Throughout the manuscript, functions with (resp. without) a hat will be associated to the step set $\mathcal S$ (resp. to the step set after change of variable $\varphi$). We have $$\mathcal{S}=\varphi(\widehat{\mathcal{S}})=\{(i-j,i):(i,j)\in\widehat{\mathcal{S}}\}.$$ For the reader’s convenience, we have represented on Table \[tab:transform\_models\] the effect of $\varphi$ on the symmetric models of Figures \[fig:symmetric\_models\_finite\] and \[fig:symmetric\_models\_infinite\]. We also remark on Figure \[fig:diag\_bigstep\] that the presence of anti-diagonal jumps $(-1,1)$ or $(1,-1)$ would lead to the bigger steps $(-2,-1)$ or $(2,1)$: this is the reason why they are discarded.
Group of the model
------------------
In this part and in Section \[subsec:Roots\_of\_the\_kernel\] as well, we remove the hat from our notation: indeed, the statements are valid for all step sets (with or without hat).
With our notation , the group of the walk is the dihedral group of bi-rational transformations $\langle\Phi,\Psi\rangle$ generated by the involutions $$\Phi(x,y)=\left( \frac{\widetilde{c}(y)}{\widetilde{a}(y)}\frac{1}{x},y\right) \qquad \text{and}\qquad
\Psi(x,y)=\left(x, \frac{c(x)}{a(x)}\frac{1}{y}\right).$$ It was introduced in [@Ma-72] in a probabilistic context and further used in [@FaIaMa-17; @BMMi-10]. The group $\langle\Phi,\Psi\rangle$ may be finite (of even order, larger than or equal to ${\geqslant}4$) or infinite. The order of the group for the $79$ non-equivalent quadrant models is computed in [@BMMi-10]: there are $23$ models with a finite group ($16$ of order $4$, $5$ of order $6$ and $2$ of order $8$) and $56$ models with infinite order.
For instance, the simple walk has a group of order $4$, while the three right models on Figure \[fig:symmetric\_models\_finite\] have a group of order $6$. Indeed, taking Kreweras model as an example, we have $\Phi(x,y)=(\frac{1}{xy},y)$ and $\Psi(x,y)=(x,\frac{1}{xy})$, and the orbit of $(x,y)$ under the action of $\Phi$ and $\Psi$ is $$(x,y)
\overset{\Phi}{\rightarrow}
(\textstyle{\frac{1}{xy}},y)
\overset{\Psi}{\rightarrow}
(\textstyle{\frac{1}{xy}},x)
\overset{\Phi}{\rightarrow}
(y,x)
\overset{\Psi}{\rightarrow}
(y,\textstyle{\frac{1}{xy}})
\overset{\Phi}{\rightarrow}
(x,\textstyle{\frac{1}{xy}})
\overset{\Psi}{\rightarrow}
(x,y).$$
---------------------------------------------------------------------------------------------------------------------------------------------
Model Image under $\varphi$
---------------------------------------------------------------------- ----------------------------------------------------------------------
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{N,S,E,W} \testb{E,W,NE,SW}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{S,W,NE} \testb{E,N,SW}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E,N,SW} \testb{S,W,NE}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E, NE, N, W, SW, S} \testb{E, NE, N, W, SW, S}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
---------------------------------------------------------------------------------------------------------------------------------------------
: Transformation $\varphi$ on the eight symmetric models (with finite group on the left and infinite group on the right) without the steps $(-1,1)$ and $(1,-1)$. In particular, the simple walk is related by $\varphi$ to Gessel’s model. After [@BM-16], this is another illustration that counting simple walks in three-quarter plane is related to counting Gessel walks in a quadrant[]{data-label="tab:transform_models"}
---------------------------------------------------------------------------------------------------------------------------------------------
Model Image under $\varphi$
---------------------------------------------------------------------- ----------------------------------------------------------------------
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E, NE, N, SW} \testb{NE, N, W, S}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{NE, W, SW, S} \testb{E, N, SW, S}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E, N, W, SW, S} \testb{E, NE, W, SW, S}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
$\ { $\ {
\begin{tikzpicture}[scale=0.8]{ \begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8); \coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0); \coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8); \coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0); \coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8); \coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2); \coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2); \coordinate (B2) at (-1.2,-1.2);
\testb{E, NE, N, W, S} \testb{E, NE, N, W, SW}
}\end{tikzpicture} }\end{tikzpicture}
}\ $ }\ $
---------------------------------------------------------------------------------------------------------------------------------------------
: Transformation $\varphi$ on the eight symmetric models (with finite group on the left and infinite group on the right) without the steps $(-1,1)$ and $(1,-1)$. In particular, the simple walk is related by $\varphi$ to Gessel’s model. After [@BM-16], this is another illustration that counting simple walks in three-quarter plane is related to counting Gessel walks in a quadrant[]{data-label="tab:transform_models"}
[m[2cm]{} m[1cm]{} m[2cm]{}]{} $\ {
\begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2);
\testb{NE,SE,NW,SW}
}\end{tikzpicture}
}\ $ & $\stackrel{\varphi}{\longrightarrow}$ & $\ {
\begin{tikzpicture}[scale=0.8]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2);
\testb{N, S, S2W, N2E}
}\end{tikzpicture}
}\ $
Roots and curves defined by the kernel {#subsec:Roots_of_the_kernel}
--------------------------------------
We define the discriminants in $x$ and $y$ of the kernel : $$\label{eq:discriminants}
\widetilde{d}(y)=\widetilde{b}(y)^2-4\widetilde{a}(y)\widetilde{c}(y)\qquad \text{and}\qquad d(x)=b(x)^2-4a(x)c(x).$$ The discriminant $d(x)$ (resp. $\widetilde{d}(y)$) in is a polynomial of degree three or four. Hence it admits four roots (also called branch points) $x_1, x_2, x_3, x_4$ (resp. $y_1, y_2, y_3, y_4$), with $x_4=\infty$ (resp. $y_4=\infty$) when $d(x)$ (resp. $\widetilde{d}(y)$) is of degree $3$.
\[lem:properties\_branch\_points\] Let $t\in(0,1/\vert\mathcal S\vert)$. The branch points $x_i$ are real and distinct. Two of them (say $x_1$ and $x_2$) are in the open unit disc, with $x_1 < x_2$ and $x_2>0$. The other two (say $x_3$ and $x_4$) are outside the closed unit disc, with $x_3>0$ and $x_3 < x_4$ if $x_4>0$. The discriminant $d(x)$ is negative on $(x_1,x_2)$ and $(x_3,x_4)$, where if $x_4<0$, the set $(x_3,x_4)$ stands for the union of intervals $(x_3,\infty)\cup(-\infty,x_4)$. Symmetric results hold for the branch points $y_i$.
Let $Y(x)$ (resp. $X(y)$) be the algebraic function defined by the relation $K(x,Y(x))=0$ (resp. $K(X(y),y)=0$). Obviously with and we have $$\label{eq:algebraic_expressions_Y_X}
Y(x)=\frac{-b(x)\pm\sqrt{d(x)}}{2a(x)}\qquad \text{and}\qquad X(y)=\frac{-\widetilde{b}(y)\pm\sqrt{\widetilde{d}(y)}}{2\widetilde{a}(y)}.$$ The function $Y$ has two branches $Y_0$ and $Y_1$, which are meromorphic on the cut plane $\mathbb{C} \setminus([x_1, x_2] \cup [x_3, x_4])$. On the cuts $[x_1,x_2]$ and $[x_3,x_4]$, the two branches still exist and are complex conjugate (but possibly infinite at $x_1=0$, as discussed in Lemma \[lem:properties\_curves\]). At the branch points $x_i$, we have $Y_0(x_i)=Y_1(x_i)$ (when finite), and we denote this common value by $Y(x_i)$.
Fix the notation of the branches by choosing $Y_0=Y_-$ and $Y_1=Y_+$ in . We further fix the determination of the logarithm so as to have $\sqrt{d(x)}>0$ on $(x_2,x_3)$. Then clearly with we have $$\label{eq:global_inequality}
\vert Y_0\vert {\leqslant}\vert Y_1\vert$$ on $(x_2,x_3)$, and as proved in [@FaIaMa-17 Thm. 5.3.3], the inequality holds true on the whole complex plane and is strict, except on the cuts, where $Y_0$ and $Y_1$ are complex conjugate.
A key object is the curve $\mathcal L$ defined by $$\label{eq:curve_L}
\mathcal L =Y_0([x_1,x_2])\cup Y_1([x_1,x_2])=\{y\in \mathbb C:
K(x,y)=0 \text{ and } x\in[x_1,x_2]\}.$$ By construction, it is symmetric with respect to the real axis. We denote by $\mathcal G_\mathcal L$ the open domain delimited by $\mathcal L$ and avoiding the real point at $+\infty$. See Figures \[fig:curves\] and \[fig:curve\_Kreweras\] for a few examples. Furthermore, let $\mathcal{L}_0$ (resp. $\mathcal{L}_1$) be the upper (resp. lower) half of $\mathcal{L}$, i.e., the part of $\mathcal L$ with non-negative (resp. non-positive) imaginary part, see Figure \[fig:conformal-functions\]. Likewise, we define $\mathcal M =X_0([y_1,y_2])\cup X_1([y_1,y_2])$.
\[lem:properties\_curves\] The curve $\mathcal L$ in is symmetric in the real axis. It intersects this axis at $Y(x_2)>0$.
If $\mathcal L$ is unbounded, $Y(x_2)$ is the only intersection point. This occurs if and only if neither $(-1,1)$ nor $(-1,0)$ belong to $\mathcal S$. In this case, $x_1=0$ and the only point of $[x_1, x_2]$ where at least one branch $Y_i(x)$ is infinite is $x_1$ (and then both branches are infinite there). Otherwise, the curve $\mathcal L$ goes through a second real point, namely $Y(x_1) {\leqslant}0$.
Consequently, the point $0$ is either in the domain $\mathcal G_\mathcal L$ or on the curve $\mathcal L$. The domain $\mathcal G_\mathcal L$ also contains the (real) branch points $y_1$ and $y_2$, of modulus less than $1$. The other two branch points, $y_3$ and $y_4$, are in the complement of $\mathcal G_\mathcal L \cup\mathcal L$.
[cccc]{}
plot\[smooth\] file[./figures/Y0.dat]{}; plot\[smooth\] file[./figures/Y1.dat]{}; (0,-10) – (0,10); (-12,0) – (13,0); (0,0) circle (0.15 cm); (-0.9,-1.2) node; (0.5,0) circle (0.15 cm); (0.9,-1.2) node; (12,0) circle (0.15 cm); (12.2,-1.2) node;
&
plot\[smooth, xshift= -290.5cm, xscale=100\] file[./figures/Y1infini2.dat]{};plot\[smooth, xshift= -290.5cm, xscale=100 \] file[./figures/Y0infini2.dat]{}; (0,-10) – (0,10); (-8,0) – (17,0); (0.1,0) circle (0.15 cm); (-0.6,-0.8) node; (0.8,0) circle (0.15 cm); (1.1,-0.8) node; (7,0) circle (0.15 cm); (7,-0.8) node; (9,0) circle (0.15 cm); (9,-0.8) node;
The step sets with jumps $\{{\sf E},{\sf N},{\sf SW}\}$, $\{{\sf E},{\sf NE},{\sf N},{\sf S}\}$ and $\{{\sf E},{\sf N},{\sf SW},{\sf S}\}$ have an unbounded contour, whereas the other models in Table \[tab:transform\_models\] have a bounded contour.
We close this section by introducing a particular conformal mapping for $\mathcal{G}_\mathcal L$, which will happen to be very useful for our study.
\[defn:CGF\] A function $w$ is said to be a conformal gluing function for the set $\mathcal{G}_\mathcal L$ if:
- $w$ is meromorphic in $\mathcal{G}_\mathcal L$ and admits finite limits on $\mathcal L$;
- $w$ is one-to-one on $\mathcal{G}_\mathcal L$;
- for all $y$ on $\mathcal L$, $w(y)=w(\overline{y})$.
For example, $w(y)=\frac{1}{2}(y+\frac{1}{y})$ is a conformal gluing function for the unit disc. See Appendix \[app:expression\_Gluing\] for a list of conformal gluing functions associated to the domains we are considering.
Expression for the generating functions {#sec:expression_GF}
=======================================
Main results and discussion
---------------------------
The first and crucial point is to prove that the diagonal $D(y)$ in satisfies a boundary value problem (BVP), in the sense of the lemma below, the proof of which is postponed to Section \[subsec:proof\_Lemma\_BVP\]. Let $\mathcal D$ denote the open unit disc and let $\widetilde d$ be the discriminant .
\[lem:BVP\] The function $D(y)$ can be analytically continued from the unit disc to the domain $\mathcal D\cup\mathcal{G}_\mathcal L$ and admits finite limits on $\mathcal{L}$. Moreover, $D(y)$ satisfies the following boundary condition, for $y\in\mathcal{L}$: $$\label{eq:bvp_RCshift_3/4}
\sqrt{\widetilde{d}(y)}D(y)-\sqrt{\widetilde{d}(\overline{y})}D(\overline{y})=y-\overline{y}.$$
In the remainder of the paper, we solve Lemma \[lem:BVP\] in two different ways, leading to the contour-integral expressions of $D(y)$ given in Theorem \[thm:first\_main\] and Theorem \[thm:second\_main\] below. Let us first remark that contrary to the usual quadrant case [@Ra-12], the prefactor $\sqrt{\widetilde{d}(y)}$ in front of the unknown $D(y)$ is not meromorphic in $\mathcal{G}_\mathcal L$, simply because it is the square root of a polynomial, two roots of which being located in $\mathcal{G}_\mathcal L$ (see Section \[subsec:Roots\_of\_the\_kernel\]). This innocent-looking difference has strong consequences on the resolution:
- Due to the presence of a non-meromorphic prefactor in , solving the BVP of Lemma \[lem:BVP\] requires the computation of an index (in the sense of Section \[sub:proof-thmbvpM1\] and Appendix \[app:RHBVP\]). This index is an integer and will be non-zero in our case, which will increase the complexity of the solutions. In Theorem \[thm:first\_main\] we solve the BVP, by taking into account this non-zero index.
- A second, alternative idea is to reduce to the case of a meromorphic boundary condition, and thereby to an index equal to $0$. To do so, we will find an analytic function $f$ with the property that $$\label{eq:decoupling_d_f}
\frac{\sqrt{\widetilde{d}(\overline{y})}}{\sqrt{\widetilde{d}(y)}}=\frac{f(\overline{y})}{f(y)}$$ for $y\in\mathcal L$, see Section \[sub:Anti-Tutte’s\] for more details. Such a function $f$ allows us to rewrite as $$\label{eq:other_way_BVP}
f(y)D(y)-f(\overline{y})D(\overline{y})=\frac{f(y)}{\sqrt{\widetilde{d}(y)}}(y-\overline{y}),$$ which by construction admits a meromorphic prefactor $f(y)$. In Theorem \[thm:second\_main\] we solve this zero-index BVP by this technique.
Although they represent the same function $D(y)$ (and so should be equal!), it will be apparent that the expressions obtained in Theorems \[thm:first\_main\] and \[thm:second\_main\] are quite different, and that the second one is simpler. However, we decided to present the two resolutions, as we think that they offer different insights on this boundary value method, and also because it is not obvious at all to be able to solve an equation of the form and thereby to reduce to the zero-index case. Recall (Section \[subsec:Roots\_of\_the\_kernel\]) that $\mathcal{L}_0$ is the upper half of the curve $\mathcal{L}$.
\[thm:first\_main\] Let $w$ be a conformal gluing function with a pole at $y_2$. For any step set $\widehat{\mathcal{S}}$ satisfying to \[main\_hyp\], the diagonal section can be written, for $y\in \mathcal{G}_{\mathcal L}$, $$D(y)=\frac{\Psi(w(y))}{2i\pi}\int_{\mathcal{L}_0}\frac{z-\overline{z}}{\sqrt{\widetilde{d}(z)}}\frac{w'(z)}{\Psi^+(w(z))(w(z)-w(y))}{\,\mathrm{d}z},$$ with $$\left\{
\begin{array}{r c l}
\Psi(y) &=& (y-Y(x_1))\exp{\Gamma(y)},\smallskip\\
\Psi^+(y) &=& (y-Y(x_1))\exp{\Gamma^+(y)},\smallskip\\
\Gamma(w(y))&=& \displaystyle\frac{1}{2i\pi}\int_{\mathcal{L}_0}\log\left(\frac{\sqrt{\widetilde{d}(\overline{z})}}{\sqrt{\widetilde{d}(z)}}\right)\frac{w'(z)}{w(z)-w(y)}{\,\mathrm{d}z}.
\end{array}
\right.$$ All quantities are computed relative to the step set $\mathcal{S}=\varphi(\widehat{\mathcal{S}})$ after the change of coordinates .
The left limit $\Gamma^+$ (and thereby $\Psi^+$) appearing in Theorem \[thm:first\_main\] can be computed with the help of Sokhotski-Plemelj formulas, that we have recalled in Proposition \[proposition:Sokhotski-Plemelj\] of Appendix \[app:RHBVP\]. We now turn to our second main result.
\[thm:second\_main\] Let $w$ be a conformal gluing function with a pole at $y_2$, with residue $r$. For any step set $\widehat{\mathcal{S}}$ satisfying to \[main\_hyp\], the diagonal section can be written, for $y\in \mathcal{G}_{\mathcal L}$, $$D(y)= \frac{-w'(y)\sqrt{r}}{\sqrt{\widetilde d'(y_2)(w(y)-w(Y(x_1)))(w(y)-w(Y(x_2)))}}\frac{1}{2i\pi}\int_{\mathcal{L}}\frac{ zw'(z)}{\sqrt{w(z)-w(y_1)}(w(z)-w(y))}{\,\mathrm{d}z}.$$ All quantities are computed according to the step set $\mathcal{S}=\varphi(\widehat{\mathcal{S}})$.
Here are some remarks about these results.
$\bullet$ First, it is important to notice that having an expression for $D(y)$ is sufficient for characterizing the complete generating function $C(x,y)$. Indeed, looking at Figure \[fig:some\_sections\] one is easily convinced that $$C(x,y)=L(\varphi^{-1}(x,y))+D(\varphi^{-1}(x,y))+L(\varphi^{-1}(y,x)),$$ with $$\left\{
\begin{array}{r c l}
L(x,y) &=&\displaystyle \frac{1}{K(x,y)}\left(c(x)L_{-0}(x)-x\left(x\widetilde{a}(y)+\frac{1}{2}\widetilde{b}(y)\right)D(y)-\frac{1}{2}xy\right),\medskip\\
L_{-0}(x) &=&\displaystyle \frac{x}{c(x)}\left( \frac{1}{2}Y_0(x)+\left( x \widetilde{a}(Y_0(x))+\frac{1}{2}\widetilde{b}(Y_0(x))\right)D(Y_0(x)) \right),\medskip\\
\varphi^{-1}(x,y)&=& (y^{-1},xy).
\end{array}
\right.$$
$\bullet$ Regarding the question of determining the algebraic nature of the diagonal series $D(y)$, the second expression is much simpler. Indeed, the integrand as well as the prefactor of the integral of Theorem \[thm:second\_main\] are algebraic functions of $y$, $z$, $t$ and $w$ (and its derivative) evaluated at various points. In addition, let us recall from [@Ra-12 Thm. 2] that $w$ is algebraic if and only if the group is finite, and non-D-finite in the infinite group case. See Table \[tab:recap\] for some implications. On the contrary, based on the exponential of a D-finite function, the integrand in Theorem \[thm:first\_main\] is a priori non-algebraic.
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
Model Nature of $w$ Nature of $Q(x,y)$ Nature of $C(x,y)$
---------------------------------------------------------------------- -------------------- ---------------------- ----------------------------------------------------
${ rational [@Ra-12] D-finite [@BMMi-10] D-finite by [@BM-16] and Thm. \[thm:second\_main\]
\begin{tikzpicture}[scale=0.4]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2);
\testb{N,S,E,W}
}\end{tikzpicture}
}$
${ algebraic [@Ra-12] algebraic [@BMMi-10] D-finite by Thm. \[thm:second\_main\]; algebraic?
\begin{tikzpicture}[scale=0.4]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2);
\testb{S,W,NE}
}\end{tikzpicture}
}\ {
\begin{tikzpicture}[scale=0.4]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2);
\testb{E,N,SW}
}\end{tikzpicture}
}\ {
\begin{tikzpicture}[scale=0.4]{
\coordinate (O) at (0,0); \coordinate (N) at (0,0.8);
\coordinate (NE) at (0.8,0.8); \coordinate (E) at (0.8,0);
\coordinate (SE) at (0.8,-0.8); \coordinate (S) at (0,-0.8);
\coordinate (SW) at (-0.8,-0.8);\coordinate (W) at (-0.8,0);
\coordinate (N2E) at (1.6,0.8);\coordinate (S2W) at (-1.6,-0.8);
\coordinate (NW) at (-0.8,0.8); \coordinate (B1) at (1.2,1.2);
\coordinate (B2) at (-1.2,-1.2);
\testb{E, NE, N, W, SW, S}
}\end{tikzpicture}
}$
non-D-finite in $t$ [@Mu-19];
non-D-finite in $x,y$?
-----------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Algebraic nature of the conformal mapping $w$, the quadrant generating function $Q(x,y)$ and the three-quarter plane counting function $C(x,y)$[]{data-label="tab:recap"}
$\bullet$ Lemma \[lem:BVP\] entails that the function $D(y)$ can be analytically continued to the domain $\mathcal D\cup\mathcal{G}_\mathcal L$. This is apparent on the first statement (using properties of contour integrals). This is a little bit less explicit on Theorem \[thm:second\_main\], because of the prefactor.
$\bullet$ Theorem \[thm:first\_main\] (resp. Theorem \[thm:second\_main\]) will be proved in Section \[sub:proof-thmbvpM1\] (resp. Sections \[sub:Anti-Tutte’s\] and \[subsec:proof\_thm\_3/4\_index0\]).
Simplification and series expansion in the reverse Kreweras case
----------------------------------------------------------------
In this part we apply Theorem \[thm:second\_main\] to reverse Kreweras walks in the three-quarter plane: we first make explicit all quantities appearing in the statement of Theorem \[thm:second\_main\], then we explain how to deduce the series expansion $$\label{eq:series_expansion_reverse_Kreweras}
D(0)=1+4\,{t}^{3}+46\,{t}^{6}+706\,{t}^{9}+12472\,{t}^{12}+239632\,{t}^{15}+4869440\,{t}^{18}+102995616\,{t}^{21}+O \left( {t}^{24} \right),$$ obtained here by direct enumeration. Let us recall that the coefficients in front of $t^{n}$ are the $c_{0,0}(n)$, which count the numbers of reverse Kreweras walks of length $n$, starting and ending at $(0,0)$ and confined to the three-quarter plane.
This symmetric model has the step set $\widehat{\mathcal S}=\{ (1,0), (0,1), (-1,-1) \}$, see Figure \[fig:symmetric\_models\_finite\]. The change of variable $\varphi$ defined in transforms it into Kreweras step set, see Figure \[fig:symmetric\_models\_finite\] and Table \[tab:transform\_models\], with $\mathcal S=\{ (1,1), (-1,0), (0,-1)\}$.
#### Computation of various quantities.
The kernel is $
K(x,y)=xy(t(xy+x^{-1}+y^{-1}) -1),
$ and with the notations and , we have $$a(x)=tx^{2},\quad b(x)=t-x,\quad c(x)=tx, \quad d(x)=(t-x)^{2}-4t^{2}x^{3},$$ and by symmetry $\widetilde a=a$, $\widetilde b=b$, $\widetilde c=c$ and $\widetilde d=d$. The branch points $x_{1}$ and $x_{2}$ are the roots of $d$ in the open unit disc, such that $x_{1}<x_{2}$. We easily obtain $$\left\{
\begin{array}{r c l}
x_{1}\ =\ y_1&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle t-2\,{t}^{5/2}+6\,{t}^{4}-21\,{t}^{11/2}+80\,{t}^{7}-{\frac {1287}{4}}{t}^{17/2}+O ( {t}^{10} ),\medskip\\
x_{2}\ =\ y_2&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle t+2\,{t}^{5/2}+6\,{t}^{4}+21\,{t}^{11/2}+80\,{t}^{7}+{\frac {1287}{4}}{t}^{17/2}+O ( {t}^{10} ).
\end{array}\right.$$ We further have $$\widetilde{d}'(y_2) = 2\,{t}^{5/4}-\frac{3}{2}\,{t}^{{17/4}}-8\,{t}^{{23/4}}-{\frac {603}{16}{t}^{{29/4}}}-174\,{t}^{{35/4}}+O \left( {t}^{{41/4}} \right).$$ We finally need to compute $Y(x_{1})$ and $Y(x_{2})$. By these quantities may be simplified as $$Y(x_{1})=-\sqrt{\frac{c(x_{1})}{a(x_{1})}}=-\sqrt{\frac{1}{x_{1}}} \qquad \text{and}\qquad Y(x_{2})=\sqrt{\frac{c(x_{2})}{a(x_{2})}}=\sqrt{\frac{1}{x_{2}}}.$$
#### Expression of the conformal gluing function.
As we shall prove in Lemma \[lem:list\_conformal\_gluing\_functions\], the following is a suitable conformal mapping: $$w(y) = \left(\frac{1}{y}-\frac{1}{W}\right)\sqrt{1-yW^2},$$ where $W$ is the unique power series solution to $W=t(2+W^3)$. As Theorem \[thm:second\_main\] is stated for a conformal gluing function with a pole at $y_{2}$ and not at $0$, we should consider instead $w_{y_{2}}=\frac{1}{w-w(y_{2})}$. We will need the following expansions: $$\left\{
\begin{array}{r c l}
W&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle 2\,t+8\,{t}^{4}+96\,{t}^{7}+1536\,{t}^{10}+O \left( {t}^{11} \right),\medskip\\
w_{y_{2}}(y_1)&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle \,{\frac {1}{4}}t^{-1/2}-\frac{3}{8}\,{t}^{5/2}-{\frac {97}{32}}{t}^{11/2}-{\frac {2611}{64}}{t}^{17/2}+O \left( {t}^{23/2} \right),\medskip\\
w(y_2)&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle \frac{1}{2}t^{-1}-2\, {t}^{1/2}-{t}^{2}-3\,{t}^{7/2}-7\,{t}^{5}-{\frac {115}{4}}{t}^{13/2}-90\,{t}^{8}-{\frac {3247}{8}}{t}^{19/2}+O \left( {t}^{11} \right),\medskip\\
w_{y_{2}}(Y(x_1))&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle -t-2\,{t}^{4}-18\,{t}^{7}+O \left( {t}^{10} \right),\medskip\\
w_{y_{2}}(Y(x_2))&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle -t-4\,{t}^{5/2}-18\,{t}^{4}-86\,t^{11/2}-418\,{t}^{7}-{\frac {4131}{2}}{t}^{17/2}+O \left( {t}^{10} \right),\medskip\\
w'(y_2)&\hspace{-1.5mm}=\hspace{-1.5mm}&\displaystyle {t}^{-1}-2\,{t}^{1/2}-5/2\,{t}^{2}-6\,{t}^{7/2}-{\frac {169}{8}}{t}^{5}-75\,{t}^{13/2}-{\frac {4957}{16}}{t}^{8}-1251\,{t}^{19/2}+O \left( {t}^{11} \right).
\end{array}\right.$$
#### Explicit expression of $D(y)$.
We apply now Theorem \[thm:second\_main\] and obtain$$\begin{gathered}
D(y)= \frac{-w_{y_{2}}'(y)}{\sqrt{(w_{y_{2}}(y)-w_{y_{2}}(-1/\sqrt{x_1}))(w_{y_{2}}(y)-w_{y_{2}}(1/\sqrt{x_2}))\widetilde{d}'(y_2)w'(y_2)}}
\\
\frac{1}{2i\pi}\int_{\mathcal{L}}\frac{ zw_{y_{2}}'(z)}{\sqrt{w_{y_{2}}(z)-w_{y_{2}}(y_1)}(w_{y_{2}}(z)-w_{y_{2}}(y))}{\,\mathrm{d}z},\end{gathered}$$ where $\mathcal{L}$ is the contour defined in , represented on Figure \[fig:curve\_Kreweras\]. Since $w_{y_{2}}(0)=0$ and $w_{y_{2}}'(0)=-1$ (remember that $w$ has a pole at $y=0$), evaluating at $y=0$ the expression above yields $$\label{eq:expression_D(0)}
D(0)= \frac{1}{\sqrt{w_{y_{2}}(-1/\sqrt{x_1})w_{y_{2}}(1/\sqrt{x_2})\widetilde{d}'(y_2)w'(y_2)}}
\frac{1}{2i\pi}\int_{\mathcal L}\frac{ zw_{y_{2}}'(z)}{\sqrt{w_{y_{2}}(z)-w_{y_{2}}(y_1)}w_{y_{2}}(z)}{\,\mathrm{d}z}.$$ The integrand in the right-hand side of the above equation is analytic on $\mathcal G_\mathcal L\setminus [y_1,y_2]$. Hence by Cauchy’s integral theorem, the contour $\mathcal L$ may be replaced by the unit circle $\mathcal C(0,1)$.
plot\[smooth\] file[./figures/KrewY0.dat]{}; plot\[smooth\] file[./figures/KrewY1.dat]{}; (0,0) circle (1cm); (-3,0) – (3,0); (0,-3) – (0,3); (1,0) circle (0.05 cm); (1.1,-0.4) node; (0.1,0) circle (0.05 cm); (0.18,-0.4) node; (0.25,0) circle (0.05 cm); (0.5,-0.4) node; (-2.6,0) circle (0.05 cm); (-3.2,-0.4) node; (2.26,0) circle (0.05 cm); (3,-0.4) node; (1.2, -1) node; (1.8, -2) node;
#### Expression of $D(0)$ as a function of $W$.
We could directly make a series expansion of $D(0)$ in $t$. However, for greater efficiency of the series expansion computation, we will first express $D(0)$ in terms of $W$, expand this integral in a series of $W$ and finally get back to a series in $t$. The generating function of excursions $D(0)$ can be written as $$\label{eq:D0-fct-W}
D(0)=-\frac{\sqrt{w(y_{1})-w(y_{2})}}{\sqrt{w_{y_{2}}(-1/\sqrt{x_{1}})w_{y_{2}}(1/\sqrt{x_{2}})\widetilde{d}'(y_{2})w'(y_{2})}}\frac{1}{2i\pi}\int_{\mathcal L}\frac{zw'(z)}{\sqrt{P-Sw(z)+w(z)^{2}}}{\,\mathrm{d}z},$$ with $$\label{eq:P-S}
\left\{
\begin{array}{l c l c l }
S &=& w(y_{1})+w(y_{2}) &=& \displaystyle\sqrt{2P-\frac{1}{4W^{2}}\left(W^{6}-20W^{3}-8 \right)},
\\
P &=& w(y_{1})w(y_{2}) &=& \displaystyle\frac{(1-W^{3})^{3/2}}{W^{2}}.
\end{array}
\right.$$ In order to derive , we start by writing the integrand of in terms of $w$: $$\int_{\mathcal L}\frac{zw'_{y_{2}}(z)}{\sqrt{w_{y_{2}}(z)-w_{y_{2}}(y_{1})}w_{y_{2}}(z)}{\,\mathrm{d}z}=-\sqrt{w(y_{1})-w(y_{2})}\int_{\mathcal L}\frac{zw'(z)}{\sqrt{(w(z)-w(y_{1}))(w(z)-w(y_{2}))}}{\,\mathrm{d}z}.$$ Then, note that $\widetilde{d}(y)=-4t^{2}(y-y_{1})(y-y_{2})(y-y_{3})=-4t^{2}(y-y_{1})(y-y_{2})\left(y-\frac{1}{W^{2}}\right)$. By identification we have $y_{1}+y_{2}=\frac{1}{4t^{2}}-\frac{1}{W^{2}}$ and $y_{1}y_{2}=\frac{W^{2}}{4}$. On the one hand, we can deduce that $$P=\left(\frac{1}{y_{1}}-\frac{1}{W}\right)\left(\frac{1}{y_{2}}-\frac{1}{W}\right)\sqrt{(1-y_{1}W^{2})(1-y_{2}W^{2})}=\frac{-(W-2t)(W-3t)}{W^{5}t^{2}}\sqrt{W^{6}t^{2}-W^{2}+8t^{2}}.$$ On the other hand, $$S^{2}=\left(1-y_{1}W^{2}\right)\left(\frac{1}{y_{1}^{2}}-\frac{2}{y_{1}W}+\frac{1}{W^{2}}\right)+\left(1-y_{2}W^{2}\right)\left(\frac{1}{y_{2}^{2}}-\frac{2}{y_{2}W}+\frac{1}{W^{2}}\right)+2P.$$ Both equations can be simplified into , using several times the minimal polynomial of $W$.
#### Series expansion.
Let us first expand in $t$ the factor in front of the integral in ; we get $$-\frac{\sqrt{w(y_{1})-w(y_{2})}}{\sqrt{w_{y_{2}}(-1/\sqrt{x_1})w_{y_{2}}(1/\sqrt{x_2})\widetilde{d}'(y_2)w'(y_2)}}=-\frac{1}{t}+O \left( {t}^{{10}} \right).$$ (One could even prove that the left-hand side of the above equation is identically equal to $-\frac{1}{t}$.) Then the factor in the integral in may be written as $$\begin{aligned}
\frac{zw'(z)}{\sqrt{P-Sw(z)+w(z)^{2}}}=&-\frac{1}{2z}\,{W}+ \left( -\frac{1}{4z^{2}}\,+\frac{z}{4} \right) {W}^{2}-\frac{1}{8{z}^{3}}\,{W}^{3}+ \left( -\frac{1}{16{z}^{4}}\,+\frac{3z^{2}}{16}\,\right) {W}^{4}
\\
&+ \left( -\frac{z}{32}-\frac{1}{16z^{2}}\,-\frac{1}{32z^{5}}\, \right) {
W}^{5}+ \left( -{\frac {3}{32\,{z}^{3}}}-{\frac {1}{64\,{z}^{6}}}+{
\frac {5\,{z}^{3}}{32}} \right) {W}^{6}
\\
&+ \left( -\frac{z^{2}}{32}\,+{\frac
{1}{64\,z}}-{\frac {3}{32\,{z}^{4}}}-{\frac {1}{128\,{z}^{7}}}
\right) {W}^{7}
\\
&+ \left( {\frac {1}{64\,{z}^{2}}}-{\frac {1}{256\,{z}^
{8}}}-{\frac {5}{64\,{z}^{5}}}-{\frac {z}{128}}+{\frac {35\,{z}^{4}}{
256}} \right) {W}^{8}
\\
&+ \left( -{\frac {1}{512\,{z}^{9}}}-{\frac {15}{
256\,{z}^{6}}}-{\frac {15\,{z}^{3}}{512}} \right) {W}^{9}
+O \left( {W}^{{10}} \right)\end{aligned}$$ and when we integrate the latter on the unit circle. Coming back to a series in $t$ we obtain $$\begin{gathered}
\frac{1}{2i\pi}\int_{\mathcal C(0,1)}\frac{zw'(z)}{\sqrt{P-Sw(z)+w(z)^{2}}}{\,\mathrm{d}z}=\\-t-4\,{t}^{4}-46\,{t}^{7}-706\,{t}^{10}-12472\,{t}^{13}
-239632\,{t}^{
16}-4869440\,{t}^{19}+O \left( {t}^{21} \right).\end{gathered}$$ Finally, putting every ingredients in order, we deduce .
Proof of Lemma \[lem:BVP\] {#subsec:proof_Lemma_BVP}
--------------------------
Assuming that $D(y)$ may be continued as in the statement of Lemma \[lem:BVP\], it is easy to prove the boundary condition . We evaluate the functional equation at $Y_0(x)$ for $x$ close to $[x_1,x_2]$: $$\label{cancel_kernel_3/4}
-\frac{1}{2}xY_0(x)+c(x)L_{-0}(x)-x(x\widetilde{a}( Y_0(x))+\frac{1}{2}\widetilde{b}(Y_0(x)) )D(Y_0(x))+\frac{1}{2}txD(0)=0.$$ We obtain two new equations by letting $x$ go to any point of $[x_1,x_2]$ with a positive (resp. negative) imaginary part. We do the subtraction of the two equations and obtain .
We now prove the analytic continuation. Note that similar results are obtained in [@FaIaMa-17 Thm. 3.2.3], [@Ra-12 Thm. 5] and [@BeBMRa-17 Prop. 19]. We follow the same idea as in [@Ra-12 Thm. 5]. Starting from we can prove that $$2c(X_0(y))L_{-0}(X_0(y))+X_0(y)\sqrt{\widetilde{d}(y)}D(y)-X_0(y)y=0$$ for $y\in \{y\in \mathbb{C} : \vert X_0(y)\vert<1\} \cap \mathcal{D}$, and then $$2c(X_0(y))\sum_{n{\geqslant}0,j{\geqslant}0}c_{0,-j-1}(n)X_0(y)^j t^n+\sqrt{\widetilde{d}(y)}D(y)-y=0$$ for $y\in \{y\in \mathbb{C} : |X_0(y)|<1 \text{ and } X_0(y)\neq 0\} \cap \mathcal{D}$ which can be continued in $ \mathcal{G}_{\mathcal{L}} \cup \mathcal{D}$. Being a power series, $D(y)$ is analytic on $\mathcal{D}$ and on $(\mathcal{G}_{\mathcal{L}} \cup \mathcal{D})\setminus \mathcal{D}$, $D(y)$ may have the same singularities as $X_0$ and $\sqrt{\widetilde{d}(y)}$, namely the branch cuts $[y_1,y_2]$ and $[y_3,y_4]$. But none of these segments belong to $(\mathcal{G}_{\mathcal{L}} \cup \mathcal{D})\setminus \mathcal{D}$, see Lemma \[lem:properties\_curves\]. Then $D(y)$ can be analytically continued to the domain $\mathcal{G}_{\mathcal{L}} \cup \mathcal{D}$. Using the same idea, we can prove that $D(y)$ has finite limits on $\mathcal{L}$. From , it is enough to study the zeros of $x\widetilde{a}( Y_0(x))+\frac{1}{2}\widetilde{b}(Y_0(x))$ for $x$ in $[x_1,x_2]$. Using the relation $X_0(Y_0(x))=x$ valid in $\mathcal G_\mathcal M$ (see [@FaIaMa-17 Cor. 5.3.5]) shows that it recurs to study the zeros of $\widetilde{d}(y)$ for $y\in(\mathcal{G}_{\mathcal{L}} \cup \mathcal{D})\setminus \mathcal{D}$. None of these roots ($y_1, y_2, y_3, y_4$) belong to the last set, then $D$ has finite limits on $\mathcal{L}$.
Proof of Theorem \[thm:first\_main\] {#sub:proof-thmbvpM1}
------------------------------------
The function $\sqrt{\widetilde{d}(y)}D(y)$ satisfies a BVP of Riemann-Carleman type on $\mathcal{L}$, see Lemma \[lem:BVP\]. Following the literature [@FaIaMa-17; @Ra-12], we use a conformal mapping to transform the latter into a more classical Riemann-Hilbert BVP. Throughout this section, we shall use notation and results of Appendix \[app:RHBVP\].
More precisely, let $w$ be a conformal gluing function for the set $\mathcal{G}_{\mathcal{L}}$ in the sense of Definition \[defn:CGF\], and let $\mathcal{U}$ denote the real segment $$\mathcal{U} = w(\mathcal{L}).$$ (With this notation, $w$ is a conformal mapping from $\mathcal{G}_{\mathcal{L}}$ onto the cut plane $\mathbb{C} \setminus \mathcal{U}$.) The segment $\mathcal{U}$ is oriented such that the positive direction is from $w(Y(x_2))$ to $w(Y(x_1))$, see Figure \[fig:conformal-functions\].
Define $v$ as the inverse function of $w$. The latter is meromorphic on $\mathbb{C} \setminus \mathcal{U}$. Following the notation of Appendix \[app:RHBVP\] and [@FaIaMa-17], we denote by $v^+$ and $v^-$ the left and right limits of $v$ on $\mathcal{U}$. The quantities $v^+$ and $v^-$ are complex conjugate on $\mathcal{U}$, and more precisely, since $w$ preserves angles, we have for $u\in\mathcal{U}$ and $y\in \mathcal{L}_0$ $$\left\{
\begin{array}{l c l c l}
v^+(u) &=& v^+(w(y)) &=& y, \\
v^-(u) &=& v^-(w(y)) &=& \overline{y},
\end{array}
\right.$$ see Figure \[fig:conformal-functions\] for an illustration of the above properties.
Then may be rephrased as the following new boundary condition on $\mathcal{U}$: $$\label{eq:pbl_RH}
D(v^+(u))=\frac{\sqrt{\widetilde{d}( v^-(u))}}{\sqrt{\widetilde{d}( v^+(u))}} D(v^-(u)) + \frac{v^+(u) - v^-(u)}{\sqrt{\widetilde{d}( v^+(u))}}.$$ As explained in Appendix \[app:RHBVP\] (see in particular Definition \[def:index\]), the first step in the way of solving the Riemann-Hilbert problem with boundary condition is to compute the index of the BVP.
\[prop:computation\_index\] The index of $\frac{\sqrt{\widetilde{d}( v^-(u))}}{\sqrt{\widetilde{d}( v^+(u))}}$ along the curve $\mathcal U$ is $-1$.
First of all, let us recall that when $\mathcal{L}$ is a closed curve of interior $\mathcal{G}_{\mathcal{L}}$ and $G$ is a non-constant, meromorphic function without zeros or poles on $\mathcal{L}$, then $$\operatorname{ind}_\mathcal{L} G=\frac{1}{2i\pi}\int_\mathcal{L}\frac{G'(z)}{G(z)}{\,\mathrm{d}z}=Z-P,$$ where $Z$ and $P$ are respectively the numbers of zeros and poles of $G$ in $\mathcal{G}_{\mathcal{L}}$, counted with multiplicity.
Applying this result to the function $d(y)$, which in $\mathcal{G}_{\mathcal L}$ has no pole and exactly two zeros (at $y_1$ and $y_2$—remember that $y_3$ and $y_4$ are also roots of $d(y)$ but are not in $\mathcal{G}_{\mathcal L}$), we have $\operatorname{ind}_{\mathcal{L}}\widetilde{d}(y)=2$, see Figure \[fig:index\_Gessel\] for an illustration.
We get then $$\begin{aligned}
\operatorname{ind}_{\mathcal{U}}\frac{\sqrt{\widetilde{d}(v^-(u))}}{\sqrt{\widetilde{d}(v^+(u))}}&=
\operatorname{ind}_{\mathcal{U}}\sqrt{\widetilde{d}(v^-(u))} - \operatorname{ind}_{\mathcal{U}}\sqrt{\widetilde{d}(v^+(u))}
=-\operatorname{ind}_{\mathcal{L}_1}\sqrt{\widetilde{d}(y)}-\operatorname{ind}_{\mathcal{L}_0}\sqrt{\widetilde{d}(y)}
\\
&=-\operatorname{ind}_{\mathcal{L}}\sqrt{\widetilde{d}(y)}
=-\frac{1}{2}\operatorname{ind}_{\mathcal{L}}\widetilde{d}(y)
=-1.\qedhere\end{aligned}$$
plot\[smooth\] file[./figures/R0.dat]{}; plot\[smooth\] file[./figures/R1.dat]{}; (0,-140) – (0,150); (-100,0) – (170,0);
With Theorem \[thm:Sol-BVP\], we deduce a contour-integral expression for the function $D(v(u))$, namely $$D(v(u))=\frac{\Psi(u)}{2i\pi}\int_{\mathcal{U}}\frac{v^+(s) - v^-(s)}{\sqrt{\widetilde{d}\left( v^+(s)\right)}}\frac{1}{\Psi^+(s)(s-u)}{\,\mathrm{d}s}.$$ With the changes of variable $u=w(y)$ and $s=w(z)$, we easily have the result of Theorem \[thm:first\_main\].
Anti-Tutte’s invariant {#sub:Anti-Tutte's}
----------------------
Our aim here is to find a function $f$ satisfying to the decoupling condition , namely $$\frac{\sqrt{\widetilde{d}(\overline{y})}}{\sqrt{\widetilde{d}(y)}}=\frac{f(\overline{y})}{f(y)},\qquad \forall y\in \mathcal L.$$ Indeed, such a function is used in a crucial way in Theorem \[thm:second\_main\].
Before giving a systematic construction of a function $f$ as above, we start by an example. For Gessel’s model, we easily prove that the function $$g(y) = \frac{y}{t(y+1)^2}$$ satisfies $g(Y_0)g(Y_1)=1$, and so for $x\in[x_1,x_2]$ the condition announced in the introduction. By the same reasoning as in the proof of Theorem \[thm:solving\_decoupling\_condition\] below, we deduce that $$f(y)=\frac{g(y)}{g'(y)}=\frac{y(y+1)}{y-1}$$ satisfies the decoupling condition .
However, a simple rational expression of $f$ as above does not exist in general. Instead, our general construction consists in writing $f$ in terms of a conformal mapping. Our main result is the following.
\[thm:solving\_decoupling\_condition\] Let $g$ be any conformal mapping from $\mathcal{G}_\mathcal L$ onto the unit disc $\mathcal D$, with the property that $g(\overline{y})=\overline{g(y)}$. Then the function $f$ defined by $$f=\frac{g}{g'}$$ satisfies the decoupling condition . Moreover, $f$ is analytic in $\mathcal{G}_\mathcal L$ and has finite limits on $\mathcal L$.
Finally, defining $h(z)=-z+\sqrt{z^2-1}$ and letting $w$ be a conformal gluing function as in Definition \[defn:CGF\], one can choose $$\label{eq:expression_g}
g(y)=h\left(\frac{2}{w(Y(x_2))-w(Y(x_1))}\left(w(y)-\frac{w(Y(x_1))+w(Y(x_2))}{2}\right)\right),$$ see Figure \[fig:conformal-functions\].
To obtain the expression of $g$ in for a given model, we refer to the list of conformal mappings $w$ provided in Appendix \[app:expression\_Gluing\].
plot\[smooth\] file[./figures/Y1.dat]{}; plot\[smooth\] file[./figures/Y0bis1.dat]{}; plot\[smooth\] file[./figures/Y0bis2.dat]{}; (0,-10) – (0,10); (-12,0) – (8,0); (-9.9,0) circle (0.15 cm) node\[below left\]; (5.8,0) circle (0.15 cm) node\[below right\]; (-4, -1.7) node; (-6,9) node; (-6,-9) node; (2,7) circle (0.15 cm) node\[above\]; (2,-7) circle (0.15 cm) node\[below\]; (2, 4) – (2, 6.7); (2, -4) – (2, -6.7);
(2,4) to\[out=20,in=160\] (12,4); (7,6) node; (30,4) to\[out=20,in=160\] (40,4); (3.5,-4) to\[out=350,in=190\] (38.5,-4) ; (21,-7.5) node;
(0,0) – (5,0); (5,0) – (16,0); (17.5,0) node; (10,4) node; (10,-4) node; (10, 3) – (10, 0.1); (10, -3) – (10, -0.1); (0,0) circle (0.15 cm) node\[below\]; (16,0) circle (0.15 cm) node\[below\];
(-1.5,0) – (1.5,0); (0,-1.5) – (0,1.5); (0:0) arc (0:180:1); (0:0) arc (0:120:1); (0:0) arc (0:-180:1); (1,0) circle (0.0225 cm) node\[below right\]; (0.2,-0.2) node;
We first prove that if $g$ is a conformal mapping from $\mathcal{G}_\mathcal L$ onto the unit disc $\mathcal D$ with the property that $g(\overline{y})=\overline{g(y)}$, then $f=\frac{g}{g'}$ satisfies the decoupling condition . First, for $x\in[x_1,x_2]$ one has $$g(Y_0(x))g(Y_1(x))=g(Y_0(x))g(\overline{Y_0(x)})=g(Y_0(x))\overline{g(Y_0(x))}=\vert g(Y_0(x))\vert^2=1.$$ Differentiating the identity $g(Y_0(x))g(Y_1(x))=1$, one finds on $[x_1,x_2]$ $$\frac{f(Y_0(x))}{f(Y_1(x))}= - \frac{Y_0'(x)}{Y_1'(x)}.$$ To conclude the proof, we show that on $[x_1,x_2]$ $$\label{eq:decoupling_sqrt_d}
\frac{\sqrt{\widetilde d(Y_0(x))}}{\sqrt{\widetilde d(Y_1(x))}}= - \frac{Y_0'(x)}{Y_1'(x)}.$$ To that purpose, let us first consider $x\in\mathcal G_\mathcal M\setminus [x_1,x_2]$. Differentiating the identity $K(x,Y_0(x))=0$ in yields $$\label{eq:first_id}
Y'_0(x)(2a(x)Y_0(x)+b(x))=-(a'(x)Y_0(x)^2+b'(x)Y_0(x)+c'(x)).$$ First, it follows from Section \[subsec:Roots\_of\_the\_kernel\] that $2a(x)Y_0(x)+b(x)=- \sqrt{d(x)}$. Moreover, differentiating in $x$ and using the relation $X_0(Y_0(x))=x$ valid in $\mathcal G_\mathcal M$ (see [@FaIaMa-17 Cor. 5.3.5]) shows that the right-hand side of satisfies $$a'(x)Y_0(x)^2+b'(x)Y_0(x)+c'(x)=-\sqrt{\widetilde{d}(Y_0(x))}.$$ Then for $x\in\mathcal G_\mathcal M\setminus [x_1,x_2]$, Equation becomes $$-\sqrt{d(x)}Y_0'(x)=\sqrt{\widetilde{d}(Y_0(x))}.$$ To complete the proof of , we let $x$ converge to a point $x\in[x_1,x_2]$ from above and then from below, and we compute the ratio of the two identities so-obtained. The minus sign in comes from that $$\lim_{x\downarrow[x_1,x_2]} \sqrt{d(x)}=-\lim_{x\uparrow[x_1,x_2]} \sqrt{d(x)},$$ see Section \[subsec:Roots\_of\_the\_kernel\].
Our second point is to show that the function $g$ in is a conformal mapping from $\mathcal{G}_\mathcal L$ onto the unit disc $\mathcal D$, which in addition is such that $g(\overline{y})=\overline{g(y)}$. This is obvious from our construction , since as illustrated on Figure \[fig:conformal-functions\], $g=h\circ \widehat w$ is the composition of the conformal mapping $h$ from the cut plane $\mathbb C\setminus [-1,1]$ onto the unit disc, by the conformal mapping $$\label{eq:def_hat_w}
\widehat w=\frac{2}{w(Y(x_1))-w(Y(x_2))}\left(w-\frac{w(Y(x_1))+w(Y(x_2))}{2}\right)$$ from $\mathcal G_\mathcal L$ onto the same cut plane.
The third item is to prove that $f$ has finite limits on $\mathcal L$, for any initial choice of conformal mapping $g$. We may propose two different proofs of this fact. First, we could prove that the function $f$ constructed from the particular function $g$ in has the desired properties (this follows from a direct study). Then as any two suitable conformal mappings $g_1$ and $g_2$ are necessarily related by a linear fractional transformation $$g_1=\frac{\alpha g_2+\beta}{\gamma g_2+\delta},$$ it is easily seen that all functions have indeed the good properties.
The second idea is to use a very general statement on conformal mapping. Namely, any conformal mapping which maps the unit disc onto a Jordan domain (the domain $\mathcal G_\mathcal L$) with analytic boundary (our curve $\mathcal L$) can be extended to a univalent function in a larger disc, see [@Du-83 Sec. 1.6]. As the extension is univalent, it becomes obvious that the derivative $g'$ in the denominator of $f$ cannot vanish.
Proof of Theorem \[thm:second\_main\] {#subsec:proof_thm_3/4_index0}
-------------------------------------
Our main idea here is to reformulate the initial boundary condition as , with the help of a function $f$ which is analytic in $\mathcal{G}_\mathcal L$, admits finite limits on $\mathcal L$ and satisfies on $\mathcal L$ the decoupling condition . Using Lemma \[lem:BVP\] and Theorem \[thm:solving\_decoupling\_condition\], we deduce that $f(y)D(y)$ is analytic in $\mathcal{G}_{\mathcal L}$ and has finite limits on $\mathcal L$. As a consequence, $f(y)D(y)$ satisfies a Riemann-Carleman BVP with index zero (in the sense of Definition \[def:index\]). Similarly to Section \[sub:proof-thmbvpM1\] and using again a conformal gluing function, we transform the latter BVP into a Riemann-Hilbert BVP on an open contour, whose solution is $$\label{eq:resolution_BVP0}
D(y)f(y)= \frac{1}{2i\pi}\int_{\mathcal{L}}\frac{zf(z)}{\sqrt{\widetilde{d}(z)}}\frac{w'(z)}{w(z)-w(y)}{\,\mathrm{d}z}+c,$$ where $c$ is constant in $y$, but may depend on $t$ (as recalled in Theorem \[thm:Sol-BVP\] from Appendix \[app:RHBVP\], the solutions to a BVP of index zero are determined up to one constant). Notice that $f$ cancels at $y_2$ (the unique pole of $w$) and the integral in the right-hand side of as well, it follows that $c=0$.
We now simplify the integrand in . First, noting that $h$ satisfies the simple differential equation $h'=\frac{-h}{\sqrt{z^2-1}}$, we obtain with our notation $$f=\frac{g}{g'}=\frac{h(\widehat{w})}{\widehat{w}'h'(\widehat{w})}=-\frac{\sqrt{\widehat{w}^2-1}}{\widehat{w}'}=-\frac{\sqrt{(w-w(Y(x_1)))(w-w(Y(x_2)))}}{w'}.$$ Furthermore, the conformal gluing function $w$ satisfies the following differential equation $$\label{eq:differential-eq-w}
\widetilde{d}(z)w'(z)^2=(w(z)-w(Y(x_1)))(w(z)-w(Y(x_2)))(w(z)-w(y_1)),$$ see [@FaIaMa-17 Sec. 5.5.2.2]. Taking the square root of in the neighborhood of $[y_2,y_3]\cap \mathcal G_\mathcal L$ gives $$-\sqrt{\widetilde{d}(z)}w'(z)=\sqrt{(w(z)-w(Y(x_1)))(w(z)-w(Y(x_2)))(w(z)-w(y_1))},$$ as $w$ is decreasing on $[y_2,y_3]\cap \mathcal G_\mathcal L$. It follows that $$\frac{f(z)}{\sqrt{\widetilde{d}(z)}}=\frac{1}{\sqrt{w(z)-w\left(y_1\right)}}.$$ The proof of Theorem \[thm:second\_main\] is complete.
\[rem:pole\_y2\] The differential equation is only true for the conformal gluing function $w$ whose expression is given in , with a pole at $y_2$. If instead we have at hand a function $w$ with a pole at $y_0\neq y_2$ (for example $y_0=0$, as in Lemma \[lem:list\_conformal\_gluing\_functions\]), we can consider $w_{y_{2}}=\frac{1}{w-w(y_2)}$, which instead of satisfies the differential equation $$\widetilde{d}(z)w_{y_{2}}'(z)^2=\widetilde{d}'(y_2)w'(y_2)(w_{y_{2}}(z)-w_{y_{2}}(Y_0(x_1))) (w_{y_{2}}(z)-w_{y_{2}}(Y_0(x_2))) (w_{y_{2}}(z)-w_{y_{2}}(y_1)).$$
Expression and properties of conformal gluing functions {#app:expression_Gluing}
=======================================================
A crucial ingredient in our main results (Theorems \[thm:first\_main\] and \[thm:second\_main\]) is the function $w(y)$, which we interpret as a conformal mapping from the domain $\mathcal G_\mathcal L$ onto a complex plane cut along an interval, see Section \[subsec:Roots\_of\_the\_kernel\]. In this appendix, we recall from [@Ra-12; @BeBMRa-17] an explicit expression as well as some analytic properties of this function, first in the finite group case, then for infinite group models.
Let us recall that if $w$ is a suitable mapping, then any $\frac{\alpha w+\beta}{\gamma w+\delta}$ is also a suitable mapping, as soon as $\alpha\delta-\beta\gamma\neq 0$. Therefore, all expressions hereafter are given up to such a fractional linear transform.
Finite group models
-------------------
We start by giving an expression of the conformal mapping $w(y)$ for the Kreweras trilogy of Figure \[fig:symmetric\_models\_finite\]. Let $W=W(t)$ (resp. $Z=Z(t)$) be the unique power series (resp. the unique power series with no constant term) satisfying $$\label{eq:W_Z}
W=t(2+W^3)\qquad \text{and}\qquad Z=t\frac{1-2Z+6Z^2-2Z^3+Z^4}{(1-Z)^2}.$$
\[lem:list\_conformal\_gluing\_functions\] Let $W$ and $Z$ as in . The function $$ w(y) = \left(\frac{1}{y}-\frac{1}{W}\right)\sqrt{1-yW^2}$$ is a conformal mapping for Kreweras model. Likewise, a conformal mapping for reverse Kreweras model is given by $$ w(y) = \frac{-ty^3+y^2+t}{2yt} -\frac{2y^2-yW^2-W}{2yW}\sqrt{1-yW(W^3+4)/4+y^2W^2/4}.$$ Finally, a conformal mapping for double Kreweras model is $$\begin{gathered}
w(y) = \sqrt{1-2yZ(1+Z^2)/(1-Z)^2+Z^2y^2}\frac{(Z(1-Z)+2yZ-(1-Z)y^2)}{2yZ(1-Z)(1+y)}\\
+\frac{Z(1-Z)^2-Z^2(-1+2Z+Z^2)y+(1-2Z+7Z^2-4Z^3)y^2-Z(1-Z)^2y^3}{2y(1+y)Z(1-Z)^2}.\end{gathered}$$
Notice that the functions $w$ given in Lemma \[lem:list\_conformal\_gluing\_functions\] all have a pole at $y=0$.
Expressions for $w$ are given in [@Ra-12 Thm. 3 (iii)], but some quantities in the latter statement (namely $\alpha$, $\beta$, $\delta$ and $\gamma$, all depending on $t$) are not totally explicit. So to derive the above expressions of $w$, we will rather use a combination of the works [@BMMi-10] and [@BeBMRa-17]. Indeed, algebraic expressions of $Q(0,y)$ in terms of $y$ and $t$ are obtained in [@BMMi-10] for the three Kreweras models (see Prop. 13, Prop. 14 and Prop. 15 there). On the other hand, an alternative formulation of $Q(0,y)$ as a rational function of $w(y)$, $y$ and $t$ is derived in [@BeBMRa-17] (see Thm. 23 and Table 8 there). The formulas of Lemma \[lem:list\_conformal\_gluing\_functions\] are obtained by equating the two expressions.
An expression for $w(y)$ for Gessel’s model is obtained in [@KuRa-11 Thm. 7].
Infinite group models
---------------------
In the infinite group case, the function $w$ is not algebraic anymore (it is even non-D-finite, see [@Ra-12 Thm. 2]). As $\mathcal{L}$ is a quartic curve [@FaIaMa-17 Thm. 5.3.3 (i)], $w$ can be expressed in terms of Weierstrass’ elliptic functions (see [@FaIaMa-17 Sec. 5.5.2.1] or [@Ra-12 Thm. 6]):
\[lem:conformal\_gluing\_functions\_infinite\_group\] The function $w$ defined by $$\label{eq:expression_gluing}
w(y)=\wp_{1,3}\Big(-\frac{\omega_1+\omega_2}{2}+\wp_{1,2}^{-1}(f(y))\Big)$$ is a conformal mapping for the domain $\mathcal G_\mathcal L$, and has in this domain a unique (and simple) pole, located at $y_2$. The function $w$ admits a meromorphic continuation on $\mathbb C\setminus [y_3,y_4]$. It is D-algebraic in $y$ and in $t$.
The differential algebraicity is shown in [@BeBMRa-17 Thm. 33]. The remaining properties stated in Lemma \[lem:conformal\_gluing\_functions\_infinite\_group\] come from [@FaIaMa-17; @Ra-12], see e.g. [@Ra-12 Thm. 6 and Rem. 7].
Let us now comment on the expression , following the discussion in [@BeBMRa-17 Sec. 5.2]. First, $f(y)$ is a rational function of $y$ whose coefficients are algebraic functions of $t$: $$f(y) = \left\{\begin{array}{ll}
\displaystyle \frac{\widetilde d''(y_4)}{6}+\frac{\widetilde d'(y_4)}{y-y_4} & \text{if } y_4\neq \infty,\medskip\\
\displaystyle\frac{\widetilde d''(0)}{6}+\frac{\widetilde d'''(0)y}{6}& \text{if } y_4=\infty,
\end{array}\right.$$ where $\widetilde d(y)$ is the discriminant and $y_4$ is one of its roots.
The next ingredient in is Weierstrass’ elliptic function $\wp$, with periods $\omega_1$ and $\omega_2$: $$\wp(z)= \wp(z, \omega_1, \omega_2) = \frac 1 {z^2} +\sum_{(i,j) \in \mathbb Z^2
\setminus\{(0,0)\}} \left( \frac 1
{(z-i\omega_1-j\omega_2)^2}-\frac1{(i\omega_1+j\omega_2)^2}\right).$$ Then $\wp_{1,2}(z)$ (resp. $\wp_{1,3}(z)$) is the Weierstrass function with periods $\omega_1$ and $\omega_2$ (resp. $\omega_1$ and $\omega_3$) defined by: $$\omega_1 = i\int_{y_1}^{y_2} \frac{\text{d} y}{\sqrt{- \widetilde d(y)}},\qquad
\omega_2 = \int_{y_2}^{y_3} \frac{\text{d} y}{\sqrt{ \widetilde d(y)}},\qquad
\omega_3 = \int_{Y(x_1)}^{y_1} \frac{\text{d} y}{\sqrt{ \widetilde d(y)}}.$$ These definitions make sense thanks to the properties of the $y_i$’s and $Y(x_i)$’s (see [@BeBMRa-17 Sec. 5.1]). If $Y(x_1)$ is infinite (which happens if and only if neither $(-1,0)$ nor $(-1,1)$ are in $\mathcal S$), the integral defining $\omega_3$ starts at $-\infty$. Note that $\omega_1\in i\mathbb R_+$ and $\omega_2,\omega_3\in \mathbb R_+$.
Finally, as the Weierstrass function is not injective on $\mathbb C$, we need to clarify our definition of $\wp_{1,2}^{-1}$ in . The function $\wp_{1,2}$ is two-to-one on the fundamental parallelogram $[0,\omega_1)+[0,\omega_2)$ (because $\wp(z)=\wp(-z+\omega_1+\omega_2)$), but is one-to-one when restricted to a half-parallelogram—more precisely, when restricted to the open rectangle $(0,\omega_1)+(0,
\omega_2/2)$ together with the three boundary segments $[0, \omega_1/2]$, $[0, \omega_2/2]$ and $\omega_2/2+[0, \omega_1/2]$. We choose the determination of $\wp_{1,2}^{-1}$ in this set.
Riemann-Hilbert BVP {#app:RHBVP}
===================
In the way of proving our main results (Theorems \[thm:first\_main\] and \[thm:second\_main\]), a crucial ingredient is the BVP with shift of Lemma \[lem:BVP\]. It is solved by reduction to a more classical Riemann BVP (Sections \[sub:proof-thmbvpM1\] and \[subsec:proof\_thm\_3/4\_index0\]). In this appendix we present the main formulas used to solve the latter, so as to render our paper self-contained. Our main references are the books of Gakhov [@Ga-90 Chap. 2] and Lu [@Lu-93 Chap. 4].
plot \[smooth\] coordinates [(-5,-3) (-3,1) (1,1) (4,4)]{}; (-3,1)–(-2.9,1.1); (-5,-3) circle (0.15 cm) node\[below\]; (4,4) circle (0.15 cm) node\[below\]; (-4.75,0.5) node; (-0.45,2.45)–(0.9,1.1); (1,2.5) node; (2.45,-0.45)–(1.1,0.9); (1,-0.5) node;
Suppose that $\mathcal{U}$ is an open, smooth, non-intersecting, oriented curve from $a$ to $b$, see Figure \[fig:BVP-open\] for an example. Throughout, for $z\in\mathcal U$, we will denote by $\Phi^+(z)$ (resp. $\Phi^-(z)$) the limit of a function $\Phi$ as $y\to z$ from the left (resp. right) of $\mathcal{U}$, see again Figure \[fig:BVP-open\].
\[def:BVP\] Let $\mathcal{U}$ be as above. A function $\Phi$ satisfies a BVP on $\mathcal{U}$ if:
- $\Phi$ is sectionally analytic, i.e., analytic in $\mathbb{C}\setminus\mathcal{U}$;
- $\Phi$ has finite degree at $\infty$ (the only singularity at $\infty$ is a pole of finite order), and $\Phi$ is bounded in the vicinity of the extremities $a$ and $b$;
- $\Phi$ has left limits $\Phi^+$ and right limits $\Phi^-$ on $\mathcal U$;
- $\Phi$ satisfies the following boundary condition $$\label{eq:RBVP}
\Phi^+(z)=G(z)\Phi^-(z)+g(z), \quad z\in\mathcal{U},$$ where $G$ and $g$ are Hölder functions on $\mathcal{U}$, and $G$ does not vanish on $\mathcal{U}$.
Let us recall the so-called Sokhotski-Plemelj formulas, which represent a crucial tool to solve the BVP of Definition \[def:BVP\].
\[proposition:Sokhotski-Plemelj\] Let $\mathcal{U}$ be as above, and let $f$ be a Hölder function on $\mathcal{U}$. The contour integral $$F(z)=\frac{1}{2i\pi} \int_{\mathcal{U}} \frac{f(u)}{u-z}{\,\mathrm{d}u}$$ is sectionally analytic on $\mathbb C\setminus \mathcal U$. Its left and right limit values $F^+$ and $F^-$ are Hölder functions on $\mathcal{U}$ and satisfy, for $z\in\mathcal{U}$, $$F^\pm(z) =\pm \frac{1}{2}f(z) +\frac{1}{2i\pi}\int_\mathcal{U}\frac{f(u)}{u-z}{\,\mathrm{d}u},$$ where the very last integral is understood in the sense of Cauchy-principal value, see [@Ga-90 Chap. 1, Sec. 12]. This is equivalent to the following equations on $\mathcal U$: $$\left\{
\begin{array}{l r l l|}
F^+(z) -F^-(z)&\hspace{-2mm}=&\hspace{-2mm}f(z),\smallskip\\
F^-(z) +F^-(z)&\hspace{-2mm}=&\hspace{-2mm}\displaystyle\frac{1}{i\pi}\int_\mathcal{L}\frac{f(u)}{u-z}{\,\mathrm{d}u}.
\end{array}
\right.$$
We also define the following important quantity:
\[def:index\] Let $\mathcal{U}$ be as above and let $G$ be the function (continuous on $\mathcal{U}$) as in . The index $\chi$ of the BVP of Definition \[def:BVP\] is $$\chi=\operatorname{ind}_{\mathcal U}G=\frac{1}{2\pi}[\arg G]_\mathcal{U}=\frac{1}{2i\pi}[\log G]_\mathcal{U}=\frac{1}{2i\pi}\int_{\mathcal U} \frac{G'(u)}{G(u)}{\,\mathrm{d}u}.$$
Plainly, $\chi$ represents the variation of argument of $G(u)$, when $u$ moves along the contour $\mathcal{U}$ in the positive direction.
The main result is the following, see [@Lu-93 Chap. 4, Thm. 2.1.2]:
\[thm:Sol-BVP\] Let $\mathcal U$ be as above. The solution of the BVP of Definition \[def:BVP\] is given by, for $z\notin \mathcal{U}$, $$\Phi(z)=
\left\{
\begin{array}{l l}
X(z)\psi(z)+X(z)P_\chi(z) & \text{ if } \chi{\geqslant}0, \\
X(z)\psi(z) & \text{ if } \chi=-1, \\
X(z)\psi(z) & \text{ if } \chi <-1 \text{ and if the solvability conditions below hold:}
\end{array}
\right.$$ $$ \frac{1}{2i\pi}\int_\mathcal{U}\frac{g(u)u^{k-1}}{X^+(u)}{\,\mathrm{d}u}=0, \quad k=1,\ldots,-\chi-1,$$ where $P_\chi$ is an arbitrary polynomial of degree $\chi$, and $$\left\{
\begin{array}{r c l}
X(z)&=&\displaystyle (z-b)^{-\chi}\exp \Gamma(z),\medskip\\
X^+(z)&=&\displaystyle(z-b)^{-\chi}\exp \Gamma^+(z),\medskip\\
\Gamma(z)&=&\displaystyle\frac{1}{2i\pi}\int_\mathcal{U}\frac{\log G(u)}{u-z}{\,\mathrm{d}u},\medskip\\
\psi(z)&=&\displaystyle\frac{1}{2i\pi}\int_\mathcal{U}\frac{g(u)}{X^+(u)(u-z)}{\,\mathrm{d}u}.
\end{array}
\right.$$
Proof of Lemma \[lem:functional\_equation\_sym\] {#app:proof_lem1}
================================================
The decomposition in expresses $C(x,y)$ as a sum of three generating functions. Thanks to the symmetry of the step set and the fact that the starting point lies on the diagonal, $\widehat{U}(x,y)=\widehat{L}(y,x)$ and $C(x,y)$ is written as the sum $\widehat{L}(x,y)+\widehat{D}(x,y)$ of two unknowns. We further introduce the generating functions $$\widehat{D}^{\ell}(x,y)=\sum_{n{\geqslant}0,i{\geqslant}0}c_{i,i-1}(n)x^{i}y^{i-1}t^n\quad\text{and}
\quad
\widehat{D}^{u}(x,y)=\sum_{n{\geqslant}0,i{\geqslant}0}c_{i-1,i}(n)x^{i-1}y^it^n,$$ which respectively count walks ending on the lower (resp. upper) diagonal, see Figure \[fig:some\_sections\]. In this section, we consider walks starting on the diagonal and ending anywhere in the three-quadrant $\mathcal C$.
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (0,-1); (0,-1) – (2,-1); (2,-1) – (2,2); (2,2) – (-2,2); (-2,2) – (-2,1); (-2,1) – (-3,1); (-3,1) – (-3,4); (-3,4) – (1,4); (1,4) – (1,5); (1,5) – (2,5); (2,5) – (2,3); (2,3) – (4,3); (4,3) – (4,-1); (4,-1) – (3,-1); (3,-1) – (3,-2); (3,-2) – (4,-2); (4,-2) – (4,-3); (4,-3) – (2,-3); (2,-3) – (2,-4); (6.5,0) node[$=$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (-1,0); (-1,0) – (-1,1); (-1,1) – (-2,1); (-2,1) – (-2,2); (-2,2) – (-3,2); (-3,2) – (-3,4); (-3,4) – (2,4); (2,4) – (2,-2); (2,-2) – (1,-2); (2,-2) – (3,-2); (2,-2) – (2,-3); (2,-2) – (2,-1); (6.5,0) node[$+$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (2,0); (2,0) – (2,2); (2,2) – (-3,2); (-3,2) – (-3,4); (-3,4) – (4,4); (4,4) – (5,4); (4,4) – (4,3); (6.5,0) node[$-$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (-1,0); (-1,0) – (-1,2); (-1,2) – (1,2); (1,2) – (1,-2); (1,-2) – (4,-2); (4,-2) – (4,3); (4,3) – (4,4); (4,3) – (3,3); (6.5,0) node[$-$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (-1,0); (-1,0) – (-1,2); (-1,2) – (1,2); (1,2) – (1,-2); (1,-2) – (2,-2); (2,-2) – (2,-3); (2,-3) – (0,-3); (0,-3) – (-1,-3); (6.5,0) node[$+$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (1,0); (1,0) – (1,3); (1,3) – (2,3); (2,3) – (2,2); (2,2) – (3,2); (3,2) – (3,-2); (3,-2) – (1,-2); (1,-2) – (1,-1); (1,-1) – (0,-1); (0,-1) – (-1,-1);
Thereafter, $c_{i,j}(n)$ is counting walks from $(i_0,i_0)$ to $(i,j)$ in $n$ steps. Classically [@BMMi-10], we construct a walk by adding a new step at the end of the walk at each stage. We first derive a functional equation for $\widehat{L}(x,y)$ by taking into account all possibilities of ending in the lower part:
- we may add a step from $\widehat{\mathcal{S}}$ (recall that $\widehat{\mathcal{S}}$ is the step set before the change of variable $\varphi$) to walks ending in the lower part, yielding in the term $t(\sum_{(i,j)\in\widehat{\mathcal{S}}} x^i y^j )\widehat{L}(x,y)$, see the second picture on Figure \[fig:diagtolowerSW\] in the particular case of the simple walk;
- walks coming from the diagonal also need to be counted up, giving rise in to the term $t(\delta_{1,0}x+\delta_{0,-1}y^{-1})\widehat{D}(x,y)$ (third picture on Figure \[fig:diagtolowerSW\]);
- on the other hand, walks going out of the three-quarter plane need to be removed, yielding the terms $t( \delta_{-1,0}x^{-1}+\delta_{0,1}y) \widehat{D}^\ell(x,y)$ (the lower diagonal) and $t(\delta_{-1,0}x^{-1}+\delta_{-1,-1}x^{-1}y^{-1}) \widehat{L}_{0-}(y^{-1})$ (negative $y$-axis), see the fourth and fifth pictures on Figure \[fig:diagtolowerSW\];
- we finally add the term $t\delta_{-1,0}x^{-1}\sum_{n{\geqslant}0}c_{0,-1}(n)y^{-1}t^n$ which was subtracted twice, corresponding to the rightmost picture on Figure \[fig:diagtolowerSW\].
We end up with a first functional equation: $$\begin{gathered}
\label{eq:fct-eq-L(x,y)}
\widehat{L}(x,y)= t\sum_{(i,j)\in\widehat{\mathcal{S}}} x^i y^j \widehat{L}(x,y)
+ t(\delta_{1,0}x+\delta_{0,-1}y^{-1})\widehat{D}(x,y)
-t( \delta_{-1,0}x^{-1}+\delta_{0,1}y) \widehat{D}^\ell(x,y)
\\
-t( \delta_{-1,0}x^{-1}+\delta_{-1,-1}x^{-1}y^{-1}) \widehat{L}_{0-}(y^{-1})
+t( \delta_{-1,0}x^{-1})\sum_{n{\geqslant}0}c_{0,-1}(n)y^{-1}t^n.\end{gathered}$$ We now prove the second equation $$\begin{gathered}
\label{eq:fct-eq-D(x,y)}
\widehat{D}(x,y)=x^{i_0}y^{i_0}
+t( \delta_{1,1}xy+\delta_{-1,-1}x^{-1}y^{-1})\widehat{D}(x,y)-t\delta_{-1,-1}x^{-1}y^{-1}\widehat{D}(0,0)\\
+2t( \delta_{-1,0}x^{-1}+\delta_{0,1}y) \widehat{D}^\ell(x,y)
-2t\delta_{-1,0}x^{-1}\sum_{n{\geqslant}0}c_{0,-1}(n)y^{-1}t^n,\end{gathered}$$ and remark that by plugging in into we get , thereby completing the proof of Lemma \[lem:functional\_equation\_sym\].
This second equation is obtained by writing all possibilities of ending on the diagonal, as illustrated on Figure \[fig:diagtodiagSW\] for simple walks:
- we first count the empty walk, giving the term $x^{i_0}y^{i_0}$;
- we add the walks remaining on the diagonal $t( \delta_{1,1}xy+\delta_{-1,-1}x^{-1}y^{-1})\widehat{D}(x,y)$, the walks ending on the diagonal coming from the upper part $t(\delta_{0,-1}y^{-1}+\delta_{1,0}x)\widehat{D}^u(x,y)$ and those coming from the lower part $t(\delta_{-1,0}x^{-1}+\delta_{0,1}y) \widehat{D}^\ell(x,y)$;
- finally, walks going out of the domain need to be removed, giving $t\delta_{-1,-1}x^{-1}y^{-1}\widehat{D}(0,0)$, $t\delta_{0,-1}y^{-1}\sum_{n{\geqslant}0}c_{-1,0}(n)x^{-1}t^n$ and $t\delta_{-1,0}x^{-1}\sum_{n{\geqslant}0}c_{0,-1}(n)y^{-1}t^n$.
Thanks to the symmetry of the step set, the number of walks coming from the upper part is the same as the number of walks coming from the lower part.
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (0,-2); (0,-2) – (1,-2); (1,-2) – (1,2); (1,2) – (-3,2); (-3,2) – (-3,4); (-3,4) – (2,4); (2,4) – (2,1); (2,1) – (4,1); (4,1) – (4,3); (4,3) – (5,3); (5,3) – (5,5); (6.5,0) node[$=$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) circle (0.2 cm); (6.5,0) node[$+$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (-1,0); (-1,0) – (-1,1); (-1,1) – (1,1); (1,1) – (1,-3); (1,-3) – (2,-3); (2,-3) – (2,-2); (2,-2) – (4,-2); (4,-2) – (4,2); (4,2) – (2,2); (2,2) – (2,4); (2,4) – (3,4); (3,4) – (4,4); (3,4) – (3,3); (6.5,0) node[$-$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (0,-1); (0,-1) – (2,-1); (2,-1) – (2,3); (2,3) – (3,3); (3,3) – (3,4); (3,4) – (-1,4); (-1,4) – (-1,2); (-1,2) – (-3,2); (-3,2) – (-3,1); (-3,1) – (-1,1); (-1,1) – (-1,0); (-1,0) – (-1,-1); (6.5,0) node[$+$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (1,0); (1,0) – (1,1); (1,1) – (-1,1); (-1,1) – (-1,3); (-1,3) – (2,3); (2,3) – (2,2); (2,2) – (4,2); (4,2) – (4,3); (4,3) – (3,3); (4,3) – (4,4); (6.5,0) node[$-$]{};
(0,-1) – (5.5,4.5) – (5.5,-5.5) – (0,-5.5); (-1,0) – (-5.5,0) – (-5.5,5.5) – (4.5,5.5); (0,-5.5) grid (5.5,5.5); (-5.5,0) grid (0,5.5); (0,0) – (5.5,5.5); (0,-5.5) – (0,5.5); (-5.5,0) – (5.5,0); (0,0) – (0,2); (0,2) – (-1,2); (-1,2) – (-1,3); (-1,3) – (-3,3); (-3,3) – (-3,4); (-3,4) – (3,4); (3,4) – (3,3); (3,3) – (4,3); (4,3) – (4,2); (4,2) – (2,2); (2,2) – (2,-3); (2,-3) – (1,-3); (1,-3) – (1,-2); (1,-2) – (0,-2); (0,-2) – (0,-1); (0,-1) – (-1,-1);
A step set containing the jumps $(-1,1)$ and $(1,-1)$ would lead to two additional terms in the functional equations, namely $$\delta_{-1,1}x^{-1}y\sum_{n,i{\geqslant}0}c_{i,i-2}(n)x^i y^{i-2}t^n\qquad
\text{and}
\qquad \delta_{1,-1}xy^{-1} \sum_{n,j{\geqslant}0}c_{j-2,j}(n)x^{j-2}y^jt^n,$$ making the resolution much more complicated (not to say impossible, by our techniques!). Likewise, considering an asymmetric step set and/or a starting point out of the diagonal would lead to other terms in the functional equation.
[^1]: [*Keywords.*]{} Lattice walks in cones; Generating function; Boundary value problem; Conformal mapping
[^2]: CNRS, Institut Denis Poisson, Université de Tours, France; `[email protected]`
[^3]: This project has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme under the Grant Agreement No 759702.
[^4]: Department of Mathematics, Simon Fraser University, Canada & Institut Denis Poisson, Université de Tours, France; `[email protected]`
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Inferring a decision tree from a given dataset is one of the classic problems in machine learning. This problem consists of buildings, from a labelled dataset, a tree such that each node corresponds to a class and a path between the tree root and a leaf corresponds to a conjunction of features to be satisfied in this class. Following the principle of parsimony, we want to infer a minimal tree consistent with the dataset. Unfortunately, inferring an optimal decision tree is known to be NP-complete for several definitions of optimality. Hence, the majority of existing approaches relies on heuristics, and as for the few exact inference approaches, they do not work on large data sets. In this paper, we propose a novel approach for inferring a decision tree of a minimum depth based on the incremental generation of Boolean formula. The experimental results indicate that it scales sufficiently well and the time it takes to run grows slowly with the size of dataset.'
author:
- Florent Avellaneda
bibliography:
- 'bibi.bib'
title: Learning Optimal Decision Trees from Large Datasets
---
Introduction
============
In machine learning, the problem of classification consists of inferring a model from observations (also called *training examples*) that make it possible to identify which class of a set of classes a new observation belongs to. When the training examples used to infer a model are assigned to the classes to which they belong, it is called supervised learning. Many machine learning models exist to solve this problem, such as support vector machines [@vapnik1995nature], artificial neural network [@haykin1994neural], decision trees [@quinlan1986induction], etc. Because inferring such models is complicated and the number of training examples is generally very large, most existing inference algorithms are heuristic in the sense that the algorithms infer models without any guarantee of optimality.
Although these heuristic-based techniques generally work well, there are always cases where a new example is not correctly recognized by the model. In the context of critical systems in which errors are not allowed, such models are difficult to use. This requirement often goes together with the demand that models must also be understandable. Known as eXplainable AI (XAI), this area consists of inferring models capable of explaining their own behaviour. XAI has been the subject of several studies in recent years [@goebel2018explainable; @li2018deep; @van2004explainable; @angelino2017learning], as well as several events [@ACMFAT; @IJCAIXAI]. One approach to obtain an explainable model is to use decision trees because the reasons for classification are clearly defined [@quinlan1986induction].
In order to obtain accurate and explainable models, we are interested in the inference of optimal decision trees. The optimality of a decision tree is generally defined by the simplicity of the tree based on the principle of parsimony. Our chosen simplicity criteria are the depth of the tree and the number of nodes. In particular, for a fixed maximum depth of the tree, we want to infer a decision tree with a minimum number of nodes that is consistent with the training examples.
Although decision tree inference is a well-studied classic problem, the majority of known algorithms are heuristic and try to minimize the number of nodes without guaranteeing any optimality [@quinlan1986induction; @salzberg1994c4; @kass1980exploratory; @breiman1984classification]. It is because the problem is known to be NP-complete for several definitions of optimality [@laurent1976constructing; @hancock1996lower]. In addition, the first algorithms to infer optimal decision trees were ineffective in practice [@rokach2008data]. However, in recent years, several studies have focused on improving the performance of optimal decision tree inference algorithms.
The first series of studies was carried out on a similar problem to ours: inferring optimal decision tree with a given depth such that the total classification error on the training examples is minimized [@bertsimas2007classification; @bertsimas2017optimal; @verwer2019learning]. For this problem, Verwer and Zhang [@verwer2019learning] propose a binary linear programming formulation that infers optimal decision trees of depths four in less than ten minutes.
The studies closest to ours are those of Bessier et al. [@bessiere2009minimising] and Narodytska et al. [@narodytska2018learning]. These authors were interested in a particular case of our problem: inferring decision trees with a minimal number of nodes without trying to minimize the depth. Bessier et al. propose a SAT formulation, but experiments show that the method only works for small models, i.e., trees of about fifteen nodes. The authors also propose a method based on constraint programming to minimize the number of nodes, but without necessarily reaching the optimal. Narodytska et al. propose a new SAT formulation that greatly improves the practical performance of optimal decision tree inference. Thus, with their new formulation, Narodytska et al. were able to build, for the “Mouse" dataset, a decision tree with a minimum number of nodes in 13 seconds, while this required 577 seconds with the SAT formulation of Bessier et al. The authors claim that to the best of their knowledge, their paper is the first presentation of an optimal decision tree inference method based on well-known datasets.
In this paper, we propose an even more efficient method than the last one. Our benchmarks show that we can process the “Mouse" dataset in only 75 milliseconds. Moreover, well-known datasets that were considered too large to infer optimal decision trees from them can now be processed by our algorithm.
The paper is organized as follows. In the next section, we provide definitions related to decision trees needed to formalize the approach. Section 3 provides a new Boolean formulation for passive inference of a decision tree from a set of training examples. We propose in Section 4 an incremental way of generating the Boolean formulas which ensures that the proposed approach scales to large datasets. Section 5 reports several experiments comparing our approach to others. Finally, we conclude in Section 6.
Definitions {#section:def}
===========
Let $\mathcal{E} = \{e_0, ..., e_{n-1}\}$ be a set of *training examples*, that is, Boolean valuations of a set $\mathcal{F} = \{f_0, f_1, ..., f_{m-1}\}$ of *features*, and let $\mathcal{E}_0, \mathcal{E}_1, ..., \mathcal{E}_{c-1}$ be a partition of $\mathcal{E}$ into classes. Note that even if we only consider binary features, we can easily handle non-binary features by encoding them in a binary way [@bartnikowski2014effects]. Features that belong to $x$ categories can be represented by $x$ Boolean features where each one represents the affiliation to one category. If the categories are ordered, then each Boolean feature can represent the affiliation to a smaller or equal category (see Example \[codageNum\]). The second encoding provides constraints of type $\leq$ on numerical features for example. In the following, we denote by $e[f]$ the Boolean valuation of the feature $f \in \mathcal{F}$ in example $e \in \mathcal{E}$. A *decision tree* is a binary tree where each node is labelled by a single feature and each leaf is labelled by a single class. A decision tree is said to be *perfect* if all leaves have the same depth and all internal nodes have two children. Formally, we denote a perfect decision tree by $T = (V, V')$ where $V \in \mathcal{F}^{2^k-1}$ is the set of internal nodes and $V' \in \{0, 1, ..., c-1\}^{2^k}$ is the set of leaves and where $k$ is the depth of the tree. We denote by $V[i]$ the $i^{\text{th}}$ node in the tree $T$ and by $V[1]$ the root of the tree. Then we define $V[i \times 2]$ as the left child of $V[i]$ and $V[i \times 2 + 1]$ as the right child. In a similar way, if $i \geq 2^{k-1}$, we define the leaf $V'[(i-2^{k-1}) \times 2]$ as the left child of $V[i]$ and the leaf $V'[(i-2^{k-1}) \times 2 + 1]$ as the right child. An illustration of this encoding is depicted by Figure \[codingIndex\].
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$V[1]$]{} ; (Q2) at (-2,-1) [$V[2]$]{} ; (Q3) at (2,-1) [$V[3]$]{} ;
(Q4) at (-3,-2) [$V[4]$]{} ; (Q5) at (-1,-2) [$V[5]$]{} ; (Q6) at (1,-2) [$V[6]$]{} ; (Q7) at (3,-2) [$V[7]$]{} ;
(Q8) at (-3.5,-3) [$V'[0]$]{} ; (Q9) at (-2.5,-3) [$V'[1]$]{} ; (Q10) at (-1.5,-3) [$V'[2]$]{} ; (Q11) at (-0.5,-3) [$V'[3]$]{} ; (Q12) at (0.5,-3) [$V'[4]$]{} ; (Q13) at (1.5,-3) [$V'[5]$]{} ; (Q14) at (2.5,-3) [$V'[6]$]{} ; (Q15) at (3.5,-3) [$V'[7]$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
(Q4) edge node\[left\] [$false$]{} (Q8); (Q4) edge node\[right\] [$true$]{} (Q9);
(Q5) edge node\[left\] [$false$]{} (Q10); (Q5) edge node\[right\] [$true$]{} (Q11);
(Q6) edge node\[left\] [$false$]{} (Q12); (Q6) edge node\[right\] [$true$]{} (Q13);
(Q7) edge node\[left\] [$false$]{} (Q14); (Q7) edge node\[right\] [$true$]{} (Q15);
This way of associating a number to every node and leaf may appear complicated, but it will be useful for our Boolean encoding. We will use the semantics associated with binary coding of node indexes to obtain compact SAT formulas.
If $T$ is a decision tree, and $\mathcal{E}$ is a set of training examples, we say that $T$ is *consistent* with $\mathcal{E}$, denoted $\mathcal{E} \subseteq T$, if each example $e \in \mathcal{E}$ is correctly classified by $T$.
\[codageNum\] Let $\mathcal{E}$ be a set of training examples where each example has a single integer feature $f$. Let $\mathcal{E}_0 = \{ (1), (3) \}$ and $\mathcal{E}_1 = \{ (4), (5) \}$ be the partition of $\mathcal{E}$ into two classes. Then, we can transform $\mathcal{E}_0$ and $\mathcal{E}_1$ into $\mathcal{E}'_0$ and $\mathcal{E}'_1$ such that each example has four Boolean features $f'_0, f'_1, f'_2, f'_3$. If the feature $f'_0$ is true, it means that the example is smaller or equal to $1$ for feature $f$; if the feature $f'_1$ feature is true, it that the example is smaller or equal to $3$ for feature $f$, etc. Thus, with this transformation we obtain $\mathcal{E}'_0 = \{ (true, false, false, false), (true, true, false, false) \}$ and $\mathcal{E}'_1 = \{ (true, true, true, false), (true, true, true, true) \}$.
Passive Inference
=================
There are two types of methods for solving the problem of inferring a decision tree from examples.
One group constitutes heuristic methods which try to find a relevant feature for each internal node in polynomial time [@breiman1984classification; @salzberg1994c4; @quinlan1986induction; @kass1980exploratory]. They are often used in practice because of their efficiency; however, they provide no guarantee of optimality. Thus, better choices in the feature order can lead to smaller decision trees that are consistent with training examples.
Another group includes exact algorithms to determine a decision tree with a minimal number of nodes. It is a much more complicated problem, as it is NP-complete [@laurent1976constructing]. There are essentially two works that focus on this problem [@narodytska2018learning; @bessiere2009minimising].
We propose a SAT formulation that differs from them. Our approach has two main steps. In the first step, we seek a perfect decision tree of a minimal depth. In a second step, we add constraints to reduce the number of nodes in order to potentially obtain an imperfect decision tree.
Inferring perfect decision trees of fixed maximum depth {#minimize depth}
-------------------------------------------------------
Our SAT encoding to infer decision trees of a fixed maximum depth is based on the way the nodes are indexed. As mentioned in Section \[section:def\], the index of a node depends on its position in the tree. In particular, the root node corresponds to the node $V[1]$, and for each node $V[i]$, the left child corresponds to $V[i \times 2]$ and the right child to $V[i \times 2 + 1]$. This coding has a useful capability of providing precise information on the position of a node based on the binary coding of its index. Indeed, reading the binary coding of a node from the highest to the lowest weight bit indicates which branches to take when moving from the root to the node. For example, if the binary coding of $i$ is 1011, then the node $V[i]$ is reached by taking the right branch of the root, then the left branch, and finally twice the right branch. Note that if a node $V[i]$ is a descendant of the right branch of a node $V[j]$ then we say that $V[j]$ is a right ancestor of $V[i]$.
The idea of our coding consists in arranging the training examples in the leaves of the tree while respecting the fact that all training examples placed in the same leaf must belong to the same class. Moreover, if an example is placed in a leaf, then all the right ancestors of that leaf can only be labelled by features true for that example (and conversely for the left ancestor).
The encoding idea is formalized using the following types of Boolean variables.
- $X_{i, j}$: If the variable $X_{i, j}$ is true, it means that the example $e_i$ is assigned to a leaf that is a right ancestor of a node located at depth $j$. If $X_{i, j}$ is false, then $e_i$ is assigned to a leaf that is a left ancestor of that node. Note that with this semantics on the variables $X_{i, j}$, we have the property that the binary coding $(X_{i,0} X_{i,1} ... X_{i,k-1})$, denoted $X_i$, corresponds to the index of the leaf where the example $e_i$ belongs, i.e., if $X_i = v$, then the example $e_i$ is assigned to the leaf $V'[v]$. We also denote by $X_i[..a]$ the number formed by the binary coding of $(X_{i,0} X_{i,1} ... X_{i,a})$.
- $F_{i, j}$: If $F_{i,j}$ is true, it means that the node $V[i]$ is labelled by the feature $f_j$.
- $C_{i, j}$: If $C_{i, j}$ is true, it means that the leaf $V'[i]$ is labelled by the class $j$.
We then use the following set of clauses to formulate constraints that a perfect decision tree of depth $k$ should satisfy.
For each $i \in [1, 2^k-1]$, we have the clauses: $$\label{equation 1}
F_{i, 0} \vee F_{i, 1} \vee ... \vee F_{i, m-1}$$ These clauses mean that each node should have at least one feature.\
For each $i \in [1, 2^k-1]$ and every features $f_1, f_2$ such that $0 \leq f_1 < f_2 < m$, we have the clauses: $$\label{equation 2}
\neg F_{i, f_1} \vee \neg F_{i, f_2}$$ These clauses mean that each node has at most one feature.\
For every $i$ and $f$ such that $e_i[f] = 0$, and each $j \in [0, k-1]$, we have: $$\label{equation 3}
X_{i, j} \Rightarrow \neg F_{X_i[..j], f}$$ And for every $i$ and $f$ such that $e_i[f] = 1$, and each $j \in [0, k-1]$, we have: $$\label{equation 3 bis}
X_{i, j} \Rightarrow F_{X_i[..j], f}$$ These formulas add constraints that some features cannot be found in certain nodes depending on where the training examples are placed in the decision tree. We use the binary coding of the index of a leaf to determine which nodes in the tree are its parents. Thus, all the parent nodes for which the left branch has been taken cannot be labelled by features that must be true, and vice versa. Note that it is not trivial to translate these formulas into clauses, but we show in Algorithm \[algoGen3\] how it could be done. This algorithm performs a depth-first search of the perfect decision tree in a recursive way. The variable $q$ corresponds to the index of the current node and $\neg clause$ constraints such that $X_i$ correspond to the index of a $q$ successor. Each time the algorithm visits a state $q$, it adds constraints on the features that can be labeled by this node based on where $e_i$ is placed in the left or right branch of $q$. If $e_1$ is placed in the left branch, then $q$ cannot contain a feature that is true for $e_i$ and vice versa with the right branch.\
For each $e_i \in \mathcal{E}_a$ with $a \in [0, c-1]$ and each integer $v \in [0, 2^k-1]$, we have: $$\label{equation 4}
X_i = v \Rightarrow C_{v, a}$$ And for each $e_i \in \mathcal{E}_a$, each integer $v \in [0, 2^k-1]$, and $a' \neq a$ we have: $$\label{equation 4 bis}
X_i = v \Rightarrow \neg C_{v, a'}$$ These formulas assign the classes to the leaves according to the places of the training examples in the decision tree. Again, since it is not trivial to translate these formulas into clauses, we show in Algorithm \[algoGen4\] how it could be performed. This algorithm performs a depth-first search of the perfect decision tree such that when it reaches a leaf, $\neg clauses$ corresponds to the index of that leaf. After that, for each leaf, the algorithm generates the constraints that if $e_i$ is present in that leaf, then that leaf must have the same class as $e_i$.
[**Input:** A new example $e_i$, a clause $clause$, an index node $q$, the depth of the tree $lvl$ already considered. Initially, $clause=\emptyset$, $q=1$ and $lvl=0$.\
**Output:** Clauses for formulas (\[equation 3\]) and (\[equation 3 bis\]) when we consider a new example $e_i$]{}
$result = \emptyset$
$result$
$result := result \wedge (clause \vee X_{i, lvl} \vee \neg F_{q, f} )$
$result := result \wedge GenerateFeatureConstraints(e_i, (clause \vee X_{i, lvl}), q \times 2, lvl + 1)$
$result := result \wedge (clause \vee \neg X_{i, lvl} \vee F_{q, f} )$
$result := result \wedge GenerateFeatureConstraints(e_i, (clause \vee \neg X_{i, lvl}), q \times 2 + 1, lvl + 1)$
$result$
[**Input:** A new example $e_i \in \mathcal{E}_a$, a clause $clause$, a node number $q$, an integer $lvl$ and an integer $lvlMax$. Initially, $clause = \emptyset$, $q=0$ and $lvl=0$.\
**Output:** Clauses for formulas (\[equation 4\]) and (\[equation 4 bis\]) when we consider a new example $e_i$]{}
$result = (clause \vee C_{q, a})$ $result := result \wedge (clause \vee \neg C_{q, a'})$ $result$
$GenerateClassConstraints(e_i, clause \vee X_{i, lvl}, q \times 2, lvl+1, lvlMax ) \cup GenerateClassConstraints(e_i, clause \vee \neg X_{i, lvl}, q \times 2 + 1, lvl+1, lvlMax )$
Minimizing the number of nodes {#minimize nodes}
------------------------------
Perfect decision trees are often considered as unnecessarily too large, i.e. they can contain too many nodes. For example, there may be an imperfect decision tree consistent with training examples, with the same depth $k$ as the perfect tree, but with fewer nodes.
In order to find a tree with a minimum number of nodes, we show in this section how we can add constraints to set a maximum number of nodes of the tree. The idea is to limit the number of leaves that can be assigned to a class in the perfect tree. Indeed, if a leaf is not assigned to a class, then the parent of this leaf can be replaced by its other child. By applying this algorithm recursively until all leaves are assigned to a class, we get a decision tree with exactly one leaf more than the internal nodes. Thus, limiting the number of nodes to $MaxNodes$ has the same effect as limiting the number of leaves to $\lfloor MaxNodes / 2 \rfloor + 1$.
To add the constraint of the maximum number of leaves that can be assigned to a class, we add two types of additional variables. The variables $U_{i}$ which are true if a class is assigned to the leaf $i$, and the variables $H_{i,0}, H_{i,1}, ..., H_{i,MaxNodes+1}$ which will be used to count, with unary coding, the number of leaves labelled by a class. The variable $H_{i+1, j}$ will be true if there are at least $j$ leaves labelled by a class among the $i$ first leaves.
The clauses encoding the new constraint are as follows:
For each $i \in [0, 2^{k}-1]$ and each class $a \in [0, c-1]$, we have the clauses: $$\label{equation A}
\neg C_{i, a} \vee U_{i}$$ These clauses assign $U_i$ to true if the leaf $i$ is labelled by a class.\
For each $i \in [0, 2^{k}-1]$ and each class $j \in [0, MaxNodes+1]$, we have the clauses: $$\label{equation B}
\neg H_{i, j} \vee H_{i+1, j}$$ These clauses propagate the fact that if $H_{i, j}$ is true, then $H_{i+1, j}$ is also true.
For each $i \in [0, 2^{k}-1]$ and each class $j \in [0, MaxNodes+1]$, we have the clauses: $$\label{equation C}
\neg U_{i} \vee \neg H_{i, j} \vee H_{i+1, j+1}$$ These clauses increase the value of $H_{i+1}$ by one if $U_{i}$ is true. Thus $H_{i+1, j}$ is true if there is at least $j$ leaves labelled by class among the $i$ first leaves.\
Finally, we assign the start and end of the counter $H$ : $$\label{equation D}
\neg H_{2^{k+1}, \lfloor MaxNodes / 2 \rfloor + 2} \wedge H_{0,0}$$ The first assignment prohibits having more than $ \lfloor MaxNodes / 2 \rfloor + 1 $ leaves, so $MaxNodes$ nodes. The second assignment sets the counter to $0$.
The formula for inferring a decision tree of depth $k$ (and a specific number of nodes) from $n$ training examples with $m$ features classed in $c$ classes require $O(2^k \times (n + m + c))$ literals, and $O( 2^k \times (m^2 + m \times n + c) )$ clauses.
It can be noted that the number of literals and clauses whether the maximum number of nodes is specified or not is of the same orders of magnitude. However, finding a tree with a minimum number of nodes will take more time because it requires us to search for this number by a dichotomous search.
Incremental Inference
=====================
To alleviate the complexity associated with large sets of training examples, we propose an approach which, instead of attempting to process all the training examples $\mathcal{E}$ at once, iteratively infers a decision tree from their subset (initially it is an empty set) and uses active inference to refine it when it is not consistent with one of the training examples. While active inference usually uses an oracle capable of deciding to which class an example belongs, we assign this role to the training examples $\mathcal{E}$. Even if such an oracle is restricted since it cannot guess the class for all possible input features, nevertheless, as we demonstrate, it leads to an efficient approach for passive inference from training examples. The approach is formalized in Algorithm \[algoLDT\].
[**Input:** The maximum depth $k$ of the tree to infer, the maximal number $MaxNodes$ of nodes of the tree to infer, the set of training examples $\mathcal{E} = \{\mathcal{E}_0, \mathcal{E}_1, ..., \mathcal{E}_{c-1}\}$.\
**Output:** A decision tree consistent with $\mathcal{E}$ with at most $MaxNodes$ nodes and with maximum depth $k$ if it exists.]{}
$C := $ formulas (1) and (2) Let $T$ be a decision tree of a solution of $C$ $T$ Let $e \in \mathcal{E}_a$ be an example mislabelled by $T$ $C := C \wedge GenerateFeatureConstraints(e, \emptyset, 1, 0) \wedge GenerateClassConstraints(e, \emptyset, 0, 0, k, a) $ $C := C \wedge C'$ where $C'$ is clauses described by formulas (\[equation A\]), (\[equation B\]), (\[equation C\]) and (\[equation D\]). “No solution"
An illustration of the execution of this algorithm is given in Appendix on a simple example.
Benchmarks
==========
In this paper, we have presented two algorithms to solve two different problems.
The first algorithm, denoted $DT\_depth$, finds a perfect decision tree of minimal depth. It uses Algorithm \[algoLDT\] without defining $MaxNodes$. We initially set $k=1$ and while the algorithm is not finding a solution, we increase the value of $k$.
The second algorithm, denoted $DT\_size$, minimizes the depth of the tree and the number of nodes. It starts by applying $DT\_depth$ to leanr the minimum depth $k$ required to find a decision tree consistent with the training examples. Then it performs a dichotomy search on the number of nodes allowed between $1$ and $2^{k+1}-1$ to find a decision tree with a minimal number of nodes.
We compare our algorithms with the one of Bessiere et al. [@bessiere2009minimising], denoted $DT2$, and the one of Naradytska et al. [@narodytska2018learning], denoted $DT1$. The main metric we will compare is the execution time and accuracy. The accuracy is calculated with a *$k$-fold cross-validation* defined as follows. We divide the dataset into $k$ equal parts (plus or minus one element). Then $k-1$ of these parts are used to infer a decision tree, the last part is used to calculate the percentage of its elements correctly classified by the decision tree. This operation is performed $k$ times to try all possible combinations among these ten parts and the average percentage is calculated.
The prototype was implemented in C++ calling the SAT solver MiniSAT [@minisat] and we run the prototype on Ubuntu in a computer with 12GB of RAM and i7-2600K processor.
The “Mouse" dataset
-------------------
Our first experiment is performed on the “Mouse" dataset that the authors Bessiere et al. shared with us. This dataset has the advantage of having been used with both algorithm $DT1$ and $DT2$. In Table \[mouse table\], we compare the time and accuracy for different algorithms. Each entry in rows $DT\_size$ and $DT\_depth$ corresponds to the average over 100 runs. The first four columns correspond to inferring a decision tree from the whole dataset. The last column corresponds the 10-fold cross-validations.
By analyzing Table \[mouse table\], we can notice that our incremental approach is very effective on this dataset. Only 37 examples for $DT\_size$ and 33 examples for $DT\_depth$ were used to build an optimal decision tree consistent with the entire dataset. Thus, thanks to our incremental approach and an efficient SAT formulation, our algorithms are much faster than $DT2$ and $DT1$. We could not compare the accuracy because this data is missing in the two respective papers for $DT2$ and $DT1$.
The “Car" dataset
-----------------
Another data set provided to us and used by the authors Bessiere et al. and Naradytska et al. is “Car". This dataset is much more complicated and to the best of our knowledge, no algorithm has been able to infer an optimal decision tree consistent with the entire dataset. The authors Bessiere et al. process this dataset using linear programming (denoted $DT3$) to minimize the number of nodes. However, they do not guarantee that the decision tree they find is optimal. The approach used by the authors Naradytska et al. simplifirs the dataset by considering only 10% of the data. Thus, they can infer an optimal decision tree consistent with the 10% of the data selected in 684 sec. Table \[car table\] compares the results of different algorithms. Each entry in rows $DT\_size$ and $DT\_depth$ corresponds to the average over ten runs. The first four columns correspond to inferring a decision tree from the whole dataset. The last column corresponds the 10-fold cross-validations.
We can see in Table \[car table\] that of the $1729$ examples in the “Car" dataset, our incremental approach uses less than half of it. Although this number is still much higher than the number of examples used by the algorithm $DT1$, we can see that our algorithms run faster. Moreover, since our algorithms ensure that the resulting decision trees are consistent with all training examples, we can see that the accuracy remains very high compared to the $DT1$ algorithm which randomly considers only $10\%$ of training examples. Note that the algorithm $DT3$ performs better that our algorithm but their algorithm is a heuristic that infers a decision tree without any guarantee of optimality. In addition, because $DT3$ do not have an optimality constraint, it seek to minimize the number of nodes of a general decision tree without constraint on the depth of the tree. This way, $DT3$ find trees with fewer nodes, but deeper than ours.
Other datasets
--------------
As we mentioned in the introduction of the paper, there is a series of algorithms that addresses a different but very similar problem to ours: inferring optimal decision tree with a given depth such that the total classification error on the training examples is minimized. In this section, we compare our results against these algorithms. The datasets we use are extracted from the paper of Verwer and Zhang [@verwer2019learning] and are available at *https://github.com/SiccoVerwer/binoct*. Each dataset corresponds to a 5-fold cross-validation. In their paper, Verwer and Zhang compare their approach $BinOCT^*$ to two other approaches. The first one is $CART$ [@breiman1984classification], run from sciki-learn with its default parameter setting but with a fixed maximum depth of the trees generated, and the second one is OCT from Bertsimas and Dunn [@bertsimas2017optimal]. The time limit used is 10 minutes for $BinOCT^*$ and 30 minutes to 2 hours for $OCT$. The depth of tree used is between $2$ and $4$, but we report in Table \[other table\] the best value among the three depths tried.
It should be noted that only 6 of the 16 datasets present in the Verwer and Zhang paper could be executed. The reason is that the decision trees consistent with some datasets are too large and deep to be inferred. In contrast to the algorithms to which we compare ours, we cannot set a maximum tree depth value because all examples must be correctly classified by the tree we infer.
Note that in Table \[other table\], our algorithms $DT\_depth$ and $DT\_size$ are very fast even when the trees to be inferred are large. In fact, for the dataset “balance-scale", our algorithms infer decision trees of depth $8$ in a few minutes while the other algorithms require more time for trees of depth $4$.
Concerning the accuracy of the trees we infer, it seems that when the depth is small ($< 5$) accuracy is equal for all approaches. However, when the depth becomes bigger, then our algorithms get higher accuracy. The most obvious example is the dataset balance-scale where we get $92.6\%$ accuracy compared to $78.9\%$ for $BinOCT^*$.
Artificial dataset
------------------
A last series of experimentations was carried out in order to see how the execution times of our two algorithms changes with different parameters of the datasets. The parameters we evaluate are the depth $k$ of the tree to infer, the number $f$ of features, the number $c$ of classes and the number $n$ of training examples. In order to perform the experiment, we randomly generate decision trees with the characteristics of depths, number of features and number of classes desired, then we randomly generate training examples from such trees. In Figures \[chartC\], \[chartK\], \[chartF\] and \[chartN\], we set three of these parameters and vary the remaining one to observe the effect on execution time.
![Chart of the average time over 10000 runs for $k=4$, $c=2$ and $f=10$.\[chartN\][]{data-label="fig:my_label"}](chartC2.pdf){width="100.00000%"}
![Chart of the average time over 10000 runs for $k=4$, $c=2$ and $f=10$.\[chartN\][]{data-label="fig:my_label"}](chartK.pdf){width="100.00000%"}
![Chart of the average time over 10000 runs for $k=4$, $c=2$ and $f=10$.\[chartN\][]{data-label="fig:my_label"}](chartF2.pdf){width="100.00000%"}
![Chart of the average time over 10000 runs for $k=4$, $c=2$ and $f=10$.\[chartN\][]{data-label="fig:my_label"}](chartN.pdf){width="100.00000%"}
In Figure \[chartC\], time seems to grow linearly. The algorithm $DT\_size$ appears to have a coefficient of $1$ which means that the fact that the algorithm is able to infer a decision tree that contains several classes does not bring any gain (or loss) in performance compared to the method consisting of inferring only decision trees with two classes and that would infer $c$ decision trees (one for each class). However, the algorithm $DT\_depth$ has a coefficient lower than $1$. Thus, the algorithm’s ability to infer decision trees with multiple classes provides a performance gain in this case. In Figure \[chartK\], time grows exponentially with $k$. However, a tree has generally exponentially more nodes than the depth of the tree. So, the figure indicates that the inference time increases almost polynomially with the number of nodes of the inferred tree. In Figure \[chartF\], time grows almost exponentially with the number of features. It is thus the number of features that seems to have the most impact on the inferring time. Then one way to improve our method is to try to reduce this impact. In Figure \[chartN\], we observe that time is growing rapidly until it reaches its peak. This peak corresponds to the number of examples that the algorithm needs to infer a decision tree consisting with all training examples. Thus, adding more examples will not affect the inference time unlike the previous approaches using SAT solver.
Conclusion
==========
We have presented a method that can infer an optimal decision tree for two definitions of optimality. The first definition a decision tree of minimal depth and consistent with the training examples is optimal. The second definition of optimality adds the constraint that the tree, in addition to having a minimum depth, must also have a minimum number of nodes. Although this optimal decision tree inference problem is known to be NP-complete [@laurent1976constructing; @hancock1996lower], we proposed an effective method to solve it.
Our first contribution is an effective SAT formulation that allows us to infer perfect decision trees for a fixed depth consistent with training examples. We have also shown how to add constraints in order to set the maximum number of nodes. In this case, the inferred decision tree will no longer necessarily be a perfect tree.
Our second contribution addresses the scalability issue. Indeed, the previous approach using SAT solver has the disadvantage that the execution time increases significantly with the number of training examples [@narodytska2018learning; @bessiere2009minimising]. Thus, we proposed an approach which does not process all the examples at once, instead it does it incrementally. The idea of processing a set of traces incrementally is to consider one example at a time, generate a decision tree and verify that it is consistent with the remaining examples. If it is not, choose an example that is incorrectly classified by the decision tree, i.e., a counterexample, and use it to refine the decision tree.
We evaluated our algorithms using various experiments and compared the execution time and quality of decision trees with other optimal approaches. Experimental results show that our approach performs better than other approaches, with shorter execution times, better prediction accuracy and better scalability. In addition, our algorithms have been able to process datasets for which, to the best of our knowledge, there are no other inference methods able of producing optimal models consistent with these datasets.
Appendix {#appendix .unnumbered}
========
Illustration of the inferring algorithm {#illustration-of-the-inferring-algorithm .unnumbered}
---------------------------------------
We present here an illustration of the inferring algorithm of a decision tree for depth 2. We use the following dataset $\mathcal{E}_0 = \{(false, false, true, false),$ $(false, false, false, true),$ $(true, false, true, true),$ $(true, true, true, false)\}$ and $\mathcal{E}_1 = \{(false, true, false, true),$ $(false, true, true, false),$ $(true, false, false, false),$ $(true, true, false, true)\}$
**Initialization:**
--------------- -----------------------------------------------------------------
Formula (1): $(\neg F_{1, 0} \vee \neg F_{1, 1}) \wedge
(\neg F_{1, 0} \vee \neg F_{1, 2}) \wedge
(\neg F_{1, 0} \vee \neg F_{1, 3}) \wedge $
$(\neg F_{1, 1} \vee \neg F_{1, 2}) \wedge
(\neg F_{1, 1} \vee \neg F_{1, 3}) \wedge
(\neg F_{1, 2} \vee \neg F_{1, 3}) \wedge $
$(\neg F_{2, 0} \vee \neg F_{2, 1}) \wedge
(\neg F_{2, 0} \vee \neg F_{2, 2}) \wedge
(\neg F_{2, 0} \vee \neg F_{2, 3}) \wedge $
$(\neg F_{2, 1} \vee \neg F_{2, 2}) \wedge
(\neg F_{2, 1} \vee \neg F_{2, 3}) \wedge
(\neg F_{2, 2} \vee \neg F_{2, 3}) \wedge $
$(\neg F_{3, 0} \vee \neg F_{3, 1}) \wedge
(\neg F_{3, 0} \vee \neg F_{3, 2}) \wedge
(\neg F_{3, 0} \vee \neg F_{3, 3}) \wedge $
$(\neg F_{3, 1} \vee \neg F_{3, 2}) \wedge
(\neg F_{3, 1} \vee \neg F_{3, 3}) \wedge
(\neg F_{3, 2} \vee \neg F_{3, 3})$
--------------- -----------------------------------------------------------------
--------------- ----------------------------------------------------------------
Formula (2): $(F_{1, 0} \vee F_{1, 1} \vee F_{1, 2} \vee F_{1, 3}) \wedge $
$(F_{2, 0} \vee F_{2, 1} \vee F_{2, 2} \vee F_{2, 3} )\wedge $
$(F_{3, 0} \vee F_{3, 1} \vee F_{3, 2} \vee F_{3, 3})$
--------------- ----------------------------------------------------------------
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$f_1?$]{} ; (Q2) at (-1,-1) [$f_0?$]{} ; (Q3) at (1,-1) [$f_0?$]{} ;
(Q4) at (-1.5,-2) [$1$]{} ; (Q5) at (-0.5,-2) [$1$]{} ; (Q6) at (0.5,-2) [$1$]{} ; (Q7) at (1.5,-2) [$1$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
**Add example: $e_0 = (0, 0, 1, 0) \in \mathcal{E}_0$**
--------------- --------------------------------------------------------------------------------------------------------------------------------------------
Formula (3): $(X_{0, 0} \vee \neg F_{1, 2}) \wedge (X_{0, 0} \vee X_{0, 1} \vee \neg F_{2, 2}) \wedge (\neg X_{0, 0} \vee X_{0, 1} \vee \neg F_{3, 2})$
--------------- --------------------------------------------------------------------------------------------------------------------------------------------
--------------- -----------------------------------------------------------------------------------------------------------------------------------------------
Formula (4): $(X_{0, 0} \vee \neg X_{0, 1} \vee \neg F_{2, 0}) \wedge (X_{0, 0} \vee \neg X_{0, 1} \vee \neg F_{2, 1}) \wedge$
$(X_{0, 0} \vee \neg X_{0, 1} \vee \neg F_{2, 3}) \wedge (\neg X_{0, 0} \vee \neg F_{1, 0}) \wedge (\neg X_{0, 0} \vee \neg F_{1, 1}) \wedge$
$(\neg X_{0, 0} \vee \neg F_{1, 3}) \wedge (\neg X_{0, 0} \vee \neg X_{0, 1} \vee \neg F_{3, 0}) \wedge$
$(\neg X_{0, 0} \vee \neg X_{0, 1} \vee \neg F_{3, 1}) \wedge (\neg X_{0, 0} \vee \neg X_{0, 1} \vee \neg F_{3, 3})$
--------------- -----------------------------------------------------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------
Formula (5): $(X_{0, 0} \vee X_{0, 1} \vee C_{0, 0}) \wedge (X_{0, 0} \vee \neg X_{0, 1} \vee C_{1, 0}) \wedge$
$(\neg X_{0, 0} \vee X_{0, 1} \vee C_{2, 0}) \wedge (\neg X_{0, 0} \vee \neg X_{0, 1} \vee C_{3, 0})$
--------------- -------------------------------------------------------------------------------------------------------
--------------- -----------------------------------------------------------------------------------------------------------------
Formula (6): $(X_{0, 0} \vee X_{0, 1} \vee \neg C_{0, 1}) \wedge (X_{0, 0} \vee \neg X_{0, 1} \vee \neg C_{1, 1})$
$(\neg X_{0, 0} \vee X_{0, 1} \vee \neg C_{2, 1}) \wedge (\neg X_{0, 0} \vee \neg X_{0, 1} \vee \neg C_{3, 1})$
--------------- -----------------------------------------------------------------------------------------------------------------
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$f_1?$]{} ; (Q2) at (-1,-1) [$f_0?$]{} ; (Q3) at (1,-1) [$f_0?$]{} ;
(Q4) at (-1.5,-2) [$0$]{} ; (Q5) at (-0.5,-2) [$-$]{} ; (Q6) at (0.5,-2) [$-$]{} ; (Q7) at (1.5,-2) [$-$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
**Add example: $e_1 = (1, 0, 1, 1) \in \mathcal{E}_0$**
--------------- ------------------------------------------------------------------------------------------------------------------
Formula (3): $(X_{1, 0} \vee \neg F_{1, 0}) \wedge (X_{1, 0} \vee \neg F_{1, 2}) \wedge (X_{1, 0} \vee \neg F_{1, 3}) \wedge$
$(X_{1, 0} \vee X_{1, 1} \vee \neg F_{2, 0}) \wedge (X_{1, 0} \vee X_{1, 1} \vee \neg F_{2, 2}) \wedge $
$(X_{1, 0} \vee X_{1, 1} \vee \neg F_{2, 3}) \wedge (\neg X_{1, 0} \vee X_{1, 1} \vee \neg F_{3, 0}) \wedge$
$ (\neg X_{1, 0} \vee X_{1, 1} \vee \neg F_{3, 2}) \wedge (\neg X_{1, 0} \vee X_{1, 1} \vee \neg F_{3, 3})$
--------------- ------------------------------------------------------------------------------------------------------------------
--------------- -----------------------------------------------------------------------------------------------------
Formula (4): $(X_{1, 0} \vee \neg X_{1, 1} \vee \neg F_{2, 1}) \wedge (\neg X_{1, 0} \vee \neg F_{1, 1}) \wedge$
$(\neg X_{1, 0} \vee \neg X_{1, 1} \vee \neg F_{3, 1})$
--------------- -----------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------
Formula (5): $(X_{1, 0} \vee X_{1, 1} \vee C_{0, 0}) \wedge (\neg X_{1, 0} \vee X_{1, 1} \vee C_{2, 0}) \wedge$
$(X_{1, 0} \vee \neg X_{1, 1} \vee C_{1, 0}) \wedge (\neg X_{1, 0} \vee \neg X_{1, 1} \vee C_{3, 0})$
--------------- -------------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------------------
Formula (6): $(X_{1, 0} \vee X_{1, 1} \vee \neg C_{0, 1}) \wedge (X_{1, 0} \vee \neg X_{1, 1} \vee \neg C_{1, 1}) \wedge$
$ (\neg X_{1, 0} \vee X_{1, 1} \vee \neg C_{2, 1}) \wedge (\neg X_{1, 0} \vee \neg X_{1, 1} \vee \neg C_{3, 1}) $
--------------- -------------------------------------------------------------------------------------------------------------------
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$f_1?$]{} ; (Q2) at (-1,-1) [$f_0?$]{} ; (Q3) at (1,-1) [$f_0?$]{} ;
(Q4) at (-1.5,-2) [$0$]{} ; (Q5) at (-0.5,-2) [$0$]{} ; (Q6) at (0.5,-2) [$-$]{} ; (Q7) at (1.5,-2) [$-$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
**Add example: $e_2 = (1, 1, 1, 0) \in \mathcal{E}_0$**
--------------- --------------------------------------------------------------------------------------------------------------------
Formula (3): $ (X_{2, 0} \vee \neg F_{1, 0}) \wedge (X_{2, 0} \vee \neg F_{1, 1}) \wedge (X_{2, 0} \vee \neg F_{1, 2}) \wedge $
$ (X_{2, 0} \vee X_{2, 1} \vee \neg F_{2, 0}) \wedge (X_{2, 0} \vee X_{2, 1} \vee \neg F_{2, 1}) \wedge $
$ (X_{2, 0} \vee X_{2, 1} \vee \neg F_{2, 2}) $
--------------- --------------------------------------------------------------------------------------------------------------------
--------------- -----------------------------------------------------------------------------------------------------
Formula (4): $(X_{2, 0} \vee \neg X_{2, 1} \vee \neg F_{2, 3}) \wedge (\neg X_{2, 0} \vee \neg F_{1, 3}) \wedge$
$(\neg X_{2, 0} \vee \neg X_{2, 1} \vee \neg F_{3, 3})$
--------------- -----------------------------------------------------------------------------------------------------
--------------- ---------------------------------------------------------------------------------------------------------
Formula (5): $ (X_{2, 0} \vee X_{2, 1} \vee C_{0, 0}) \wedge (X_{2, 0} \vee \neg X_{2, 1} \vee C_{1, 0}) \wedge $
$ (\neg X_{2, 0} \vee X_{2, 1} \vee C_{2, 0}) \wedge (\neg X_{2, 0} \vee \neg X_{2, 1} \vee C_{3, 0}) $
--------------- ---------------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------------------
Formula (6): $ (X_{2, 0} \vee X_{2, 1} \vee \neg C_{0, 1}) \wedge (X_{2, 0} \vee \neg X_{2, 1} \vee \neg C_{1, 1}) \wedge $
$ (\neg X_{2, 0} \vee X_{2, 1} \vee \neg C_{2, 1}) \wedge (\neg X_{2, 0} \vee \neg X_{2, 1} \vee \neg C_{3, 1}) $
--------------- -------------------------------------------------------------------------------------------------------------------
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$f_1?$]{} ; (Q2) at (-1,-1) [$f_0?$]{} ; (Q3) at (1,-1) [$f_0?$]{} ;
(Q4) at (-1.5,-2) [$0$]{} ; (Q5) at (-0.5,-2) [$0$]{} ; (Q6) at (0.5,-2) [$-$]{} ; (Q7) at (1.5,-2) [$0$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
**Add example: $e_3 = (0, 1, 0, 1) \in \mathcal{E}_1$**
--------------- --------------------------------------------------------------------------------------------------------------
Formula (3): $ (X_{3, 0} \vee \neg F_{1, 1}) \wedge (X_{3, 0} \vee \neg F_{1, 3}) \wedge$
$ (X_{3, 0} \vee X_{3, 1} \vee \neg F_{2, 1}) \wedge (X_{3, 0} \vee X_{3, 1} \vee \neg F_{2, 3}) \wedge $
$ (\neg X_{3, 0} \vee X_{3, 1} \vee \neg F_{3, 1}) \wedge (\neg X_{3, 0} \vee X_{3, 1} \vee \neg F_{3, 3}) $
--------------- --------------------------------------------------------------------------------------------------------------
--------------- ------------------------------------------------------------------------------------------------------------------------
Formula (4): $ (X_{3, 0} \vee \neg X_{3, 1} \vee \neg F_{2, 0}) \wedge (X_{3, 0} \vee \neg X_{3, 1} \vee \neg F_{2, 2} )\wedge $
$ (\neg X_{3, 0} \vee \neg F_{1, 0}) \wedge (\neg X_{3, 0} \vee \neg F_{1, 2}) \wedge $
$ (\neg X_{3, 0} \vee \neg X_{3, 1} \vee \neg F_{3, 0}) \wedge (\neg X_{3, 0} \vee \neg X_{3, 1} \vee \neg F_{3, 2}) $
--------------- ------------------------------------------------------------------------------------------------------------------------
--------------- ---------------------------------------------------------------------------------------------------------
Formula (5): $ (X_{3, 0} \vee X_{3, 1} \vee C_{0, 1}) \wedge (X_{3, 0} \vee \neg X_{3, 1} \vee C_{1, 1}) \wedge $
$ (\neg X_{3, 0} \vee X_{3, 1} \vee C_{2, 1)} \wedge (\neg X_{3, 0} \vee \neg X_{3, 1} \vee C_{3, 1}) $
--------------- ---------------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------------------
Formula (6): $ (X_{3, 0} \vee X_{3, 1} \vee \neg C_{0, 0}) \wedge (X_{3, 0} \vee \neg X_{3, 1} \vee \neg C_{1, 0}) \wedge $
$ (\neg X_{3, 0} \vee X_{3, 1} \vee \neg C_{2, 0}) \wedge (\neg X_{3, 0} \vee \neg X_{3, 1} \vee \neg C_{3, 0}) $
--------------- -------------------------------------------------------------------------------------------------------------------
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$f_1?$]{} ; (Q2) at (-1,-1) [$f_0?$]{} ; (Q3) at (1,-1) [$f_0?$]{} ;
(Q4) at (-1.5,-2) [$0$]{} ; (Q5) at (-0.5,-2) [$0$]{} ; (Q6) at (0.5,-2) [$1$]{} ; (Q7) at (1.5,-2) [$0$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
**Add example: $e_4 = (1, 0, 0, 0) \in \mathcal{E}_1$**
--------------- ----------------------------------------------------------------------------------------------------------------------------------------------
Formula (3): $ (X_{4, 0} \vee \neg F_{1, 0}) \wedge (X_{4, 0} \vee X_{4, 1} \vee \neg F_{2, 0}) \wedge (\neg X_{4, 0} \vee X_{4, 1} \vee \neg F_{3, 0}) $
--------------- ----------------------------------------------------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------------------------------------------------
Formula (4): $ (X_{4, 0} \vee \neg X_{4, 1} \vee \neg F_{2, 1}) \wedge (X_{4, 0} \vee \neg X_{4, 1} \vee \neg F_{2, 2}) \wedge $
$ (X_{4, 0} \vee \neg X_{4, 1} \vee \neg F_{2, 3}) \wedge (\neg X_{4, 0} \vee \neg F_{1, 1}) \wedge (\neg X_{4, 0} \vee \neg F_{1, 2}) \wedge $
$ (\neg X_{4, 0} \vee \neg F_{1, 3}) \wedge (\neg X_{4, 0} \vee \neg X_{4, 1} \vee \neg F_{3, 1}) \wedge $
$ (\neg X_{4, 0} \vee \neg X_{4, 1} \vee \neg F_{3, 2}) \wedge (\neg X_{4, 0} \vee \neg X_{4, 1} \vee \neg F_{3, 3}) $
--------------- -------------------------------------------------------------------------------------------------------------------------------------------------
--------------- ---------------------------------------------------------------------------------------------------------
Formula (5): $ (X_{4, 0} \vee X_{4, 1} \vee C_{0, 1}) \wedge (X_{4, 0} \vee \neg X_{4, 1} \vee C_{1, 1}) \wedge $
$ (\neg X_{4, 0} \vee X_{4, 1} \vee C_{2, 1}) \wedge (\neg X_{4, 0} \vee \neg X_{4, 1} \vee C_{3, 1}) $
--------------- ---------------------------------------------------------------------------------------------------------
--------------- -------------------------------------------------------------------------------------------------------------------
Formula (6): $ (X_{4, 0} \vee X_{4, 1} \vee \neg C_{0, 0}) \wedge (X_{4, 0} \vee \neg X_{4, 1} \vee \neg C_{1, 0}) \wedge $
$ (\neg X_{4, 0} \vee X_{4, 1} \vee \neg C_{2, 0}) \wedge (\neg X_{4, 0} \vee \neg X_{4, 1} \vee \neg C_{3, 0}) $
--------------- -------------------------------------------------------------------------------------------------------------------
=\[circle,fill=black!10,draw=black!75,minimum size=25pt,inner sep=0pt\]
(Q1) at (0,0) [$f_0?$]{} ; (Q2) at (-1,-1) [$f_1?$]{} ; (Q3) at (1,-1) [$f_2?$]{} ;
(Q4) at (-1.5,-2) [$0$]{} ; (Q5) at (-0.5,-2) [$1$]{} ; (Q6) at (0.5,-2) [$1$]{} ; (Q7) at (1.5,-2) [$0$]{} ;
(Q1) edge node\[left\] [$false$]{} (Q2); (Q1) edge node\[right\] [$true$]{} (Q3);
(Q2) edge node\[left\] [$false$]{} (Q4); (Q2) edge node\[right\] [$true$]{} (Q5);
(Q3) edge node\[left\] [$false$]{} (Q6); (Q3) edge node\[right\] [$true$]{} (Q7);
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Jayadev Acharya[^1]'
- 'Ashkan Jafarpour[^2]'
- 'Alon Orlitksy[^3]'
- 'Ananda Theertha Suresh[^4]'
title: |
Universal Compression of Envelope Classes:\
Tight Characterization via Poisson Sampling[^5]
---
abstract introduction preliminary rest iid\_small power\_law exponential\_law expected
[^1]: [email protected]. Part of the work done while author was a student at UCSD.
[^2]: [email protected]
[^3]: [email protected]
[^4]: [email protected]
[^5]: A part of the paper will appear in IEEE International Symposium on Information Theory 2014
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Federated learning (FL) is an emerging paradigm for distributed training of large-scale deep neural networks in which participants’ data remains on their own devices with only model updates being shared with a central server. However, the distributed nature of FL gives rise to new threats caused by potentially malicious participants. In this paper, we study targeted data poisoning attacks against FL systems in which a malicious subset of the participants aim to poison the global model by sending model updates derived from mislabeled data. We first demonstrate that such data poisoning attacks can cause substantial drops in classification accuracy and recall, even with a small percentage of malicious participants. We additionally show that the attacks can be targeted, i.e., they have a large negative impact only on classes that are under attack. We also study attack longevity in early/late round training, the impact of malicious participant availability, and the relationships between the two. Finally, we propose a defense strategy that can help identify malicious participants in FL to circumvent poisoning attacks, and demonstrate its effectiveness.'
author:
- Vale Tolpegin
- Stacey Truex
- Mehmet Emre Gursoy
- Ling Liu
bibliography:
- 'labelflip\_fl.bib'
title: |
Data Poisoning Attacks Against\
Federated Learning Systems
---
Introduction {#sec:intro}
============
Preliminaries and Attack Formulation {#sec:attack_eval}
====================================
Analysis of Label Flipping Attacks in FL {#sec:experiments}
========================================
Related Work {#sec:related_work}
============
Conclusion {#sec:conclusion}
==========
**Acknowledgements**. This research is partially sponsored by NSF CISE SaTC 1564097. The second author acknowledges an IBM PhD Fellowship Award and the support from the Enterprise AI, Systems & Solutions division led by Sandeep Gopisetty at IBM Almaden Research Center. Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the views of the National Science Foundation or other funding agencies and companies mentioned above.
DNN Architectures and Configuration {#appendix:a}
===================================
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We report long-slit spectroscopic observations of the dust-lane polar-ring galaxy obtained with the Southern African Large Telescope (SALT) during its performance-verification phase. The observations target the spectral region of the H$\alpha$, \[\] and \[\] emission-lines, but show also deep [NaI]{} stellar absorption lines that we interpret as produced by stars in the galaxy. We derive rotation curves along the major axis of the galaxy that extend out to about 8 kpc from the center for both the gaseous and the stellar components, using the emission and absorption lines. We derive similar rotation curves along the major axis of the polar ring and point out differences between these and the ones of the main galaxy.
We identify a small diffuse object visible only in H$\alpha$ emission and with a low velocity dispersion as a dwarf galaxy and argue that it is probably metal-poor. Its velocity indicates that it is a fourth member of the galaxy group in which belongs.
We discuss the observations in the context of the proposal that the object is the result of a major merger and point out some observational discrepancies from this explanation. We argue that an alternative scenario that could better fit the observations may be the slow accretion of cold intergalactic gas, focused by a dense filament of galaxies in which this object is embedded.
Given the pattern of rotation we found, with the asymptotic rotation of the gas in the ring being slower than that in the disk while both components have approximately the same extent, we point out that may be a galaxy in which a dark matter halo is flattened along the galactic disk and the first object in which this predicted behaviour of polar ring galaxies in dark matter haloes is fulfilled.
author:
- |
Noah Brosch,$^{1,2}$[^1] Alexei Y. Kniazev,$^{2,3}$ David Buckley,$^{2}$ Darragh O’Donoghue,$^{2}$ Yas Hashimoto,$^{2}$ Nicola Loaring,$^{2}$ Encarni Romero,$^{2}$ Martin Still,$^{2}$ Petri Vaisanen,$^{2}$ Eric B. Burgh,$^{4}$ Kenneth Nordsieck$^{4}$\
$^{1}$The Wise Observatory and the School of Physics and Astronomy, the Raymond and Beverly Sackler Faculty of Exactâ Sciences,\
Tel Aviv University, Tel Aviv 69978, Israel\
$^{2}$South African Astronomical Observatory, Observatory Road, Cape Town, South Africa\
$^{3}$Special Astrophysical Observatory, Nizhnij Arkhyz, Karachai-Circassia, 369167, Russia\
$^{4}$Space Astronomy Laboratory, University of Wisconsin, Madison, WI 53706, USA
date: 'Accepted 2007 April ??. Received 2007 March ??; in original form 2007 March ??'
title: 'The polar ring galaxy revisited [^2] '
---
\[firstpage\]
galaxies: ring galaxies — galaxies: evolution — galaxies: individual: — galaxies: dark matter — galaxies: galaxy haloes
Introduction {#introduction .unnumbered}
============
Ring galaxies posed significant astronomical interest since @LT76 modelled the Cartwheel galaxy as the result of a small galaxy passing through a larger one. While such events probably happen and produce some of the ring galaxies, in other instances different mechanisms might be at work. A particularly interesting kind of ring galaxy is the polar ring galaxy (PRG) where a flattened disk galaxy exhibits an outer ring of stars and interstellar matter that rotate in a plane approximately perpendicular to the central disk. An extensive catalog of PRGs was produced by @Whietal90.
The issue of PRGs was reviewed by @Co06. She reviewed a number of formation mechanisms for PRGs: minor or major mergers, tidal accretion events, or direct cold gas accretion from filaments of the cosmic web. @Co06 proposed that these objects can be used to probe the three-dimensional shape of dark matter (DM) haloes, provided the PRG is in equilibrium in the gravitational potential.
The well-known Spindle Galaxy (NGC 2685), an archetypal PRG, exhibits two sets of rings: an outer one visible only on HI maps and which might be in the plane of the galaxy, and an inner one that is helix-shaped, is perpendicular to the main axis of the galaxy, is optically bright, shows embedded present-day star formation, and is associated with prominent dust lanes. Shane (1980) explained the system as consisting of a lenticular galaxy that recently accreted an HI gas cloud that formed the inner ring, while the outer gas ring might be a remnant of the formation of the galaxy. Hagen-Thorn (2005) found that the stellar population of the inner system of dust and gas, arranged in a spiral around the ”spindle” but really in a disk, is 1.4$\times10^9$ years old.
In a different ring galaxy, NGC 660, Karataeva (2004) detected red and blue supergiants belonging to the ring system. They showed that the age of the youngest stars there is only $\sim$7 Myr; thus star formation is currently taking place. N660 is special in that both the disk and the polar ring contain stars, gas and dust. @Resh04, who analyzed three other ring galaxies, showed that their rings result from ongoing interactions or mergers where the main galaxy is a spiral and the rings are currently forming stars.
Other claims of interactions being at the origin of the rings and of the star formation taking place therein have been put forward by Mayya & Korchagin (2001, revised 2006). On the other hand, others claimed that rings are formed as a dynamical event in a larger disk galaxy (e.g., Mazzuca 2001). It is clear that more studies of ring galaxies, and in particular such investigations that can time the ring and star formation events, can help understand the particular instances when a galaxy-galaxy interaction took place, when a ring is formed, and when the event does trigger the SF process. There is also the possibility that careful tracing of the polar ring and of the galaxy itself, and their kinematic properties, might reveal the DM halo shape and properties, as advocated by @Co06. This singles out PRGs as valuable targets for DM studies.
In this paper we analyze new observations of the polar-ring galaxy , a PRG with an optical redshift of 11649$\pm$10 km sec$^{-1}$ located at l=341.02, b=-28.73, also identified as PRC B-18 in Whitmore (1990). The object was recently studied by @Resh06, who showed that this is a giant galaxy in a compact triplet, together with PGC 400092 (classified Sd/Irr:) and PGC 399718 (classified SBc:) at approximately the same redshift. The authors used the 1.6-meter telescope of the Pico dos Dias Observatory in Brazil for imaging in BVRI, the CTIO 1.5-meter telescope to collect spectral observations, and included data from IRAS and 21-cm line observations. However, most of their conclusions about the nature of the object rely on the morphological appearance of the galaxy.
@Resh06 modelled using an N-body code that includes gas dynamics using sticky particles and star formation. They concluded that the best-fitting model is of a major merger, whereby a gas-rich galaxy transferred a sizable amount of matter to during a parabolic encounter. The matter subsequently relaxed and now forms a complete ring of stars, gas, and dust around whereas the donor galaxy is one of the two other galaxies in the same group.
The reason to revisit this object was the availability of high-quality spectra obtained with the effectively 8-meter diameter Southern African Large Telescope (SALT) telescope. We derive, for the first time, rotation curves for the ionized gas and for the stellar components of both the main galaxy and the polar ring. Since PRGs might make good test cases for the properties of dark matter haloes in and around galaxies, as argued by @Co06, the more observational data collected on these objects and with higher quality, the better.
Very few PRG observations obtained with large telescopes have been published. A noticeable one is by Swaters & Rubin (2003), with the Baade 6.5-meter telescope on Las Campanas, tracing the dynamics of the stellar component of the prototype PRG NGC 4650A where they showed that the polar ring is actually a polar disk, an extended feature rather than a narrow gas disk. They favour a scenario by which the ring/disk was formed by the polar merger to two similar disks, as previously suggested by Iodice (2002). Iodice (2006) observed the gaseous component in the ring of N4650A with ESO’s FORS2 on UT4 and concluded that a scenario by which it could be formed was through slow gas accretion from the cosmic web filaments. We propose that the same situation could be taking place for .
This paper is organized as follows: § \[txt:Obs\_and\_Red\] gives a description of all the observations and data reduction. In § \[txt:results\] we present our results, analyze them in § \[txt:disc\], and present our interpretation in § \[txt:interp\]. The conclusions drawn from this study are summarized in § \[txt:summ\].
------------ -------------- ------------- ------ ------------------ ---------
Date Exp.time Spec. Range Slit [1]{}[c]{}[PA]{} Disp.
(sec) (Å) () ($^\circ$) (Å/pix)
16.07.2006 2$\times$600 3650–6740 1.5 140 0.98
16.07.2006 1$\times$600 3650–6740 1.5 35 0.98
20.09.2006 2$\times$900 6050–7315 1.5 140 0.40
20.09.2006 1$\times$750 6050–7315 1.5 27 0.40
21.09.2006 3$\times$900 6050–7315 1.5 27 0.40
------------ -------------- ------------- ------ ------------------ ---------
: Details of the RSS observations[]{data-label="t:Obs"}
[ ]{}
[ ]{}
[ ]{}
[ ]{}
Observations and data reduction {#txt:Obs_and_Red}
===============================
SALT was described by Buckley (2006) and by O’Donoghue (2006), its Robert Stobie Spectrograph (RSS) was described by Burgh (2003) and Kobulnicky et al. (2003), and the first scientific papers based on its observations were published by Woudt (2006) and by O’Donoghue (2006). We used the SALT and RSS to observe . The observations of were obtained during the Performance Verification (PV) phase of the SALT telescope with the RSS spectrograph and are described in Table \[t:Obs\].
The July 2006 spectra (see Table \[t:Obs\]) were obtained during unstable weather conditions (high humidity, seeing worse than 5), without fully stacking the SALT mirrors. They cover the range 3650–6740 Å with a spectral resolution of $\sim$1.0 Å pixel$^{-1}$ or a FWHM of 6–7 Å. These spectra do not show strong and extended emission lines but were used to measure equivalent widths (EWs) of absorption lines in that spectral range following observations.
The spectra obtained on the nights of September 2006 were taken during stable weather conditions with seeing $\sim$15. They cover the range from $\sim$6050Å to $\sim$7300Å with a spectral resolution of 0.4 Å pixel$^{-1}$ or 2.4 Å FWHM. All data were taken with a 15 wide slit and a final scale along the slit of 0258 pixel$^{-1}$ (after binning the CCDs by a factor of two). Each exposure was broken up into 2–3 sub-exposures to allow the removal of cosmic rays. Spectra of a Cu–Ar comparison lamp were obtained after the science exposures to calibrate the wavelength scale.
The September 2006 data include two spectra obtained at position angle 140$^{\circ}$ centered on extending about four arcmin along the galaxy’s major axis and at a shallow angle to the dust lane, where the northern part passes also through the “northwest companion” PGC 400092 [@Resh06], and three spectra centered on the same position but obtained at position angle 27$^{\circ}$, along the major axis of the ”polar ring” described by @Resh06. We emphasize that the sampling of the major axis spectra was at PA=140$^{\circ}$, not at 130$^{\circ}$ as done by @Resh06, since 140$^{\circ}$ is closer to the position angle of the disk as given by Reshetnikov (148$^{\circ}$) and allows for a moderate degree of disk warping.
Although the observations discussed here are mostly spectroscopic, one image of the galaxy was obtained with a two-sec exposure in the V filter with the SALTICAM camera (O’Donoghue 2006) prior to the spectrometer observations in order to adjust the slit orientation, and is shown here as Figure \[fig:AM\_direct\]. The $\sim$15 seeing during the observations, and the problematic image quality SALT exhibited at that time, which can be evaluated from the stellar images on Figure \[fig:AM\_direct\_masked\] (see below), caused the images far from the good-quality $\sim$3 arcmin region to assume complicated shapes. The full SALTICAM image is $\sim$10 arcmin across with 0.28 arcsec/pixel (after binning on-chip by a factor of two).
The data for each RSS chip were bias and overscan subtracted, gain corrected, trimmed and cross-talk corrected, sky-subtracted and mosaiced. All the primary reduction was done using the IRAF[^3] package [*salt*]{}[^4] developed for the primary reduction of SALT data. Cosmic ray removal was done with the FILTER/COSMIC task in MIDAS.[^5] We used the IRAF software tasks in the [*twodspec*]{} package to perform the wavelength calibration and to correct each frame for distortion and tilt. One-dimensional (1D) spectra were then extracted using the IRAF APALL task.
Figures \[fig:AM\_2D\_130\] and \[fig:AM\_2D\_27\] show parts of fully reduced and combined spectral images for PA=140$^{\circ}$ and PA=27$^{\circ}$, respectively. Figure \[fig:AM\_1D\_130\] shows the spectrum of the central part of . The $\sim$40Å missing sections at $\sim \lambda \lambda$ 6500 and 6930Å are produced by small gaps between the three CCDs of the RSS. The noisy region of the RSS images shown in Figs. \[fig:AM\_2D\_130\] and \[fig:AM\_2D\_27\] near $\sim$6685Å is a subtraction artifact of laser light scattered into the RSS from SALT’s interferometric auto-collimating system. Figure \[fig:AM\_1D\_130\] shows the 1D spectra of the central part of extracted from the 2D spectra. Figure \[fig:PGC40092\_1D\] shows the 1D spectrum of the galaxy PGC 400092 extracted from the 2D spectrum observed at PA=140.
The derived internal errors for the 2D wavelength calibrations were small and did not exceed 0.04 Å for a resolution of 0.4 Å pixel$^{-1}$, or $<$2 km s$^{-1}$ at the wavelength of redshifted H$\alpha$ line. To exclude systematic shifts originating from known RSS flexure, we calculated line-of-sight velocity distributions along the slit for both emission and absorbtion lines using a suite of MIDAS programs described in detail in @Zasov00. These programs allow the use of additional correction factors derived from tracing nearby night-sky lines whose accurate wavelengths are very well known to correct the observed wavelengths of the [NaI]{}D, H$\alpha$ \[\] $\lambda$6583 and \[\] $\lambda$6716 emission lines. After implementing the night-sky line corrections, the line-of-sight velocity distributions are accurate to $\sim$1.5 km s$^{-1}$. Most of the calculated velocity distributions are shown in Figures \[fig:AM\_rot\_130a\]–\[fig:AM\_rot\_27\]. All velocities derived with this procedure are heliocentric.
All emission lines were measured with the MIDAS programs described in detail in @SHOC [@Sextans]. These programs determine the location of the continuum, perform a robust noise estimation, and fit separate lines with single Gaussian components superposed on the continuum-subtracted spectrum. Nearby lines, such as the H$\alpha$ and \[\] $\lambda\lambda$6548, 6583 lines on the one hand, the \[\] $\lambda\lambda$6716, 6731 lines on the other, and [NaI]{}D $\lambda\lambda$5890, 5896 absorption doublet were fitted simultaneously as blends of two or more Gaussian features.
[ ]{}
[ ]{}
Results {#txt:results}
=======
Spectra of and PGC 400092
--------------------------
A cursory inspection of the spectra obtained at PA=140$^{\circ}$ (see Figure \[fig:AM\_2D\_130\]) shows rotation detectable in the same amount and behaviour exhibited by the H$\alpha$, \[\] $\lambda\lambda$6548,6583 and \[\] $\lambda\lambda$6716,6731 emission lines, and rotation as almost a solid body exhibited by the [NaI]{} $\lambda\lambda$5890,5896 doublet lines. The NE extension of the spectrum, away from and crossing the companion galaxy PGC 400092, shows that the same emission lines seen in are produced by the NE companion; the rotation there is much slower and the [NaI]{} doublet is not visible, even though the continuum there is visible. In addition, the spectrum of PGC 400092 shows also weak \[\] $\lambda$6300 and HeI $\lambda$5876 in emission, while the spectrum of shows \[\] $\lambda$6300 emission only in the central part.
------------------- ------------------
Absorption Line Equivalent Width
[1]{}[c]{}[(Å)]{} (Å)
CaII H 8.9$\pm$1.5
CaII K 10.3$\pm$1.8
H$\delta$ 6.5$\pm$2.1
H$\gamma$ 5.8$\pm$2.4
H$\beta$ 6.4$\pm$2.5
[MgI]{}b 3.5$\pm$0.8
[NaI]{}D 5.8$\pm$0.7
------------------- ------------------
: EWs of absorption lines in spectra of []{data-label="t:EW_abs"}
The short-wavelength spectra obtained in June 2006 (top panel of Fig. \[fig:AM\_1D\_130\]) show the blend of the [NaI]{} doublet as a single line (due to the lower resolution of this setup), and the H$\beta$, H$\gamma$ and H$\delta$ lines in absorption. The CaII H and K doublet is seen in absorption at the blue end of the spectrum. The spectra also show very weak \[\] $\lambda\lambda$4959, 5007 emission lines. In this figure and in the following plots we describe as ”intensity” the raw counts extracted from the spectra. Since our data have not been spectrophotometrically calibrated, this is in reality ”relative intensity”. The equivalent widths of the main absorption lines were measured for the central part of the galaxy and are shown in Table \[t:EW\_abs\]. Measurements of lines detected in more than one spectrum were averaged.
[ ]{}
[ ]{}
[ ]{}
[ ]{}
[ ]{}
The rotation curve of along the major axis, derived from the two-spectra combination shown in Figure \[fig:AM\_2D\_130\], is shown in Figures \[fig:AM\_rot\_130a\] and \[fig:AM\_rot\_130b\]. Figure \[fig:AM\_rot\_130a\] show the velocity-position plot and Figure \[fig:AM\_rot\_130b\] shows the galacto-centric velocity-distance plot. In general, the emission-line rotation curve derived here corresponds with that shown in Figure 5 of @Resh06, except that ours is better sampled, has a higher signal-to-noise, and the rotation curves derived from the different emission lines practically coincide, as can be estimated from the formal 1$\sigma$ error bars plotted in the figures and from the scatter of the individual points. Figure \[fig:AM\_rot\_130b\] shows also a comparison of our measurements with those of @Resh06.
Deriving the rotation curves shown in Figures \[fig:AM\_rot\_130a\] and \[fig:AM\_rot\_130b\] we found that the systemic radial velocity of is 11663$\pm$3 km sec$^{-1}$, formally higher by some 14 km sec$^{-1}$ than the value given by @Resh06 in their Table 3 but consistent with their value within the quoted uncertainties. This offset might be the result of a slightly different definition of the systemic velocity; we chose the value for which the NW branch of the rotation curve matched best that for the SE branch and by this procedure also found the rotation center of the galaxy. Independently, we found that this location on the velocity curve is also the central point for the linear fitting of all the measurements for the [NaI]{}D lines seen in absorption, as shown in Figure \[fig:AM\_rot\_130a\]. We obtained a best-fit line following the relation: $${\rm V_r = (11663 \pm 2) + (15.2 \pm 0.4) \times {\rm R} } $$ where $\rm R$ is the distance in arcsec from the point where the radial velocity of , defined using the emission lines, equals 11663 km sec$^{-1}$ and we adopt this location as the kinematic centre of the galaxy. The different symbols indicate the H$\alpha$ velocity (black squares), the \[\] $\lambda$6583 line velocity (red squares), and the \[\] $\lambda$6716 line velocity (blue triangles). The stellar rotation along the same slit position on the major axis, as derived from an average of the two [NaI]{}absorption lines, is depicted as filled black circles.
We detected a discrepant systemic velocity 11680$\pm$10 km sec$^{-1}$ for the NW companion PGC 400092 as well, where our value is significantly lower than the 11735$\pm$6 km sec$^{-1}$ given in @Resh06. Since the velocity discrepancies for and for PGC400092 are in opposite directions, we can probably rule out a systematic shift between our velocity scale and the one of @Resh06. This is confirmed also by the plot in Figure \[fig:AM\_rot\_130b\] where their derived velocity curve points are plotted over our results. The shift between our data for PGC400092 and that from @Resh06 could be the result of the slit position for $\rm PA = 140\degr$ used here that did not cross exactly the physical center of that galaxy.
We could also derive the velocity dispersion of the H$\alpha$ line along the slit for $\rm PA = 140\degr$; this is shown in the bottom panel of Figure \[fig:AM\_rot\_130b\]. The dispersion is shown as the FWHM of the line after correcting for the intrinsic spectrometer line width. The corrected H$\alpha$ line FWHM=5–7 Å found for the central part ($\pm$3 arcsec) of indicates internal motions of 200–300 km s$^{-1}$. The corrected FWHM=$<$1 Å measured for the H$\alpha$ line of PGC400092 indicates internal motions slower than 45 km s$^{-1}$.
The rotation curve along the polar ring axis, at $\rm PA =
27\degr$, is shown in Figure \[fig:AM\_rot\_27\] as a velocity-position plot. This, as already mentioned, relies mostly on the emission lines since the [NaI]{} absorptions are visible only in the central part of the spectrum, and is therefore more limited in extent. The spectra for $\rm PA = 27\degr$ show a linearly increasing rotation for $\sim$7 SW of the galaxy centre outwards, where the center position is that derived for the major axis. Since the NE and SW branches of the ring’s major axis show very different behaviour from that observed along the galaxy’s major axis, the method used previously to find the rotation center by matching the two branches could not be used in this case, thus we do not show a folded and combined velocity curve for the major axis of the ring.
The NE branch shows an approximately flat rotation from $\sim$2 away from the centre, as derived from the emission lines, with some oscillations from the center to the periphery at 10 arcsec from the center. These oscillations are evident in both H$\alpha$ and \[\] $\lambda$6583; they may be caused by the overlap of the emission lines from the ring with those from the main body of the galaxy. The plot in the top panel of Fig. \[fig:AM\_rot\_27\] shows that the strongest H$\alpha$ emission is encountered close to the location of the most intense continuum contribution (compare the solid and the dashed lines).
Our spectra along $\rm PA = 27\degr$ show a completely different kinematic behaviour than the one described by @Resh06. Their Fig. 7 shows a $\sim$50 km sec$^{-1}$ difference between the velocity of the \[\] $\lambda$6583 and H$\alpha$ at the galaxy centre that increases to $\sim$100 km sec$^{-1}$ at the SW end of the ring. We, on the other hand, see no difference between the velocities of these two lines. Moreover, the \[\] lines in our observed spectrum also show the same behavior as the \[\] $\lambda$6583 and H$\alpha$ lines. We also note that the extent to which the rotation is defined and measurable for this position angle and using the emission lines is practically the same as for the major axis of , some 8 kpc from the center (at 167 Mpc).
Similar to the case of the major axis, $\rm PA = 140\degr$, we see here also a straight-line behaviour with galacto-centric distance of the [NaI]{} absorption lines. We find a formal linear fit of the form $${\rm V_r = (11662 \pm 2) + (14.9 \pm 0.8) \times {\rm R}} $$ The [NaI]{} rotation curve is linear from 15 SW of the centre to $\sim$5" NE of the kinematic centre. Note that the value found for the slope at this position angle is virtually identical with that for the major axis in equation (1).
A comparison of the two panels of Fig. \[fig:AM\_rot\_27\], the lower one which is a velocity-position plot for $\rm PA = 27\degr$ and the upper one which is a plot of the line intensity vs. position along the slit, shows that the region where most of the line emission is produced is about 4 to the NE of the kinematic center of and that the emission is practically only along the NE part of the ring.
As for $\rm PA = 140\degr$, we derive the velocity dispersion for this position angle as the FWHM of the H$\alpha$ line vs. galacto-centric distance. This is shown in the bottom panel of Fig. \[fig:AM\_rot\_27\_all\] after correction for the intrinsic width of the lines using the night sky spectrum. The corrected FWHM=7 Å for the redshifted H$\alpha$ indicates internal motions of $\sim$300 km s$^{-1}$.
Although not spectrophotometrically calibrated, our spectra allow the derivation of a few physical parameters of the gas using line ratios. The good signal-to-noise of the spectra allows the derivation of these ratios along the slit, as shown in Figs. \[fig:AM\_140\_cond\] and \[fig:AM\_27\_cond\]. The ratios plotted in Fig. \[fig:AM\_140\_cond\] allow a derivation along the galaxy major axis and for its NW companion. Since these ratios are based on the very closely located emission lines, they practically do not depend on whether the spectral data were corrected for sensitivity or not. For the red spectral range, using the sensitivity curve cannot change these ratios by more that a few percent; this is less than the displayed errors.
Creating these ratios we took into account the possible stellar absorption in the H$\alpha$ line. Checking Table \[t:EW\_abs\], and considering the Balmer spectra of @Rosa99 we suggest that EW$_{abs}$(H$\alpha$)=6 Å with a constant value along the slit. Since EW(H$\alpha$)$\approx$15Å for the emission line at the center of , decreasing to the galaxy’s edges, this correction is very important and without it the ratios of \[\]/H$\alpha$ and \[\]/H$\alpha$ would increase from the centre to the edges. That could be interpreted as an increase in of metallicity with galacto-centric distance, which is not correct. With a measured line ratio for the central part of ($\pm$2 arcsec) $\rm ([\ionn{N}{II}]
\lambda6583/H\alpha)$=0.54$\pm$0.02, the metallicity in the center of is 12+log(O/H)=8.92$\pm$0.06 dex [@Den02] and drops down to 8.81$\pm$0.07 in the outer parts of the galaxy. The figure indicates that along the major axis of the galaxy n$_e\simeq$50 cm$^{-3}$. The measurements for detected part of PGC 400092 give 12+log(O/H) = 8.45$\pm$0.12 dex and n$_e \simeq$ 500 cm$^{-3}$.
In a similar way, we derive the gas properties along the major axis of the ring (see Fig. \[fig:AM\_27\_cond\]). With the line ratios measured in the central part of ($\pm$2 central arcsec) $\rm ([\ionn{N}{II}] \lambda6583/H\alpha)$=0.51$\pm$0.04, the metallicity in the center of is 12+log(O/H)=8.91$\pm$0.06 [@Den02], essentially the same value found from the major axis measurement. From the measured \[\] lines ratio we obtain the same value found previously: n$_e\simeq$ 50 cm$^{-3}$.
Newly detected H$\alpha$ emission knot
--------------------------------------
An isolated H$\alpha$ emission knot was detected at ; , some 78 away from the main body of the galaxy to the NE and on the extension of the ring’s major axis. This knot is real and was detected on all spectra observed at $\rm PA = 27\degr$ taken on 2006 September 20 and 21. The velocity distribution with distance is shown in the top panel of Figure \[fig:AM\_rot\_27\_all\]. It is evident that the line emitting knot is fairly isolated and is very distant from the galaxy, yet its radial velocity is close to that of the systemic velocity. The measured velocity for the knot is 11645$\pm$5 km sec$^{-1}$; this is more than three standard deviations away from the systemic velocity of and very many standard deviations away from the recession velocity measured for H$\alpha$ at the NW tip of the galaxy. It is also very different from the velocity of PGC 400092, the NW companion of , or from that of PGC 399718, the other companion in the triplet.
Our observations do not show a significant velocity dispersion of the H$\alpha$ line observed from the knot, as shown in the bottom panel of Fig. \[fig:AM\_rot\_27\_all\]; a formal measurement indicates that this H$\alpha$ line has the same FWHM ($\sim$2.4 Å) as the reference night-sky line. The corrected FWHM$\leq$1 Å for the redshifted H$\alpha$ from the knot indicates internal motions slower than 40 km s$^{-1}$. The size of the line-emitting region is only $\sim$5 arcsec; small, but well-resolved by our observations. The very weak continuum is detected; this allows a measurement of EW(H$\alpha$) = 120$\pm$15 Å. No additional emission lines are visible in the spectrum.
Analysis {#txt:disc}
========
The image of the field displayed in Figure \[fig:AM\_direct\] shows not only but also its two companion galaxies. Fig. \[fig:AM\_direct\] is a V-band image of the field obtained with SALTICAM in the same night as the spectroscopic observations on September 21. The image of the three galaxies allows one to note that (a) the region around the target contains many diffuse, low surface brightness (LSB) images that might be parts of galaxies or LSB dwarfs at the same redshift, or distant objects in the background, and (b) the appearance of the companion galaxy PGC 400092 to the NW is that of a Sd galaxy with a similar overall size to that of . The LSB objects are also visible on Digitized Sky Survey images of the region.
We performed unsharp masking of Figure \[fig:AM\_direct\] to emphasize the dust lane; this is shown in Figure \[fig:AM\_direct\_masked\] and, contrary to the claim of @Resh06 that the dust lane is split and embraces the galaxy nucleus from SE and NW, indicates that the lane is fairly straight, passes south and west of the brightest part of the galaxy, and is probably not split at all. The stars in Fig. \[fig:AM\_direct\_masked\] have the shapes of crescent moons. This arises from telescope optical problems which are being ironed out during the Performance Verification process, and have been emphasized by the unsharp masking.
The measured ratio of emission lines to corrected H$\alpha$, and the possibly very weak \[\] $\lambda$5007 emission detected in our July 2006 short-wavelength spectra, puts this object at the border between starburst nuclei (SBN) and LINERs. Norris et al. (1990) found no compact radio core in this galaxy and for this reason it should be classified as SBN; this is in agreement with the previous conclusions of @Al91.
The curves shown in Fig. \[fig:AM\_rot\_130b\] indicate that the gas rotation along the major axis has its maximum at $\sim$240 km sec$^{-1}$ and not at 195 km sec$^{-1}$ as given by @Resh06, and that this maximum is reached asymptotically for the NE part of the galaxy. Figire \[fig:AM\_rot\_130a\] shows that our measurements are compatible with those of @Resh06 for the regions of overlap. The last points of the rotation curve branch of the SE part of the galaxy, from galacto-centric distance of 6 to 10, drop from 200$\pm$7 km sec$^{-1}$ to 150$\pm$7 km sec$^{-1}$ in both H$\alpha$ and \[\] $\lambda$6583 lines. This drop is gradual from 6 to 8 but shows a step-like drop at this location, followed by a recovery with a similar distance-velocity gradient as for the central part of the galaxy.
A comparison of the major axis rotation curves shown in Fig. \[fig:AM\_rot\_130b\] shows clearly the difference between the kinematic behaviour of the two [NaI]{}D absorption lines and the H$\alpha$, \[\] $\lambda$6583 and \[\] $\lambda$6716 emission lines. At this point it is worth discussing the origin of the [NaI]{} absorption lines. These could be produced by stellar photospheres, or by diffuse gas in the interstellar medium of . For the case of dwarf starburst galaxies, Schwartz & Martin (2004) used giant and supergiant stars to show that the EW of the [MgI]{} triplet near 5180Å should be twice that of the [NaI]{} lines. If this would be the case for then our blue spectrum where the [MgI]{} triplet is barely visible would rule out a major [NaI]{} absorption contribution from stars.
[ ]{}
However, in giant galaxies such as the stellar populations are better represented by main sequence stars. These have stronger photospheric [NaI]{} than [MgI]{} (e.g., a M0V star from the same library as used by Schwartz & Martin (2004) has EW([MgI]{})=20Å and EW([NaI]{})=12Å. While it is not possible to separate the stellar [NaI]{} from the interstellar absorption, we can accept that at the least a fraction, and perhaps all of the observed absorption represents the stars in the galaxy. For example, in M82 Saito et al. (1984) detected [NaI]{} absorption that they attributed to stars and interpreted as solid-body rotation.
Assuming that most of the [NaI]{} absorption is photospheric, this would indicate that, while the gaseous component follows a “normal” galactic rotation law, the stellar component rotates almost like a solid body for $\sim$10 away from the centre. The maximal rotation velocity exhibited by the stellar component is only $\sim$150 km sec$^{-1}$ at 10 from the centre for both ends of the major axis.
The extent over which the emission is observed for the “polar ring” is almost the same as for the major axis, some 18overall as shown in Fig. \[fig:AM\_rot\_27\], but the derived rotation curve is completely different. The rotation curve indicates solid-body like rotation for 15 to the NE (one resolution element away from the centre, given the seeing) and for about 5 to the SW. The velocity difference between the outermost points on the slit where the absorption lines are measured is only 90 km sec$^{-1}$. The velocity gradients shown by the stellar components along the major axis of the galaxy and along the axis of the PR, in regions where a linear rotation curve can be defined, are very similar as equations (1) and (2) show. In both cases the gradients are $\sim$19 km sec$^{-1}$ kpc$^{-1}$, where we converted the observational gradients from equations (1) and (2) to physical units.
Interpretation {#txt:interp}
==============
At a distance to the object of 167 Mpc (H$_0$=70 km sec$^{-1}$ Mpc$^{-1}$) the radius of the galaxy to the outermost point where emission lines are visible is $\sim$8 kpc. We found the stellar component of a 16 kpc wide galaxy rotating as a solid body, while its gaseous component measured at the same slit position shows a smoothly increasing rotation curve which then flattens out. A ring or disk feature with an extent similar to that of the galaxy is observed at an inclination of $\sim$60$^{\circ}$ to the major axis of the galaxy. The stellar component observed with the spectrometer slit oriented along the major axis of the ring is also rotating as a solid body and with a similar velocity-distance gradient to that observed for the main body of the galaxy.
@Resh06 concluded from their photometry and spectroscopy, coupled with results of N-body modelling, that is a PRG. Their models indicate that the system might be the result of a major interaction between a ”donor” galaxy with a 17 kpc stellar disk and a 42 kpc gaseous disk, with a total mass of 3.6$\times10^{11}$ M$_{\odot}$, which encountered a 2$\times10^{11}$ M$_{\odot}$ and 14 kpc wide ”receptor” galaxy some 1.6 Gyrs ago with an impact parameter of 130 kpc and a relative velocity of 145 km sec$^{-1}$. This encounter transferred a large quantity of matter (stars, gas, and dust) from the donor to the receptor galaxy resulting in the formation of the polar ring which is inclined with respect to the galaxy disk and is warped. @Resh06 suggested that the donor galaxy survived and is PGC 399718, the southern companion in the triplet, and argued that their suggestion is supported by the reddish (B-V) colour of the galaxy and by its somewhat disturbed appearance.
In selecting this scenario in preference to those of minor mergers calculated by them, or of other possible models for the formation of ring galaxies, @Resh06 relied primarily on the morphological appearance of the galaxy. In particular, the minor merger models rejected by @Resh06 produced only partially-open rings that were not closed, whereas the preferred major merger model produced a “closed and regular ring” a few 10$^8$ years following the interaction.
Since the acceptance of the @Resh06 scenario as the explanation for the appearance of this system relies on their interpretation that the ring is closed and regular, it is worth examining whether the observations presented here support this assertion.
The specific items resulting from our observations that require understanding are:
1. Solid-body rotation is observed for stars vs. a ”regular” rotation for the gas at the same (projected) locations. No differential rotation, as expected from a stellar disk, is observed. This is true for the main body of the galaxy as well as for the ring, though with the gas showing a different distance-velocity gradient than the stars.
2. The ring is very faint and there is no evidence that it contains a considerable number of stars, as would be expected from the major merger claimed by @Resh06. Our observations of the intensity distribution along the slit at PA=27$^{\circ}$ show that the stars producing the continuum are located mostly where the HII is, namely some 2-5 NE of the centre.
3. The ring dynamics are different at its SW end, where the line and continuum emissions are very weak and the ring is more extended [@Resh06], in comparison with the other end of the ring.
4. The gas dynamics for the ring are very different from those of the gas in the galaxy. Specifically, at similar extents from the dynamical centre the gas in the ring spins much slower than the gas in the galaxy. This, while the stellar components have similar kinematic behaviours as evaluated from the velocity-distance gradients.
Apparent solid-body rotation of a galaxy could be produced, for example, by dust extinction. Baes (2003) modelled the light propagation through a dusty galactic disk and showed that, unless the disk is perfectly edge-on, no effects in the kinematics would be observable. The more the disk is edge on, and the stronger the extinction caused by the dust in the disk is, the more would the rotation curve resemble that of a solid body. Perusal of the DSS images of the object, of the image shown in Fig. 1 of @Resh06, and of our Figs. \[fig:AM\_direct\] and \[fig:AM\_direct\_masked\], shows that is not a purely edge-on galaxy and that, since the disk deviation from edge-on is definitely more than “a few degrees” but rather $\sim$25$^{\circ}$, as explained below, we should not expect to see a solid-body rotation just because of dust obscuration and light scattering. We can, therefore, reject the possibility that the solid-body rotation is an effect of dust obscuration.
Stars vs. gas in the disk
-------------------------
The key observation reported here is the difference in rotation curves between the emission lines produced by the gas and the stars as represented by the absorption lines. Such cases of different kinematic behaviour of the gas and the stars are known in the literature, Bettoni (1990), where NGC 2217 was shown to exhibit “counter-rotation” in that the gas motions in the inner parts of the galaxy indicated motions opposite those of the stars. This was interpreted there as a consequence of a warp in the disk coupled with the presence of a bar; this situation may exist for as well.
Macci[ò]{} (2006) tried to explain the origin of PRGs by accretion of cold intergalactic gas. They provide in their Fig. 4 plots of simulated velocity-position diagrams for gas and stars; the upper one, where the slit is aligned with the major axis of the galaxy, can be compared with our Figs. \[fig:AM\_rot\_130a\] and \[fig:AM\_rot\_130b\]. It seems that the presence of a stellar bar in could be producing the linearly-rising stellar rotation curve, whereas the rotation curve for the gas fits the simulation quite well.
Since none of our observations are of photometric-quality, we rely on parameters derived by @Resh06 to characterize the galaxy. In particular, we adopt their photometric disk parameters: a disk exponential scale length h(B)=5“.1$\pm$0”.3=3.8 kpc and their scaling to other bandpasses: h(B)/h(V)=1.18$\pm$0.11 and h(B)/h(R)=1.25$\pm$0.12. The R-band disk scale length is, therefore, 4.8$\pm$0.5 kpc. This is useful when comparing with properties of other galaxies or of model galaxies.
To compare with the rotational properties of other galaxies, we use the observations of edge-on galaxy disks from Kregel (2004) for the stellar kinematics and from Kregel & van der Kruit (2004) for the gas kinematics. Fig. 6 in Kregel shows that the stellar rotation curve can be almost linear with galacto-centric distance for about 1.5 disk scale lengths and this for galaxies earlier than Sbc. Note that this galaxy sample does not include barred galaxies, though Kregel mention that some do show boxy or peanut-shaped bulges. The gas in none of their galaxies (Kregel & van der Kruit 2004) rotates with as small a gradient with distance from the center as observed in .
It is also possible to compare both the imaged galaxy and its stellar kinematics with the diagnostic plots calculated by Bureau & Athanassoula (2005). Inspection of their Figs. 1 and 4 indicates that a good fit with could be obtained for an intermediate or strong bar viewed at least at 45$^{\circ}$ to the bar or even edge-on, and at a disk inclination of at least 80$^{\circ}$ to the line of sight. The conclusion is that does probably have a fairly strong bar that is almost side-on to our line of sight, and its disk is seen almost edge-on.
Another comparison for our rotation curve is with the collection of template rotation curves of Catinella (2006) who, however, studied normal galaxies, not PRGs. They normalize the rotation curves between 2 and 3 disk radii; applying this to , with the peak rotation derived from the curve, indicates that the galaxy should have an absolute I-band magnitude brighter than –23 mag. Indeed, using the photometry from @Resh06, with a measured M$_B \simeq$–21 mag and a color index (B-I)=2.06, the absolute I magnitude of is –23.06 mag. This confirms the assumption that, in analyzing the gaseous rotation curve along the major axis, it is a valid assumption to adopt the rotation pattern of a regular galaxy, not that of a PRG, since the presence of the polar ring does not affect significantly the kinematics of the galaxy.
HI vs. other kinematic indicators
---------------------------------
The HI in a number of PRGs, including , was studied by van Driel (2002) with the Parkes radio telescope. This observation produced a puzzling and troublesome result for ; van Driel reported the HI line at a heliocentric velocity of 11282$\pm$24 km sec$^{-1}$ with a full-width at half-maximum of the two-horned profile of 193 km sec$^{-1}$. Note that their data were taken with the Parkes multibeam system, which implies a beam width of 144 FWHM. The 12500 km sec$^{-1}$ bandwidth was centered at 10000 km sec$^{-1}$ and the channel separation was 6.6 km sec$^{-1}$.
If the HI would have been associated with , we would expect to find the neutral hydrogen line at a similar systemic velocity to that measured here, that in @Resh01, or that measured by @Resh06. We would also expect a much wider HI profile than quoted by van Driel (2002), since the H$\alpha$ kinematics indicate a width of $\sim$450 km sec$^{-1}$ along the major axis, as befitting a major galaxy given its bright absolute magnitude of M$_B$=–21.1 measured by @Resh06. The very wide Parkes beam implies that all three objects were included in the measurement, and probably many outlying HI clouds that may exist in this neighbourhood as well, but does not explain the velocity discrepancy since all three galaxies should have appeared on the red shoulder of the HI profile shown by van Driel
Another indication that something is wrong with the HI measurement comes from applying the Tully-Fisher relation to . @Co06 gives a Tully-Fisher diagram for PRGs in Fig. 2 of her paper; these galaxies seem to follow the T-F relation for spirals and S0 galaxies and it is worthwhile to check where fits in this diagram. Adopting the HI width given in van Driel (2002) indicates that should have an M$_B\simeq$–18 mag, completely different from the magnitude measured by @Resh06. Adopting a velocity width as measured by us albeit from the emission lines and not from the HI profile, namely 450 km sec$^{-1}$, yields the proper value of M$_B\simeq$–21 mag.
Irrespective of the explanation regarding the HI redshift discrepancy, it is possible that extended HI is present in the system. The possibility that such HI clouds or other gas-rich galaxies might be present is supported by our discovery of the H$\alpha$ knot (see below), and by the presence of a few low surface brightness (LSB) extended objects in the immediate vicinity. These resemble LSBs the nearby Universe that are often found to be very gas-rich. In addition, there are a few very blue star-like objects that stand out in comparisons of the Second Digitized Sky Survey images in different bands.
We do not have redshifts for these LSB objects but the fact that they are of similar sizes to the main galaxies in the group hints that they might be group members; such companions are seen in other groups as well (Grossi 2007) and could have interacted with in the past. We predict that once HI synthesis observations will be obtained for and its neighbours, for example with the ATNF, at least some of these candidates and in particular the H$\alpha$ knot discovered by us will prove to be actually gas-rich members of this group.
The H$\alpha$ knot
------------------
The H$\alpha$ knot reported above, which is $\sim$78 arcsec away to the NE from the galaxy center but almost at the same velocity, is in reality $\sim$630 kpc away in projected distance. Its detectable H$\alpha$ emission, combined with a lack of \[\], \[\] and only weak continuum emissions, argue that this is probably a metal-poor dwarf galaxy that belongs to the same group as . Such objects are known as ”HII galaxies” (Sargent & Searle 1970) since they show an HII region spectrum with negligible continuum and have considerable redshifts.
Our fitting procedure to the emission lines, used for the galaxies and for the ring, allows the derivation of an upper limit for the \[\] $\lambda$6583 flux that can be used to obtain an upper limit to the metal abundance. With a measured upper limit line ratio of $\rm log([\ionn{N}{ii}] \lambda6583/H\alpha) =
-1.46$ the metallicity upper limit is 12+log(O/H) $<$ 8.05 [@Den02]. The knot appears to be somewhat metal-poor, though we cannot set a definite upper limit on its metal abundance, and we have shown that it also shows a very low internal velocity dispersion as befits a dwarf galaxy.
Brosch et al. (2006) identified a considerable number of H$\alpha$-emitting knots in the neighbourhoods of a few star-forming galaxies qualified as “dwarfs” (M$_B\geq$–18) and located in some very under-dense regions of the nearby Universe. The study revealed these prospective neighbour galaxies through the presence of H$\alpha$ emission at or near the central galaxy redshift. It is possible that the knot found here is a similar type of object.
Differences between the two slit positions
------------------------------------------
Spectroscopy of the ring in NGC 4650A has been reported by Swaters & Rubin (2003). They found a ring rotation curve that seems to flatten out from the center to the North, but which is steadily increasing from the center to 20 arcsec on the South side and then flattens out. This is more pronounced for the stellar component of the ring than for its gaseous component. The galaxy itself, an S0 as most PRGs are, shows solid-body-like stellar rotation from the center to $\sim$15 arcsec out while the emission lines show a different pattern of constant velocity. This they interpret as due to the galaxy being devoid of gas while the line emission is produced only in the ring. Comparisons of the rotational properties of polar rings and of galaxy disks are valuable in understanding PRGs.
We return now to the appearance of the stellar rotation curves observed at $\rm PA = 140\degr$ and at $\rm PA = 27\degr$. These curves, derived from the [NaI]{} absorption lines, are very similar. They appear linear for a considerable distance and their velocity-distance gradients are $\sim$19 km sec$^{-1}$ kpc$^{-1}$.
We point the reader back to Fig. \[fig:AM\_2D\_27\] where the extent of the continuum that allows the detection and measurement of the [NaI]{}D lines is considerably narrower than that for $\rm PA
= 140\degr$. The discrepancy could be resolved by assuming that the absorption lines, and therefore most of the continuum, would not be produced by stars in the ring, as implicitly assumed in the previous sections, but by stars in the main galaxy, perhaps in a stellar disk or in a strong bar. The inclination can be derived from the axial ratio of the galaxy given in @Resh06: i$\simeq 63
^{\circ}$.
In this case, the angle difference between the two slit positions, $\rm 67\degr$, would explain the difference in the extent of the linear rotation curves at the two position angles as a combination of foreshortening and obscuration by the dust lane. The dust lane produces about one magnitude of extinction, as the intensity profiles along the slit in Fig. 4c of @Resh06 show. The sudden disappearance of the absorption lines only 15 SW of the dynamical center could be explained by the crossing of the dark lane by the slit at this position angle. The weak intensity of the underlying continuum of the H$\alpha$ line, plotted with a short-dashed line in the top panel of Fig. \[fig:AM\_rot\_27\], supports this interpretation.
The ring would, in this case, be composed mostly of gas, would be located between us and the disk with its dark lane, and would necessarily be much less massive than assumed by @Resh06. The emission lines measured within $\pm$5 of the kinematic centre (see e.g., Fig \[fig:AM\_rot\_27\]) would then be produced primarily in the disk, while those for $\rm PA = 27\degr$ but measured at a galacto-centric distance of more than 65 would originate in the ring.
The arguments presented above indicate that the model proposed by @Resh06 to explain as a major merger, with the donor galaxy being PGC 399718, might not fit the observations. We therefore propose another alternative, that the unsettled disk or ring around the galaxy was formed by accretion of cold gas from a cosmic filament, one of the possibilities accounting for ring galaxy formation put forward by @Co06. The presence of anomalous HI redshifts in the region, our discovery of an apparent dwarf HII galaxy in the group, and the circumstantial detection of large but low surface brightness galaxies in the immediate vicinity of , albeit lacking redshifts at present, argue in favour of this interpretation.
Perusing the large-scale structures identified in this region by Radburn-Smith (2006), specifically those in Panel 6 of their Figure 4, indicates that the location of , at $\rm l \simeq 341.02$, $\rm b\simeq -28.73$, corresponds to the tip of a galaxy filament extending out of the Zone of Avoidance. This might be a distant structure related to the Centaurus wall and the Norma and Pavo II clusters of galaxies at lower redshifts, through which intergalactic matter is accreted by the galaxy and forms the ring.
Models of cold gas accretion from cosmic filaments by Macci[ò]{} (2006) show how a ring galaxy, such as NGC 4650A or for that matter could be formed by such a process. Their simulations show that the accreted gas is not completely cold but rather at 15,000K due to its collapse within the gravitational potential of the filamentary structure. Moreover, they mention that some of the gas might also be shock-heated by the halo potential.
A similar process could take place in . There is no clear-cut evidence that the ring is closed or relaxed, or that it has a substantial stellar component. Its disturbed appearance at its SW end is more similar to that of an assemblage of diffuse gas clouds, not of a coherent and relaxed structure. The NE part is smaller and sharper; it is possible that accreted gas collides there with itself, becomes compressed and shocked, and reaches higher temperatures that produce the enhanced line emission. At this location the accreted gas could perhaps enter a circular or quasi-circular orbit.
An alternative could be that in the case we are indeed witnessing a merger with a gas-rich galaxy, which takes place in a polar configuration. This is, in a way, similar to the major merger scenario of @Resh06 with the exception that the ”donor” galaxy would now be the ring itself. The argument reducing the likelihood of this explanation is the lack of a significant stellar continuum from the ring, indicating its low mass.
The shape of the dark matter halo of
--------------------------------------
Considering the two gas rotation curves, the one along the galaxy’s major axis and the other along the ring’s major axis, one observation is in order. The two rotation curves derived from the emission lines extend a similar distance from the galaxy’s kinematic centre, are presumably in the same dark matter potential well if is indeed a PRG, yet show a completely different full amplitude. While the galaxy major axis rotation curve has a full end-to-end amplitude of $\sim$450 km sec$^{-1}$, that for the ring has a full amplitude of only $\sim$240 km sec$^{-1}$. The asymptotic rotation of the ring is slower than the asymptotic rotation of the galaxy.
The formation of PRGs has been studied by Bournaud & Combes (2003) via N-body simulations. They discussed, in particular, cases when both the galaxy disk and the ring contain gas. Their argument was that in such cases the polar ring must, by necessity, be wider than the galaxy. If this is not the case, the gas in the ring would interact with the gas in the disk and one of the components would join the other. Two orthogonal, or almost orthogonal gas rings, can coexist in the same galaxy only if they have different radii and do not cross each other. Such crossing presumably occurs in NGC 660, where both the disk and the ring contain gas; the N660 system is unstable and according to Bournaud & Combes did not have sufficient time to dissolve the ring since its formation.
The specific question of the DM halo shape in PRGs was studied by Iodice (2003). They explained that the ring material would move slower than the gas in the disk if the gravitational potential would be oblate, like the flattened disk galaxy. In this case the ring would be elliptical and would show a lower observed velocity than the disk at its outermost locations (see their Fig. 3). In the case of , since the ring and the galaxy appear to have the same size but the ring must be wider in order to avoid crossing the disk, a possible conclusion would be that the ring is elliptical with its major axis close to our line of sight to the object and its minor axis seen almost perpendicular to the disk. This way, the ring could indeed be larger than the galaxy, the gas in the ring and that in the galaxy would not cross, and the velocities at the apo-galactic ring locations would be slower than in the galaxy disk. In this case the outermost visible ring segments would correspond to locations near the ends of the minor axis and ring material should show there its highest orbital speed, larger than that of the galactic disk. As this is not observed, we conclude that the disk and the ring in are of similar sizes, their contents do cross, and the system is unstable.
With the additional kinematic information now available, could be considered a test case for DM gravitational potential tracing. The discussion of PRGs by @Co06 was based on the hope that PRGs would prove to be useful probes of the DM potential in which a galaxy and its polar ring find themselves. @Co06 found that the rings in observed PRGs show faster rotation than the maximal velocity observed in the host galaxy. The theoretical prediction is in the opposite direction to the observations, namely rings in PRGs devoid of DM halos or with spherical halos should be rotating slower than their galaxies. According to @Co06, this effect should be accentuated for flattened or oblate DM halos.
We find that fulfills the theoretical predictions for non-spherical oblate haloes in that the polar ring does rotate slower than its host galaxy. Any DM halo of , if it exists at all, would have to be flattened along the barred disk, but this configuration could not be stable on the long run because the ring would cross the disk. Resolving this possibility and deriving more constraints on the existence and shape of a possible DM halo for would require detailed modelling and further observations that are not within the scope of this paper.
Conclusions {#txt:summ}
===========
We presented observations obtained with SALT and RSS during their performance verification phase that emphasize the long-slit capabilities of the RSS for galaxy observations. We traced the stellar and gaseous rotation curves for the major axis of the galaxy and for the major axis of a polar ring-like feature almost perpendicular to the disk of the galaxy. We showed that, while the gas rotates regularly when sampled along the galaxy major axis, the stellar component shows rotation like a solid body, supporting an interpretation that this is an object with a strong bar viewed almost side-on.
The ionized gas rotation along the major axis of the ring was found to be much less regular than along the major axis of the galaxy and shows a somewhat shallower gradient with galacto-centric distance. The [NaI]{} stellar rotation from the $\sim$6.5 kpc ring segment where the lines are measurable shows a similar distance dependence to that seen along the galaxy’s major axis. The systemic velocity derived by us for differs from previously published values. We propose that the discrepancy of rotation curves along the two position angles can be resolved by recognizing that the absorption lines are probably produced only by the main galaxy or its bar, and not by the ring where only emission lines are produced.
We discovered a small H$\alpha$ knot at a projected distance of about 700 kpc from but at a similar velocity, which we interpret as a fourth member of this compact group of galaxies, presumably a metal-poor dwarf galaxy. The lack of continuum emission for this object while only the H$\alpha$ line is detected indicates that it might be forming stars for the first time. The low velocity dispersion measured from the knot indicates its low mass.
We argue that a more plausible explanation to the major merger scenario proposed by @Resh06 to explain could be the slow accretion of cold cosmic gas along a galaxy filament directed to the region. In the cold gas accretion case the flow is probably towards the galaxy from the South-West and becomes more compressed at the NE end of the polar ring feature. We point out that the kinematic properties we measured follow the theoretical predictions for PRGs in a dark matter halo that is not spherical, but is flattened along the plane of the galaxy.
Acknowledgments {#acknowledgments .unnumbered}
===============
This paper was written while NB was a sabbatical visitor at the South African Astronomical Observatory in Cape Town; NB is grateful for this opportunity offered by the SAAO management. We are grateful for the generous allocation of SALT observing time during the PV phase to complete this project. We acknowledge a private communication from Vladimir P. Reshetnikov concerning this galaxy. We acknowledge the use of products of the second Digitized Sky Survey produced at the Space Telescope Science Institute under U.S. Government grant NAG W-2166. The images are based on photographic data obtained using the UK Schmidt Telescope. The UK Schmidt Telescope was operated by the Royal Observatory Edinburgh, with funding from the UK Science and Engineering Research Council (later the UK Particle Physics and Astronomy Research Council), until 1988 June, and thereafter by the Anglo-Australian Observatory. The blue plates of the southern Sky Atlas and its Equatorial Extension (together known as the SERC-J), as well as the Equatorial Red (ER), and the Second Epoch \[red\] Survey (SES) were all taken with the UK Schmidt. An anonymous referee provided some insightful comments that improved the clarity of the presentation.
[99]{}
Allen at al. 1991, [MNRAS]{}, 248, 528
Baes M., et al., 2003, MNRAS, 343, 1081
Bekki K., 1998, ApJ, 499, 635
Bettoni D., Fasano G., Galletta G., 1990, AJ, 99, 1789
Bournaud F., Combes F., 2003, [A&A]{}, 401, 817
Brosch, N., Bar-Or, C., and Malka, D. 2006, [MNRAS]{}, 368, 864
Brosch, N. 1985, A&A, 153, 199
Brosch, N. 1987, Mercury, 16, 174
Buckley D. A. H., Swart G. P., Meiring J. G., 2006, SPIE, 6267,
Bureau M., Athanassoula E., 2005, ApJ, 626, 159
Burgh E. B., Nordsieck K. H., Kobulnicky H. A., Williams T. B., O’Donoghue D., Smith M. P., Percival J. W., 2003, SPIE, 4841, 1463
Combes, F. 2006, EAS, 20, 97
Catinella B., Giovanelli R., Haynes M. P., 2006, ApJ, 640, 751
Denicoló, G., Terlevich, R., & Terlevich, E. 2002, [MNRAS]{}, 330, 69
González-Delgado, R.M., Leitherer, C., & Heckman, T.M. 1999, [ApJS]{}, 125, 489
Grossi M., Disney M. J., Pritzl B. J., Knezek P. M., Gallagher J. S., Minchin R. F., Freeman K. C., 2007, MNRAS, 374, 107
Kregel M., van der Kruit P. C., 2004, MNRAS, 352, 787
Kregel M., van der Kruit P. C., Freeman K. C., 2004, MNRAS, 351, 1247
Hagen-Thorn, V.A., Shalyapina, L.V., Karataeva, G.M., Yakovleva, V.A., Moiseev, A.V., and Burenkov, A.N. 2005, Astron. Reports, 49, 958
Iodice E., Arnaboldi M., Bournaud F., Combes F., Sparke L. S., van Driel W., Capaccioli M., 2003, ApJ, 585, 730
Karataeva, G.M., Tikhonov, N.A., Galazutdinova, O.A., Hagen-Thorn, V.A., and Yakovleva, V.A. 2004, A&A, 421, 833
Kobulnicky H. A., Nordsieck K. H., Burgh E. B., Smith M. P., Percival J. W., Williams T. B., O’Donoghue D., 2003, SPIE, 4841, 1634
Kniazev, A.Y., Pustilnik, S.A., Grebel, E.K., Lee, H., & Pramskij, A.G. 2004, ApJS, 153, 429
Kniazev, A.Y., Grebel, E.K., Pustilnik, S.A., Pramskij, A.G., & Zucker, D. 2005, [AJ]{}, 130, 1558
Iodice E., Arnaboldi M., De Lucia G., Gallagher J. S., III, Sparke L. S., Freeman K. C., 2002, AJ, 123, 195
Iodice E., et al., 2006, ApJ, 643, 200
Leitherer, C. 1999, ApJS, 123, 3
Lynds, R. & Toomre, A. 1976, ApJ, 209, 382
Macci[ò]{} A. V., Moore B., Stadel J., 2006, ApJ, 636, L25
Mayya, Y.D. & Korchagin, V. 2001, ApSSS, 277, 339 (on-line revision in 2006)
Mazucca, L.M., Knapen, J.H., Regan, M.W. & Böker, T. 2001, in [*The Central kiloparsec of Starbursts and AGN*]{}, J.H. Knapen , eds. San Francisco: ASP, 573
Norris R. P., Kesteven M. J., Troup E. R., Allen D. A., Sramek R. A., 1990, ApJ, 359, 291
O’Donoghue D., et al., 2006, MNRAS, 372, 151
Persic M., Salucci P., Stel F., 1996, MNRAS, 281, 27 (erratum 1996, MNRAS, 283, 1102)
Radburn-Smith D. J., Lucey J. R., Woudt P. A., Kraan-Korteweg R. C., Watson F. G., 2006, MNRAS, 369, 1131
Reshetnikov V. P., Fa[ú]{}ndez-Abans M., de Oliveira-Abans M., 2001, [MNRAS]{}, 322, 689
Reshetnikov, V.P. 2004, A&A, 416, 889
Reshetnikov, V., Bournaud, F., Combes, F., Fa[ú]{}ndez-Abans, M., and de Oliveira-Abans, M.: 2006, [*Astron. Astroph.*]{} 446, 447.
Sargent W. L. W., Searle L., 1970, ApJ, 162, L155
Schweizer, F., Ford, W.K. Jr., Jederzejewsky, R. & Giovanelli, R. 1987, ApJ, 320, 454
Shane, W.W. 1980, A&A 82, 314
Swaters R. A., Rubin V. C., 2003, ApJ, 587, L23
van Driel W., Combes F., Arnaboldi M., Sparke L. S., 2002, A&A, 386, 140
Whitmore B. C., Lucas R. A., McElroy D. B., Steiman-Cameron T. Y., Sackett P. D., Olling R. P., 1990, AJ, 100, 1489
Zasov, A., Kniazev, A., Pustilnik, S., et al. 2000, A&AS, 144, 429
Woudt P. A., et al., 2006, MNRAS, 371, 1497
\[lastpage\]
[^1]: E-mail: [email protected] (NB); [email protected] (AYK); [email protected] (DB); [email protected] (DOD); [email protected] (YH); [email protected] (NL); [email protected] (ER); [email protected] (MS); [email protected] (PV); [email protected] (EBB); [email protected] (KN)
[^2]: Based on observations obtained with the Southern African Large Telescope (SALT).
[^3]: IRAF: the Image Reduction and Analysis Facility is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, In. (AURA) under cooperative agreement with the National Science Foundation (NSF).
[^4]: See http://www.salt.ac.za/partners-login/partners/data-analysis-software for more information.
[^5]: MIDAS is an acronym for the European Southern Observatory package – Munich Image Data Analysis System.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In the light of the recent LHC data on proton-proton and lead-lead collisions we examine the question of the multiplicity scaling of HBT radii in relativistic nuclei and particle interactions. Within the UrQMD transport approach we study a large variety of system sizes at different beam energies and extract the HBT radii. In the calculation, we find a good scaling of the radii as a function of charged particle multiplicity, if the change in the multiplicity is caused by a change of centrality at the same energy. However, the scaling is only approximate when the energy, $\sqrt{s}$, is changed and breaks down when comparing pp to AA reactions.'
author:
- Gunnar Gräf
- Qingfeng Li
- Marcus Bleicher
title: 'Examination of scaling of Hanbury-Brown–Twiss radii with charged particle multiplicity'
---
Introduction
============
The properties of strongly interacting matter are described by the theory of Quantum-Chromo-Dynamics (QCD). To explore the details of QCD matter under extreme conditions, one needs to compress and heat up QCD matter to regimes present microseconds after the Big Bang. Today these conditions can only be found in the interior of neutron stars or created in heavy-ion collisions at relativistic energies. Over the last decade the experimental programs at the SPS (e.g. with the NA49, CERES and NA50/NA60 experiments) and at RHIC (e.g. PHENIX, STAR, PHOBOS, and BRAHMS) have provided exciting pioneering data on the equation of state, the transport properties of the matter created and its spatial distributions [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez]. These programs are currently extended into a system size scan with NA61 at SPS and a systematic beam energy scan with the RHIC-BES initiative. In addition, at the high energy frontier unprecedented data from the Large Hadron Collider (LHC) for (high multiplicity) proton-proton and Pb+Pb reactions up to $\sqrt{s_{NN}}=7$ TeV has become available (see [@Aamodt:2011kd; @Aamodt:2011mr] for HBT related results). Particle correlations, i.e. Hanbury-Brown Twiss correlation (HBT) or femtoscopy allow to gain deeper insights into the emission patterns and coherence regions of the matter created [@Pratt:1984su; @Sinyukov:1989xz; @Hama:1987xv; @nucl-th/9901094]. One generally assumes that the observed HBT radii scale with the charged particle density (or number of participants) as the charged particle density should be a good proxy for the final state volume [@Lisa:2005dd]. However, the interferometry volume may not only depend on multiplicity, but also on the initial size of the colliding system [@Sinyukov:2011mw]. Indeed, one of the surprising LHC results concerns the scaling violation observed in pp reactions as compared to AA reactions at lower energies at the same charged particle density. In this paper, we want to explore the spatial structure of the source created in collisions of various heavy ions at different energies and centralities to shed light on the observed scaling violation when going from proton-proton to AA collisions at the LHC. Other investigations on the charged particle yield scaling can be found in [@Adamova:2002wi; @Adamova:2002ff; @Akkelin:2003kp; @Akkelin:2005ms]. Results for PbPb and pp reactions at the LHC within the same model can be cound in [@Graef:2012za; @Li:2012ta].
Model and HBT calculation
=========================
For the present study we employ the UrQMD [@Bass:1998ca; @Bleicher:1999xi] transport model in version 3.3 (for details of version 3.3 see [@Petersen:2008kb; @Petersen:2008dd]). The model can be downloaded from [@urqmd-webpage]. For earlier HBT results from UrQMD see [@Li:2006gp; @Li:2007yd; @Li:2008qm; @Li:2008ge]. UrQMD is a microscopic non-equilibrium transport model. It models the space-time evolution of nucleus-nucleus collisions from the beginning of the collision until the kinetic freeze-out. Particles are produced via hard collisions, string excitation and fragmentation and via resonance excitation and decay.\
For the calculation of the HBT radii we use the pion freeze-out distribution from UrQMD. Then we calculate the HBT correlation function by [@nucl-th/9901094; @Lisa:2005dd] $$C({\bf q},{\bf K}) = 1 + \int d^4x\ cos(q\cdot x)\ d(x,K)\ , \label{eqn:correlation}$$ where $C$ is the correlation function, $q$ is the four-momentum distance of the correlated particles, $K=(p_1+p_2)/2$ is the pair momentum, $x$ is the particle separation four-vector and $d$ is the normalized pion freeze-out separation distribution, which is an even function of $x$. For the analysis in this paper all values are taken in the pair longitudinal comoving system (LCMS). Since UrQMD generates a discrete set of freeze-out points, the integral in Eq. \[eqn:correlation\] is substituted by a sum.\
The HBT radii $R_{ij}$ are obtained by fitting the function $$C({\bf q},{\bf K}) = 1 + \lambda({\bf K}) exp \left [ - \sum_{i,j=o,s,l} q_iq_jR_{ij}^2({\bf K}) \right ]$$ to the calculated three-dimensional correlation functions. For the analysis in this paper the correlation functions are fitted over a range $|{\rm q_i}| < $ 800 MeV/c for proton-proton collisions, $|{\rm q_i}| < $ 300 MeV/c for carbon-carbon collisions and $|{\rm q_i}| < $ 150 MeV/c for all other collisions. The difference in the momentum ranges is motivated by the fact that the width of the peak in the correlation function gets broader for smaller systems. Thus, the fit range is bigger for proton-proton and carbon-carbon, than it is for lead-lead collisions.
Scaling of the HBT radii
========================
![(Color online) The three HBT radii $R_{out}$, $R_{side}$ and $R_{long}$ as a function of the charged particle multiplicity at midrapidity, $(dN_{ch}/d\eta)^{1/3}$ and fixed $k_T=300-400$ MeV. The lines with symbols are the simulation results. The gray triangles, the black circles, the red squares and the green crosses are for lead-lead collisions at $\sqrt{s}=$ 2760, 200, 130, 62.4 GeV (in the same order) at 0-5%, 5-20%, 20-50% and 50-80% centrality for the different points. The pink crosses are results for carbon-carbon at $\sqrt{s}=$ 200 GeV for the same centrality classes and the beige diamonds show results for various multiplicity classes from proton-proton collisions [@Graef:2012za]. Blue circles and brown squares depict results for central copper-copper events at $\sqrt{s}=$ 200 GeV and central lead-lead collisions at $E_{lab}=$ 158 GeV. The green stars are experimental results for central gold and lead collisions at $k_T =$ 300 GeV/c taken from [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez; @Aamodt:2011kd; @Aamodt:2011mr]. \[fig:radii\]](fig01.eps){width=".9\textwidth"}
Fig. \[fig:radii\] shows the three HBT radii $R_{out}$, $R_{side}$ and $R_{long}$ as a function of the charged particle multiplicity at midrapidity ($|\eta|<1.2$ for pp and $|\eta|<0.8$ for all other classes), $(dN_{ch}/d\eta)^{1/3}$ and fixed $k_T=300-400$ MeV. The lines with symbols are simulation results for lead-lead collisions at $\sqrt{s}=$ 2760, 200, 130, 62.4 GeV for 0-5%, 5-20%, 20-50% and 50-80% centrality, for carbon-carbon at $\sqrt{s}=$ 200 GeV in the same centrality classes, for proton-proton at $\sqrt{s}=$ 7 TeV with different $dN_{ch}/d\eta$ classes, for central copper-copper collisions at $\sqrt{s}=$ 200 GeV and for central lead-lead collisions at $E_{lab}=$ 158 GeV. The green stars are experimental results taken from [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez; @Aamodt:2011kd; @Aamodt:2011mr]. For nucleus-nucleus reactions one observes a rather linear scaling with $(dN_{ch}/d\eta)^{1/3}$. The scaling is very good if the change in $(dN_{ch}/d\eta)^{1/3}$ is caused by a change of centrality at a fixed energy. However, a small offset on the order of 2 fm - 3 fm is visible for different system sizes, if the radii are extrapolated to $N_{ch}\rightarrow 0$. This is expected due to the finite size of the nuclei in AA reactions [@Sinyukov:2011mw]. In contrast, increasing the center-of-mass energy leads to a reduction of the radii at a given fixed $N_{ch}$-bin. The scaling of the source size with $(dN_{ch}/d\eta)^{1/3}$ for different centralities is a hint that the underlying physics, e.g. pion production via resonance decay versus production via string fragmentation, is nearly unchanged by changes in the collision geometry. A change in $\sqrt{s}$ on the other hand results not only in different weights of the production mechanisms, but also in changed expansion dynamics towards a more violent expansion with increased energy. Qualitatively, one expects a scaling of the length of homogeneity as $R =
R_{geom} / \sqrt{1 + \langle v_\perp^2 \rangle m_\perp/2T}$ [@Sinyukov:2011mw; @Akkelin:1995gh], where $R_{geom}$ is the geometric size of the collision region, $v_\perp$ is the transverse flow velocity and T is the freeze-out temperature. I.e. the increase in transverse flow leads to a decrease of the observed radii with increasing energy as observed in the model. This combination leads to a deviation from the $( dN_{ch}/d\eta)^{1/3}$ scaling of the HBT radii. The proton-proton calculation (and the data) show significantly smaller radii and a different slope from what is expected from nucleus-nucleus results. This behaviour is attributed to the strongly different particle production mechanisms in AA and pp. I.e., bulk emission vs. string/jet dominated emission which is also in line with the theoretically observed dependence of the HBT radii on the formation time of the hadrons from the jet fragmentation and string decay [@Graef:2012za].\
Since the $K_\perp$ dependence of the HBT radii tells us much about the expansion of the source [@Pratt:1984su; @Hama:1987xv], let us next investigate how a variation of $dN_{ch}/d\eta$ is reflected in the differential HBT radii as recently discussed in [@Truesdale:2012zz]. Fig. \[fig:kt\] shows the three HBT radii $R_{out}$, $R_{side}$ and $R_{long}$ at fixed charged particle multiplicity at midrapidity as a function of $k_T$. The shown calculations are chosen so that they fall roughly into two $\left <
dN_{ch}/d\eta \right >$ classes. The first class contains calculations with $\left < dN_{ch}/d\eta \right > \approx$ 600 (exact values are 670 for Pb+Pb at $\sqrt{s}=2760$ GeV, 20-50% centrality and 665, 595 and 509 for Pb+Pb at $\sqrt{s}=$ 200, 130, 62.4 GeV, 0-5% centrality). The second class contains calculations for $\left < dN_{ch}/d\eta \right > \approx$ 25 (exact values are 23 for C+C at $\sqrt{s} = 200$ GeV, 0-5% centrality and 32, 28 and 23 for Pb+Pb at $\sqrt{s}=$ 200, 130, 62.4 GeV and 50-80% centrality).
A very similar slope in K$_\perp$ is observed for all UrQMD results. This leads to the conclusion that the observed HBT radii dependence on the radial flow in the model is weaker than observed in the data. The shift in magnitude of the radii is related to the magnitude differences already observed in Fig. \[fig:radii\] that are mainly dominated by geometry and $\sqrt{s}$ effects.
![(Color online) The $k_\perp$ dependence of $R_{out}$, $R_{side}$ and $R_{long}$. The black dots are calculations at $\sqrt{s}=$ 2760 GeV and 20-50% centrality, the red squares, the green crosses and the pink crosses are lead-lead for 0-5% centrality at $\sqrt{s}=$ 200, 130, 62.4 GeV. They have $\left < dN_{ch}/d\eta
\right > \approx$ 670, 665, 595, 509. The other presented calculations are carbon-carbon at $\sqrt{s}=$ 200 GeV for 0-5% centrality (blue circles) and lead-lead at $\sqrt{s}=$ 200, 130, 62.4 GeV (beige diamonds, grey triangles, blue triangles) all for 50-80% centrality. These collisions have $\left <
dN_{ch}/d\eta \right > \approx$ 23, 32, 28, 23. The green stars represent ALICE lead-lead data for central collisions at $\sqrt{s}=$ 2760 GeV [@Aamodt:2011mr]. The blue diamonds are experimental results for central gold-gold collisions at $\sqrt{s}=$ 200 GeV from the STAR collaborator. [@Adams:2004yc] \[fig:kt\]](fig02.eps){width=".8\textwidth"}
Volume and freeze-out time
==========================
![(Color online) Two definitions of the volume of homogeneity as a function of energy for various systems. In the left plot the volume is defined as $R_{out} R_{side} R_{long}$ and in the right plot the volume is defined as $R_{side}^2 R_{long}$. The gray triangles, black circles, red squares and green crosses depict UrQMD results for lead-lead collisions at (in this order) $\sqrt{s}=$2760, 200, 130, 62.4 GeV for the centralities 0-5%, 5-20%, 20-40%, 40-80%. The pink crosses are carbon-carbon calculations at $\sqrt{s}=$ 200 GeV for the same centralities, the blue circles are central copper-copper collisions at $\sqrt{s}=$ 200 GeV and the brown squares are central lead-lead collisions at $E_{lab}=$ 158 AGeV. The beige diamonds depict proton-proton results at $\sqrt{s}=$ 7 TeV for different $(dN_{ch}/d\eta)^{1/3}$ classes. The green stars show experimental results taken from [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez; @Aamodt:2011kd; @Aamodt:2011mr]. \[fig:volume\]](fig03.eps){width=".85\textwidth"}
Next, let us investigate the energy and system size dependence of the homogeneity volume. Fig. \[fig:volume\] shows the volume of homogeneity as a function of $dN_{ch}/d\eta$ for various systems. Lead-lead calculations are shown for $\sqrt{s}=$ 2760, 200, 130, 62.4 GeV (grey triangles, black circles, red squares, green crosses) in the centrality classes 0-5%, 5-20%, 20-40% and 40-80%. The pink crosses show $\sqrt{s}=$ 200 GeV carbon-carbon results for the same centralities, and the beige diamonds represent proton-proton calculations at $\sqrt{s}=$ 7 TeV for different $dN_{ch}/d\eta$ bins. Blue circles and brown squares depict results for central copper-copper events at $\sqrt{s}=$ 200 GeV and central lead-lead events at $E_{lab}=158$ GeV. These results are compared to experimental data [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez; @Aamodt:2011kd; @Aamodt:2011mr] which is represented by green stars. In line with the experimental data, a strong increase in the volume proportional to the charged particle multiplicity is observed. A good agreement between experiment and theory is observed for the quantity $R_{side}^2R_{long}$ while the experimental results for $R_{out} R_{side} R_{long}$ are slightly overestimated. This is due to a too large $R_{out}$ in the calculations. The overestimation of $R_{out}$ is common for hadronic cascade models and can be explained by a lack of pressure in the early stage of the heavy ion collision [@Li:2007yd; @Pratt:2009hu]. While the volume of the homogeneity region for each individual energy scales very well with $dN_{ch}/d\eta$ Fig. \[fig:volume\] shows a steeper slope with decreasing energy. The calculations also hint to an offset for AA reactions on the order of $25$ fm$^3$ ($R_{side}^2 R_{long}$) and 50 fm$^3$ ($R_{out} R_{side} R_{long}$).\
![(Color online) The freeze-out time as a function of energy for various systems. The gray triangles, black circles, red squares and green crosses depict UrQMD results for lead-lead collisions at (in this order) $\sqrt{s}=$2760, 200, 130 62.4 GeV for the centralities 0-5%, 5-20%, 20-40%, 40-80%. The pink crosses are carbon-carbon calculations at $\sqrt{s}=$ 200 GeV for the same centralities, the blue circles are central copper-copper collisions at $\sqrt{s}=$ 200 GeV and the brown squares are central lead-lead collisions at $E_{lab}=$ 158 AGeV. The beige diamonds depict proton-proton results at $\sqrt{s}=$ 7 TeV for different $(dN_{ch}/d\eta)^{1/3}$ classes. The green stars show experimental results taken from [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez; @Aamodt:2011mr]. \[fig:tau\]](fig04.eps){width=".7\textwidth"}
Finally, we explore the apparent freeze-out times $\tau_f$. The results are obtained by fitting the hydrodynamically motivated Eq. \[eq:tau\] [@Aamodt:2011mr; @Makhlin:1987gm] to the $k_\perp$ dependence of $R_{long}$ in the interval $K_\perp=$ 200-800 MeV/c. For this purpose the pion freeze-out temperature is assumed to be $T=$ 120 MeV. $$R_{long}^2=\tau_f^2 \frac{T}{m_\perp} \frac{K_2(m_\perp/T)}{K_1(m_\perp/T)},
\label{eq:tau}$$ where $m_\perp = \sqrt{m_\pi^2+k_\perp^2}$ and $K_i$ are the integer order modified Bessel functions. Fig. \[fig:tau\] shows the freeze-out time as a function of $dN_{ch}/d\eta$ for various systems. The grey triangles, the black circles, the red squares and the green crosses are calculations of lead-lead collisions at $\sqrt{s}=$ 2760, 200, 130, 62.4 GeV (in the same order) for the centralities 0-5%, 5-20%, 20-40%, 40-80%. The pink crosses are carbon-carbon collisions at $\sqrt{s}=$ 200 GeV for the same centralities. The blue circles are calculations for central copper-copper collisions at $\sqrt{s}=$ 200 GeV and central lead-lead collisions at $E_{lab}=$ 158 AGeV. Experimental results [@Adams:2004yc; @Lisa:2000no; @Alt:2007uj; @Afanasiev:2002mx; @Adamova:2002wi; @Abelev:2009tp; @Back:2004ug; @Back:2005hs; @Back:2002wb; @Abelev:2008ez; @Aamodt:2011mr] are depicted by green stars. As for all the other observables, there is scaling for each energy individually. As anticipated from the calculations of $R_{long}$ the decoupling time $\tau_f$ increases with decreasing energy. This confirms the idea of a shorter decoupling time with increased energy. The offset in $\tau_f$ for $dN_{ch}/d\eta
\rightarrow 0 $ seems to hint towards a minimal decoupling time $\tau^{min}_f
\sim 4-8 $fm/c in AA reactions and $\tau_f^{min} < 2$ fm/c in pp.
Summary and outlook
===================
In the light of recent LHC data on pp and AA collisions, which indicate a modification of the multiplicity scaling of the HBT radii, we have explored the $N_{ch}$ scaling for a large variety of systems and energies. We find good scaling of the radii with $dN_{ch}/d\eta$ within a given system and energy. While the radii decrease slightly with increasing beam energy, they have a similar slope when plotted versus $(dN_{ch}/d\eta)^{1/3}$ at all energies. When analyzing the freeze-out volume versus $dN_{ch}/d\eta$ the increasing steepness of the slope for decreasing energies becomes visible. For all observables the scaling of the results for pp collisions differ strongly from the nucleus-nucleus results. We relate this observation to the different particle emission patterns (bulk vs. strings) in AA and pp.
Acknowledgements {#acknowledgements .unnumbered}
================
This work was supported by the Helmholtz International Center for FAIR within the framework of the LOEWE program launched by the State of Hesse, GSI, and BMBF. G.G. thanks the Helmholtz Research School for Quark Matter Studies (H-QM) for support. Q.L. thanks the financial support by the key project of the Ministry of Education (No. 209053), the NNSF (Nos. 10905021, 10979023), the Zhejiang Provincial NSF (No. Y6090210), and the Qian-Jiang Talents Project of Zhejiang Province (No. 2010R10102) of China.
References {#references .unnumbered}
==========
[36]{}
J. Adams [*et al.*]{} \[STAR Collaboration\], Phys. Rev. C [**71**]{}, 044906 (2005). M. A. Lisa, et al., E895 Collaboration, Phys. Rev. Lett. 84 (2000) 2798–2802.
C. Alt, et al., NA49 Collaboration, Phys. Rev. C77 (2008) 064908.
S. V. Afanasiev, et al., NA49 Collaboration, Phys. Rev. C66 (2002) 054902.
D. Adamová, et al., CERES Collaboration, Nucl. Phys. A714 (2003) 124–144.
B. I. Abelev, et al., STAR Collaboration, Phys. Rev. C80 (2009) 024905.
B. B. Back, et al., PHOBOS Collaboration, Phys. Rev. C73 (2006) 031901.
B. B. Back, et al., PHOBOS Collaboration, Phys. Rev. C74 (2006) 021901.
B. B. Back, et al., PHOBOS Collaboration, Phys. Rev. Lett. 91 (2003) 052303.
B. I. Abelev, et al., STAR Collaboration, Phys. Rev. C79 (2009) 034909.
K. Aamodt [*et al.*]{} \[ ALICE Collaboration \], \[arXiv:1101.3665 \[hep-ex\]\].
K. Aamodt [*et al.*]{} \[ALICE Collaboration\], Phys. Lett. B [**696**]{}, 328 (2011). S. Pratt, Phys. Rev. Lett. [**53**]{}, 1219 (1984). Y. .M. Sinyukov, Nucl. Phys. A [**498**]{}, 151C (1989). Y. Hama and S. S. Padula, Phys. Rev. D [**37**]{}, 3237 (1988). U. A. Wiedemann and U. W. Heinz, Phys. Rept. [**319**]{}, 145 (1999). M. A. Lisa, S. Pratt, R. Soltz and U. Wiedemann, Ann. Rev. Nucl. Part. Sci. [**55**]{}, 357 (2005). Y. .M. Sinyukov and I. .A. Karpenko, Phys. Part. Nucl. Lett. [**8**]{}, 896 (2011). S. V. Akkelin and Y. .M. Sinyukov, Phys. Lett. B [**356**]{}, 525 (1995).
D. Adamova [*et al.*]{} \[CERES Collaboration\], Phys. Rev. Lett. [**90**]{}, 022301 (2003). S. V. Akkelin and Y. .M. Sinyukov, nucl-th/0310036. S. V. Akkelin and Y. .M. Sinyukov, Phys. Rev. C [**73**]{}, 034908 (2006). G. Graef, Q. Li and M. Bleicher, arXiv:1203.4421 \[nucl-th\]. Q. Li, G. Graef and M. Bleicher, arXiv:1203.4104 \[nucl-th\]. S. A. Bass, M. Belkacem, M. Bleicher, M. Brandstetter, L. Bravina, C. Ernst, L. Gerland, M. Hofmann [*et al.*]{}, Prog. Part. Nucl. Phys. [**41**]{}, 255-369 (1998). M. Bleicher, E. Zabrodin, C. Spieles, S. A. Bass, C. Ernst, S. Soff, L. Bravina, M. Belkacem [*et al.*]{}, J. Phys. G [**G25**]{}, 1859-1896 (1999). H. Petersen, M. Bleicher, S. A. Bass, H. Stocker, \[arXiv:0805.0567 \[hep-ph\]\].
H. Petersen, J. Steinheimer, G. Burau, M. Bleicher, H. Stocker, Phys. Rev. [**C78**]{}, 044901 (2008). Download the most recent UrQMD source code from http://urqmd.org/
Q. Li, M. Bleicher and H. Stoecker, Phys. Rev. C [**73**]{}, 064908 (2006). Q. Li, M. Bleicher and H. Stocker, Phys. Lett. B [**659**]{}, 525 (2008). Q. Li, J. Steinheimer, H. Petersen, M. Bleicher and H. Stocker, Phys. Lett. B [**674**]{}, 111 (2009). Q. Li, M. Bleicher and H. Stocker, Phys. Lett. B [**663**]{}, 395 (2008). D. Truesdale and T. J. Humanic, J. Phys. G G [**39**]{}, 015011 (2012). S. Pratt, Nucl. Phys. A [**830**]{}, 51C (2009). A. N. Makhlin, Y. .M. Sinyukov, Z. Phys. [**C39**]{}, 69 (1988).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Let $G$ be an unmixed bipartite graph of dimension $d-1$. Assume that $K_{n,n}$, with $n\ge 2$, is a maximal complete bipartite subgraph of $G$ of minimum dimension. Then $G$ is Cohen-Macaulay in codimension $d-n+1$. This generalizes a characterization of Cohen-Macaulay bipartite graphs by Herzog and Hibi and a result of Cook and Nagel on unmixed Buchsbaum graphs. Furthermore, we show that any unmixed bipartite graph $G$ which is Cohen-Macaulay in codimension $t$, is obtained from a Cohen-Macaulay graph by replacing certain edges of $G$ with complete bipartite graphs. We provide some examples.'
address:
- |
Hassan Haghighi\
Department of Mathematics, K. N. Toosi University of Technology, Tehran, Iran.
- |
Siamak Yassemi\
School of Mathematics, Statistics & Computer Science, University of Tehran, and School of Mathematics, Institute for Research in Fundamental Sciences (IPM), Tehran, Iran.
- |
Rahim Zaare Nahandi\
School of Mathematics, Statistics & Computer Science, University of Tehran, Tehran, Iran.
author:
- Hassan Haghighi
- Siamak Yassemi
- Rahim Zaare Nahandi
title: 'Cohen-Macaulay-ness in codimension for bipartite graphs'
---
[^1]
Introduction
============
Cohen-Macaulay simplicial complexes are among central research topics in combinatorial commutative algebra. While characterization of such complexes is a far reaching problem, one appeals to study specific families of Cohen-Macaulay simplicial complexes. Flag complexes are among important families of complexes recommended to study [@S1995 page 100]. However, it is known that a simplicial complex is Cohen-Macaulay if and only if its barycentric subdivision is a Cohen-Macaulay flag complex. Therefore, a characterization of Cohen-Macaulay flag complexes is equivalent to a characterization of Cohen-Macaulay simplicial complexes. Nevertheless, after all, the ideal of a flag complex is generated by quadratic square-free monomials, which are simpler compared with arbitrary square-free monomial ideals. Furthermore, it seems that, expressing many combinatorial properties in terms of graphs are more convenient. As some evidences, the characterization of unmixed bipartite graphs by Villarreal [@V2007] and Cohen-Macaulay bipartite graphs by Herzog and Hibi [@HH2005] are well expressed in terms of graphs.
On the other hand, in the hierarchy of families of graphs with respect to Cohen-Macaulay property, Buchsbaum complexes appear right after Cohen-Macaulay ones. Unmixed bipartite Buchsbaum graphs were characterized by Cook and Nagel [@CN2012] (also by the authors [@HYZ2010]). Natural families of graphs in this hierarchy are bipartite ${\mbox{CM}\,}_t$ graphs, i.e., graphs that their independence complexes are pure and Cohen-Macaulay in codimension $t$. The concept of ${\mbox{CM}\,}_t$ simplicial complexes were introduced in [@HYZ2012] which is the pure version of simplicial complexes Cohen-Macaulay in codimension $t$ studied by Miller, Novik and Swartz [@MNS2010]. In this note, we give characterizations of unmixed bipartite ${\mbox{CM}\,}_t$ graphs in terms of its dimension and the minimum dimension of its maximal nontrivial complete bipartite subgraphs. Cook and Nagel showed that the only non-Cohen-Macaulay unmixed bipartite graphs are complete bipartite graphs [@CN2012 Theorem 4.10] and [@HYZ2010 Theorem 1.3]. Our results are generalizations of this fact to unmixed bipartite graphs which are Cohen-Macaulay in arbitrary codimension. In the next section we gather necessary definitions and known results to be used in the rest of the paper. In Section 3 we improve some results on joins of simplicial complexes and disjoint unions of graphs with respect to the ${\mbox{CM}\,}_t$ property. Section 4 is devoted to two characterizations of bipartite ${\mbox{CM}\,}_t$ graphs and some examples.
Preliminaries
=============
For basic definitions and general facts on simplicial complexes we refer to the book of Stanley [@S1995]. By a complex we will always mean a simplicial complex. Let $G = (V, E)$ be a simple graph with vertex set $V$ and edge set $E$. The [*inclusive neighborhood*]{} of $v\in V$ is the set $N[v]$ consisting of $v$ and vertices adjacent to $v$ in $G$. The [*independence complex*]{} of $G = (V, E)$ is the complex Ind$(G)$ with vertex set $V$ and with faces consisting of independent sets of vertices of $G$, i.e., sets of vertices of $G$ where no two elements of them are adjacent. These complexes are called [*flag complexes*]{}, and their Stanley-Reisner ideal is generated by quadratic square-free monomials. By [*dimension*]{} of a graph $G$ we mean the dimension of the complex Ind$(G)$. A graph $G$ is said to be [*unmixed*]{} if Ind$(G)$ is pure.\
For an integer $t \ge 0$, a complex $\Delta$ is called ${\mbox{CM}\,}_t$ if it is pure and for every face $F\in\Delta$ with $\#(F) \ge t$, ${\mbox{link}\,}_{\Delta}(F)$ is Cohen-Macaulay. This is the same as pure complexes which are Cohen-Macaulay in codimension $t$. Accordingly, ${\mbox{CM}\,}_0$ and ${\mbox{CM}\,}_1$ complexes are precisely Cohen-Macaulay and Buchsbaum complexes, respectively. Clearly, a ${\mbox{CM}\,}_t$ complex is ${\mbox{CM}\,}_r$ for all $r\ge t$ and a complex of dimension $d-1$ is always ${\mbox{CM}\,}_{d-1}$. One uses the convention that for $t<0$, ${\rm CM}_t$ would mean ${\rm CM}_0$. A graph $G$ is called ${\mbox{CM}\,}_t$ if Ind$(G)$ is ${\mbox{CM}\,}_t$. A basic tool for checking ${\mbox{CM}\,}_t$ property of complexes is the following lemma.
\[basiclemma\]([@HYZ2012 Lemma 2.3]) Let $t\ge 1$ and let $\Delta$ be a nonempty complex. Then the following are equivalent:
- $\Delta$ is a ${\rm CM}_t$ complex.
- $\Delta$ is pure and ${\mbox{link}\,}_\Delta(v)$ is ${\rm CM}_{t-1}$ for every vertex $v\in\Delta$.
By the straightforward identity ${\mbox{link}\,}_{{\rm Ind}(G)}(v) = {\rm
Ind}(G\setminus N[v])$, the counter-part of this lemma for graphs would be the following:
\[lemmagraphs\] Let $t\ge 1$ and let $G$ be a graph. Then the following are equivalent:
- $G$ is a ${\rm CM}_t$ graph.
- $G$ is unmixed and $G\setminus N[v]$ is a ${\rm CM}_{t-1}$ graph for every vertex $v \in G$.
We recall some basic relevant facts on bipartite graphs. A graph $G
= (V,E)$ is called [*bipartite*]{} if $V$ is a disjoint union of a partition $V_1$ and $V_2$ and $E \subset V_1 \times V_2$. If $\#(V_1) = m$ and $\#(V_2) = n$ and $E = V_1 \times V_2$, then $G$ is the [*complete*]{} bipartite graph $K_{m,n}$. We will be interested in unmixed complete bipartite graphs $K_{n,n}$.\
Unmixed bipartite graphs are characterized by Villarreal in the following result.
\[Villarreal\] [@V2007 Theorem 1.1] Let $G$ be a bipartite graph without isolated vertex. Then $G$ is unmixed if and only if there is a partition $V_1 = \{x_1, \cdots, x_n\}$ and $V_2 = \{y_1, \cdots,
y_n\}$ of vertices of $G$ such that
- $x_iy_i$ is an edge in $G$ for $1\le i \le n$ and
- If $x_iy_j$ and $x_jy_k$ are edges in $G$, for some distinct $i$, $j$ and $k$, then $x_iy_k$ is an edge in $G$.
In this case, such a partition and ordering is called a [*pure order*]{} of $G$. The edges $x_iy_i$, $i =1,\cdots, n$ are called a [*perfect matching*]{} edges of $G$. A pure order is said to have a [*cross*]{} if, for some $i\ne j$, $x_iy_j$ and $x_jy_i$ are both edges in $G$. Otherwise, the order is called [*cross-free*]{} (see [@CN2012 $\S$ 4]). For unmixed bipartite graphs, being cross-free is independent of an ordering of vertices of $G$. More precisely, if $G$ has a cross in some pure ordering, it has a cross in every pure ordering [@CN2012 Lemma 4.5].\
An immediate consequence of Theorem \[Villarreal\] is the following useful lemma.
\[compneighb\] Let $G$ be an unmixed bipartite graph with pure order of vertices $(\{x_1,\cdots, x_d\},
\{y_1,\cdots, y_d\})$ and let $K_{n,n}$ be a complete bipartite subgraph of $G$ on $(\{x_{i_1},\cdots, x_{i_n}\}, \{y_{i_1},\cdots,
y_{i_n}\})$.
- If $x_jy_{i_k}$ is an edge in $G$ for some $j$ and $k$, then $x_jy_{i_l}$ is an edge in $G$ for all $l=1,\cdots, n$.
- If $x_{i_k}y_{j}$ is an edge in $G$ for some $k$ and $j$, then $x_{i_l}y_j$ is an edge in $G$ for all $l=1,\cdots, n$.
The assertion $(i)$ is immediate by Theorem \[Villarreal\] because $x_{i_k}y_{i_l}$ is an edge in $K_{n,n} \subset G$ for all $l=1,\cdots, n$. Also $(ii)$ follows because $x_{i_l}y_{i_k}$ is an edge in $K_{n,n} \subset G$ for all $l=1,\cdots, n$.
There are also at least two nice characterization of Cohen-Macaulay bipartite graphs.
\[HerHib\] [@HH2005 Theorem 3.4] Let $G$ be a bipartite graph without isolated vertices. Then $G$ is Cohen-Macaulay if and only if there is a pure ordering $V_1 = \{x_1, \cdots, x_n\}$ and $V_2 =
\{y_1, \cdots, y_n\}$ of vertices of $G$ such that $x_iy_j$ being in $G$ implies $i \le j$.
The ordering in Theorem \[HerHib\] is called a [*Macaulay order*]{} of vertices of $G$.
\[cookNag\] [@CN2012 Proposition 4.8] Let $G$ be a bipartite graph. Then $G$ is Cohen-Macaulay if and only if $G$ has a cross-free pure order.
Bipartite Buchsbaum graphs are also classified. First recall that a complex is Buchsbaum if and only if it is pure and the link of each vertex is Cohen-Macaulay [@Sch1981]. Thus, a graph is Buchsbaum if and only if $G$ is unmixed and for each vertex $v\in G$, $G\setminus N[v]$ is Cohen-Macaulay. For bipartite graphs there is a sharper result. Complete bipartite graphs are well-known to be Buchsbaum (e.g., see [@ZZ2004 Proposition 2.3]). But indeed, the converse is also true.
\[HYZBuch\] (see [@CN2012 Theorem 4.10] or [@HYZ2010 Theorem1.3]) Let $G$ be a bipartite graph. Then $G$ is Buchsbaum if and only if $G$ is a complete bipartite graph $K_{n,n}$ for some $n\ge 2$, or $G$ is Cohen-Macaulay.
Joins of ${\mbox{CM}\,}_t$ complexes and disjoint unions of ${\mbox{CM}\,}_t$ graphs
====================================================================================
It is known that the join of two complexes is Cohen-Macaulay if and only if they are both Cohen-Macaulay (see [@S1985] and [@F1988]). If $\Delta$ is a ${\mbox{CM}\,}_r$ complex of dimension $d-1$ and $\Delta'$ is a ${\mbox{CM}\,}_{r'}$ complex of dimension $d'-1$, then their join $\Delta \ast \Delta'$ is a ${\mbox{CM}\,}_t$ complex where $t={\rm
max}\{d+r', d'+r\}$ [@HYZ2012 Proposition 2.10]. However, if one of the complexes is Cohen-Macaulay, this result could be strengthened. Below we combine this with relevant known results.
\[join\] Let $\Delta$ and $\Delta'$ be two complexes of dimensions $d-1$ and $d'-1$, respectively. Then
- The join complex $\Delta \ast \Delta'$ is Cohen-Macaulay if and only if both $\Delta$ and $\Delta'$ are so.
- If $\Delta$ is Cohen-Macaulay and $\Delta'$ is ${\mbox{CM}\,}_{r'}$ for some $r'\ge 1$, then $\Delta \ast \Delta'$ is ${\mbox{CM}\,}_{d+r'}$ (independent of $d'$). This is sharp, i.e., if $\Delta'$ is not ${\mbox{CM}\,}_{r'-1}$, then $\Delta \ast \Delta'$ is not ${\mbox{CM}\,}_{d+r'-1}$. In particular, a cone on $\Delta'$ is ${\mbox{CM}\,}_{r'+1}$.
- If $\Delta$ is ${\mbox{CM}\,}_r$ and $\Delta'$ is ${\mbox{CM}\,}_{r'}$ for some $r, r' \ge 1$, then $\Delta \ast \Delta'$ is ${\mbox{CM}\,}_t$ where $t={\rm max}\{d+r', d'+r\}$. Conversely, if $\Delta \ast \Delta'$ is ${\mbox{CM}\,}_t$, then $\Delta$ is ${\mbox{CM}\,}_{t-d'}$ and $\Delta'$ is ${\mbox{CM}\,}_{t-d}$.
The statement in $(i)$ is proved by Sava [@S1985] and Fröberg [@F1988]. The assertion $(iii)$ is proved in [@HYZ2012 Theorem 2.10]. We prove $(ii)$ using induction on $d+r' \ge 2$. Let $d+r'=2$, i.e., $d=1$ and $r'=1$. Then $\Delta = \{v\}$ is a singleton. Thus ${\mbox{link}\,}_{\Delta \ast \Delta'}(v) = \Delta'$, which is ${\mbox{CM}\,}_1$. For $v \in \Delta'$, ${\mbox{link}\,}_{\Delta \ast \Delta'}(v) = \Delta
\ast {\mbox{link}\,}_{\Delta'}(v)$, which is Cohen-Macaulay by $(i)$. Thus by Lemma \[basiclemma\], $\Delta \ast \Delta'$ is ${\mbox{CM}\,}_2$. Now let $d+r'\ge 2$. Let $v\in \Delta$. Then, ${\mbox{link}\,}_{\Delta \ast \Delta'}(v)
= {\mbox{link}\,}_\Delta (v)\ast \Delta'$. But ${\mbox{link}\,}_\Delta (v)$ is Cohen-Macaulay of dimension less than $d-1$, and $\Delta'$ is ${\mbox{CM}\,}_{r'}$. Thus by induction hypothesis ${\mbox{link}\,}_{\Delta\ast \Delta'}
(v)$ is ${\mbox{CM}\,}_{d-1+r'}$. If $v\in \Delta'$, then ${\mbox{link}\,}_{\Delta \ast
\Delta'}(v) = \Delta' \ast {\mbox{link}\,}_{\Delta'}(v)$. But ${\mbox{link}\,}_{\Delta'}(v)$ is ${\mbox{CM}\,}_{r'-1}$ and hence ${\mbox{link}\,}_{\Delta\ast \Delta'} (v)$ is again ${\mbox{CM}\,}_{d+r'-1}$. Therefore, $\Delta \ast \Delta'$ is ${\mbox{CM}\,}_{d+r'}$. To prove that this result is sharp, proceed by induction on $d\ge 1$. Indeed, in this case, for any $v\in \Delta$, ${\mbox{link}\,}_\Delta(v)$ has dimension less than $d-1$ and hence by induction hypothesis, ${\mbox{link}\,}_{\Delta \ast \Delta'}(v) = {\mbox{link}\,}_\Delta(v) \ast \Delta$ is not ${\mbox{CM}\,}_{d+r'-2}$. Therefore, $\Delta \ast \Delta'$ is not ${\mbox{CM}\,}_{d+r'-1}$
Let $G\sqcup G'$ denote the disjoint union of graphs $G$ and $G'$. By the fact that Ind$(G\sqcup G') =$ Ind$(G)\ast$Ind$(G')$, the counter-part of Theorem \[join\] for graphs will be the following.
\[joingraphs\] Let $G$ and $G'$ be two graphs on disjoint sets of vertices and of dimensions $d-1$ and $d'-1$, respectively. Then
- The graph $G\sqcup G'$ is Cohen-Macaulay if and only if both $G$ and $G'$ are so.
- If $G$ is Cohen-Macaulay and $G'$ is ${\mbox{CM}\,}_{r'}$ for some $r'\ge 1$, then $G\sqcup G'$ is ${\mbox{CM}\,}_{d+r'}$. If $G'$ is not ${\mbox{CM}\,}_{r'-1}$, then $G\sqcup G'$ is not ${\mbox{CM}\,}_{d+r'-1}$.
- If $G$ is ${\mbox{CM}\,}_r$ and $G'$ is ${\mbox{CM}\,}_{r'}$ for some $r, r' \ge 1$, then $G\sqcup G'$ is ${\mbox{CM}\,}_t$ where $t={\rm
max}\{d+r', d'+r\}$. Conversely, if $G\sqcup G'$ is ${\mbox{CM}\,}_t$, then $G$ is ${\mbox{CM}\,}_{t-d'}$ and $G'$ is ${\mbox{CM}\,}_{t-d}$.
Two characterizations of bipartite ${\mbox{CM}\,}_t$ graphs
===========================================================
We now restrict to the case of bipartite graphs. Since Cohen-Macaulay bipartite graphs are characterized by Herzog and Hibi [@HH2005 Theorem 3.4], and also in a different version by Cook and Nagel [@CN2012 Proposition 4.8], we consider the non-Cohen-Macaulay case.
\[main\] Let $G$ be an unmixed bipartite graph of dimensions $d-1$. Let $K_{n,n}$, with $n\ge 2$, be a maximal complete bipartite subgraph of $G$ of minimum dimension. Then $G$ is ${\mbox{CM}\,}_{d-n+1}$ but it is not ${\mbox{CM}\,}_{d-n}$.
We prove both assertions by induction on $d \ge 2$. If $d = 2$ then $G=K_{2,2}$ which is ${\mbox{CM}\,}_1$ but it is not Cohen-Macaulay. Assume that $d
> 2$. We show that for every $v\in G$, $G\setminus N[v]$ is ${\mbox{CM}\,}_{n-d}$ and for some $v\in G$ it is not ${\mbox{CM}\,}_{n-d-1}$. Let $(\{x_1, \cdots, x_d\}, \{y_1, \cdots, y_d\})$ be a pure order of $G$. Let $x_i$ be a vertex of some maximal bipartite subgraph $K_{m,m}$ with $m\ge n$. Then $G\setminus N[x_i]$ is a disjoint union of $c\ge m-1$ isolated vertices and an unmixed bipartite graph $H$ of dimension $d-c-2$. The graph $H$ is unmixed because Ind$(G\setminus N[x_i]) = {\mbox{link}\,}_{x_i}({\rm Ind}(G))$, and any link of a pure complex is pure. But $G\setminus N[x_i] = \{x_{i_1}, \cdots,
x_{i_c}\} \sqcup H$ is unmixed if and only if $H$ is so. Observe that if $y_{j_0}$ is a vertex of a maximal bipartite subgraph of $G$ and $y_{j_0}\in N[x_i]$, then by Lemma \[compneighb\], all $y_j$ vertices of this subgraph belong to $N[x_i]$. Thus if $H$ has no crosses, by Proposition \[cookNag\] it is Cohen-Macaulay. Otherwise, the minimum dimension of maximal complete bipartite subgraphs of $H$ will not be less than the minimum dimension of such subgraphs in $G$. Hence by the induction hypothesis $H$ is ${\mbox{CM}\,}_{d-c-n}$ and by Theorem \[joingraphs\](ii), $G\setminus
N[x_i]$ is ${\mbox{CM}\,}_{n-d}$. If $x_i$ does not belong to any maximal bipartite subgraph of $G$ of positive dimension, then $G\setminus
N[x_i]$ is a disjoint union of $c\ge 0$ isolated vertices and an unmixed bipartite graph $H$ of dimension $d-c-2$. Hence $H$ is ${\mbox{CM}\,}_{d-c-n}$ and by Theorem \[joingraphs\](ii), $G\setminus
N[x_i]$ is ${\mbox{CM}\,}_{d-n}$. A similar argument reveals that for any $y_i\in G$, the graph $G\setminus N[y_i]$ is ${\mbox{CM}\,}_{d-n}$. Therefore, by Lemma \[lemmagraphs\], $G$ is ${\mbox{CM}\,}_{d-n+1}$. We now proceed the induction step to show that this result is sharp. Let $d>2$ and let $K_{n,n}$, $n\ge 2$, be a maximal bipartite subgraph of $G$ of minimum dimension. Take $x_i \in G\setminus K_{n,n}$. First assume that $x_i$ is not adjacent to any vertex in $K_{n,n}$ and consider $G\setminus N[x_i]$. Let $G\setminus N[x_i]$ be the disjoint union of $c\ge 0$ isolated vertices and an unmixed bipartite graph $H$ of dimension $d-c-2$. Then $H$ contains $K_{n,n}$ and hence by induction hypothesis $H$ is sharp ${\mbox{CM}\,}_{d-c-n}$ and $G\setminus
N[x_i]$ is sharp ${\mbox{CM}\,}_{d-n}$. Therefore, $G$ can not be ${\mbox{CM}\,}_{d-n}$. Now assume that $x_iy_j \in G$ for some $j$ with $y_j\in K_{n,n}$. Then by purity of the order, all $y_k \in K_{n,n}$ is adjacent to $x_i$. But then $y_i$ is not adjacent to any vertex of $K_{n,n}$, because otherwise, $ K_{n,n}$ will not be maximal. In this case, consider $G\setminus N[y_i]$ and proceed similar to the previous case.
As a second characterization of bipartite ${\mbox{CM}\,}_t$ graphs, we show that any ${\mbox{CM}\,}_t$ graph is obtained from a Cohen-Macaulay graph $H$ by replacing the perfect matching edges of $H$ by complete bipartite graphs. This statement will be more precise in the next theorem. But first we provide a definition and a lemma.
\[replacement\] Let $H$ be an unmixed bipartite graph with pure order $$(\{x_1, \cdots, x_r\},\{y_1, \cdots, y_r\}).$$ For a fixed $i$, by [*replacing*]{} the edge $x_iy_i\in H$ with a complete bipartite graph $$K_{n_i,n_i} = \{x_{i1}, \cdots, x_{in_i}\}\times\{y_{i1}, \cdots, y_{in_i}\}$$ we mean a bipartite graph $H'$ with vertex set $$(\{x_1, \cdots,
x_{i-1}, x_{i1}, \cdots, x_{in_i}, x_{i+1}, \cdots, x_r\}, \{y_1,
\cdots, y_{i-1}, y_{i1}, \cdots, y_{in_i}, y_{i+1}, \cdots,
y_r\}),$$ preserving all adjacencies, i.e.,
- $x_sy_t\in H'$ for all $s, t \ne i$ if and only if $x_ty_s\in H$,
- $x_{i_k}y_j\in H'$ for all $k$ if and only if $x_iy_j\in H$,
- $x_jy_{i_k}\in H'$ for all $k$ if and only if $x_jy_i\in H$.
\[unmixedreplace\] Let $G$ be an unmixed bipartite graph with pure order on the vertex set $V(G) = V\cup W$ where $V =
\{x_1, \cdots, x_d\}$ and $W = \{y_1, \cdots, y_d\}$. Let $n_1,\cdots, n_d$ be any positive integers. Let $G' = G(n_1,\cdots,
n_d)$ be the graph obtained by replacing each edge $x_iy_i$ with the complete bipartite graph $K_{n_i,n_i} = \{x_{i1}, \cdots,
x_{in_i}\}\times \{y_{i1}, \cdots, y_{in_i}\}$ for all $i = 1,\dots,
d$. Then $G'$ is also unmixed.
Let $K_{n_i,n_i} = \{x_{i1}, \cdots, x_{in_i}\}
\times \{y_{i1}, \cdots, y_{in_i}\}$. Then $$V(G') = (\{x_{11},
\cdots, x_{1n_1}, \cdots, x_{d1}, \cdots, x_{dn_d}\}, \{y_{11},
\cdots, y_{1n_1}, \cdots, y_{d1}, \cdots, y_{dn_d}\})$$ is a pure order of $G'$. In fact, for all $i, r$, $x_{ir}y_{ir} \in G'$. Also if $x_{ir}y_{js} \in G'$ and $x_{js}y_{kt} \in G'$, then $x_iy_j \in
G$ and $x_jy_k \in G$, and hence, $x_iy_k \in G$. Thus by the construction of $G'$, $x_{ir}y_{kt} \in G'$.
\[macaulayreplace\] Let $G$ be a Cohen-Macaulay bipartite graph with a Macaulay order on the vertex set $V(G) = V\cup W$ where $V = \{x_1, \cdots, x_d\}$ and $W = \{y_1, \cdots, y_d\}$. Let $n_1,
\cdots, n_d$ be any positive integers with $n_i \ge 2$ for at list one $i$. Let $G' = G(n_1, \cdots, n_d)$ be the graph obtained by replacing each edge $x_iy_i$ with the complete bipartite graph $K_{n_i,n_i}$ for all $i = 1,\dots, d$. Let $n_{i_0} = {\rm min}
\{n_i > 1: i = 1, \cdots, d\}$, $n= \sum_{i=1}^{d}n_i$. Then $G'$ is exclusively a ${\mbox{CM}\,}_{n-n_{i_0}+1}$ graph. Furthermore, any bipartite ${\mbox{CM}\,}_t$ graph is obtained by such a replacement of complete bipartite graphs in a unique bipartite Cohen-Macaulay graph.
The first claim follows by Lemma \[unmixedreplace\] and Theorem \[main\]. We settle the second claim. Let $G$ be a bipartite ${\mbox{CM}\,}_t$ graph with a pure order of vertices. Let $K_{n_1,n_1}, \cdots, K_{n_d,n_d}$ be the maximal bipartite subgraphs of $G$, where $n_i \ge 1$ for all $i$. Observe that, by maximality, these complete subgraphs of $G$ are disjoint. Choose one edge $x_{i1}y_{i1}$ from each subgraph $K_{n_i,n_i}$ for all $i = 1,
\cdots, d$. Let $H$ be the induced subgraph of $G$ on the vertex set $(\{x_{11}, \cdots, x_{d1}\},\{y_{11}, \cdots, y_{d1}\})$. By Lemma \[compneighb\], $H$ is independent of the choice of particular edge $x_{i1}y_{i1}$ from $K_{n_i,n_i}$ and hence $H$ is unique. Since the ordering of vertices of $G$ is a pure order, its restriction to $H$ is also pure. Thus, $H$ is an unmixed bipartite graph. But by the maximality of the complete bipartite subgraphs $K_{n_i,n_i}$, and the construction of $H$, it is cross-free. Therefore, by Proposition \[cookNag\], $H$ is Cohen-Macaulay. Now any edge $x_{i1}y_{i1}$ replace in $H$ with $K_{n_i,n_i}$ for all $i
= 1, \cdots, d$, preserving all other adjacencies. Let $H'$ be the resulting graph. Then by the construction, $G = H'$, as required.
\[remark\] Let $H$ be a bipartite Cohen-Macaulay graph and let $G
= H'$ be a bipartite ${\mbox{CM}\,}_t$ graph obtained from $H$ by the replacing process described above. Assume that $G$ is is not ${\mbox{CM}\,}_{t-1}$ and $t \ge 2$. Using the the results of this section, the following observations are immediate.
First of all, $1 \le {\rm dim} H \le t-1$. Because if ${\rm dim} H
\ge t$ and we replace just one $K_{n,n}$ with $n\ge 2$, then $G$ is strictly ${\mbox{CM}\,}_r$ with $r\ge t+1$. On the other hand, if ${\rm dim} H
= 0$, then $G$ is ${\mbox{CM}\,}_1$.
If ${\rm dim} H = t-1$, then only one $K_{n,n}$ with $n\ge 2$ can be replaced. Because replacing at least two $K_{n,n}$ with $n\ge 2$, $G$ is strictly ${\mbox{CM}\,}_r$ with $r \ge t+1$.
If ${\rm dim} H = t-1$, for replacing just one $K_{n,n}$, $n$ is arbitrary and hence $G$ is of dimension $n+t-2$.
If ${\rm dim} H \le t-2$, the number of replacements should be at least $2$. Again because if with one replacement of $K_{n,n}$, $n
\ge 2$, $G$ would be ${\mbox{CM}\,}_r$ with $r \le t-1$.
When ${\rm dim} H \le t-2$, the maximum number of replacements of $K_{n,n}$, $n \ge 2$, is at most $t - {\rm dim} H$ which may occur replacing $K_{2,2}$’s.
For ${\rm dim} H \le t-2$, the maximum size of $K_{n,n}$ to be replaced is also $n = t - {\rm dim} H$ which may occur when we have two replacements.\
Using these remarks we may easily distinguish all bipartite ${\mbox{CM}\,}_t$ graphs for $t = 2, 3, 4$.\
\[exam1\] Bipartite ${\mbox{CM}\,}_2$ graphs which are not Buchsbaum. Using the notation of Remark \[remark\] we have ${\rm dim} H = 1$. There are just two non-isomorphic bipartite Cohen-Macaulay graphs of dimension one. By replacing process, they produce two types of bipartite ${\mbox{CM}\,}_2$ graphs which are not Buchsbaum. They are of arbitrary dimensions. More precisely, one such graph is the disjoint union of an edge $x_1y_1$ with $K_{n_2,n_2} = \{x_{21},\cdots,
x_{2n_2}\}\times\{y_{21},\cdots, y_{2n_2}\}$, $n_2\ge 2$, and the other one consists of the first graph together with the edges $x_1y_{2i}$ for all $i = 1,\cdots, n_2$. The second graph with $n_2
= 3$ could be depicted in Figure 1.
(100,60)(-100,30) (0,0)[(0,1)[50]{}]{} (0,50)[(1,-1)[50]{}]{} (50,50)[(1,-1)[50]{}]{} (50,50)[(2,-1)[100]{}]{} (100,50)[(1,-1)[50]{}]{} (50,0)[(0,1)[50]{}]{} (100,0)[(0,1)[50]{}]{} (150,0)[(0,1)[50]{}]{} (100,0)[(1,1)[50]{}]{} (50,0)[(2,1)[100]{}]{} (50,0)[(1,1)[50]{}]{} (0,50)[(2,-1)[100]{}]{}(0,50)[(3,-1)[150]{}]{} (0,0) (0,50) (50,0) (50,50) (100,0) (100,50) (150,0) (150,50) (-4,55)[$x_1$]{} (46,55)[$x_{21}$]{} (96,55)[$x_{22}$]{} (146,55)[$x_{23}$]{}
(-4,-10)[$y_1$]{} (46,-10)[$y_{21}$]{} (96,-10)[$y_{22}$]{} (146,-10)[$y_{23}$]{}
$$Figure \ 1$$
Bipartite ${\mbox{CM}\,}_3$ graphs which are not ${\mbox{CM}\,}_2$. For these graphs ${\rm dim} H = 1, 2$.\
If ${\rm dim} H = 1$, by Example \[exam1\], there are just two bipartite ${\mbox{CM}\,}_3$ graphs by replacing two edges of a perfect matching by $K_{2,2}$’s. In this case, ${\rm dim} G = 3$. (see Figure 2, and Figure 3).\
If ${\rm dim} H = 2$, then there are $4$ non-isomorphic bipartite Cohen-Macaulay graphs of dimension $2$. By replacing one perfect matching edge with $K_{n,n}$ of arbitrary size in each Cohen-Macaulay graph, they produce $7$ types of bipartite ${\mbox{CM}\,}_3$ graphs which are not ${\mbox{CM}\,}_2$. Note that depending on the choice of the edge to be replaced in each case, we may get non-isomorphic bipartite graphs. In this case ${\rm dim} G = n + 1$.
(100,60)(-100,30) (0,0)[(0,1)[50]{}]{} (0,50)[(1,-1)[50]{}]{} (100,50)[(1,-1)[50]{}]{} (50,0)[(0,1)[50]{}]{} (0,0)[(1,1)[50]{}]{} (100,0)[(0,1)[50]{}]{} (150,0)[(0,1)[50]{}]{} (100,0)[(1,1)[50]{}]{} (0,0) (0,50) (50,0) (50,50) (100,0) (100,50) (150,0) (150,50) (-4,55)[$x_{11}$]{} (46,55)[$x_{12}$]{} (96,55)[$x_{21}$]{} (146,55)[$x_{22}$]{}
(-4,-10)[$y_{11}$]{} (46,-10)[$y_{12}$]{} (96,-10)[$y_{21}$]{} (146,-10)[$y_{22}$]{}
$$Figure \ 2$$
(100,60)(-100,30) (0,0)[(0,1)[50]{}]{} (0,50)[(1,-1)[50]{}]{} (50,50)[(2,-1)[100]{}]{} (100,50)[(1,-1)[50]{}]{} (50,0)[(0,1)[50]{}]{}(0,0)[(1,1)[50]{}]{} (100,0)[(0,1)[50]{}]{} (150,0)[(0,1)[50]{}]{} (100,0)[(1,1)[50]{}]{} (0,50)[(2,-1)[100]{}]{} (0,50)[(3,-1)[150]{}]{} (50,50)[(1,-1)[50]{}]{} (0,0) (0,50) (50,0) (50,50) (100,0) (100,50) (150,0) (150,50) (-4,55)[$x_{11}$]{} (46,55)[$x_{12}$]{} (96,55)[$x_{21}$]{} (146,55)[$x_{22}$]{}
(-4,-10)[$y_{11}$]{} (46,-10)[$y_{12}$]{} (96,-10)[$y_{21}$]{} (146,-10)[$y_{22}$]{}
$$Figure \ 3$$
Bipartite ${\mbox{CM}\,}_4$ graphs which are not ${\mbox{CM}\,}_3$. For these graphs ${\rm dim} H = 1, 2, 3$.\
If ${\rm dim} H = 1$, there are two bipartite ${\mbox{CM}\,}_4$ graphs obtained by replacing two edges of a perfect matching by $K_{3,3}$’s. In this case, ${\rm dim} G = 5$. And, similarly, there are two others obtained by replacing one edge with $K_{2,2}$ and another edge with $K_{3,3}$. In this case, ${\rm dim} G = 5$.\
If ${\rm dim} H = 2$, then while there are $4$ non-isomorphic bipartite Cohen-Macaulay graphs of dimension $2$, by replacing two perfect matching edges with $K_{2,2}$’s in each Cohen-Macaulay graph, they produce $7$ bipartite ${\mbox{CM}\,}_4$ graphs which are not ${\mbox{CM}\,}_3$. They all have dimension $4$.\
If ${\rm dim} H = 3$, then there are $10$ non-isomorphic bipartite Cohen-Macaulay graphs of dimension $3$. Replacing one perfect matching edge with $K_{n,n}$, $n\ge 2$, in each Cohen-Macaulay graph, they produce $25$ bipartite ${\mbox{CM}\,}_4$ graphs which are not ${\mbox{CM}\,}_3$. They all have dimension $n+2$. Out of all 36 bipartite ${\mbox{CM}\,}_4$ graphs, 21 graphs are connected.\
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors would like to thank J. Herzog and U. Nagel for some fruitful discussions on the subject of this paper. This work has been supported by Center for International Studies & Collaboration (CISSC) and French Embassy in Tehran in the framework of the Gundishapur project 27462PL on the Homological and Combinatorial Aspects of Commutative Algebra and Algebraic Geometry. The research of R. Zaare-Nahandi has been partially supported by research grant no. of University of Tehran. The research of S. Yassemi was in part supported by a grant from IPM (No. 91130214).
[CHT99]{}
D. Cook and U. Nagel, [*Cohen-Macaulay graphs and face vectors of flag complexes,*]{} SIAM J. Discrete Math. **26** (2012), no. 1, 89–101.
R. Fröberg, [*A note on the Stanley-Reisner ring of a join and of a suspension,*]{} Manuscripta Math. **60** (1988), no. 1, 89–91.
H. Haghighi, S. Yassemi and R. Zaare-Nahandi, [*Bipartite $S_2$ graphs are Cohen-Macaulay,*]{} Bull. Math. Soc. Sci. Roumanie. **53** (2010), 125–132.
H. Haghighi, S. Yassemi and R. Zaare-Nahandi, [*A generalization of $k$-Cohen-Macaulay simplicial complexes*]{}, Ark. Mat. **50** (2012), 279–290.
J. Herzog and K. Hibi, [*Distributive lattices, bipartite graphs and Alexander duality,*]{} J. Algebraic Combin. **22** (2005), 289–302.
E. Miller, I. Novik and E. Swartz, [*Face rings of simplicial complexes with singularities,*]{} Math. Ann. (2011), 351:857–-875.
H. Sabzrou, M. Tousi and S. Yassemi, [*Simplicial join via tensor products*]{}, Manuscripta Math. **126** (2008), 255–272.
C. Sava, [*On the Stanley-Reisner ring of a join*]{}, An. Sti. Univ. “Al. l. Cuza”, lasi, Tom XXXI, Ser. Ia, (1985), 145–148.
P. Schenzel, [*On the number of faces of simplicial complexes and the purity of Frobenius*]{}, Math. Z. **178** (1981), 125–142.
R. Stanley, [*Combinatorics and Commutative Algebra,*]{} Second Edition, Birkhäuser, Boston, 1995.
R. H. Villarreal, [*Unmixed bipartite graphs,*]{} Rev. Colombiana Mat. **41** (2007), 393–395.
R. Zaare-Nahandi and R. Zaare-Nahandi, [*The minimal free resolution of a class of square-free monomial ideals*]{}, J. Pure Appl. Algebra **189** (2004) 263–278.
[^1]: Emails: [email protected], [email protected], [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Berezin integration over fermionic degrees of freedom as a standard tool of quantum field theory is analysed from the viewpoint of noncommutative geometry. It is shown that among the variety of contradictory integration prescriptions existing in the current literature, there is only one unique minimal set of consistent rules, which is compatible with Connes’ normalized cyclic cohomology of the Gra[ß]{}mann algebra.'
author:
- |
G. Grensing and M. Nitschmann\
Institut fűr Theoretische Physik und Astrophysik\
Universita̋t Kiel\
D-24118 Kiel, Germany
date: January 2004
title: Berezin integration over anticommuting variables and cyclic cohomology
---
Introduction {#introduction .unnumbered}
============
That branch of mathematics in which a $\mathds{Z}/2$-grading plays dominant role is, somewhat emphatically, also called ‘supermathematics’. It is of fundamental importance in physics since it underlies the treatment of fermionic particles, and as such it is part of the standard repertoire in quantum field theory (see, e.g. [@Itzy80; @Ryde85]). Of course, these tools get involved also in supersymmetric theories [@Nieu81; @Wess83], but for its justification there is no need to take recourse to supersymmetry since already for conventional theories containing fermions one must make use of what is known as the differential and integral calculus on a Gra[ß]{}mann algebra. After preparatory work of Schwinger [@Schw53; @Schw70], Matthews $\&$ Salam [@Matt55] and others this calculus was created by the Russian mathematician Berezin [@Bere66] in the year 1966. Since then this theory has undergone revision, also by Berezin himself (see [@Bere79; @Bere87]), and nowadays a variety of conflicting rules is given in the literature, e.g., some authors use anticommuting differentials for the Grassmann variables, whereas others use commuting ones, and the like. Hence, as regards the present status of the art concerning the Berezin calculus, the situation is controversial, to say the least.
In all, it seems that a more fundamental understanding of the Berezin rules is desirable. As we believe, with the advent of noncommutative geometry à la Connes [@Conn94] new avenues have been opened to reconsider these problems. In this context, a universal differential calculus over a general noncommutative algebra is supplied for, where the algebra is a substitute for the commutative algebra of functions over a manifold in the classical situation; furthermore, the notion of Connes’ characters provides for abstract integral calculi over such a ‘virtual manifold’. Whence, this whole machinery can serve as an ideal vehicle to gain deeper insight into the origin of Berezin integration, which is the main motivation for the present work.
The Gra[ß]{}mann algebra is certainly noncommutative, but in a mild form since it is graded commutative. Thus, the extension of the universal differential algebra to the $\mathds{Z}/2$-graded situation is needed; this is available since the year 1988 with the book of Kastler [@Kast88]. His treatment essentially relies on the Karoubi [@Karo82] approach, which allows for a drastic simplification of the formalism. It also delivers a rather natural set of sign factors that get involved in the definition of the Hochschild boundary operator and the cyclicity operator. But the latter differ from those obtained by means of the sign rules ascribed to Koszul, Milnor or Quillen, depending on ones preferences. In the subsequent work of Kastler and collaborators [@Coqu90; @Coqu91; @Coqu95], the original sign factors were forgotten in favour of the customary Koszul-Milnor-Quillen signs. Though one would expect the choice of a sign to be a matter of convention, we will show that only the sign factors obtained by means of the Karoubi approach are compatible with the rules governing the Connes’ characters of the Gra[ß]{}mann algebra.
Further progress was made in 1995 by Coquereaux $\&$ Ragoucy [@Coqu95] with their work on the cohomology of the Gra[ß]{}mann algebra. They introduced the Gra[ß]{}mann analogue of de Rham currents, which in the classical situation over a compact manifold were shown by Connes [@Conn85] to be in canonical one-to-one correspondence with skewsymmetrized Hochschild cohomology. One expects something similar to be valid in the Gra[ß]{}mann case, and the above authors were partially successful in obtaining an analogous result. We shall sharpen their arguments and prove that Gra[ß]{}mann currents compute the normalized cyclic Hochschild cohomology of the Gra[ß]{}mann algebra.
In defining a Gra[ß]{}mann current, use is made of the Berezin rules, which thus have to be given beforehand. They are of completely algebraic origin and are specified by a linear map $J:G^n\to\mathds{R}$, where $G^n$ denotes the Gra[ß]{}mann algebra with $n$ generators $\xi^i$, which we require to be translational invariant. The explicit form of this condition will be specified later; it selects one unique map amongst this multitude of possibilities, the one that singles out the top component of an algebra element $f(\xi)\in G^n$, i.e. $J(f)=f_{1\dots n}$ where the right hand side denotes the coefficient of the term of highest degree. It has become customary to write this unique element in the suggestive form of an integral, namely $$J(f)=\int d^{\,n}\xi\,f(\xi).$$ In this version, however, the meaning of the symbol $d^{\,n}\xi$ is left open, and up to the present day it awaits a proper definition. This is the main source for the many conflicting Berezin integration rules existing in the literature. It is the purpose of the present paper to justify the suggestive symbolic notation, and for this the abstract apparatus of noncommutative geometry is needed, in particular, the universal differential calculus and its associated cyclic cohomology (see, e.g., [@Land97; @Grac01]).
We understand the Berezin rules as a purely algebraic recipe, and only use it in this sense in the course of the development. This conception permits to ask the question whether Connes’ characters have something to tell us about the deeper origin of the Berezin rules; the answer will be given at the very end of the present paper, where it is shown that there is one unique normalized cyclic cocycle that entails those Berezin rules, we except as valid. Furthermore, it provides for a proper definition of the ‘volume element’ $d^{\,n}\xi$ that enters the formal Berezin integral.
The paper is organized as follows: In the first Section, the universal differential graded algebra for a $\mathds{Z}/2$-graded algebra is constructed along with the definition of its normalized cyclic cohomology; afterwards, the necessary modifications for a graded commutative algebra are given. The second Section supplies for the necessary basic facts about Gra[ß]{}mann algebras; in particular, we give a proof that the request for translational invariance determines the Berezin rules uniquely. Furthermore, Fourier transformation is discussed and used to endow the Grassmann algebra with two hermitean inner products, both of which are nondegenerate. One of these is positive definite and provides for a Hilbert space structure, whereas the second one is indefinite; we shall construct a unitary grading operator, which relates these two inner products. They give rise to two operators, that are dual to the exterior differential operator, and these in turn are instrumental for a proof of the Poincaré lemma and the Hodge decomposition theorem. We then turn to Connes’ characters over the Gra[ß]{}mann algebra in the third Section and demonstrate that compatibility requires the use of the sign rules deriving from the Karoubi approach, whereas the Koszul-Milnor-Quillen conventions lead to contradictions. The indefinite inner product mentioned above gives rise to a nondegenerate pairing between the space of $p$-forms and $p$-currents; it is used in the fourth Section to compute the normalized cyclic cohomology of the Gra[ß]{}mann algebra by means of currents, where we use results derived earlier by Kassel [@Kass86]. These tools are applied in the fifth Section to single out among the possible Connes’ characters one single element, which may legitimately be called a ‘volume character’. On the one hand, it embodies the Berezin rules in an encoded form, and on the other hand it ultimately provides for a rigorous justification for the rewriting of the formal Berezin integral as an integral. We end in the last Section with some concluding remarks, being devoted to the impact of the above results on supersymmetric theories.
Connes’ cyclic cohomology for $\mathds{Z}/2$-graded algebras
============================================================
We give a short introduction to the differential graded algebra of a $\mathds{Z}/2$-graded unital algebra $A$ over a commutative ring $k$, and the associated Connes’ characters; we mainly follow the treatment given by Kastler [@Kast88; @Kast90], being based on the Karoubi approach [@Karo82]. This will prove to be essential in avoiding the nightmare to invent the correct Koszul-Milnor-Quillen signs, at least in part.
The universal differential calculus
-----------------------------------
So let $A$ be an associative algebra with unit $e$ over a commutative ring $k$; it is also assumed to be $\mathds{Z}/2$-graded, with $\varepsilon$ denoting the grading automorphism. The graded differential algebra, written $\Omega^{\displaystyle\cdot}A$, is constructed as follows through generators and relations. We begin by defining the space of $1$-forms $\Omega_{}^1A$ on associating with every element $a\in A$ a symbol $da$; they generate the left $A$-module $\Omega_{}^1A$, and so we have a $k$-linear map $d:A\to \Omega_{}^1A$ which is required to obey the graded Leibniz rule $$d(aa')=(da)a'+(-1)^{|d|\,|a|}a(da')\qquad\qquad\qquad:\,a,a'\in A$$ where $|a|$ denotes the parity. In this way, $\Omega_{}^1A$ is equipped also with a right $A$-structure. Since $|ea|=|e|+|a|$ implies $|e|=0$, the map $d$ annihilates the unit. The $A$-bimodule $\Omega^pA$ of $p$-forms is defined as the tensor product $\Omega^pA=\overset{p}{\otimes}^{}_A\Omega_1A$, and its general element $\omega_p$ can thus be written as a finite linear combination of monomials $a_0da_1\cdots da_p$ with the tensor product sign being omitted; the structure of $\Omega^pA$ as a right $A$-module is determined by the easily verifiable identity $$(-1)^pa_0da_1\cdots da_p\,a_{p+1}=$$ $$\nonumber
(-1)^{\sum\limits_{j=1}^p|a_j|}a_0a_1da_2\cdots da_{p+1}
+\sum_{i=1}^p(-1)^{i+\sum\limits_{j=i+1}^p|a_j|}a_0da_1\cdots d(a_ia_{i+1})\cdots da_{p+1}.$$ Furthermore, the operator $d$ of exterior differentiation is defined by $d(a_0da_1\cdots da_p)=da_0da_1\cdots da_p$, which gives $d(da_1\cdots da_p)=0$ and so we have $d^{\,2}=0$ on $\Omega^{\displaystyle\cdot}A=\underset{p}{\oplus}\Omega^pA$. Next we need $$d(a_0da_1\cdots da_p\cdot a'_0da'_1\cdots da'_p)=d(a_0da_1\cdots da_p)a'_0da'_1\cdots da'_p+
(-1)^{(p+|a_0|+\cdots+|a_p|)}a_0da_1\cdots da_p\,d(a'_0da'_1\cdots da'_p)$$ and thus the parity in the $\mathds{Z}/2$-graded case must be identified as $$|a_0da_1\cdots da_p|=p+\sum_{i=0}^p|a_i|\qquad\qquad\qquad:\,|d|=1.$$ Hence, the product rule is $$d(\omega\,\omega')=d\omega\,\omega'+(-1)^{|\omega|}\omega\,d\omega'$$ which implies $d^{\,2}(\omega\,\omega')=0$, and so the assignment of the parity $|d|=1$ to the exterior derivation $d$ is indeed correct. Finally, the universality property of $\Omega^{\displaystyle\cdot}A$ is proven in much the same way as in the trivially graded case (see, e.g. [@Land97]).
We may then address the problem of defining a boundary operator on the universal differential graded algebra. This is done, instead of starting from scratch, on taking the Karoubi form ([@Karo82], see also [@Brod98]) of the ungraded case as guiding principle. Accordingly, the definition simply is $$b(\omega\,da)=(-1)^{|\omega|}[\omega,a]\qquad\qquad\qquad\qquad b(a)=0$$ where on the right hand side the graded commutator is understood; with the notation $a_0da_1\cdots da_{p}=(a_0,a_1,\ldots a_{p})$ the explicit form reads $$\label{boundopvarphi}
b(a_0,a_1,\ldots,a_{p})=$$ $$\nonumber
\sum_{i=0}^{p-1}(-1)^{i+\sum\limits_{j=0}^i|a_j|}(a_0,\ldots,a_ia_{i+1},\ldots,a_{p})
-(-1)^{(1+|a_{p}|)((p-1)+\sum\limits_{j=0}^{p-1}|a_j|)}(a_{p}\,a_0,a_1,\ldots,a_{p-1}).$$ It reduces to the correct boundary operator in the trivially graded case and as well obeys $b^{\,2}=0$ in the present situation.
Let us also introduce a mutilated version of the boundary operator, denoted $b'$, in which only the crossover term is absent, viz. $$b'(\omega\,da)=(-1)^{|\omega|}\omega\,a$$ and $b'(a)=0$. It also obeys $b^{\,\prime 2}=0$, and its anticommutator with $d$ is $$\label{anticdbprimeequunity}
b^{\,\prime}\,d+d\,b^{\,\prime}=1.$$ Hence $b'$ is a contracting homotopy so that the complex $(\Omega^{\displaystyle\cdot}A,d)$ is acyclic.
In order to find the generalization of the cyclicity operator, we pass to the dual situation. A cochain $\varphi$ is a $k$-linear form on $\Omega^{\displaystyle\cdot}A$; it is called closed if $\varphi\circ d=0$, i.e. vanishes on exact forms, and graded if $\varphi\circ\text{ad}(\omega)=0$ for all $\omega\in\Omega^{\displaystyle\cdot}A$, i.e. vanishes on graded commutators. So let $\varphi$ denote a both closed and graded trace; from $\varphi([\omega,da])=0$ we infer $$\begin{aligned}
\varphi(\omega\,da)&=(-1)^{|\omega|(1+|a|)}\varphi(da\,\omega)\\
&=(-1)^{|\omega|(1+|a|)}(\varphi(d(a\,\omega))-(-1)^{|a|}\varphi(a\,d\omega))\\
&=(-1)^{(1+|\omega|)(1+|a|)}\varphi(a\,d\omega)\\
&=\varphi(\lambda(a\,d\omega)))\end{aligned}$$ where the cyclicity operator acting on $\Omega^{\displaystyle\cdot}A$ is read off to be $$\label{cyclopvarphi}
\lambda(\omega\,da)=(-1)^{(1+|\omega|)(1+|a|)}a\,d\omega$$ and we thus define it on cochains by the rule $$\label{deflambdagng}
\lambda\,\varphi=\varphi\circ\lambda.$$ Analogously, for the boundary operator, the definition is $$\label{defboundopgng}
b\,\varphi=\varphi\circ\,b.$$ In particular, if $\varphi$ is a graded trace, we thus have $b\,\varphi=0$. Finally, the definition of a character $\tau(a_0,a_1,\ldots a_{n})$ of an $n$-dimensional cycle $(\Omega^{(n)}A,d,\int)$ is the same as in the trivially graded case; hence, the cyclicity property for a character reads $$\label{cyclopkdefgncg}
\tau(a_0,\ldots,a_n)=(-1)^{(1+|a_n|)(n+\sum\limits_{i=0}^{n-1}|a_i|)}\tau(a_n,a_0,\ldots,a_{n-1})$$ and, according to the previous remark, the boundary of $\tau$ vanishes. Furthermore, a slight variant of the proof in the trivially graded case (see, e.g. [@Grac01]) shows that the correspondence between characters and normalized cyclic Hochschild cocycles is one-to-one; we recall, a $p$-cochain $\varphi$ is called cyclic if $\lambda\,\varphi=\varphi$, and normalized if $\varphi(a^{}_0,\ldots,a_p)=0$ in case that $a_i=e$ for some $i\in\{0,\ldots,p\}$. It is in this way that Connes’ cyclic cohomology can naturally be extended to the $\mathds{Z}/2$-graded case.
As for our notational conventions, the normalized cyclic cohomology is denoted by $HC^{\displaystyle\cdot}(A)$, whereas $H^{\displaystyle\cdot}_{\lambda}(A)$ signifies the cyclic cohomology.
The Koszul-Milnor-Quillen sign conventions
------------------------------------------
The definition of the boundary operator and the cyclicity operator in the $\mathds{Z}/2$-graded case, as given throughout in the literature (see [@Conn85]), differs from the one we have obtained above. In order to relate these two versions (cf. also [@Kast88]), the heuristic motive consists in distributing $p$ factors of $d$ in front of the product $a_0a_1\cdots a_p$ such that one ends up with the $p$-form $a_0da_1\cdots da_p$; this gives the sign factor $$\epsilon_p(a_0,a_1,\ldots,a_p)=(-1)^{\sum\limits_{i=0}^{p-1}(p-i)|a_i|}$$ (where the dependence on the parities of the $a$s is suppressed below) and we thus define $$\phi(a_0da_1\cdots da_p)=\epsilon_p\,\varphi(a_0da_1\cdots da_p).$$ The new cochain $\phi$ inherits the cyclicity operator $\lambda\phi(a_0da_1\cdots da_p)=\epsilon_p\,\lambda\varphi(a_0da_1\cdots da_p)$, denoted by the same symbol; its explicit form, on using eq. , is obtained to be $$\label{KMQsignrulelambda}
\lambda\phi(a_0,a_1,\ldots,a_p)=(-1)^{p
+|a_p|\sum\limits_{i=0}^{p-1}|a_i|}\phi(a_p,a_0,a_1,\ldots,a_{p-1}).$$ Analogously, for the boundary operator we set $b\,\phi(a_0da_1\cdots da_p)=\epsilon_p\,b\,\varphi(a_0da_1\cdots da_p)$; by means of eq. , the computation yields $$\label{KMQsignruleb}
b\,\phi(a_0,a_1,\ldots,a_{p+1})=\sum_{i=0}^p(-1)^i\phi(a_0,\ldots,a_ia_{i+1},\ldots,a_{p+1})
+(-1)^{p+1+|a_{p+1}|\sum\limits_{j=0}^{p}|a_j|}\phi(a_{p+1}a_0,a_1,\ldots,a_{p}).$$ For example, the latter form of the boundary operator on cochains is rather similar in appearance to that in the trivially graded case since it only differs by an additional contribution to the sign factor in the crossover term; but the disadvantage in using the $\phi$s lies in the crucial fact that the simplicity in the definition $b\,\varphi=\varphi\circ b$ of the boundary operator for the $\varphi$s (see ) then gets lost.
The graded commutative case {#Thegradedcommutativecase}
---------------------------
We shall have need for the graded commutative case; then a further reduction of $\Omega^1A$ is required (cf. also [@Kast88; @Loda92]). This comes about from the additional relations $$\label{abeliancaserelonegng}
db\,a=(-1)^{|a|(1+|b|)}a\,db$$ that arise as follows. Let $A$ be a $\mathds{Z}/2$-graded algebra and $M$ a $\mathds{Z}/2$-graded left $A$-module with compatible gradings; if $A$ is graded commutative, we define a compatible right action by $v\,a=(-1)^{|a|\,|v|}a\,v$, which explains the above relation. On passing to $\Omega_{}^2A$, and also higher values of $p$, the computation $$d^{\,2}(a\,b)=d(da\,b+(-1)^{|a|}a\,db)=d((-1)^{(1+|a|)|b|}b\,da+(-1)^{|a|}a\,db)=(-1)^{(1+|a|)|b|}db\,da+(-1)^{|a|}da\,db=0$$ shows that the further relations $$\label{abeliancasereltwogng}
db\,da=(-1)^{(1+|a|)(1+|b|)}da\,db$$ must be guaranteed; they demonstrate that in the trivially graded case the differentials anticommute, whereas in the nontrivially graded case for, e.g., $|a|=1=|b|$ they commute. The relations and must be respected on defining the universal (bi)graded differential algebra.
For this purpose, let us first consider the general non commutative $\mathds{Z}/2$-graded case; the elements $da$ of $\Omega^1A$ can be defined by $$\label{defofdgradedcasegng}
da=e\otimes a-(-1)^{|a|}a\otimes e$$ on the tensor product $A\otimes A$, which is understood to be graded with the product $a\otimes b\cdot c\otimes d=(-1)^{|b|\,|c|}ac\otimes bd$; with this definition, the graded Leibniz rule is easily verified. Then $\Omega^1A$ may be characterized as the kernel of the modified multiplication map $\tilde{\mu}=\mu\circ(\text{id}\otimes\varepsilon)$ where $\varepsilon$ denotes the grading operator, i.e. $$\label{defmutildegng}
\tilde{\mu}(a\otimes b)=(-1)^{|b|}a\,b.$$ Furthermore, the $A$-bimodule of $p$-forms is $\Omega^pA=\overset{p}{\otimes}^{}_A\Omega^1A$, where again the tensor product, now over $A$, is understood to be graded.
Suppose then $A$ is graded commutative; as to property , we observe that the following identity holds $$a\,db-(-1)^{|a|(1+|b|)}db\,a=(-1)^{|a|}(e\otimes a-(-1)^{|a|}a\otimes e)(e\otimes b-(-1)^{|b|}b\otimes e)$$ for the verification of which the gradation of the tensor product and graded commutativity is used. Hence we have $[a,db]\in(\Omega^1A)^2$, where $(\Omega^1A)^2$ is a subbimodule of the graded tensor product $A\otimes A$, and this is also contained in $\Omega^1A$ by construction. The above identity instructs us to pass to the quotient $$\Omega^1(A):=\Omega^1A/(\Omega^1A)^2$$ the elements of which thus obey $[a,db]=0$; again, the space of $1$-forms $\Omega^1(A)$ is distinguished by the universality property, as is easily verified.
Turning to higher values of $p$ in the trivially graded commutative case, one can simply set $$\Omega^p(A)=\overset{p}{\wedge}^{}_A\Omega^1(A)$$ and $\Omega^{\displaystyle\cdot}(A)=\underset{p\geq 0}{\oplus}\Omega^1(A)$; the skewsymmetric tensor product is as usual given by $$a^{}_0da_1\wedge\cdots\wedge da_p=\frac{1}{\sqrt{p\,!}}\sum_{\sigma\in S_p}\varepsilon(\sigma)
a^{}_0da_{\sigma^{-1}(1)}\otimes\cdots\otimes da_{\sigma^{-1}(p)}$$ where $\varepsilon(\sigma)$ denotes the sign of the permutation $\sigma$; with this definition, the relations are guaranteed.
In the nontrivially graded commutative situation, however, the standard symmetrization does not work because the relations $db\,da=(-1)^{|da|\,|db|}da\,db$ depend on the parities. Hence, since $\Omega^1(A)$ is $\mathds{Z}/2$-graded with $\varepsilon(a\,db)=(-1)^{|a|+|db|}a\,db$, a new concept is needed.
To begin with, consider a nontrivially graded vector space $V$ of dimension $n$ and its graded tensor product $V\otimes V$; from this we can construct the symmetrized graded symmetric tensor product $V\vee V$ and the graded antisymmetric tensor product $V\wedge V$ defined by $$v\vee w=v\otimes w+(-1)^{|v|\,|w|}w\otimes v\qquad\qquad\qquad\qquad:\,v,w\in V$$ and $$v\wedge w=v\otimes w-(-1)^{|v|\,|w|}w\otimes v\qquad\qquad\qquad\qquad:\,v,w\in V$$ on simple tensors, and extended to $V\otimes V$ through linearity. These products obey $w\vee v=(-1)^{|v|\,|w|}v\vee w$ and $w\wedge v=-(-1)^{|v|\,|w|}v\wedge w$, and so they are graded commutative and graded anticommutative, respectively; in the trivially graded case, one regains the standard versions. For higher powers, the general construction is obvious now; the prefactor of a particular summand is the sign $(\pm1)^{|\sigma|}$ of the permutation $\sigma$, times the corresponding Koszul-Milnor-Quillen sign.
Applied to the case at hand, let $M$ be a $\mathds{Z}/2$-graded $A$-bimodule and $M\otimes^{}_AM$ the graded tensor product over $A$; we then have the product $M\vee^{}_AM$ available. The $A$-bimodule $\Omega^p(A)$ of $p$-forms for a nontrivially graded commutative algebra $A$ can thus be defined as $$\Omega^p(A)=\overset{p}{\vee}^{}_A\,\Omega^1(A).$$ Its decomposable elements read as $$a^{}_0\,da^{}_1\vee\cdots\vee da_p\qquad\qquad\qquad:\,a^{}_0,a^{}_1,\cdots,a_p\in A$$ where again the tensor product over $A$ is understood.
Berezin Integration and Graded Cyclic Cohomology
================================================
In the present section we apply the general machinery, having been developed for the differential calculus and the integral calculi on a $\mathds{Z}/2$-graded algebra, to one particular example; this is the Gra[ß]{}mann algebra, which figures in quantum field theories with fermions. What we aim at is to investigate whether Connes’ characters over a Gra[ß]{}mannn algebra have something to tell us about the Berezin integral calculus over anticommuting variables. If the question can be answered in the affirmative, then we expect that the general theory of Connes’ characters should support (at least some of) the Berezin rules so that they are put on a firmer ground. If not, then this no go result would say, noncommutative geometry is not capable to include supersymmetric ideas; a disappointing result. As we shall demonstrate, however, Connes’ approach is indeed capable to give us valuable insights into the origin of the Berezin integration rules; similar ideas were enounced by Kastler, but his introductory remark in [@Kast88] does not go beyond the one-dimensional case, which is almost trivial.
Differential calculus on the Gra[ß]{}mann algebra
-------------------------------------------------
Let us recall that the Gra[ß]{}mann algebra, denoted $G^n$, is the associative unital algebra over the reals with $n$ generators $\xi^i$ and relations $$\label{bcch1}
\xi^i\xi^j+\xi^j\xi^i=0\qquad\qquad\qquad\qquad:\, i,j\in\{1,\ldots,n\}.$$ It is isomorphic to the exterior algebra over an $n$-dimensional real vector space, but this isomorphism is not natural; at any rate, for our purposes the definition through generators and relations is the one we need. As a $2^n$-dimensional basis of the underlying vector space one can choose the elements $$\label{bcch2}
\xi^{j_1}\cdots\xi^{j_q}\qquad\qquad\qquad\qquad:\,j_1< \cdots <j_q$$ with $0\leq q\leq n$; the case $q=0$ is meant to signify the unit element, denoted by $1$ in the present case. The algebra $G^n$ is obviously $\mathds{Z}/2$-graded commutative, and so we must specialize the general theory of the last section to the graded commutative situation. Accordingly, the generators $\xi^i$ all have parity $|\xi^i|=1$, and the general element $f\in G^n$ can be written as $$\label{bcch3}
f(\xi)=\sum\limits_{q=0}^n \frac{1}{q\,!}\,f_{j_1\cdots j_q}\xi^{j_1}\cdots\xi^{j_q}$$ with the coefficients $f_{j_1\cdots j_q}\in\mathds{R}$ being totally skew in their indices. We also endow the Gra[ß]{}mann algebra with a $\ast$-structure, though $G^n$ is considered as an algebra over $\mathds{R}$; it is a linear involutive automorphism defined by $(\xi^i)^{\ast}=\xi^i$ and $$\label{bcch3a}
(\xi^{\,j_1}\cdots\xi^{\,j_q})^{\ast}=\xi^{\,j_q}\cdots\xi^{\,j_1}$$ so that its only effect is to reverse the order of factors. Note that most authors use a different convention; first of all, the algebra $A$ is taken over $\mathds{C}$ so that the involution is antilinear, and in the $\mathds{Z}/2$-graded case one sets $(ab)^{\ast}=(-1)^{|a|\,|b|}b^{\ast}a^{\ast}$ for all $a,b\in A$. If this definition is extended to the real case, the $\ast$-involution acts as the identity for the case at hand. Our version, which is also used by DeWitt [@DeWi84], will help to simplify many of the calculations to follow.
For Taylor expansion, one needs the finite difference $\delta f(\xi)=f(\xi+\delta\xi)-f(\xi)$, where $\xi+\delta\xi$ with $\delta\xi^i=\epsilon\xi^i$ is an ‘infinitesimal’ basis transformation, i.e. $\epsilon\ll 1$; whence, the $\delta\xi$ are odd quantities. For the computation it is convenient to introduce the formal partial derivatives $$\label{bcch4}
\frac{\partial}{\partial\xi^j}(\xi^{j_1}\cdots\xi^{j_q})=
\sum_{r=1}^q(-1)^{r-1}\xi^{j_1}\cdots\delta^{j_r}{}_{j}\cdots\xi^{j_q}.$$ which are elements of $\text{Der}(G^n,G^n)$ of parity $|\partial_i|=1$, and one obtains $\delta=\delta\xi^{j}\,\partial/\partial\xi^j$, which is an even operator.
Let us turn to general elements $X$ of $\text{Der}(G^n,G^n)$; they are all outer derivations since $G^n$ is graded commutative. For a proper treatment one needs translations, and for these one must introduce a second Gra[ß]{}mann algebra $G^n$ with the same number of generators, denoted by $\eta^i$; on introducing the graded tensor product $G^n\otimes G^n$, we can identify the $\eta$s and $\xi$s with the elements $\eta^i\otimes 1$ and $1\otimes\xi^i$, respectively, and the graded tensor product guarantees that they anticommute. In this way it makes sense to consider the translated generators $\xi^{\,\prime i}=\xi^i+s^i\eta^i$ with $s^i\in\mathds{R}$, and also the translated element $f(\xi^{\,\prime})\in G^n\otimes G^n$, which is manipulated as follows $$\begin{aligned}
f(\xi^{\,\prime})&=f(\xi)+\int\limits_0^1dt\,\frac{\partial}{\partial t}\,f(\xi+t(\xi-\xi^{\,\prime}))\\
&=f(\xi)+(\xi-\xi^{\,\prime})^i\int\limits_0^1dt\,(\partial_if)(\xi+t(\xi-\xi^{\,\prime}))\\
&=f(\xi)+(\xi-\xi^{\,\prime})^i\,g^{}_i(\xi^{\,\prime}).\end{aligned}$$ We thus obtain $$(Xf)(\xi^{\,\prime})=X\xi^{\,\prime i}\,g^{}_i\,(\xi^{\,\prime})+(-1)^{|X|}
(\xi-\xi^{\,\prime})^iXg^{}_i(\xi^{\,\prime})$$ and letting $\xi^{\,\prime}\to\xi$, we arrive at $$\label{bcch5}
(Xf)(\xi)=X^i(\xi)\partial_if(\xi)\qquad\qquad\qquad\qquad:\,X^i(\xi)=X(\xi^i).$$ We thus have proven that the derivations $X\in\text{Der}(G^n,G^n)$, called Gra[ß]{}mann vector fields, form a free $\mathds{R}$-module of rank $n$, with the partial derivatives $\partial_i$ as basis. Since the Gra[ß]{}mann algebra is graded commutative, the derivations $X$ form a left $G^n$-module, denoted by $V(G^n)$, which can be given the structure of a $\mathds{Z}/2$-graded Lie algebra.
Finite transformations of Gra[ß]{}mann variables are obtained on exponentiating vector fields $X$ of parity $|X|=0$, i.e. all the components $X^i$ must be odd. Since the ordinary (ungraded) commutator obeys $[X,X]=0$, the standard property $e^{X}e^{-X}=e^{-X}e^{X}=1$ of the exponential remains intact, and so the transformation $$\xi^{\prime i}=e^{X(\xi)}\xi^{i}=\xi^{i}+X^i(\xi)+\frac{1}{2}\,X^j(\xi)\partial_jX^{i}(\xi)+\cdots$$ is invertible. Special cases are translations with the vector field $$\label{bcch7}
X_{\eta}(\xi)=\eta^i\frac{\partial}{\partial\xi^i}\qquad\qquad\qquad:\,e^{X_{\eta}(\xi)}\xi=\xi+\eta$$ where the $\eta$s are the generators of a second $G^n$ as discussed above, and for a homogeneous transformation the generating vector field is $$\label{bcch8}
X_{\alpha}(\xi)=\alpha^{i}{}_j\xi^j\frac{\partial}{\partial\xi^i}\qquad\qquad\qquad:\,e^{X_{\alpha}(\xi)}\xi^i=A^i{}_j\xi^j$$ with $\alpha^i{}_j\in\mathds{R}$ and $A(\alpha)=\exp\alpha$ an element of the general linear group in $n$ dimensions. In particular, for $\alpha^{i}{}_j=\lambda\,\delta^{i}{}_j$ one has $$\label{bcch9}
X_{\lambda}(\xi)=\lambda\,\xi^i\frac{\partial}{\partial\xi^i}\qquad\qquad\qquad:
\,e^{X_{\lambda}(\xi)}\xi^{i}=e^{\lambda}\xi^{i}$$ so that one obtains the Gra[ß]{}mann analogue of a dilatation; this was used above for the Taylor expansion.
The importance of the affine transformations $\xi\mapsto A\xi+\eta$ resides in the fact that they constitute the automorphism group of the defining relations, i.e. the property remains intact.
Berezin integration
-------------------
The discussions of the Berezin integration rules on a Gra[ß]{}mann algebra as given in the physical literature are mostly heuristic; if not, they are more or less oriented towards the supersymmetric situation (see [@Bere87] and [@Leit80; @DeWi84; @Vlad84; @Cons94; @Deli99]). In the mathematical literature, treatments of this theme are rare; we are aware of [@Guil99]. The approach given below is related, but different from the one described in [@Vlad84].
A priori, a Berezin ‘integral’ is an $\mathds{R}$-linear map $J:G^n\to\mathds{R}$ only; as such, it is an element of the algebraic dual $(G^n)^{\prime}$, and so the latter linear space has the same dimension as the original $G^n$ itself. The multitude of possible maps $J$ is further narrowed down by the request for translational invariance. So let $s\eta^i$ denote a translation of the generators $\xi^i$ of $G^n$, as explained in the previous paragraph; for notational simplicity, the additional dependence on the real parameter $s$ is suppressed. For $f\in G^n$, the translated algebra element is then defined by $$T(\eta)f(\xi)=f(\xi-\eta).$$ A Berezin ‘integral’ $J$ is said to be translational invariant if the condition $J(T(\eta)f)=J(f)$ is fulfilled, where both $f$ and the translation are arbitrary. Actually, such a condition makes not much sense since the left hand side is $\eta$-valued, and we have not yet defined what this means; the correct version will be given below.
For this purpose, let $A$ and $B$ be two $\mathds{Z}/2$-graded algebras, and $A\otimes B$ their graded tensor product; the latter may be viewed as a left $A$-module under the action $a^{\prime}(a\otimes b)=a^{\prime}a\otimes b$, and also $A$ may trivially be regarded as a left $A$-module. For the case at hand, $A=G^n$ is generated by the $\eta$s and $B=G^n$ by the $\xi$s, the original generators. Consider then an $A$-linear map $J^{\,\prime}:A\otimes B\to A$; this means that $J^{\,\prime}(a\otimes b)=a\,J^{\,\prime}(e^{}_A\otimes b)$ and so $J^{\,\prime}$ is uniquely determined by the values it takes on $B$. Also observe that $J^{\,\prime}|_{B}:B\to A$ is an $\mathds{R}$-linear map with values in $A$. We now impose the further condition $$J^{\,\prime}(e^{}_A\otimes b)=e^{}_A\,J(b)$$ where $J:A\to\mathds{R}$ is a given $\mathds{R}$-linear map, which thus determines $J^{\,\prime}$ uniquely. A proper definition of the invariance condition then reads $$\label{bcch9a}
J^{\,\prime}(T(\eta)f)=J(f)$$ which is a further condition on the admissible maps $J^{\,\prime}$, saying that $J^{\,\prime}$ is independent of translations.
We now claim, the above invariance requirement determines the Berezin ‘integral’ $J$ uniquely; the proof is straightforward. By means of the expansion $$f(\xi)=\sum_If^{}_I\xi_{}^I$$ where the $\xi_{}^I$ denote the basis with $I$ a (lexicographically ordered) multi-index, we have $$J^{\,\prime}(T(\eta)f)=\sum_{I,K\atop |I|+|K|\leq n}\eta_{}^Kf^{}_{K,I}J(\xi_{}^I)$$ on using $G^n$-linearity with respect to the first factor. Consequently, all terms with $|I|+|K|< n$ must vanish, and for $|I|+|K|=n$ only the term with $K=0$ can survive. In total, translational invariance enforces $J(\xi_{}^I)=0$ for $I<n$, or $$\label{bcch9b}
J(\xi^{i_1}\cdots\xi^{i_p})\,=\,\begin{cases}0&:\,p<n\\\varepsilon^{i_1\cdots i_n}&:\,p=n\end{cases}$$ with a suitable normalization; an equivalent form is $$\label{bcch9b'}
J(f)\,=\,\frac{\partial^n}{\partial\xi^n\cdots\partial\xi^1}f(\xi)=f_{1\cdots n}.$$ For a homogeneous transformation $T(A)f(\xi)=f(A^{-1}\xi)$ with $A$ a real $n\times n$-matrix, this entails the transformation law $$\label{bcch9b''}
J(T(A)f)\,=\,\text{det}(A)^{-1}J(f).$$ The defining property and its consequence are the only Berezin integration rules that we accept as valid.
Recall, the map $J$ is simply a special element of the algebraic dual of $G^n$; but it has become customary, and this is the weak point, to rewrite it as a kind of integral: $$\label{bcch9c}
J(f)=\int d^{\,n}\xi\,f(\xi).$$ This rewriting, however, is void of any definite meaning since a precise definition of the ‘volume element’ $d^{\,n}\xi$ is missing. It is the ultimate purpose of the present paper to provide for a rigorous justification of this formal notation.
Fourier transformation and inner products on the Gra[ß]{}mann algebra
---------------------------------------------------------------------
Before turning to the problem just raised, some more technique is needed. In what follows we shall make use of the notational convention , which simply amounts to the instruction to sort out the coefficient $f_{1\cdots n}$ of the term of highest degree, and no more. As an immediate consequence of the rules one obtains the further property $$\label{bcch9d}
\int d^{\,n}\xi\,\partial_if(\xi)=0$$ saying that a ‘boundary integral’ vanishes. We shall also have need for Fourier transformation, defined by $$\label{bcch9e}
Ff(\xi)=\int d^{\,n}\xi^{\,\prime}\,e^{\xi\cdot\xi^{\,\prime}}f(\xi^{\,\prime})$$ where $\xi\cdot\xi^{\,\prime}=\xi^i\delta_{ij}\xi^{\,\prime j}$ with the $\xi^{\,\prime}$s the generators of a further $G^n$; hence, the underlying structure is again the graded tensor product $G^n\otimes G^n$ so that the two sets of generators anticommute. Fourier transformation obeys $F^2=(-1)^{{n+1\choose 2}}\,1$; furthermore, it is related to the conventional Hodge dual, as follows from the explicit form $$\label{bcch9f}
Ff(\xi)=\sum_q\frac{1}{\bar{q}\,!}\,(-1)^{\bar{q}+{\bar{q}\choose 2}}\frac{1}{q\,!}f_{j_1\cdots j_q}\varepsilon^{j_1\cdots j_qj_{q+1}\cdots j_n}\xi_{j_{q+1}}\cdots\xi_{j_n}$$ where $\bar{q}=n-q$ for short, and $\varepsilon^{j_1\cdots j_n}$ the Levi-Civita symbol. We are now ready to introduce an inner product on $G^n$, viewed as a linear space, defined by $$(f|g)=(-1)^{{n\choose 2}}\int d^{\,n}\xi^{\,\prime}(Ff^{\ast})(-\xi^{\,\prime})\,g(\xi^{\,\prime})$$ in which also the algebra structure enters. In the form of a twofold integral it reads $$\label{bcch9g}
(f|g)=(-1)^{{n\choose 2}}\int d^{\,n}\xi^{\,\prime}\int d^{\,n}\xi\,e^{\,\xi\cdot\xi^{\,\prime}}
\,f(\xi)^{\ast}\,g(\xi^{\,\prime})$$ and its explicit version is obtained to be $$\label{bcch9h}
(f|g)=\sum_{q=0}^n\frac{1}{q\,!}\,\bar{f}_{j_1\cdots j_q}\,g^{j_1\cdots j_q}$$ where the raising and lowering of indices is performed with the Kronecker metric. Since in the present context the complexification of $G^n$ is considered, complex conjugation denoted by an overbar gets involved. As is obvious now, the inner product is both hermitean and positive definite.
But it is not the only inner product we shall have need for; another even more natural one is obtained on simply defining $$\label{bcch9i}
\langle f|g\rangle=i^{{n\choose 2}}\int d^{\,n}\xi\,f(\xi)^{\ast}\,g(\xi).$$ It is nondegenerate, and the prefactor is designed so as to guarantee hermiticity, as follows from the explicit form $$\label{bcch9j}
\langle f|g\rangle=i^{{n\choose 2}}\sum_{q=0}^n\frac{(-1)^{{q\choose 2}}}{q\,!\,\bar{q}\,!}\,
\bar{f}_{j_1\cdots j_q}\,\varepsilon^{j_1\cdots j_qj_{q+1}\cdots j_n}g_{j_{q+1}\cdots j_n}.$$ But as opposed to the former, the present inner product, though being nondegenerate, is indefinite.
Both these structures can be subsumed as follows under a common heading. Let us define a grading operator, which should not be confused with the $\mathds{Z}/2$-grading induced by the $\mathds{N}$-grading, through $$\label{bcch9k}
(\gamma f_q)^{j_{q+1}\cdots j_n}=(-i)^{{n\choose 2}}\,\frac{(-1)^{\bar{q}+{\bar{q}\choose 2}}}{q\,!}\,
\varepsilon^{j_{q+1}\cdots j_nj_1\cdots j_q}f_{j_1\cdots j_q}$$ or, using the Hodge dual, by $$\label{bcch9l}
(\gamma f_q)(\xi)=(-i)^{{n\choose 2}}\,(-1)^{{q\choose 2}}(\ast f_q)(\xi).$$ As is straightforward to verify, it indeed obeys $\gamma_{}^2=1$; beyond this, it is unitary with respect to the positive definite inner product , i.e. $$\label{bcch9m}
(\gamma f|\gamma g)=(f|g).$$ The same property is valid for the indefinite inner product , which may now be written by means of the grading operator in terms of the positive definite inner product as $$\label{bcch9n}
\langle f|g\rangle=(f|\gamma g).$$ Hence, through the inner product $(\,\cdot\,|\,\cdot\,)$ the Gra[ß]{}mann algebra $G^n$ can be equipped with a Hilbert space structure; on the other hand, the inner product $\langle\,\cdot\,|\,\cdot\,\rangle$ is indefinite, which becomes manifest in the orthogonal decomposition $G^n=(G^n)_{+}\oplus(G^n)_{-}$ with $\gamma f_{\pm}=\pm f_{\pm}$ since $$\langle f|g\rangle=(f_{+}| g_{+})-(f_{-}| g_{-}).$$ The above structure, namely, a separable Hilbert space on which a unitary grading operator is defined that gives rise for an indefinite inner product according to eq. , is sometimes called a Krein space (see [@Bogn74]). It underlies the construction of an even Fredholm module [@Conn94], and also makes its appearance in the context of ghost fermions [@Gren02] which get involved in the Batalin-Vilkovisky approach to the Becchi-Rouet-Stora-Tyutin quantization of gauge theories (see, e.g., [@Henn92]).
We conclude with the following observation. Let us introduce the operators $\eta^i=\xi^i$ and $\zeta_i=\partial_i$, which both act from the left. They obey the anticommutation relations $$\label{bcch9o}
[\zeta_i,\eta^j]_{+}=\delta_{i}{}^j$$ and may also be viewed as the operators of external and internal multiplication. With respect to the inner product $\langle\,\cdot\,|\,\cdot\,\rangle$, these operators are selfadjoint, i.e. $$\label{bcch9p}
\langle f|\zeta_ig\rangle=\langle\zeta_i f|g\rangle\qquad\qquad\qquad\qquad
\langle f|\eta^ig\rangle=\langle\eta^i f|g\rangle.$$ This property relies on the identity $$\label{bcch9q}
\langle f|\partial_ig\rangle=\langle f|\partial_ig\rangle$$ being valid for any elements $f,g\in G^n$, i. e. without any restriction to homogeneous elements. By contrast, with respect to the inner product $(\,\cdot\,|\,\cdot\,)$, the operators $\eta^i=\xi^i$ and $\zeta_i=\partial_i$ are adjoint to one another, i.e. $$\label{bcch9r}
(f|\zeta^ig)=(\eta^if|g)\qquad\qquad\qquad\qquad
\langle f|\eta^ig\rangle=\langle\zeta^i f|g\rangle$$ where $\zeta^i=\delta^{ij}\zeta_j$. In the latter case, we introduce the further operators $$\label{bcch9s}
\gamma^i=\eta^i+\delta^{ij}\zeta_j$$ which form a Clifford algebra, i.e. $\gamma^i\,\gamma^j+\gamma^j\,\gamma^i=2\delta^{ij}$; also, the $\gamma$s are selfadjoint operators. Restricting ourselves to the case of an even $n$, the representation of the $\gamma$s on the (complexified) Gra[ß]{}mann algebra is not irreducible; it is equivalent to the Dirac representation. But it acts irreducibly on the subspace of even elements. For the indefinite inner product $\langle\,\cdot\,|\,\cdot\,\rangle$, we choose $$\label{bcch9t}
\gamma^i=\zeta^i\qquad\qquad\qquad\gamma^{n+i}=\eta^i$$ with $$\label{bcch9u}
([\gamma^a,\gamma^b]_{+})_{a,b=1,\ldots,2n}=\begin{pmatrix} & 1_n\\1_n& \end{pmatrix}.$$ They obey $\langle f|\gamma^ag\rangle=\langle f|\gamma^ag\rangle$ and form a Clifford algebra with $2n$ generators, with an indefinite metric of signature zero; in this latter case, the representation is irreducible.
Finally, returning to the Clifford algebra with generators , we can devise a Dirac type operator $$\label{bcch9x}
D=\gamma^i\frac{\partial}{\partial\xi^i}$$ which reduces to $D=(\xi^i+\delta^{ij}\partial_j)\partial^{}_i=\xi^i\partial^{}_i$ and thus coincides (cf. ) with the Gra[ß]{}mannian version of Euler’s dilatation operator.
The universal differential bigraded Gra[ß]{}mann algebra
--------------------------------------------------------
Let us return to the problem raised in the last but one subsection, namely, whether the ‘volume element’ $d^{\,n}\xi$ that enters the formal Berezin integral, can be given a precise meaning. We approach this problem on taking recourse to the general construction of the universal differential graded algebra of a general noncommutative algebra; this will supply for the necessary abstract tools to investigate these questions.
We have discussed the universal differential calculus for a graded commutative algebra $A$ in Subsec. \[Thegradedcommutativecase\]; the results obtained there are now applied to the situation where $A$ is defined through generators and relations, as it is the case for the Gra[ß]{}mann algebra.
As we know, in the graded commutative case one must pass to the $G^n$-bimodule $\Omega_{}^1(G^n)$ of $1$-forms. This is a free module of rank $n$, for which the differentials $d\xi^i$ form basis; their parity is zero. As should be noted, there is no conflict with the fact that finite differences have parity one; the $\delta\xi^i$ and the differentials $d\xi^i$ are essentially different constructs. Differential forms of order $p$ are obtained on forming the graded symmetrized product $\Omega_{}^p(G^n)=\overset{p}{\vee}_{G^n}\Omega_{}^1(G^n)$ so that the $d\xi^i$ commute: $$\label{bcch10}
d\xi^i\,d\xi^j=d\xi^j\,d\xi^i.$$ This assignment guarantees the operator of exterior differentiation to have parity $|d|=1$; for $f\in G^n$ we then obtain $$d^{\,2}f(\xi)=d(\partial_jf\,d\xi^j)=\partial_i\partial_jf\,d\xi^id\xi^j=0$$ since the partial differentiations anticommute and the differentials commute, which is the result one wants. Hence, for the construction of the basis for the module of $p$-forms we must select the graded symmetrized tensor product $$\label{bcch11}
d\xi^{i_1}\vee\cdots\vee d\xi^{i_p}=\frac{1}{\sqrt{p\,!}}\,\sum_{\sigma\in S_n}
d\xi^{i_{\sigma(1)}}\otimes\cdots\otimes d\xi^{i_{\sigma(p)}}$$ that here coincides with the standard symmetrization. But the graded version is needed for a general $p$-form, which is a finite sum $$\label{bcch12}
\omega_p=\sum f^{}_0\,df^{}_1\vee\cdots\vee df_p$$ where $f_i\in G^n$; the symbol $\vee$ signifying the symmetrization will often be suppressed. We thus can write down the general $p$-form $$\label{bcch13}
\omega_p(\xi)=\frac{1}{p\,!}\,\omega_{i_1\cdots i_p}(\xi)\,d\xi^{i_1}\vee\cdots\vee d\xi^{i_p}$$ with $$\label{bcch14}
\omega^{}_{i_1\cdots i_p}(\xi)=\frac{1}{q\,!}\,\omega_{i^{}_1\cdots i_p,j^{}_1\cdots j_q}\,\xi^{j^{}_1}\cdots
\xi^{j_q}$$ where $\omega_{i^{}_1\cdots i_p,j^{}_1\cdots j_q}$ is completely symmetric in the indices $i^{}_1,\ldots,i_p$, and completely antisymmetric in the indices $j^{}_1\ldots,j_q$; furthermore, the operator of exterior differentiation $d$ acts on $\omega_p$ via $$\label{bcch15}
d\omega_p(\xi)=\frac{1}{p\,!}\partial_{i_0}\omega_{i_1\cdots i_p}(\xi)\,d\xi^{i_0}\vee d\xi^{i_1}\vee\cdots\vee d\xi^{i_p}$$ and obeys $d\circ d=0$ by construction. To resume, we now have available the universal differential algebra $\Omega^{\displaystyle\cdot}(G^n)$; the general element of $\Omega^p(G^n)$ may be written as a finite sum of terms $f_0\,df_1\vee\cdots\vee df_p$ with parity $$\label{bcch16}
|f_0\,df_1\vee\cdots\vee df_p|=p+\sum_{i=0}^p|f_p|.$$ Furthermore, $\Omega^{\displaystyle\cdot}(G^n)$ is graded commutative, i.e. $[\omega^{}_{p},\omega^{\,\prime}_{p^{\,\prime}}]=0$ for all $\omega^{}_{p}\in
\Omega^p(G^n)$ and $\omega^{\,\prime}_{p^{\,\prime}}\in
\Omega^{p^{\,\prime}}(G^{n})$, since $G^n$ is.
Let us compare the present situation with that of conventional differential forms over a smooth manifold $M$; in this case, where $$\omega_p(x)=\frac{1}{p\,!}\omega_{i_1\cdots i_p}(x)\,dx^{i_1}\wedge\cdots\wedge dx^{i_p}$$ with $x\in M$, the coefficient functions $\omega_{i_1\cdots i_p}(x)$ contain arbitrary powers of the $x^i$, whereas the $\omega_p$ vanish for all $p>n$ because the differentials anticommute. By contrast, in the Gra[ß]{}mann case the expansion of the coefficient functions terminate for powers higher than $n$, but there is no restriction on the degree $p$ of the differential forms since the differentials $d\xi^i$ commute.
We have already discussed the left $G^n$-module $V(G^n)$ of vector fields, which is free of rank $n$, with the partial derivatives $\partial_i$ forming a basis. Also in this case we can form tensor products, which we denote by $\Omega_p(G^n)$, with $\Omega^{}_0(G^n)=G^n$ and $\Omega^{}_1(G^n)=V(G^n)$; but before doing so we must specify the parity of the partial derivatives. We expect the relevant construction to be the graded symmetrization $\Omega_p(G^n)=\overset{p}{\vee}_{G^n}\Omega_1(G^n)$; in order to have $\partial_i\vee\partial_j=\partial_j\vee\partial_i$, this requirement enforces the assignment $|\partial_i|=0$, contrary to naive expectation, so that $$\label{bcch17}
\frac{\partial}{\partial\xi^{i_1}}\vee\cdots\vee\frac{\partial}{\partial\xi^{i_p}}=\frac{1}{\sqrt{p\,!}}\,\sum_{\sigma\in S_n}
\frac{\partial}{\partial\xi^{i_{\sigma(1)}}}\otimes\cdots\otimes
\frac{\partial}{\partial\xi^{i_{\sigma(p)}}}.$$ For the general element $\chi^p\in\Omega_p{G^n}$, called a Gra[ß]{}mann *current*, we are then guaranteed that the $G^n$-valued tensor components $\chi^{i^{}_1\cdots i_p}$ in the expansion $$\label{bcch18}
\chi^p=\frac{1}{p\,!}\,\chi^{i^{}_1\cdots i_p}\partial_{i^{}_1}\vee\cdots\vee\partial_{i_p}$$ are symmetric in their indices.
The construction of the module of $p$-forms $\Omega^p(G^n)$ and the space of $p$-currents $\Omega_p(G^n)$ is quite different; it is for that reason why we have not defined the differentials $d\xi^i$ as the dual basis of the $\partial_i$, the partial derivatives. After the fact, however, we can define a pairing between these spaces. For this, let $A$ be an involutive $\mathds{Z}/2$-graded algebra; given two left $A$-modules $M$ and $N$, both being equipped with compatible $\mathds{Z}/2$-gradings, we define an A-valued pairing of $M$ and $N$ as a sesquilinear map $\langle\,\cdot\,|\,\cdot\,\rangle:M\times N\to A$, subject to the condition $$\label{bcch19}
\langle av|bw\rangle=a^{\ast}\langle v|w\rangle b\qquad\qquad\qquad:\,a,b\in A;\,\,v\in M,\,\,w\in N.$$ In case that $M=N$, one can then define a hermitean structure (see, e.g. [@Grac01]). For the construction of a $G^n$-valued pairing $\langle\,\cdot\,|\,\cdot\,\rangle:\Omega_p(G^n)\times\Omega^p(G^n)\to G^n$, we begin with $p=1$ and set $\langle \partial_i|d\xi^j\rangle=\delta_i{}^j$ by definition. The extension to values of $p\geq 1$ is thus obtained to be $$\label{bcch20}
\langle\frac{\partial}{\partial\xi^{i_1}}\vee\cdots\vee\frac{\partial}{\partial\xi^{i_p}}|
\,d\xi^{j_1}\vee\cdots\vee d\xi^{j_q}\rangle=\delta_p{}^q\sum_{\sigma}\delta_{i_{\sigma(1)}}{}^{j_1}\cdots
\delta_{i_{\sigma(p)}}{}^{j_p}$$ where the normalization made in the eqs. and proves to be essential. In this way, the $G^n$-valued pairing is given by $$\label{bcch21}
\langle\chi^p(\xi)|\omega_p(\xi)\rangle=\frac{1}{p\,!}\chi^{i_1\cdots i_p}(\xi)^\ast\,
\omega_{i_1\cdots i_p}(\xi)$$ which can be made a nondegenerate $\mathds{R}$-valued pairing $\langle\,\cdot\,|\,\cdot\,\rangle:\Omega_p(G^n)\times\Omega^p(G^n)\to \mathds{R}$ by means of the Berezin integral: $$\label{bcch22}
\langle\chi^p|\omega_p\rangle=\int d^{\,n}\xi\,\langle\chi^p(\xi)|\omega_p(\xi)\rangle.$$ As opposed to the inner product , we here omit the prefactor since from now on $G^n$ is considered as an algebra over the reals. The basic identity , i.e. $\langle\chi^p|\partial_i\omega_p\rangle=\langle\partial_i\chi^p|\omega_p\rangle$ remains valid, by means of which many of the computations to follow are drastically simplified. Hence, the space of $p$-currents is the dual of the space of $p$-forms.
In addition, the space of $p$-forms $\Omega^p(G^n)$ can be equipped with a positive definite inner product. For this, we introduce a metric on $\Omega^p(G_{}^1)$ by $$\label{bcch23}
\langle d\xi^i|d\xi^j\rangle=\delta^{ij}$$ with the extension to the basis of $\Omega^p(G^n)$ being obtained by similar manipulations as above; in this way, we can freely raise and lower indices. We now define the scalar product by $$\label{bcch24}
(\omega_p|\lambda_p)=(-1)^{{n\choose 2}}\int d^{\,n}\xi^{\,\prime}\int d^{\,n}\xi\,
e^{\,\xi\cdot\xi^{\,\prime}}
\frac{1}{p\,!}\,\chi^{i_1\cdots i_p}(\xi)^\ast\,
\omega_{i_1\cdots i_p}(\xi^{\,\prime})$$ for $\omega_p,\lambda_p\in\Omega^p(G^n)$, which is to be compared with eq. above. Note that the decoration with additional indices does not alter the conclusions drawn there, but the relation is no longer available.
The pairing and the inner product give rise to two operations, deriving from the operator of exterior differentiation. Beginning with the pairing , we define the dual $\bar{d}$ of $d$ by $$\label{bcch25}
\langle\bar{d}\chi^p|\omega_p\rangle=\langle\chi^p|d\omega_p\rangle.$$ which is calculated to be $$\label{bcch26}
\bar{d}\chi^p(\xi)=\frac{1}{(p-1)!}\,\partial_{i_1}\chi^{i^{}_1i^{}_2\dots i_p}(\xi)
\,\partial_{i^{}_2}\vee\cdots\vee \partial_{i_p}.$$ This operator, a divergence which is also denoted below as $\partial=\bar{d}$, obeys $\bar{d}^{\,2}=0$ and is of degree minus one. Turning to the inner product , the adjoint $d^{\ast}$ of $d$ is defined by $$\label{bcch28}
(\omega_{p}|d\lambda_{p-1})=(d^{\ast}\omega_{p}|\lambda_{p-1})$$ and obeys $(d^{\ast})^2=0$; the calculation yields $$\label{bcch29}
d^{\ast}\omega_{p}(\xi)=\frac{1}{(p-1)!}\,\xi^{i^{}_1}
\omega_{i^{}_1i^{}_2\cdots i_p}(\xi)\,d\xi^{i^{}_2}\vee\cdots\vee d\xi^{i_p}.$$ We write in the form of a contraction, i.e. $\iota^{}_D=d^{\ast}$, with $D=\xi^i\partial/\partial\xi^i$ the Euler vector field generating dilatations. For $X$ a general vector field of zero parity, let us introduce the contraction $$\label{bcch31}
\iota^{}_X\omega_{p+1}(\xi)=\frac{1}{p\,!}\,X^{i^{}_0}\omega_{i^{}_0i^{}_1\cdots i_p}(\xi)
\,d\xi^{i^{}_1}\vee\cdots\vee d\xi^{i_p}.$$ We then define the Lie derivative $L^{}_X$ by the Cartan formula $L^{}_X=\iota^{}_Xd+d\iota^{}_X$, the explicit form of which is $$\label{bcch32}
L^{}_X\omega_{p}=\left(\frac{1}{p\,!}\,X^{i}\partial_i\,\omega_{i^{}_1\cdots i_p}+
\frac{1}{(p-1)!}\,\partial_{i^{}_1}X^i\omega_{ii^{}_2\dots i_p}\right)
d\xi^{i^{}_1}\vee\cdots\vee d\xi^{i_p}.$$ In particular, for the Euler vector field $D$ this gives $$\label{bcch33}
L^{}_D\omega_{p}=\left(D+p\right)\omega_{p}.$$ In the conventional de Rham case, the Cartan formula is instrumental for a transparent proof of the Poincar' e lemma (see [@Bott82]); this is as well the case in the present situation. The strategy of the proof consists in extracting a homotopy operator from the Cartan formula; the calculation is almost the same as in the classical situation and yields that an exact $p$-form $\omega_p$ can be written as $\omega_p=d\alpha_{p-1}$, with $$\label{bcch34}\alpha_{p-1}(\xi)=\int_0^1\,dt\frac{t^{p-1}}{(p-1)!}
\,\xi^{i^{}_1}\,\omega_{i^{}_1i^{}_2\cdots i_p}(t\xi)\,d\xi^{i^{}_2}\vee\cdots\vee d\xi^{i_p}.$$ Whence, the complex $(\Omega^{\displaystyle\cdot}(G^n),d)$ is acyclic.
Observe that the classical case over the manifold $\mathds{R}^n$ and the Grassmann case have some similarities, but are also essentially different. This can be seen on having a closer look at the operator $\bar{d}$, defined in , which comes from the pairing ; it can also be interpreted as acting on $\Omega^p(G^n)$ by means of the metric. Together with $d$ it gives rise to a rather unique candidate for a Laplacean type of operator, namely the square of the Dirac-Hodge-Kähler like operator $\not{\!d}=\bar{d}+d$; but $\bar{d}\,d+d\,\bar{d}$ vanishes since it is equal to $g^{ij}\partial_i\partial_j$ so that in the Gra[ß]{}mann case all $p$-forms are trivially harmonic. Turning to the positive definite inner product with the adjoint operator $d^{\ast}$, it permits to introduce the quadratic operator $d^{\ast}d+d\,d^{\ast}$; as we have seen, however, this is the Lie derivative $L^{}_D$, which has not much in common with a Laplace operator. Nevertheless, we can prove the following analogue of the Hodge decomposition theorem $$\Omega^{p}(G^n)=d\,\Omega^{p-1}(G^n)\oplus d^{\ast}\Omega^{p+1}(G^n)$$ where, similar to the classical situation over the flat manifold $\mathds{R}^n$, the contribution of the harmonic forms is absent. However, whereas the conventional proof requires elliptic operator theory [@Tayl96], it is purely algebraic in the present case and follows standard arguments, which we omit.
In concluding this section, let us note that one can also define an external multiplication on $p$-forms by $\varepsilon(d\xi^i)=d\xi^i\vee$ which together with the internal multiplication $(\iota^{}_X=\iota(X))$ operator $\iota(\partial/\partial\xi^i)$ obey the commutation relations $$[\iota\left(\frac{\partial}{\partial\xi^i}\right),\varepsilon(d\xi^j)]^{}_{-}=\delta_i{}^j.$$ Hence, $\iota(\partial/\partial\xi^i)$ and $\varepsilon(d\xi^i)$ may be viewed as bosonic annihilation and creation operators; this is just the opposite situation as compared to the de Rham case, where the analogues $\iota(\partial/\partial x^i)$ and $\varepsilon(dx^i)$ are fermionic operators, with $[\iota(\partial/\partial x^i),\varepsilon(dx^j)]^{}_{+}=\delta_i{}^j$ an anticommutator.
Connes’ integral calculi on a Gra[ß]{}mann algebra
==================================================
Having available the exterior differential calculus over a Gra[ß]{}mann algebra, we can now turn to the discussion of its Connes’ characters [@Conn94]. These are maps $\int:\Omega^p(G^n)\to \mathds{R}$ subject to the conditions $$\label{bcch35b-d}
\int d\omega_{p-1}=0\qquad\qquad\qquad\text{and}\qquad\qquad\qquad
\int[\omega_r,\omega_{p-r}]=0\quad:\,p\geq r.$$ Some authors also include $\int \omega_{p}=0$ for all $p<n$ amongst the definitions, but we do not. Since $\Omega^{\displaystyle\cdot}(G^n)$ is graded commutative, the second of the conditions is void; the protection of the map $\int$ is not needed. As proven, $p$-dimensional characters are in bijective correspondence with normalized cyclic Hochschild $p$-cocycles, and so their classification can be reduced to the normalized cyclic cohomology $HC^{\displaystyle\cdot}(G^n)$; recall that we use the different notation $H^{\displaystyle\cdot}_\lambda(G^n)$ for the cyclic cohomology. We repeat the relevant definitions, particularized to the case at hand; the boundary operator (see ) acts on a cochain $\varphi_p\in C^p(G^n)$ as $$\label{bcch35e}
b\,\varphi(f_0,f_1,\ldots,f_{p})=$$ $$\nonumber
\sum_{i=0}^{p-1}(-1)^{i+\sum\limits_{j=0}^i|f_j|}\varphi(f_0,\ldots,f_if_{i+1},\ldots,f_{p})
-(-1)^{(1+|f_{p}|)((p-1)+\sum\limits_{j=0}^{p-1}|f_j|)}\varphi(f_{p}\,f_0,f_1,\ldots,f_{p-1})$$ and the cyclicity operator (see ) as $$\label{bcch35f}
\lambda\,\varphi(f_0,f_1,\ldots,f_{p})=(-1)^{(1+|f_{p}|)(p+\sum\limits_{j=0}^{p-1}|f_j|)}
\varphi(f_{p},f_0,f_1,\ldots,f_{p-1}).$$ Finally, the space of cochains $H^p(G^n)$ is equipped with a $\mathds{Z}/2$-grading; the grading operator is $$\label{bcch35g}
\varepsilon\,\varphi(f_0,f_1,\ldots,f_{p})=(-1)^{p+\sum\limits_{i=0}^{p}|f_i|)}
\varphi(f_0,f_1,\ldots,f_{p})$$ which anticommutes with both the boundary operator and the cyclicity operator. We reiterate, for emphasis, our sign conventions in and differ from those commonly used in the literature.
Let us go through some examples; for the time being, the normalization condition is dispensed. We begin with the case $p=0$ for arbitrary $n$; the $\mathds{R}$-linear maps $\varphi:G^n\to\mathds{R}$ are just the elements of the algebraic dual of $G^n$, which has the same dimension $2^n$ as $G^n$ itself. Introducing the notation $I=(i_1,\ldots,i_p)$ for the ordered $p$-tupel with $i_1 < \cdots < i_p$, the dual basis $\varepsilon_I$ is defined by $\varepsilon_I(\xi^{I^{\prime}})=\delta_I{}^{I^{\prime}}$; for the unit $1$ we set $I=0$ so that $\varepsilon_{0}(1)=1$, and zero otherwise. A particular basis element is given by $\varepsilon_{1,\ldots,n}$ with the properties $\varepsilon_{1,\ldots,n}(\xi^1,\ldots,\xi^n)=1$ and $\varepsilon_{1,\ldots,n}(\xi^{i_1},\ldots,\xi^{i_p})=0$ for all $0\leq p<n$; it reproduces the Berezin integral $J(f)$ since $$\label{bcch35h}
\varepsilon_{1,\ldots,n}(f)=f_{1\cdots n}.$$ As we know, amongst the linear maps $J:G^n\to\mathds{R}$ it is singled out by translational invariance; beyond this mere recognition, however, no new insight can be drawn from this observation.
Next, we investigate the case $n=1$ with $p=1$. As one easily verifies, there is only one single cyclic $1$-cocycle $\varphi=\varepsilon_{1,1}$; hence, it is automatically normalized. In terms of the associated character, we thus have $$\label{bcch35j}
\varphi(f,g)=\int\,f\,dg=f_1g_1$$ where $f(\xi)=f_0+f_1\xi$ and $g(\xi)=g_0+g_1\xi$, and so the outcome $$\label{bcch35k}
\int\,d\xi=0\qquad\qquad\qquad\int\,\xi\,d\xi=0$$ is identical with the Berezin rules; this observation has already been made by Kastler (see the introduction in [@Kast88]), but the case $n=1$ with $p=1$ is rather special, as will be seen below.
The next and last case we consider is $n=2$ with $p\leq 2$, from which the general strategy will then be read off. This specific example was also treated by Coquereaux $\&$ Ragoucy [@Coqu95] (cf. [@Coqu90; @Coqu91]), but our intention is different. Apart from the fact that we view $G^n$ as an algebra over $\mathds{R}$, and not over $\mathds{C}$, these authors disregard the normalization condition. The really essential point, however, is that we do not use the standard Koszul-Milnor-Quillen sign rules; though one would suspect this to make no essential difference, it will turn out to be of crucial importance, and so we are obliged to give some details. The case $p=0$ has already been commented upon, and so we can turn to $p=1$; with the notation $f(\xi)=f_0+f_1\xi^1+f_2\xi^2+f_{12}\xi^{12}$, the cyclic cocycles are: $$\begin{aligned}
\begin{split}\label{bcch35l}
\chi_{1,1} &=\varepsilon_{1,1} \\
\chi_{1,2} &=\varepsilon_{1,2}+\varepsilon_{2,1} \\
\chi_{2,2} &=\varepsilon_{2,2} \\
\chi_{1,12} &=\varepsilon_{1,12}+\varepsilon_{12,1} \\
\chi_{2,12} &=\varepsilon_{2,12}+\varepsilon_{12,2}.
\end{split}\end{aligned}$$ Why we have chosen the symbol $\chi$ for these coycles will become apparent in the next paragraph. Thus, the normalized cyclic cohomology group is $$\label{bcch35m}
HC^1(G^2)=\mathds{R}^3\oplus \mathds{R}^2$$ where the first direct summand denotes the even and the second the odd cocycles; none of these cocycles can be written as coboundaries. Proceeding to $p=2$, which we expect to be the case of interest for the purposes we have in mind, one finds the cyclic cocycles: $$\begin{aligned}
\begin{split}\label{bcch35n}
\varphi_{0,0,0} &=\varepsilon_{0,0,0} \\
\varphi_{0,0,1} &=\varepsilon_{0,0,1}-\varepsilon_{0,1,0}+\varepsilon_{1,0,0} \\
\varphi_{0,0,2} &=\varepsilon_{0,0,2}-\varepsilon_{0,2,0}+\varepsilon_{2,0,0} \\
\varphi_{0,0,12} &=\varepsilon_{0,0,12}+\varepsilon_{0,12,0}+\varepsilon_{12,0,0} \\
\chi_{1,1,1} &=\varepsilon_{1,1,1} \\
\chi_{2,2,2} &=\varepsilon_{2,2,2} \\
\chi_{1,1,2} &=\varepsilon_{1,1,2}+\varepsilon_{1,2,1}+\varepsilon_{2,1,1} \\
\chi_{2,2,1} &=\varepsilon_{2,2,1}+\varepsilon_{2,1,2}+\varepsilon_{1,2,2} \\
\chi_{1,1,12}&=\varepsilon_{1,1,12}+\varepsilon_{1,12,1}+\varepsilon_{12,1,1}\\
\chi_{2,2,12}&=\varepsilon_{2,2,12}+\varepsilon_{2,12,2}+\varepsilon_{12,2,2}\\
\chi_{1,2,12}&=\varepsilon_{1,2,12}+\varepsilon_{1,12,2}+\varepsilon_{12,1,2}+
\varepsilon_{2,1,12}+\varepsilon_{2,12,1}+\varepsilon_{12,2,1}\,.
\end{split} \end{aligned}$$ In total, there are $11$ nontrivial cocycles, i.e. $Z^2_{\lambda}(G^2)=\mathds{R}^5\oplus\mathds{R}^6$; but among these there are some that may be written as boundaries, namely $$\begin{aligned}
{5}
\varphi_{0,0,1} &=&-b\,\varphi_{0,1} \\
\varphi_{0,0,2} &=&-b\,\varphi_{0,2} \\
\varphi_{0,0,12} &=&-b\,\varphi_{0,12}\end{aligned}$$ where the $\varphi$s on the right are all elements of $C^1_{\lambda}(G^n)$. Thus the cyclic cohomology group $H^2_{\lambda}=Z^2_{\lambda}/B^2_{\lambda}$ is $H^2_{\lambda}(G^2)=\mathds{R}^4\oplus\mathds{R}^4$; note that $\varphi_{1,1,1}$ may not be written as a boundary $b\,\varphi_{1,1}$ since a cyclic $\varphi_{1,1}\in C^1_{\lambda}(G^2)$ vanishes. Passing to normalized cyclic cocycles, also $\varphi_{0,0,0}$ is eliminated, and so we end up with $$\label{bcch35o}
HC^2(G^2)=\mathds{R}^3\oplus \mathds{R}^4.$$ Observe, with our conventions all elements of $HC^p(G^2)$ with $p\leq 2$ are completely symmetric in their labels. Had we chosen instead the standard sign conventions and , as in [@Coqu95] and elsewhere, then e.g. in the last cocycle $\chi_{1,2,12}$ the contributions $\varepsilon_{1,12,2}$ and $\varepsilon_{2,12,1}$ would acquire a minus sign in front; this innocuous alteration has drastic consequences, as will be shown now.
What we claim is that only our sign rules, as opposed to those of Koszul-Milnor-Quillen, are compatible with the defining properties of a Connes’ character; this claim at least holds in the present case of a graded commutative algebra. For the proof, we look at the last normalized cyclic cochain in , which is of special relevance: $$\label{bcch35p}
\chi_{1,2,12}(f,g,h)=(f_{1}g_{2}h_{12}+f_{2}g_{1}h_{12})\pm(f_{1}g_{12}h_{2}+f_{2}g_{12}h_{1})+
(f_{12}g_{1}h_{2}+f_{12}g_{2}h_{1}).$$ Here, the upper signs are obtained with our sign rules, and the lower ones are those of Coquereaux $\&$ Ragoucy [@Coqu95] and others. We evaluate this cocycle on the basis elements $\xi^1,\xi^2$ and $\xi^{12}=\xi^1\xi^2$ in two different orderings: $$\label{bcch35q}
\chi_{1,2,12}(\xi^{12},\xi^1,\xi^2)=1\qquad\qquad\qquad\chi_{1,2,12}(\xi^{1},\xi^{12},\xi^{2})=
\pm 1$$ Now we invoke that $\chi_{1,2,12}$ may also be viewed as the character $\chi_{1,2,12}(f,g,h)=\int_{1,2,12} f\,dg\,dh$, and so we can continue with $$\label{bcch35r}
\chi_{1,2,12}(\xi^{12},\xi^1,\xi^2)=1=\int_{1,2,12}\xi^{1}\xi^{2}d\xi^1d\xi^2$$ in the first case, and in the second with $$\label{bcch35s}
\chi_{1,2,12}(\xi^{1},\xi^{12},\xi^{2})=
\pm 1=\int_{1,2,12}\xi^{1}d\xi^{12}d\xi^{2}=
\int_{1,2,12}\xi^{1}(d\xi^{1}\xi^2-\xi^{1}d\xi^2)d\xi^{2}=\int_{1,2,12}\xi^{1}\xi^2d\xi^{1}d\xi^{2}$$ since for the last term in round brackets the integrand vanishes; whence, these two results are compatible for the upper sign only, as contended.
It will prove to be important for the further development that, as eq. exhibits, the above character may also be interpreted as the Berezin integral for $n=2$; one thus expects that the Berezin integral calculus over $G^n$ will intimately be related to the theory of Connes’ characters over that algebra. Before we can turn to a closer analysis of this conjecture, however, we need a further concept to be introduced next.
Gra[ß]{}mann currents
=====================
We again assume, in an intermediate step, the (algebraic) Berezin rules to be given beforehand; the only ones we accept are (see ) $$\label{bcch36}
\int d^{\,n}\xi\,\xi^{i_1}\cdots\xi^{i_p}\,=
\begin{cases}
0&:\,p<n\\
\varepsilon^{i_1\cdots i_n}&:\,p=n
\end{cases}$$ where the position of the ‘volume element’ $d^{\,n}\xi$, which still awaits a proper definition, does not matter. Let us return to the pairing; hence, a metric is not needed. So recall, the dual $\Omega_p(G^n)$ of the space of $p$-forms $\Omega^p(G^n)$ is built from $p$-currents through (see ) $$\label{bcch37}
\langle\chi^p|\omega_p\rangle=\int\,d^{\,n}\xi\,
\frac{1}{p\,!}\chi^{i_1\cdots i_p}(\xi)^\ast\,
\omega_{i_1\cdots i_p}(\xi)$$ where the components $\chi^{i_1\cdots i_p}$ of a $p$-current are completely symmetric in their indices. In the classical situation, the dual of $\Omega^p(C^{\infty}(M))$ over a smooth manifold $M$ is the space of de Rham currents (see [@Dieu72]); in particular, for $p=0$ the dual includes the distributions. But for the case at hand such finesses are not needed. The adjoint $\bar{d}$ of the operator $d$, now being denoted as the divergence $\partial$, is $$\label{bcch38-43}
(\partial\chi^p)^{i_2\cdots i_p}=\partial_{i_1}\chi^{i_1i_2\cdots i_p}.$$ Accordingly, a current is said to be closed if $\partial\chi^p=0$, i.e. if its divergence is zero. We thus have available the dual $(\Omega^{}_{\displaystyle\cdot}(G^n),\partial)$ of the complex $(\Omega_{}^{\displaystyle\cdot}(G^n),d)$; since the latter is acyclic, also the complex of currents is. This is a general property, namely, given a cochain complex $(C_{}^{\displaystyle\cdot},d)$ over the real or complex numbers which is acyclic, then also the dual chain complex $(C^{}_{\displaystyle\cdot},\bar{d})$ is acyclic, and conversely.
We expect the latter proposition to be known to the experts; however this may be, the proof is sufficiently simple to be given. So let $V$ be linear spaces over $\mathds{R}$ or $\mathds{C}$, and $V^{\,\prime}$ its dual. For $U\leq V$ any subspace of $V$, the orthogonal complement $U^{\perp}$ of $U$ in $V^{\,\prime}$ is defined by $U^{\perp}=\{\varphi\in V^{\,\prime}|\varphi(u)=0\,\,\mathrm{for\,\,all}\,\,u\in U\}$, and similarly for a subspace $U^{\,\prime}\leq V^{\,\prime}$; one also has $U^{\perp\perp}=U$. Furthermore, if $\phi: V\to W$ is a linear mapping into a second vector space $W$ and $\phi^{\,\prime}: W^{\,\prime}\to V^{\,\prime}$ its dual, then the following identities $(\mathrm{Ker}\,\phi^{\,\prime})^{\perp}=\mathrm{Im}\,\phi$ and $(\mathrm{Ker}\,\phi)^{\perp}=\mathrm{Im}\,\phi^{\,\prime}$ hold. Suppose now that the sequence of linear maps $U\overset{\chi}{\longrightarrow}V\overset{\phi}{\longrightarrow}W$ is exact at $V$; then also the dual sequence $U^{\,\prime}\underset{\chi^{\ast}}{\longleftarrow}V^{\,\prime}\underset{\phi^{\,\prime}}{\longleftarrow}
W^{\,\prime}$ is exact at $V^{\,\prime}$. In fact, since $\mathrm{Im}\,\chi=\mathrm{Ker}\,\phi$ we have $\mathrm{Ker}\,\phi^{\,\prime}=(\mathrm{Im}\,\phi)^{\perp}=(\mathrm{Ker}\,\phi)^{\perp}=
\mathrm{Im}\,\phi^{\,\prime}$ as was to be proven.
From currents, one can construct associated Hochschild cochains in a rather obvious way (cf. also [@Coqu95]); we simply set $$\label{bcch44}
\chi(f_0,\ldots,f_p)=\chi(f_0\,df_1\vee\cdots\vee df_p)=
\int\,d^{\,n}\xi\,(\chi^{i_1\cdots i_p})^{\ast}f_0\,\partial_{i_1}f_1\cdots\partial_{i_p}f_p.$$ Here the first equality sign is meant as a definition, and the second makes sense since the left hand side is graded symmetric in the $f_i$ with $i\in\{1,\ldots,n\}$, as is the right hand side because $\chi^{i_1\cdots i_p}$ is symmetric in the ordinary sense; whence, the graded symmetric tensor product defined in Subsec. \[Thegradedcommutativecase\] enters in an essential way.
According to the definition , the boundary of the Hochschild cochain is given by $b\,\chi(\omega_p)=\chi(b(\omega))$ where $(\chi|\omega)=\chi(\omega)$ is the evaluation map, and since $\Omega^pG^n$ is graded commutative, $\chi(f_0,\ldots,f_p)$ is a cocycle. Note that this conclusion is rather immediate since our definition relies on the property , which is not shared by the alternative choice used in the literature. Hence, if a Hochschild cochain may be represented through a current, then this cochain is a cocycle.
Let us restrict our considerations to closed currents in what follows; we then show that the associated cocycles are also cyclic. As for the proof, since $\Omega^p(G^n)$ algebra is graded commutative, we have $$\begin{aligned}
\chi(f_0\,df_1\vee\cdots\vee df_p)&=(-1)^{|df_p|(|f_0|+\sum\limits_{i=1}^{p-1}|df_i|)}\chi(df_p\vee f_0\,df_1
\vee\cdots\vee
df_{p-1})\\
&=(-1)^{|df_p|(|f_0|+\sum\limits_{i=1}^{p-1}|df_i|)}\chi(d(f_pf_0)\vee df_1\vee\cdots
\vee df_{p-1}-(-1)^{|f_p|}f_pdf_0\vee df_1\vee\cdots\vee df_{p-1})\end{aligned}$$ where, since $\chi^p$ is closed, the first term vanishes. We thus end up with $$\chi(f_0\,df_1\vee\cdots\vee df_p)=(-1)^{|df_p|\sum\limits_{i=0}^{p-1}|df_i|}\chi(f_p\,df_0\vee\,df_1\vee
\cdots\vee df_{p-1})$$ and comparing this with the definition of the cyclicity operator $\lambda$, we conclude that $\lambda\chi^p=\chi^p$ as claimed.
Finally, the Hochschild cocycles associated to closed currents are also normalized. Indeed, the Berezin rules imply $\chi(1,f_1,\ldots,f_p)=0$, and cyclicity then entails the normalization property.
To summarize, via the map $\chi^p:\Omega^p(G^n)\to\mathds{R}$ a closed $p$-current determines a $p$-dimensional character over the Gra[ß]{}mann algebra.
What one would like to have then is, every character may be represented by a closed current. In the classical de Rham case over a compact manifold, Connes ([@Conn85]) has proven that this is indeed true. But in the graded commutative case, the corresponding proof does not go through (contrary to the claim made in [@Coqu95]) since in $$\chi(f_0\,df_1\vee\cdots\vee df_p)=
\int\,d^{\,n}\xi\,\,(\chi^{i_1\cdots i_p})^{\ast}\,\frac{1}{p\,!}\,\sum_{\pi\in S_p}f_0\,\partial_{i_{\pi(1)}}f_1\cdots
\partial_{i_{\pi(p)}}f_p$$ the symmetrization may not be shifted to the $f_i\,$s with $i\in\{1,\ldots,n\}$ because the latter only commute in the graded sense.
In order to gain a feeling for what one may expect, let us return to the example with $n=2$ and $p\leq 2$ of the previous paragraph; the normalized cyclic cocycles that can not be written as coboundaries are denoted as $\chi$ in eqs. and there. The assertion then is, all these cocycles may be expressed in terms of closed currents; hence, at least in this particular case the above conjecture is indeed true.
For the verification we begin with $p=0$; this is the dual $(G^2)^{\prime}$, were it not for the normalization condition. Here the request for closedness of the current is void, and with $$\label{bcch4}
\chi(f)=\int\,d^{\,2}\xi\,\chi(\xi)^{\ast}f(\xi)$$ one finds, in an obvious notation: $$\begin{aligned}
{5}
\chi_{1}(f)&=f_1\qquad\qquad&:\,&\chi_1(\xi)&=&-\xi^2 \\
\chi_{2}(f)&=f_2\qquad\qquad&:\,&\chi_2(\xi)&=&+\xi^1 \\
\chi_{12}(f)&=f_{12}\qquad\qquad&:\,&\chi_{12}(\xi)&=&+1. \end{aligned}$$ The result for $p=1$ is more interesting; in the notation of and with $(\chi^i)_{i=1,2}=\chi$ one has: $$\begin{aligned}
{5}
\chi_{1,1}(f,g)&=f_1g_1\qquad\qquad&:\,&\chi_{1,1}(\xi)&=&
\begin{pmatrix}-\xi^2\\0\end{pmatrix}\\
\chi_{2,2}(f,g)&=f_2g_2\qquad\qquad&:\,&\chi_{2,2}(\xi)&=&
\begin{pmatrix}0\\+\xi^1\end{pmatrix} \\
\chi_{1,2}(f,g)&=f_1g_2+f_2g_1\qquad\qquad&:\,&\chi_{1,2}(\xi)&=
&\begin{pmatrix}+\xi^1\\-\xi^2\end{pmatrix}\\
\chi_{1,12}(f,g)&=f_{1}g_{12}+f_{12}g_{1}\qquad\qquad&:\,&\chi_{1,12}(\xi)&=&
\begin{pmatrix}1\\0\end{pmatrix} \\
\chi_{2,12}(f,g)&=f_{2}g_{12}+f_{12}g_{2}\qquad\qquad&:\,&\chi_{2,12}(\xi)&=&
\begin{pmatrix}0\\1\end{pmatrix} \end{aligned}$$ and one checks that all these currents are closed. Finally, for $p=2$ and in matrix notation $(\chi^{ij})_{i,j=1,2}=\chi$ we obtain: $$\begin{aligned}
{5}
\chi_{1,1,1}(f,g,h)&=f_1g_1h_1\qquad&:\,&\chi_{1,1,1}(\xi)&=&
\begin{pmatrix}-\xi^2&0\\0&0\end{pmatrix}\\
\chi_{2,2,2}(f,g,h)&=f_2g_2h_2\qquad&:\,&\chi_{2,2,2}(\xi)&=&
\begin{pmatrix}0&0\\0&+\xi^1\end{pmatrix} \\
\chi_{1,1,2}(f,g,h)&=f_1g_1h_2+f_1g_2h_1+f_2g_1h_1\qquad&:\,&\chi_{1,1,2}(\xi)&=
&\begin{pmatrix}+\xi^1&-\xi^2\\-\xi^2&0\end{pmatrix}\\
\chi_{2,2,1}(f,g,h)&=f_2g_2h_1+f_2g_1h_2+f_1g_2h_2\qquad&:\,&\chi_{2,2,1}(\xi)&=
&\begin{pmatrix}0&+\xi^1\\+\xi^1&-\xi^2\end{pmatrix}\\
\chi_{1,1,12}(f,g,h)&=f_{1}g_{1}h_{12}+f_{1}g_{12}h_{1}+f_{12}g_{1}h_{1}\qquad&:
\,&\chi_{1,1,12}(\xi)&=&
\begin{pmatrix}1&0\\0&0\end{pmatrix} \\
\chi_{2,2,12}(f,g,h)&=f_{2}g_{2}h_{12}+f_{2}g_{12}h_{2}+f_{12}g_{2}h_{2}
\qquad&:\,&\chi_{2,2,12}(\xi)&=&
\begin{pmatrix}0&0\\0&1\end{pmatrix} \\
\chi_{1,2,12}(f,g,h)&=f_{1}g_{2}h_{12}+f_{2}g_{1}h_{12}+f_{1}g_{12}h_{2}+f_{2}g_{12}h_{1}+
f_{12}g_{1}h_{2}+f_{12}g_{2}h_{1}\qquad&:\,&\chi_{1,2,12}(\xi)&=&
\begin{pmatrix}0&1\\1&0\end{pmatrix}. \end{aligned}$$ Again, it is rather obvious that all these currents are closed. The last cocycle, denoted $\tau_{\text{vol}}=\chi_{1,2,12}$, will be the one of ultimate interest, and we restate it here in the form of a Berezin integral $$\label{bcch45}
\tau_{\text{vol}}(f,g,h)=
\int\,d^{\,2}\xi\,(f\,\partial_{1}g\,\partial_{2}h+f\,\partial_{2}g\,\partial_{1}h)$$ for later reference.
Let us remark, these latter results could not have been found by Coquereaux & Ragoucy [@Coqu95] since their sign rules are not compatible with the Berezin rules.
From what we have learned for $n=2$, one expects the general situation to be that all nontrivial characters may be represented in terms of currents; indeed, this expectation is correct and its verification is given below. But we have no direct proof for this claim; instead, we proceed indirectly and invoke that the dimensions of the cohomology groups $H^p_{\lambda}(G^n)$ are known. They were derived by Kassel [@Kass86] by means of a Künneth type argument and given in the form of a $\mathds{Z}/2$-graded Poincaré polynomial; but we do not state his result here since it will emerge in the course of the proof. The approach given below relies in part on ideas developed in [@Coqu95]; they require revision, however, and also our conclusions are different.
Up to now we only know that a closed current $\chi^p\in\mathrm{Ker}\,\partial_p$ determines a normalized element of $Z^p_{\lambda}(G^n)$, but we do not have under control whether there are elements of $B^p_{\lambda}(G^n)$ among these. By a counting argument we shall show that the dimension of the vector space $\mathrm{Ker}\,\partial_p$ is identical with the dimension of the space $HC^p(G^n)$; hence, these spaces are isomorphic and thus the map $\mathrm{Ker}\,\partial_p\to HC^p$ defined by is, in particular, a surjection. This final result may be paraphrased by saying that those normalized cyclic cocycles that can be represented by a nonzero current may not be written as boundaries; hence, such currents compute the normalized cyclic cohomology.
As to the proof, the strategy will be to establish a recurrence relation for the dimension of the space of closed $p$-currents. We begin with the dimension of the space $\mathcal{C}_p\doteq\Omega_p(G^n)$ of currents, which is $$\label{bcch45a}
\text{dim}\,\mathcal{C}_p=2^n{n+p-1 \choose p}\qquad\qquad\qquad:\,p\geq 0.$$ For $p>1$, the vector space $\mathcal{C}_p$ can be decomposed into the direct sum $\mathcal{C}_p=\mathcal{C}\mathcal{C}_p\oplus\not{\!\mathcal{C}}\mathcal{C}_p$ of closed currents, i.e. $\mathcal{C}\mathcal{C}_p\doteq\mathrm{Ker}\,\partial_p$, and those which are not, i.e. $\not{\!\mathcal{C}}\mathcal{C}_p\doteq\mathrm{Im}\,\partial_p$. Because $\partial_{p-1}\partial_p=0$, we thus have the inclusion $\not{\!\mathcal{C}}\mathcal{C}_p\subseteq\mathcal{C}\mathcal{C}_{p-1}$; as we have seen above, however, this is even an equality since together with the complex $\Omega^{\displaystyle\cdot}G^n$ also the dual complex of currents $\Omega_{\displaystyle\cdot}G^n$ is acyclic, and so $$\label{bcch45b}
\mathcal{C}_p=\mathcal{C}\mathcal{C}_p\oplus \mathcal{C}\mathcal{C}_{p-1}\qquad\qquad\qquad:\,p>1$$ giving the recursion $\text{dim}(\mathcal{C}\mathcal{C}_p)=
\text{dim}(\mathcal{C}_p)-\text{dim}(\mathcal{C}\mathcal{C}_{p-1})$. For $p=1$, the complementary subspace of $\mathcal{C}\mathcal{C}_1$ in $\mathcal{C}_1$ consists of those elements of $G^n$, in which the top component is absent; using the notation $\mathcal{C}\mathcal{C}_{0}\doteq\not{\!\mathcal{C}}\mathcal{C}_1$, we thus have the decomposition $\mathcal{C}_1=\mathcal{C}\mathcal{C}_1\oplus\,\mathcal{C}\mathcal{C}_0$, where $\text{dim}(\mathcal{C}\mathcal{C}_0)=2^n-1$. In this way the recursion is valid down to $p=1$, and iteration yields $$\label{bcch45c}
\text{dim}(\mathcal{C}\mathcal{C}_p)=\sum_{i=0}^p(-1)^i\,\text{dim}(\mathcal{C}_{p-i})-(-1)^p.$$ The sum can be manipulated by means of the expansion $$\frac{1}{(1+t)(1-t)^n}=\sum_{p=0}^{\infty}\,t^p
\sum_{i=0}^{p}(-1)^i{n+p-i-1\choose p-i}\qquad\qquad\qquad:\,|t|<1.$$ Whence, on using that the elements of $\mathcal{C}\mathcal{C}_p$ divide into even and odd ones, the generating function for the dimensions of the closed currents $\mathcal{C}\mathcal{C}_p$ is the Poincaré polynomial $$\label{bcch45f}
P^n(t,\theta)=\frac{2^{n-1}(1+\theta)}{(1+t)(1-t)^n}-\frac{1}{1+t}$$ where $\theta$ denotes the nontrivial generator of $\mathds{Z}/2$. In particular, for $n=2$ this confirms the values determined by direct computation in the previous paragraph.
We can now address the discussion of the results for the cyclic cohomology $H^{\displaystyle\cdot}_{\lambda}(G^n)$, as obtained by Kassel in [@Kass86]. In that work the cyclic cohomology group is written as the direct sum of two terms, the first of which is just specified by the Poincaré polynomial ; hence, according to what we have shown, this part can thus be characterized as the normalized cyclic cohomology $HC^{\displaystyle\cdot}(G^n)$. The second direct summand is the cyclic cohomology group of the real numbers $H^{\displaystyle\cdot}_{\lambda}(\mathds{R})$, which is trivial for $p$ odd and equal to $\mathds{R}$ for $p$ even (see, e.g. [@Rose94]), and so the result given in [@Kass86] can be read as $$\label{bcch45g}
H^{\displaystyle\cdot}_{\lambda}(G^n)=
HC^{\displaystyle\cdot}(G^n)\oplus H^{\displaystyle\cdot}_{\lambda}(\mathds{R}).$$ What remains is to characterize the second direct summand as (non normalized) nontrivial cyclic cocycles. This latter part we identify with the cocycles $$\varphi_{0\cdots0}(f_0,\ldots,f_p)=\varepsilon_{0\cdots0}(f_0,\ldots,f_p)$$ which, on imposing the cyclicity property, are nonvanishing only for $p$ even; of course, they can not be written in the form of a Berezin integral.
To summarize, in one aspect the above results go beyond those obtained in [@Kass86] since the cocycles belonging to the two direct summands can even be given in explicit form.
Berezin’s integral as a Connes’ character
=========================================
With this interlude on Gra[ß]{}mann currents behind us, we now forget again about Berezin integration since a precise definition of the symbol $d^{\,n}\xi$ is not yet available. So let us return to characters on $G^n$; in this setting we can give a definite meaning to that symbol, and what we want is to relate somehow such Connes’ characters to a Berezin type of integral.
Recall, for $n=2$ and $p=2$ there are seven nontrivial characters and, as we have seen, this number grows rapidly with $n$; what is more, there are also characters for $p>n$, contrary to the classical case. Amongst this multitude of available characters, we claim that there is only one which deserves to be named a ‘volume integral’.
Such a construct is uniquely singled out by three natural defining properties. The first consists in the requirement that only $n$-dimensional characters $$\label{bcch46}
\int:\Omega^n(G^n)\to \mathds{R}$$ are to be considered. At second, it should not make sense that a variable is integrated twice. Hence, a ‘volume character’ $\int$ is assumed to take nonzero values only on that subspace of $n$-forms, for which on the right hand side of $$\label{bcch47}
\int \omega_n=\int \frac{1}{n\,!}\,\omega_{i_1\cdots i_n}\,d\xi^{i_1}\cdots d\xi^{i_n}$$ all the differentials are different, i.e. if $(i_1,\ldots,i_n)$ is a permutation of $(1,\ldots,n)$; whence, formally the axiom is $$\label{bcch48}
\int\omega_{n}=\int\frac{1}{n\,!}\,\sum_{\sigma\in S_n} \omega_{\sigma(1)\cdots\sigma(n)}
\,d\xi^{1}\cdots d\xi^{n}=\int\,\omega_{1\cdots n}\,d\xi^{1}\cdots d\xi^{n}$$ since all other contributions vanish by definition, and since the differentials commute. Note that it would not suffice here to restrict the $\Omega^p(G^n)$ themselves to the corresponding subspaces since they are not invariant against exterior differentiation; instead, one must restrict the admissible maps $\int$, i.e. the characters. For $\omega_n$ of the form $f_0\,df_1\cdots df_n$, we thus arrive at $$\label{bcch49}
\tau_{\text{vol}}(f_0,\ldots,f_n)=\int\sum_{\sigma\in S_n} f_0\partial_{\sigma(1)}f_1
\cdots\partial_{\sigma(n)}f_n\,d\xi^{1}\cdots d\xi^{n}$$ which gives us a special Hochschild $n$-cochain in terms of that $n$-dimensional character.
The definition, however, is still not complete; what finally remains is to assign the right hand side of a real number, for arbitrary $f$s in $G^n$. We do this by the prescription that it be evaluated by means of the formal Berezin integration rules, which is the third requirement.
The claim then is, the construct (cf. for $n=2$) indeed defines a Connes character. This assertion, if valid, implies that the Berezin integration rules for Gra[ß]{}mann variables and the defining properties of this special $n$-dimensional Connes’ character do indeed match. After all, at the risk of being pedantic, the symbol $d^{\,n}\xi$ can then legitimately be viewed as the product $d\xi^1\cdots d\xi^n$ of the differentials; this does not mean, however, that the integral could be defined by iteration, as often claimed in the literature.
For the verification we must show, the cochain is a normalized cyclic cocycle. This is done on rewriting $\tau_{\text{vol}}$ in the form $$\label{bcch50}
\tau_{\text{vol}}(f_0,\ldots,f_n)=\int\,d^{\,n}\xi\,\chi^{i_1\cdots i_n}f_0
\partial_{i_1}f_{1}\cdots\partial_{i_n}f_n$$ where $$\label{bcch51}
\chi^{i_1\cdots i_n}=\sum_{\sigma\in S_n}\delta^{i_1}{}_{\sigma(1)}\cdots\delta^{i_n}{}_{\sigma(n)}$$ is recognized as a constant, whence closed current. Now we are back at the situation of the previous section so that we can resort to the results obtained there; hence, we are entitled to identify $\tau_{\text{vol}}$ as a normalized cyclic cocycle, i.e. a character. Furthermore, the corresponding normalized cyclic cocycle is not trivial, i.e. not a coboundary; of course, this fact is crucial. Let us also note, the property $\int d\omega_{n-1}=0$ of a character implies $\int d^{\,n}\xi\,\partial_if(\xi)=0$ for all $i\in\{1,\ldots,n\}$, which is one of the Berezin rules. The definition may be viewed as a generalization of the Berezin integral calculus on the Gra[ß]{}mann algebra, relating it to the Connes’ characters of noncommutative geometry.
Hence, eq. defines a special $n$-dimensional character if the right hand side is understood to be evaluated by means of the Berezin rules. The conventional Berezin integral over an element $f\in G^n$ is regained for $f_0=f$ and $f_i=\xi^i$ with $i\in(1,\ldots,n)$, viz. $$\label{bcch52}
\tau_{\text{vol}}(f,\xi^1,\ldots,\xi^n)=\int\,d^{\,n}\xi\,f(\xi)=f_{1\cdots n}.$$ After all, this is the final justification of the formula we began with, i.e. $J(f)=\int d^{\,n}\xi\,f(\xi)$, and so the circle of ideas closes here, at the very end.
To resume, Berezin integration over a Gra[ß]{}mann algebra fits perfectly into the scheme of noncommutative geometry in that amongst Connes’ integral calculi one unique normalized cyclic Hochschild cocycle can naturally be singled out, which reduces to the Berezin integral and is compatible with its defining rules.
Conclusion
==========
Beyond the Berezin rules or , that we consider as safe, there is a variety of other rules around in the literature. One such example, given in Berezin’s book [@Bere66] and elsewhere, is $$\int d\xi^i\,\xi^j\overset{?}{=}\delta^{ij}$$ which we refuse since it would entail that the integral over anticommuting variables could be defined iteratively. Another one, also often being taken for granted, is $$d\xi^i\,d\xi^j+d\xi^j\,d\xi^i\overset{?}{=}0$$ so that the differentials anticommute; but this condition is at variance with the demand that the exterior differentiation $d$ be of grade $+1$, i.e. be an odd operator. Anticommuting differentials are suggested also by another rule frequently being stated, namely, that the ‘volume element’ should transform under a linear transformation $\xi \mapsto \xi'=A\xi$ with $A\in\text{GL}(n,\mathds{R})$ as $$d^{\,n}\xi'\overset{?}{=}\text{det}(A)^{-1}d^{\,n}\xi.$$ This requirement is inspired by the rules and derived above - as far as one is willing at all to ascribe a transformation law to the volume element. Indeed, it requires anticommuting differentials; but instead of the determinant, its inverse appears in this latter formula, The way out being favoured in the literature is to define integration in the Gra[ß]{}mann case as an operation over the tangent space, instead of the cotangent space in the conventional case, and so one invents another volume element (see [@Bruz87], and literature cited there) $$d^{\,n}{}_{\!\!\textstyle\xi}\overset{?}{=}\frac{\partial}{\partial\xi^1}\wedge\cdots\wedge
\frac{\partial}{\partial\xi^n}$$ where the ungraded wedge product should be understood; it indeed reproduces the postulated transformation law. According to our conviction, however, deep water is reached at this point. As we believe, the idea to ascribe a transformation law to the volume element must be relinquished; actually, there is also no need to do so since Berezin integration acts as a differential operator, and the inverse of the determinant merely follows from the fact that the partial differentiations anticommute. Thus, the validity of eq. is guaranteed without further ado, and this is the only property that enters the fermionic path integral in the representation through coherent states; the transition to the complex case poses no added difficulty. Also, in the supersymmetric situation, the property is sufficient to define an integral of the form $$\int d^{\,m}x\int d^{\,n}\xi\,\, f(x,\xi)=\int d^{\,m}x\,\frac{\partial}{\partial\xi^n\cdots\partial\xi^1}\, f(x,\xi)$$ over the superspace $\mathds{R}^{m,n}$; under a change of supercoordinates, the differentiations then provide for the correct transformation law so that the superdeterminant of the Jacobian is reproduced correctly. A similar point of view is advocated by Cartier et al. [@Cart02], being almost in line with our arguments. In this way, the above apparent discrepancy is resolved, without any need to alter or supplement the rules taken for granted in the main text.
[99]{}
Itzykson C. and Zuber J. B., *Quantum Field Theory*, McGraw-Hill, New York 1980. Ryder L. H., *Quantum Field Theory*, Cambr. Univ. Press, Cambridge 1985. Nieuwenhuizen, P. van, “Supergravity”, Phys. Rep. **68** (1981) 189-398. Wess J. and Bagger J., *Supersymmetry and Supergravity*, Princeton Univ. Press, Princeton 1983. Schwinger J., “A note on the quantum dynamical principle”, Phil. Mag. **44** (1953) 1171-1179. Schwinger J., *Quantum Kinematics and Dynamics*, Benjamin, New York 1970. Matthews P. T. and Salam A., “Propagators of quantized fields”, Nuovo Cim. **2** (1955) 120-134. F. A. Berezin, *The Method of Second Quantization*, Academic Press, New York 1966. Berezin F. A., “Differential forms on supermanifolds”, Sov. J. Nucl. Phys. **30** (1979) 605-609. Berezin F. A., *Introduction to Superanalysis*, Reidel, Dordrecht 1987. Connes A., *Noncommutative Geometry*, Academic Press, London and San Diego 1994. Kastler D., *Cyclic Cohomology within the Differential Envelope*, Hermann, Paris 1988. Karoubi M., “Connexions, courbures et classes charactéristiques en K-théorie algébrique”, Canadian Math. Soc. Proc. **2** (1982) 19-27. Coquereaux R., Frappat L., Ragoucy E. and Sorba P., “Extended super-Kač-Moody algebras and their super-derivation algebras”, Comm. Math. Phys. **133** (1990) 1-35. Coquereaux R., Jadczyk A. and Kastler D., “Differential and integral calculus of Grassmann algebras”, Rev. Math. Phys. **3** (1991) 63-99. Coquereaux R. and Ragoucy E., “Currents on Grassmann algebras”, J. Geom. Phys. **15** (1995) 333-352. Connes A., “Non-commutative differential geometry”, Publ. Math. IHES **62** (1985) 41-144. Landi G., *An Introduction to Noncommutative Spaces and their Geometries*, Lecture Notes in Physics $\bf{51}$, Springer, Berlin 1997. Gracia-Bondia J. M., Várilly J. C. and Figueroa H., *Elements of Noncommutative Geometry*, Birkhäuser, Boston 2001. Kassel C., “A Künneth like formula for the cyclic cohomology of $\mathds{Z}/2$-graded algebras”, Math. Ann. **275** (1986) 683-699. Kastler D., “Introduction to entire cyclic cohomology (of $\mathds{Z}/2$-graded Banach algebras)”, in *Stochastics, Algebra and Analysis in Classical and Quantum Dynamics*, S. Albeverio, P. Blanchard and T. Testard Eds., Kluver Academic Publishers, Dordrecht 1990. Brodzki J., *An Introduction to K-theory and Cyclic Cohomology*, Polish Scientific Publishers (PWN), Warszawa 1998. Loday J. L., *Cyclic Homology*, Springer, Berlin 1992. DeWitt B. S., *Supermanifolds*, Cambridge Univ. Press, Cambridge 1984. Leites D. A., “Introduction to the theory of supermanifolds”, Russ. Math. Surv. **35** (1980) 1-64. Vladimorov V. S. and Volovich I. V., “Superanalysis I: Differential calculus”, Theor. Math. Phys. **59** (1984) 317-335; “Superanalysis II: Integral calculus”, Theor. Math. Phys. **60** (1985) 743-765. Constantinescu F. and de Groote H. F., *Geometrische und algebraische Methoden in der Physik: Supermannigfaltigkeiten und Virasoro Algebren*, Teubner, Stuttgart 1994. Deligne P., Etingof P., Freed D. S., Jeffrey L. C., Kazhdan D., Morgan J. W., Morrison D. R. and Witten E. eds., *Quantum Fields and Strings: A Course for Mathematicians*, Vol. I, Am. Math. Soc., Providence 1999. Guillemin V. W. and Sternberg S., *Supersymmetry and Equivariant de Rham Theory*, Springer, Berlin 1999. Bognar J., *Indefinite Inner Product Spaces*, Springer, Berlin 1974. Grensing G., “On ghost fermions”, Eur. Phys. J. C **23** (2002) 377-387. Henneaux M. and Teitelboim C., *Quantization of Gauge Systems*, Princeton Univ. Press, Princeton 1992. Bott R. and Tu L. W., *Differential Forms in Algebraic Topology*, Springer, New York 1982. Taylor M. E., *Partial Differential Equations* Vol. I, Springer, New York 1996. Dieudonné J., *Treatise on Analysis* Vol. III, Academic Press, New York 1972. Rosenfeld J., *Algebraic K-Theory and Its Applications*, Springer, New York 1994. Bruzzo U., “Integration on Supermanifolds”, in *General Relativity and Gravitational Physics*, U. Bruzzo, R. Cianci and E. Massa Eds., World Scientific, Singapore 1987. Cartier P., DeWitt-Morette C., Ihl M. and Sämann C., “Supermanifolds - application to supersymmetry”, math-phys/0202026.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- Tuomo Tanttu
- Alessandro Rossi
- Kuan Yen Tan
- Akseli Mäkinen
- Kok Wai Chan
- 'Andrew S. Dzurak'
- Mikko Möttönen
title: 'Three-waveform bidirectional pumping of single electrons with a silicon quantum dot'
---
Introduction {#introduction .unnumbered}
============
After a quarter of a century of development of charge pumps, we are close to redefining the International System of Units (SI) standard for the electrical current, the ampere, such that it would be based on a fixed value of the elementary charge [@Pekola2013; @Kaestner2015]. The direct experimental realization of such a quantum ampere standard is based on charge pumps which transfer accurately an integer $n$ number of electrons per cycle from the source to the drain at frequency $f$, yielding direct current $I=nef$. The parameter region where the pumped current is quantized in such a way and where it is insensitive to changes in the system parameters is referred to as a plateau. For practical realizations of the current standard and for the closure of the so-called quantum metrology triangle, it is sufficient that the pump yields a current of hundreds of picoamperes with relative accuracy of $10^{-8}$[@Feltin2009; @Drung2015; @Drung2015b].
The very first charge pumps were able to produce currents of a few picoamperes with accuracies of a few per cent [@Geerligs1990; @Geerligs1991; @Kouwenhoven1991; @Pothier1992]. After this, several different implementations of charge pumps have been proposed and tested: normal-metal tunnel junction devices [@Keller1996; @Keller1999], superconducting devices [@Niskanen2005; @Vartiainen2007; @Mottonen2008], superconductor–normal–metal hybrid turnstiles [@Pekola2008; @Maisi2009; @Kemppinen2009; @Kemppinen2009b; @Maisi2011; @Peltonen2015], and surface acoustic wave devices [@Shilton1996; @Talyanskii1997]. At the moment, the most promising candidates for the emerging quantum ampere are semiconductor quantum dots [@Blumenthal2007; @Jehl2012; @Giblin2012; @Stein2015; @Giblin2016; @Connolly2013; @Fujiwara2004; @Fujiwara2008; @Fujiwara2001; @Chan2011; @Rossi2014; @Jo2015] and single-atom impurities in semiconductors [@Yamahata2014; @Tettamanzi2014; @Lansbergen2012; @Roche2012], the state of the art being, a current of 87 pA in a GaAs/AlGaAs quantum dot with an uncertainty of less than 0.2 parts per million (ppm) [@Stein2015]. Silicon single-electron pumps have also been studied widely[@Chan2011; @Rossi2014; @Fujiwara2001; @Ono2003; @Yamahata2011; @Yamahata2014; @Yamahata2014_revB; @Yamahata2015; @Jo2015; @Tettamanzi2014; @Hollosy2015]. Benefits of silicon pumps are that they are based on technologies well-known by the industry and they may exhibit suppressed $1/f$ noise and the absence of large random charge jumps[@Zimmerman2007; @Hourdakis2008; @Zimmerman2008; @Koppinen2013; @Zimmerman2014].
The accuracy of a charge pump can be determined with a charge sensor that monitors the charge state of a reservoir island into which electrons are pumped. Several different charge sensing schemes have been demonstrated: pumping electrons into and out of the reservoir[@Keller1996; @Kautz1999; @Yamahata2011; @Yamahata2014_revB], pumping a number of electrons into the reservoir and cyclically emptying it[@Tanttu2015], monitoring multiple reservoirs interleaved with pumps in a series configuration[@Fricke2013; @Fricke2014; @Peltonen2015], and monitoring the pump dot without any storage node[@Giblin2016]. In general, pumping electrons into and out of a reservoir in semiconductor devices is highly nontrivial due to the asymmetry of the devices and pumping protocols.
In this paper, we demonstrate bidirectional electron pumping in a silicon-based quantum dot by employing a three waveform protocol, thus offering a step towards error counting based on a reservoir dot. Our technology allows simultaneous control over both barriers and the dot potential, enabling convenient switching between the pumping by changing only the phase of one driving signal. This kind of switching between the directions has been demonstrated before in silicon devices[@Ono2003; @Jehl2012] with two driving waveforms to the barrier gates[@Kouwenhoven1991] and in metallic pumps with three waveforms [@Pothier1992; @Mottonen2008]. We also show the improvement of this pumping process over a two-waveform drive using the same driving amplitudes. This is confirmed by the study of the sensitivity of the current quantization to experimental parameters such as the plunger gate (PL) dc voltage and the rf amplitude of the drive.
Results {#results .unnumbered}
=======
Pumping with three waveforms {#pumping-with-three-waveforms .unnumbered}
----------------------------
A scanning electron microscope image of our quantum dot device and a schematic measurement set-up are presented in Fig. \[sample\](a). The details of forming a two-dimensional electron gas (2DEG) and a single-electron dot isolated from the leads are discussed in Refs. . Initially, we pump with sinusoidal radio frequency (rf) waveforms applied to PL and to the barrier right gate (BR) with a phase difference of PL with respect to BR, $\phi_{\textrm{BR-PL}}=95^\circ$ at 800 mK temperature. We electrically confine the dot by setting negative bias on the confining gates C1 and C2, with gate voltages $V_\textrm{C1}$ and $V_\textrm{C2}$, respectively[@Rossi2014; @Seo2014]. Then we search in the dc parameter space for a stable pumping plateau in the positive current direction, i.e., BR corresponds to the entrance barrier and the barrier left gate (BL) to the exit barrier. Here, we define the first plateau as the parameter region for which the normalized pumped current, $I/ef$ is within 5% of unity. Subsequently, we decrease the dc potential on BL, $V_{\textrm{BL}}$, i.e., we make the exit barrier more opaque, until we measure only a narrow pumping plateau as a function of PL and BR dc voltages, $V_{\textrm{PL}}$ and $V_{\textrm{BR}}$ respectively. At this point, our pumping process is limited by the unloading process. Then we drive BL a sinusoidal waveform 180$^\circ$ phase-shifted with respect to BR so that all the waveforms are sines by $\tilde{V}_{\textrm{BR}} = A_{\textrm{BR}}\sin( \omega t)$, $\tilde{V}_{\textrm{PL}} = A_{\textrm{PL}}\sin( \omega t+\phi_{\textrm{BR-PL}})$, and $\tilde{V}_{\textrm{BL}} = A_{\textrm{BL}}\sin( \omega t+180^{\circ})$.
A schematic illustration of the pumping process with three pulses is presented in Fig. \[sample\](b). First (I) we lower the entrance barrier and the potential of the dot such that an electron tunnels into the dot. In the second phase, (II) the entrance barrier is raised and the electron is trapped in the dot. Then the electron is unloaded (III) by lowering the exit barrier and raising the dot potential. Depending on the phase of PL it is possible that electrons exit at energies above the Fermi level of the leads. Hence it is possible that the electron escapes to the source but such process unlikely due to the high opacity of the entrance barrier. The time dependence of the potentials with two different phases on PL is presented in Fig. \[sample\](c). The normalized pumped current, $I/ef$, at $f=200$ MHz is shown as a function of $V_{\textrm{PL}}$ with varying $A_\textrm{BL}$ in Fig. \[sample\](d). The length of the first plateau increases significantly with higher $A_\textrm{BL}$. Note also that with low amplitudes there is no second plateau, but with high amplitudes the second plateau corresponding to transfer of two electrons per cycle is clearly visible.
Not only does the length of the plateaus increase with increasing $A_\textrm{BL}$ but the current quantization becomes more accurate, i.e., the pumped current is closer to the expected value. The inset of Fig. \[sample\](d) shows that with low amplitude, the normalized pumped current at the first plateau is below unity by a few per cent. However, with high amplitudes the normalized current reaches unity more precisely which indicates a more robust unloading process.
Stability of the pumping process {#stability-of-the-pumping-process .unnumbered}
--------------------------------
We also study the stability of the pumping process by measuring the pumped current as a function of $V_{\textrm{PL}}$, $V_{\textrm{BR}}$, and $V_{\textrm{BL}}$. We begin by pumping with PL and BR as described above, but decrease $A_{\textrm{BR}}$ until we measure only a narrow plateau region in terms of $V_{\textrm{PL}}$, $V_{\textrm{BR}}$, and $V_{\textrm{BL}}$. We measure the pumped current as a function of $V_{\textrm{PL}}$ and $V_{\textrm{BR}}$ and as a function of $V_{\textrm{PL}}$ and $V_{\textrm{BL}}$. We repeat these scans with different $A_\textrm{BL}$ values. The results are shown in Fig. \[comparison\](a). We observe that the plateau enlarges in the parameter space ($V_{\textrm{PL}}$, $V_{\textrm{BR}}$, and $V_{\textrm{BL}}$) with increasing $A_\textrm{BL}$. The maximum length of the plateau in terms of $V_\textrm{PL}$ increases from 14 to 52 mV as $A_\textrm{BL}$ increases from 0 to 123 mV.
This experiment is repeated in the case where the exit barrier is BR, i.e., the opposite pumping direction. We pump initially with PL and BL using PL phase difference w.r.t. BL as $\phi_\textrm{BL-PL}=120^\circ$ at 100 MHz. A third waveform (complement waveform to that of BL) is applied to BR. The waveforms are: $\tilde{V}_{\textrm{BL}} = A_{\textrm{BL}}\sin (\omega t)$, $\tilde{V}_{\textrm{PL}} = A_{\textrm{PL}}\sin (\omega t+\phi_{\textrm{BL-PL}})$, and $\tilde{V}_{\textrm{BR}} = A_{\textrm{BR}}\sin (\omega t+180^{\circ})$. We observe a similar widening of the plateau in Fig. \[comparison\](b) as in Fig. \[comparison\](a). However, in this case the length of the plateau in $V_\textrm{PL}$ reaches a maximum around $A_\textrm{BR}=79$ mV, inferring that the BR gate couples to the quantum dot more strongly than the BL gate. For simplicity, we do not compensate this coupling in the experiment. Therefore the increased amplitude in BR interferes with the pumping process more than that in BL, restricting our ability to improve the pumping process beyond the observed optimal point. The maximum length of the plateau in terms of $V_\textrm{PL}$ increases from 12 mV to 25 mV as $A_\textrm{BR}$ increases from 0 to 79 mV. The maximum length of the plateau in $V_\textrm{PL}$ as a function of the exit barrier amplitude extracted from the data of Fig. \[comparison\](a) and \[comparison\](b) is presented in Fig. \[comparison\](c). In both scenarios, the length increases rather linearly as we increase the amplitude of the exit barrier, but in the case of BR, we observe a maximum.
Bidirectional pumping {#bidirectional-pumping .unnumbered}
---------------------
Let us choose the values of $V_\textrm{BR}$, $V_\textrm{BL}$, and $A_\textrm{BR}$ such that they correspond to the maximum plateau length in Fig. \[comparison\](b). Then we measure the pumped current as a function of $V_{\textrm{PL}}$ and $\phi_{\textrm{BL-PL}}$ with the results shown in Fig. \[bothdirections\](a). The pumping plateaus in both directions have the same midpoint, $V_\textrm{PL}=0.823$ V. Thus in comparison to bidirectional pumping observed with two drives in a similar sample[@Rossi2014], we can pump in both directions with fixed $V_\textrm{PL}$ value, whereas with only two drives we needed a different phase and $V_\textrm{PL}$. We examine the cross sections along $V_\textrm{PL}$ with two phase differences, $\phi_{\textrm{BL-PL}}=120^\circ$ and $\phi_{\textrm{BL-PL}}=240^\circ$. The absolute value of the pumped current in both cases is shown in Fig. \[bothdirections\](b). In the region where $V_\textrm{PL}=0.820-0.826$ V, the absolute magnitude of current in each direction is the same within the experimental uncertainty. In the middle of this region (at $V_\textrm{PL}=0.823$ V) we study the cross section of Fig. \[bothdirections\](a) along the phase difference $\phi_{\textrm{BL-PL}}$, as shown in Fig. \[bothdirections\](c). We clearly observe that the pumping plateau appears in the negative direction in the range $\phi_{\textrm{BL-PL}}=40-160^\circ$ and in the positive direction in the range $\phi_{\textrm{BL-PL}}=200-290^\circ$.
Discussion {#discussion .unnumbered}
==========
We have demonstrated robust bidirectional pumping with convenient switching between the pumping directions in a silicon single-electron pump with independent rf control over both barriers and the dot potential. Initially, we pump with the entrance barrier and PL, then we introduce a third rf waveform to the exit barrier. We study the effect of the amplitude of the exit barrier swing on the pumped current and observe that it can increase the length of the current plateau in the parameter space of applied gate potentials without the need to increase the amplitudes of the other drives. Subsequently, we studied the pumped current as a function of the plunger gate voltage and the phase difference of the plunger drive with respect to the barrier drives. We find a parameter region where we can perform pumping in positive and negative directions simply by changing the phase of the plunger drive. Furthermore, there is a parameter region where the magnitudes of the pumped currents in different directions are equal within experimental accuracy.
Our architecture, which incorporates individual control over both barriers and the quantum dot is highly flexible. It allows convenient switching between the pumping directions simply by changing the phase of the plunger gate rf drive. This property can potentially be used to extend the basic electron counting scheme demonstrated in Ref. , to a more sophisticated error counting protocol[@Keller1996], where single electrons are pumped back and forth between source and reservoir, without accumulating electrons in the reservoir. Different pumping directions may also be used to improve the signal–to–noise ratio in the current measurement by averaging over $+I$ and $-I$ rather than $+I$ and zero. A pumping scheme with three rf drives may also be used to further reduce the uncertainty of output current of the semiconductor quantum dot pumps.
![Sample and pumping protocol. (a) Scanning electron microscope image of a sample similar to the one used in the experiments together with the schematic measurement setup. The lateral position of the quantum dot is indicated by an orange oval. Blue squares indicate the ohmic contacts of source and drain to the two-dimensional electron gas. (b) Schematic potential landscapes for an electron along the red line in (a) at different stages of the pumping process for two different phases of plunger drive leading to either positive (left) or negative (right) pumped current. The whole process consists of electron loading (I), trapping (II), and unloading (III). (c) Time dependence of the electrical potentials (blue, red and green lines) in the three-waveform driving scheme with two different phases of the plunger drive with respect to BL drive. The Fermi level of the leads is shown by the black dashed line. The gray lines indicate the time instants visualized in (b). (d) Pumped current as a function of plunger gate voltage, $V_{\textrm{PL}}$, with different amplitudes of the driving voltage on the left barrier, $A_\textrm{BL}$, at 200 MHz frequency. Inset: The pumped current from the main panel in the vicinity of plateau $I/ef= 1$. The dc gate voltages defined in (a) assume the following values: $V_{\textrm{SL}}= 3.5$ V, $V_{\textrm{DL}}= 1.8$ V, $V_{\textrm{C1}}= -0.40$ V, $V_{\textrm{C2}}= -0.50$ V, $V_{\textrm{BL}}= 0.80$ V, and $V_{\textrm{BR}}= 0.72$ V. The rf driving amplitudes are $A_{\textrm{BR}}= 158$ mV, and $A_{\textrm{PL}}= 79$ mV with phase difference $\phi_\textrm{BR-PL}= 95^{\circ}$.[]{data-label="sample"}](Fig1.eps){width="130"}
 except $V_{\textrm{BL}}= 0.9$ V in left column and $V_{\textrm{BR}}= 0.725$ V in the right column and $A_{\textrm{BR}}=123 $ mV in both columns. In (b) the values are the same as in (a) except $V_{\textrm{BL}}= 0.814$ V in left column and $V_{\textrm{BR}}= 0.85$ V in the right column, $A_{\textrm{BL}}= 112$ mV and $A_{\textrm{PL}}= 205$ mV in both columns and $\phi_\textrm{BL-PL}= 120^{\circ}$. (c) Maximum plateau length in the plunger dc voltage as a function of exit barrier amplitude ($A_\textrm{PL}$ or $A_\textrm{BR}$) extracted from (a) (black squares) and (b) (red circles). The parameter values are the same as in (a) and (b) for the BL and BR, respectively. []{data-label="comparison"}](Fig2.eps){width="\linewidth"}
 except $V_\textrm{BL}=0.814$ V, $V_\textrm{BR}=0.850$ V, and $A_\textrm{BR}=79$ mV. (b) Absolute values of the cross sections along $V_\textrm{PL}$ in (a) at $\phi_\textrm{BL-PL}=120^\circ$ (blue) and 240$^\circ$ (red) with 95% confidence intervals indicated. (c) Cross section of pumped current along $\phi_{\textrm{BL-PL}}$ at $V_\textrm{PL}=0.823$ V (green dashed line in (a)).[]{data-label="bothdirections"}](Fig3.eps){width="\linewidth"}
Methods {#methods .unnumbered}
=======
Sample fabrication {#sample-fabrication .unnumbered}
------------------
Our quantum dot is based on a three-layer gate-stack silicon metal–oxide–semiconductor (MOS) technology. The sample is fabricated on a high-purity intrinsic silicon wafer. A 8-nm thick SiO$_2$ layer thermally grown in the active region to form the gate–channel oxide. Three layers of aluminum gates with thicknesses 30, 55, and 80 nm, respectively are fabricated on top of the wafer with electron beam lithography. Polymethyl methacrylate (PMMA 950k) A4 resist is used as a mask material. After patterning the mask, we develop it in a mixture of methyl isobutyl ketone and isopropanol (MIBK:IPA 1:3) solution. Each layer of Al gates is deposited with a thermal evaporator, followed by a lift-off process. The gates are oxidized on a hot plate in an ambient environment. This process is repeated in total three times to realize all the layers. The source and drain electrodes are connected to the 2DEG with phosphorus-doped regions which form the ohmic contacts. Further details of the fabrication process are discussed elsewhere [@Rossi2015].
Measurement set-up {#measurement-set-up .unnumbered}
------------------
All the experiments are performed in a self-made, torlon-based, plastic dilution refrigerator with base temperature of 100 mK submerged in 4-K helium bath. The device is cooled down to base temperature but due to the dissipation on the line impedances, the mixing chamber temperature increases during the experiments up to 800 mK. The device is mounted on a sample holder printed circuit board (PCB) with integrated bias tees with capacitance 10 nF and resistance 100 k$\Omega$. This allows simultaneously application of rf and dc voltages on the driving gates. Our silicon chip is attached to the sample stage with vacuum grease and wedge-bonded to the PCB with Al bond wires. The rf driving voltages from the room temperature set-up are connected to the PCB with coaxial cables with 10 dB attenuation at 4 K. The dc voltages are connected to the PCB with twisted-pair loom lines. The wiring of the sample and the cryostat is discussed in more detail in Ref. .
A 2DEG is induced at the interface between Si and SiO$_2$ by applying positive voltage to the Al gates. All dc gate voltages are generated by floating dc voltage sources connected to 1:5 voltage dividers. The rf waveforms are generated by an arbitrary waveform generator and synchronized with a rubidium frequency standard. These waveforms are attenuated at room temperature depending on the experiment. The output current is amplified by $10^{10}$ V/A with a transimpedance amplifier powered by a regulated battery pack. The amplified signal is optoisolated to eliminate ground loops and subsequently recorded by a digital multimeter.
[10]{} url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
*et al.* . ** ****, ().
& . ** ****, ().
& . ** ****, ().
, , , & . ** **** ().
, , & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
, , , & . ** ****, ().
, , & . ** ****, ().
, , & . ** ****, ().
, , & . ** ****, ().
, , & . ** ****, ().
, & . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ****, ().
*et al.* . ** ****, ().
, , , & . ** ().
*et al.* . ** ****, ().
*et al.* ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** **** ().
*et al.* . ** ****, ().
, , & . ** ****, ().
, & . ** ****, ().
& . ** ****, ().
& . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** **** ().
, & . ** ****, ().
, & . ** ****, ().
, & . ** ****, ().
*et al.* . ** ****, ().
, & . ** ****, ().
, & . ** ****, ().
, & . ** **** ().
*et al.* . ** ****, ().
*et al.* . ** **** ().
, & . ** **** ().
*et al.* . ** **** ().
, & . ** ****, ().
, , , & . ** ****, ().
, & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Y. Sun, M. Jenei, G. C. Tettamanzi, I. Iisakka, A. Kemppinen, J. Lehtinen, and E. Mykkänen for fruitful discussions, and M. Meschke and J. Pekola for their design and help in building the cryostat used in the experiments. The financial support from the Centre of Excellence in Computational Nanoscience (project 284621 and 251748) by the Academy of Finland (Grant Nos. 251748, 135794, 272806, and 276528), the Australian Research Council (Grant Nos. DP120104710 and DP160104923), Jenny and Antti Wihuri Foundation, The Finnish Cultural Foundation, and the Australian National Fabrication Facility are acknowledged. A. R. thanks the University of New South Wales Early Career Research Grant scheme for financial support. We acknowledge the provision of facilities and technical support by Aalto University at Micronova Nanofabrication Centre.
Author contributions statement {#author-contributions-statement .unnumbered}
==============================
T. T. and A. M. conducted the experiments with support from A. R. and K. Y. T.. T. T. wrote the manuscript with input from all authors. A. R. fabricated the sample. K. W. C. prepared the silicon substrates. A. S. D. and M. M. provided project guidance and supervision.
Competing financial interests {#competing-financial-interests .unnumbered}
=============================
The authors declare no competing financial interests.
Corresponding author {#corresponding-author .unnumbered}
====================
Correspondence to Tuomo Tanttu
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Herein we study conformal vectors of a $\mathbb{Z}$-graded vertex algebra of (strong) CFT type. We prove that the full vertex algebra automorphism group transitively acts on the set of the conformal vectors of strong CFT type if the vertex algebra is simple. The statement is equivalent to the uniqueness of self-dual vertex operator algebra structures of a simple vertex algebra. As an application, we show that the full vertex algebra automorphism group of a simple vertex operator algebra of strong CFT type uniquely decomposes into the product of certain two subgroups and the vertex operator algebra automorphism group. Furthermore, we prove that the full vertex algebra automorphism group of the moonshine module over the field of real numbers is the Monster.'
address: 'Graduate School of Mathematical Science, The University of Tokyo, 3-8-1 Komaba, Meguro-ku, Tokyo 153-8914, Japan'
author:
- YUTO MORIWAKI
title: On Classification of Conformal Vectors in Vertex Operator Algebra and the Vertex Algebra Automorphism Group
---
Introduction
============
The notions of vertex algebras and their conformal vectors were originally developed by Borcherds [@Bo1]. A vertex algebra with a distinguished conformal vector is called a vertex operator algebra (VOA) [@FLM], which is a mathematical axiomatization of a chiral algebra appearing in two-dimensional conformal field theory. Different conformal vectors in a vertex algebra generally yield different VOA structures and conformal field theories. Thus, the classification of conformal vectors in vertex algebras is a fundamental issue in the theory of VOAs.
Matsuo and Nagatomo [@MN] have completely determined the set of conformal vectors of the Heisenberg VOA of rank one, which is one of the simplest non-trivial examples of a VOA. In this case, the conformal vectors are parameterized by two variables. In general, vertex algebras have conformal vectors that are parameterized by infinitely many variables, as observed by Matsuo and Nagatomo [@MN]. As it is difficult to express all conformal vectors explicitly, we attempt to classify them under the action of the vertex algebra automorphism group, which suffices to classify VOA structures on a vertex algebra up to isomorphism. To be more precise, let $V$ be a vertex algebra, and let $\mathrm{Aut}\,V$ denote the group of vertex algebra automorphisms of $V$. A conformal vector $a \in V$ is said to be of strong CFT type if $(V,a)$ is a VOA of strong CFT type. Here, a VOA of strong CFT type, introduced in [@DM], is roughly a self-dual $\mathbb{N}$-graded VOA (see Section 2 for the precise definition). We prove that if a vertex algebra $V$ is simple, then $\mathrm{Aut}\,V$ acts on the set of conformal vectors of strong CFT type transitively (Theorem \[main\]). Hence, a simple vertex algebra has at most one self-dual VOA structure up to isomorphism. Furthermore, we determine the set of conformal vectors of CFT type for a simple VOA of strong CFT type (Theorem \[classify\]).
As an application of Theorem \[main\], we investigate the structure of the full vertex algebra automorphism group. Hereafter, $(V=\bigoplus_{n \geq 0} V_n ,\omega)$ is assumed to be a simple VOA of strong CFT type. The VOA automorphism group $\mathrm{Aut}_\omega\,V$ is defined as the stabilizer of $\omega$ in $\mathrm{Aut}\,V$. VOA automorphism groups have been studied by many authors (see for example [@DG; @DN]). In particular, the VOA automorphism group of the moonshine module $(V^\natural,\omega)$ is the Monster, i.e., the largest sporadic finite simple group [@Gr; @Bo1; @FLM]. That is one of the motivations for studying vertex algebras [@Bo1; @Bo2]. In contrast, the only example of a VOA whose full vertex algebra automorphism group is explicitly known is the above-mentioned Heisenberg VOA of rank one [@MN]. It is natural to ask what the full vertex algebra automorphism group of the moonshine module is. In this paper, we show that the full vertex algebra automorphism group uniquely decomposes into the product of certain two subgroups and the VOA automorphism group, and determine the full vertex algebra automorphism groups for the real part of unitary VOAs. In particular, the full vertex algebra automorphism group of the moonshine module $V^\natural$ over $\mathbb{R}$ is shown to be the Monster.
Let us explain the basic idea. Let $pr_k$ denote the projection of $\bigoplus_{n \geq 0} V_n$ onto $V_k$. To study the structure of the automorphism group, we introduce two subgroups of $\mathrm{Aut}\,V$, denoted by $\mathrm{Aut}^+\,V$ and $\mathrm{Aut}^-\,V$. The group $\mathrm{Aut}^- \,V$ (resp. $\mathrm{Aut}^+\,V$) consists of all vertex algebra automorphisms $f$ such that $f - \mathrm{id}$ sends $V_n$ to $\bigoplus_{k<n} V_{k}$ (resp. $\bigoplus_{k>n} V_{k}$) for all $n \in \mathbb{N}$, where $\mathrm{id}$ is the identity map. The key observation in this paper is that if two ${\mathbb{Z}}$-gradings on a vertex algebra satisfy certain properties, then by using the projections with respect to the ${\mathbb{Z}}$-gradings, we can construct a vertex algebra automorphism which gives an isomorphism between the ${\mathbb{Z}}$-graded vertex algebras (Lemma \[fundamental\]). Set $CV^+= \{a \in \bigoplus_{n \geq 2} V_n\,|\, a \; \text{is a conformal vector and } pr_2(a)=\omega \}$. If $a \in CV^+$, then two gradings given by $a$ and $\omega$ satisfy the above properties, and a vertex algebra automorphism $\psi_{a}$ is constructed. Furthermore, we can prove that $\psi_a(\omega)=a$ and $\psi_a \in \mathrm{Aut}^+\,V$ (Theorem \[mor\]). In fact, there is a one-to-one correspondence between $CV^+$ and $\mathrm{Aut}^+\,V$. In particular, any conformal vector in $CV^+$ is conjugate to $\omega$ under $\mathrm{Aut}^+\,V$.
For a conformal vector $a \in V$, it is easy to show that $pr_1(a) \in J_1(V)=\{\;a \in V_1 \; | \; a(0)=0 \; \}$. The set $J_1(V)$ is first introduced by Dong et al. [@DLMM] in their study of the radical of a VOA. Matsuo and Nagatomo have shown that if $v \in J_1(V)$, then $\exp(v(1)) \in \mathrm{Aut}\,V$ [@MN]. In this paper, we show that $\mathrm{Aut}^-\,V=\{\exp(v(1)) \,|\, v \in J_1(V)\}$ and $\exp(-pr_1(a)(1))a \in \bigoplus_{n \geq 2} V_n$ (Proposition \[Aut\_m\] and Lemma \[J1\]). Furthermore, if $a$ is of strong CFT type, then we can show that $\exp(-pr_1(a)(1))a \in CV^+$ (Lemma \[scft\]), which proves that the action of $\mathrm{Aut}\,V$ on the set of conformal vectors of strong CFT type is transitive (Theorem \[main\]). As an application, we obtain that the full vertex algebra automorphism group $\mathrm{Aut}\,V$ uniquely decomposes into the product of three subgroups, $\mathrm{Aut}^+\,V$, $\mathrm{Aut}_\omega \,V$ and $\mathrm{Aut}^-\,V$ (Theorem \[grp\]).
We assume that a VOA $V$ over $\mathbb{R}$ has a positive-definite invariant bilinear form, which is equivalent to the condition that $V \otimes_{\mathbb{R}} \mathbb{C}$ is a unitary VOA with a natural anti-involution. Then, we show that $\mathrm{Aut}^+\,V=\{1\}$ and $\mathrm{Aut}\,V \cong \mathrm{Aut}^{-}\,V \rtimes \mathrm{Aut}_{\omega} \,V$ (Corollary \[uni\]). In particular, if $J_1(V)=0$, then the full vertex algebra automorphism group coincides with the VOA automorphism group. Importantly, the space $J_1(V)$ vanishes quite often. For example, if $V$ is a rational $C_2$-cofinite simple VOA of strong CFT type, then $J_1(V)=0$ [@Ma]. Many important VOAs (e.g., simple affine VOAs of level $k \in \mathbb{Z}_{>0}$, lattice VOAs, and the moonshine module) are unitary and satisfy $J_1(V)=0$. Hence, they provide us with many examples of VOAs over $\mathbb{R}$ whose full vertex algebra automorphism groups are equal to the VOA automorphism groups.
The remainder of this paper is organized as follows. Section 2 reviews some definitions and introduces the notion of conformal vectors of (strong) CFT type, which are discussed throughout this paper. Section 3 is devoted to the construction of certain automorphisms of conformal vertex algebras that play a key role (Theorem \[mor\]). Sections 4 and 5 describe the homogeneous components of a conformal vector with respect to the grading decomposition $V =\bigoplus_{n\in \mathbb{N}} V_n$ when it is of CFT type and of strong CFT type, respectively, which proves the uniqueness of self-dual vertex operator algebra structures of a simple vertex algebra (Theorem \[main\]). Section 6 proves the decomposition theorem of the automorphism group (Theorem \[grp\]). Finally, Section 7 is devoted to determining the full vertex algebra automorphism group of a VOA over $\mathbb{R}$ with a positive-definite invariant bilinear form, especially the moonshine VOA (Theorem \[pos\]).
Preliminaries and Notation
==========================
This section provides the necessary definitions and notations for what follows.
Preliminaries
-------------
We assume that the base field $\mathbb{K}$ is $\mathbb{R} \text{ or } \mathbb{C}$ unless otherwise stated. For a vertex algebra $V$, we let $\bm{1}$ denote the vacuum vector of $V$ and $T$ denote the endomorphism of $V$ defined by setting $Ta=a(-2)\bm{1}$ for $a\in V$. A $\mathbb{Z}$-graded vertex algebra is a vertex algebra that is a direct sum of vector spaces $V_n$ ($n \in {\mathbb{Z}}$) such that $V_n(k)V_m \subset V_{n+m-k-1}$ and $\bm{1} \in V_0$.
A $\mathbb{Z}$-graded conformal vertex algebra $(V,\omega)$ is a $\mathbb{Z}$-graded vertex algebra $V=\bigoplus_{n \in \mathbb{Z}}V_n$ with a distinguished vector $\omega \in V_2$ satisfying the following conditions: the operators $L(n)=\omega(n+1)$ generate a representation of the Virasoro algebra $$[L(n),L(m)]=(n-m)L(n+m)+(1/12)\delta_{n+m,0}(n^3-n) c,$$ where $c \in \mathbb{K}$ is a constant called the central charge; further, $L(0)v=nv$ if $v \in V_n$; and $L(-1)=T$. A $\mathbb{Z}$-graded conformal vertex algebra $(V,\omega)$ is said to be a [*VOA of CFT type*]{} if $V = \bigoplus_{n \in \mathbb{N}} V_n$, $\mathrm{dim}V_n$ is finite, and $V_0=\mathbb{K}{\bm{1}}$. A VOA $(V,\omega)$ of CFT type is said to be of [*strong CFT type*]{} if $L(1)V_{1} = 0$. A simple VOA $(V,\omega)$ of CFT type is of strong CFT type precisely when $(V,\omega)$ is self-dual in the sense that the contragredient module is isomorphic to $V$ as a $V$-module.
Let $V$ be a vertex algebra. A vector $a \in V$ is said to be a [*conformal vector (resp. conformal vector of CFT type, conformal vector of strong CFT type)*]{} if $(V,a)$ is a ${\mathbb{Z}}$-graded conformal vertex algebra (resp. VOA of CFT type, VOA of strong CFT type). Let $CV$ (resp. $CV_{CFT}$, $CV_{sCFT}$) denote the set of conformal vectors (resp. the set of conformal vectors of CFT type, the set of conformal vectors of strong CFT type). Set $V_n^a= \{ v \in V \;|\; a(1)v=nv \;\}$ for $n \in \mathbb{K}$ and $a \in V$. A vector $a \in V$ is a conformal vector if and only if it satisfies the following conditions:
- OPE $$\begin{aligned}
a(1)a=2a, \\
a(3)a \in \mathbb{K}{\bm{1}}, \\
a(n)a = 0 \text{ if } n=2 \text{ or } n \geq 4;\end{aligned}$$
- derivation property $$\begin{aligned}
a(0)= T;\end{aligned}$$
- semisimplicity $$\begin{aligned}
V = \bigoplus_{n \in \mathbb{Z}}V_n^a.\end{aligned}$$
A conformal vector $a$ is of CFT type if and only if it satisfies the following conditions: $$\begin{aligned}
V = \bigoplus_{n \in \mathbb{N}}V_n^a,\\
\dim V_n^a < \infty \text{ for } n \in \mathbb{N}, \label{cft} \\
V_0^a = \mathbb{K}{\bm{1}}.\end{aligned}$$ A vertex algebra automorphism $f$ of $V$ is a linear isomorphism of $V$ which preserves all the products: $$f(a(n)b)=f(a)(n)f(b) \text{ for } a,b \in V \text{ and } n \in {\mathbb{Z}}$$ with $f({\bm{1}})={\bm{1}}$. Let $(V,\omega)$ be a VOA of CFT type. A VOA automorphism $f$ of $(V,\omega)$ is a vertex algebra automorphism of $V$ which preserves the conformal vector: $$f(\omega)=\omega.$$ Let $\mathrm{Aut}\,V$ denote the group of vertex algebra automorphisms of $V$. Then, we have the following lemma:
\[auto\_stable\] The sets $CV$, $CV_{CFT}$, and $CV_{sCFT}$ are stable under the action of $\mathrm{Aut}\,V$.
Notation
--------
In the remainder of the paper, we assume that $V=\bigoplus_{n \in {\mathbb{Z}}} V_n$ is a ${\mathbb{Z}}$-graded vertex algebra, and fix its grading. We will use the following symbols.
[ll]{} $V=\bigoplus_{n \in {\mathbb{Z}}} V_n$ & The ${\mathbb{Z}}$-graded vertex algebra. $\mathrm{pr}_n$ &The canonical projection $\bigoplus_{i \in {\mathbb{Z}}} V_i \rightarrow V_n$. $V_{\geq n}$ &$ = \bigoplus_{i \geq n} V_i$. $\mathrm{Aut}^{+}\,V$ & $= \{\, f \in \mathrm{Aut}\, V\, |\, f(V_n)
\subset \bigoplus_{k \geq n}V_k \text{ and}\; \mathrm{pr}_n \circ f |_{V_n}=id_{V_n}
\text{ for all}\, n \in \mathbb{Z} \, \}$.
$\mathrm{Aut}^{-}\,V$ & $ = \{\; f \in \mathrm{Aut}\, V\; |\; f(V_n) \subset \bigoplus_{k \leq n} V_k \text{ and}\; \mathrm{pr}_n \circ f |_{V_n}=id_{V_n} \text{ for all}\; n \in \mathbb{Z} \; \}$. $\mathrm{Aut}^{0}\,V$ & $= \{\; f \in \mathrm{Aut}\, V\; |\; f(V_n) \subset V_n\}$. $J_1(V)$ &$ =\{a \in V_1\,|\,a(0)=0 \}$.$CV$ & The set of conformal vectors.$CV_{CFT}$ & The set of conformal vectors of CFT type. $CV_{sCFT}$ & The set of conformal vectors of strong CFT type.
Construction of Automorphisms
=============================
In this section, we construct certain automorphisms of the ${\mathbb{Z}}$-graded vertex algebra $V$. The following lemma, which is clear from $V_i(k)V_j \subset V_{i+j-k-1}$, is critical to our study:
\[bot\] Let $n, m, k$ be integers, $a \in V_{\geq n}$ and $b \in V_{\geq m}$. Then, $\mathrm{pr}_{n+m-k-1}(a(k)b)=\mathrm{pr}_n(a)(k)\mathrm{pr}_m(b)$.
\[subgroup\] $\mathrm{Aut}^{+}\,V$ is a subgroup of $\mathrm{Aut}\,V$.
Let $f \in \mathrm{Aut}^{+}\,V$. First, let us prove that $f^{-1} \in \mathrm{Aut}^{+}\,V$. Let $v$ be a nonzero element of $V_n$. Set $a=f^{-1}(v)$, $a_k=\mathrm{pr}_k(a)$ for $k \in {\mathbb{Z}}$, and $k_0=\mathrm{min} \{\;k\;|\;a_k \neq 0 \;\}$. Then, we have $v = f(f^{-1}(v))=f(\sum_{k \in {\mathbb{Z}}} a_k)=\sum_{k \geq k_0} f(a_k) \in V_n.$ Since $f \in \mathrm{Aut}^{+}\,V$, we have $\mathrm{pr}_{k_0}(v)=\mathrm{pr}_{k_0}( \sum_{k \geq k_0} f(a_{k_0}) )
=\mathrm{pr}_{k_0}( f(a_{k_0}) )=a_{k_0} \neq 0$. Thus, $k_0=n$, which proves $f^{-1}(v) \in V_{\geq n}$. Furthermore, $v=\mathrm{pr}_n(f \circ f^{-1}(v))=\mathrm{pr}_n(f^{-1}(v))$. Hence, $f^{-1} \in \mathrm{Aut}^{+}\,V$. It is clear that $\mathrm{Aut}^{+}\,V$ is closed under products. Hence, $\mathrm{Aut}^{+}\,V$ is a subgroup.
Suppose that the ${\mathbb{Z}}$-graded vertex algebra $V=\bigoplus_{n \in {\mathbb{Z}}} V_n$ admits another ${\mathbb{Z}}$-grading, $V = \bigoplus_{n \in {\mathbb{Z}}} V'_n$. The following simple observation is fundamental:
\[fundamental\] If $V'_n \subset V_{\geq n}$ for all $n \in {\mathbb{Z}}$, then there exists a vertex algebra homomorphism $\psi: V \rightarrow V$ such that $\psi(V'_n) \subset V_n$ for all $n \in {\mathbb{Z}}$. Furthermore, if $V'_n \subset V_{\geq n}$ and $V_n \subset V'_{\geq n}$ for all $n \in {\mathbb{Z}}$, then there exists $\psi \in \mathrm{Aut}^+\,V$ such that $\psi(V_n) = V'_n$ for all $n \in {\mathbb{Z}}$.
Let $i_n: V'_n \rightarrow V_{\geq n}$ denote the inclusion. Define $\psi: V \rightarrow V$ by $\psi|_{V'_n}=\mathrm{pr}_n \circ i_n$. Since $\bm{1} \in V_0 \cap V'_0$, we have $\psi({\bm{1}})={\bm{1}}$. Hence, Lemma \[bot\] implies that $\psi$ is a vertex algebra homomorphism. It is clear that $\psi(V'_n) \subset V_n$.
Now, suppose that $V_n \subset V'_{\geq n}$ for all $n \in {\mathbb{Z}}$. Then, $\psi$ is injective, since $V'_n \cap V_{\geq n+1} \subset
V'_n \cap V'_{\geq n+1}=0$. Let $v \in V_n$. Since $V_n \subset V'_{\geq n}$, we can write $v = \sum_{i \geq n} v'_i$ with $v'_i \in V'_i$. Then, $v-v'_n \in V'_{\geq n+1} \subset V_{\geq n+1}$. Hence, $\psi(v'_n)=\mathrm{pr}_n(v'_n)=\mathrm{pr}_n(v)+\mathrm{pr}_n(v'_n-v)=v$. Thus, $\psi$ is a vertex algebra automorphism.
We will show that $\psi \in \mathrm{Aut}^+\,V$. Since $\mathrm{Aut}^+\,V$ is a group, it suffices to show that $\psi^{-1} \in \mathrm{Aut
}^+\,V$. Clearly, $\psi^{-1}(V_n) \subset V_{\geq n}$. Let $v \in V_n$, and set $\psi^{-1}(v)=w$. According to the definition of $\psi$, $\mathrm{pr}_n(w)=v$. Hence, $\mathrm{pr}_n \circ \psi^{-1}|_{V_n}=\mathrm{id}$, which proves that $\psi \in \mathrm{Aut}^+\,V$.
\[fil\_lem\] If $V'_n \subset V_{\geq n}$ and $V'_n \cap V_{\geq n+1}=0$ for all $n \in {\mathbb{Z}}$, then $V_n \subset V'_{\geq n}$ for all $n \in {\mathbb{Z}}$.
Let $v$ be a nonzero element of $V_n$. Then, $v=\sum_{k \in {\mathbb{Z}}}v'_k$, $v'_k \in V'_k$. Set $k_0=\mathrm{min} \{\;k\;|\;v'_k \neq 0 \;\}$. Since $V'_n \subset V_{\geq n}$, we have $\mathrm{pr}_{k_0}(v)=\mathrm{pr}_{k_0}(\sum_{k \geq k_0}v'_k)=\mathrm{pr}_{k_0}(v'_{k_0})$. Since $V'_{k_0} \cap V_{\geq k_0+1}=0$ and $v'_{k_0} \neq 0$, we have $\mathrm{pr}_{k_0}(v'_{k_0}) \neq 0$, which implies that $k_0=n$, and $V_n \subset V'_{\geq n}$.
In the remainder of the paper, we assume that $(V,\omega)$ is a $\mathbb{Z}$-graded conformal vertex algebra, and fix its conformal vector $\omega$. Let $\psi \in \mathrm{Aut}^{+}\,V$. Then, $\psi(\omega) \in CV \cap V_{\geq 2} $ and $\mathrm{pr}_2(\psi(\omega))=\omega$. Set $CV^+ \coloneqq \{a \in CV \cap V_{\geq 2} \,|\, \mathrm{pr}_2(a)=\omega \}$. Then, we have a map $\phi: \mathrm{Aut}^{+}\,V \rightarrow CV^+$ given by $f \mapsto f(\omega)$. The following lemma asserts that $\phi$ is injective.
\[auto\_unique\] If $f, g\in \mathrm{Aut}^{+}\,V$ satisfy $f(\omega)=g(\omega)$, then $f=g$.
Since $f^{-1}(g(\omega))=\omega$, $f^{-1}\circ g(V_n)=V_n$ for all $n \in {\mathbb{Z}}$. According to Lemma \[subgroup\], $f^{-1} \circ g \in \mathrm{Aut}^{+}\,V$. In other words, $f^{-1} \circ g|_{V_n}=id_{V_n}$ for all $n \in {\mathbb{Z}}$, which proves the lemma.
The following theorem asserts that $\phi$ is surjective.
\[mor\] Assume that $a \in V_{\geq 2}$ satisfies the following conditions: $\mathrm{pr}_2(a)=\omega$, $a(0)=T$, $a(1)a=2a$; $a(1)$ is semisimple on $V$; and the eigenvalues are integers. Then, there exists a unique $\psi_a \in \mathrm{Aut}^{+}\,V$ such that $\psi_a(\omega)=a$. In particular, $a$ is a conformal vector.
\[Aut\_p\] The map $\phi: \mathrm{Aut}^{+}\,V \rightarrow CV^{+}$ , $f \mapsto f(\omega)$ is a bijection. The inverse is $\psi$, $a \mapsto \psi_a$ , defined in Theorem \[mor\].
Before we prove the above theorem, we need a preliminary result. Let $a \in V_{\geq 2}$. Recall that $V_m^a=\{ v \in V \;|\; a(1)v=mv \;\}$ and $V_m^{\mathrm{pr}_2(a)}=\{ v \in V \;|\; \mathrm{pr}_2(a)(1)v=mv \;\}$ for $m \in \mathbb{K}$. Suppose that $a(1)$ is semisimple on $V$. Since $a(1)$ maps $V_{\geq n}$ onto $V_{\geq n}$, $V_{\geq n}$ decomposes into the eigenspaces of $a(1)$. That is, $V_{\geq n}= \bigoplus_{m \in \mathbb{K}} V_{\geq n}\cap V_m^a$. Let $v \in V_{\geq n}$. According to Lemma \[bot\], $\mathrm{pr}_n(a(1)v)=\mathrm{pr}_2(a)(1)\mathrm{pr}_n(v)$. That is, $\mathrm{pr}_n \circ a(1)$ is equal to $\mathrm{pr}_2(a)(1)\circ \mathrm{pr}_n$ on $V_{\geq n}$. Therefore, $\mathrm{pr}_2(a)(1)$ is semisimple on $V$, and we have a map $\mathrm{pr}_n |_{V_{\geq n} \cap V_m^a} : V_{\geq n} \cap V_m^a \rightarrow V_n \cap V_m^{\mathrm{pr}_2(a)}$. Since $\mathrm{pr}_n: V_{\geq n} \rightarrow V_n$ is surjective, $\mathrm{pr}_n |_{V_{\geq n} \cap V_m^a} : V_{\geq n} \cap V_m^a \rightarrow V_n \cap V_m^{\mathrm{pr}_2(a)} $ is also surjective. The kernel of $\mathrm{pr}_n |_{V_{\geq n} \cap V_m^a} $ is $V_{\geq n+1}\cap V_m^a$. Thus, we have the following:
\[linear\] Let $a \in V_{\geq 2}$ and assume that $a(1)$ is semisimple on $V$. Then, $\mathrm{pr}_2(a)(1)$ is semisimple. Moreover, the projections $\mathrm{pr}_n$ induce isomorphisms $(V_{\geq n} \cap V_m^a)/(V_{\geq n+1} \cap V_m^a) \rightarrow V_n \cap V_m^{\mathrm{pr}_2(a)}$ for all $n \in {\mathbb{Z}}$ and $m \in \mathbb{K}$.
According to Lemma \[auto\_unique\], $\psi_a$ is unique if it exists. First, we show that $V_n$ and $V_n^a$ satisfy the assumptions in Lemma \[fundamental\]. Since $a(0)=T$, $V=\bigoplus_{n \in {\mathbb{Z}}} V_n^a$ is a ${\mathbb{Z}}$-graded vertex algebra. By applying Lemma \[linear\] to $a(1)$ and using $\mathrm{pr}_2(a)=\omega$, we find that the maps $(V_{\geq n} \cap V_m^a) /(V_{\geq n+1} \cap V_m^a) \rightarrow V_n \cap V_m$ are linear isomorphisms for all $n,m \in {\mathbb{Z}}$. If $n \neq m$, then $V_n \cap V_m=0$, which proves $V_n^a \subset V_{\geq n}$ and $V_n^a \cap V_{\geq n+1}=0$. According to Lemma \[fil\_lem\] and Lemma \[fundamental\], there exists $\psi_a \in \mathrm{Aut}^{+}\,V$ such that $\psi_a(V_n)=V_n^a$. By $a(1)a = 2a$, we have $a \in V_2^a$. Since $pr_2(a)=\omega$ and $a \in V_2^a$, we have $\psi_a(\omega)=a$, which proves the theorem.
\[pr2\] If $a \in CV \cap V_{\geq 2}$, then $\mathrm{pr}_2(a) \in CV \cap V_2$. Moreover, $V_n^a \cong V_n^{\mathrm{pr}_2(a)}$ as vector spaces for all $n \in {\mathbb{Z}}$.
To prove $\mathrm{pr}_2(a) \in CV$, it suffices to show that $\mathrm{pr}_2(a)$ satisfies the conditions (1),...,(5). According to Lemma \[bot\], $\mathrm{pr}_{3-n}(a(n)a)=\mathrm{pr}_2(a)(n)\mathrm{pr}_2(a)$. This shows that the conditions (1), (2), and (3) hold. Using $a(0)=L(-1)$, we have the condition (4). According to Lemma \[linear\], $\mathrm{pr}_2(a)(1)$ is semisimple, and the maps $(V_m^a \cap V_{\geq n})/(V_m^a \cap V_{\geq n+1}) \rightarrow V_m^{\mathrm{pr}_2(a)} \cap V_n$ are linear isomorphisms for all $n,m \in {\mathbb{Z}}$. Thus, the condition (5) holds. Furthermore, $V_m^{\mathrm{pr}_2(a)} \cong
\bigoplus_{n \in {\mathbb{Z}}} \,V_m^{\mathrm{pr}_2(a)}\cap V_n \cong
\bigoplus_{n \in {\mathbb{Z}}} \,(V_m^a \cap V_{\geq n})/(V_m^a \cap V_{\geq n+1}) \cong V_m^a
$ for all $m \in {\mathbb{Z}}$.
\[rem\_cft\] If $a \in CV_{CFT} \cap V_{\geq 2} $, then $\mathrm{pr}_2(a) \in CV_{CFT}$.
Shape of Conformal Vectors of CFT type
======================================
Hereafter, we assume that $(V,\omega)$ is a VOA of CFT type. In this case, $V_1$ has a Lie algebra structure such that $[a, b] = a(0)b$ for $a, b \in V_1$. Set $\mathrm{Aut}_\omega\,V=\{f \in \mathrm{Aut}\, V \,|\, f(\omega)=\omega \}$.
\[zero\] Let $a \in V_n$ with $n \geq 2$. If $a(0)\omega=0$, then $a \in \mathrm{Im}\,T$.
By applying skew-symmetry to $a(0)\omega$, we have $
0=a(0)\omega = -Ta+TL(0)a-T^2\sum_{k \geq 2}(-1)^{k}T^{k-2}/k!L(k-1)a
\in (n-1)Ta+\mathrm{Im}\,T^2$. Since $n \neq 1$, $Ta \in \mathrm{Im}\,T^2$. The operator $T$ is injective on $V_{\geq 1}$, since $V$ is of CFT type. Thus, $a \in \mathrm{Im}\,T$.
The Lie algebra $J_1(V) =\{a \in V_1\,|\,a(0)=0 \}$ is a subalgebra of the center of $V_1$. The following lemma shows that $J_1(V)$ is independent of the choice of the conformal vector of CFT type.
The map $J_1(V) \rightarrow \{a \in V\,|\,a(0)=0 \}/ \{\mathrm{ker}\,T+\mathrm{Im}\,T \}$ induced by the inclusion map $J_1(V) \rightarrow \{a \in V\,|\,a(0)=0 \}$ is an isomorphism.
Since $(V,\omega)$ is of CFT type, $\mathrm{Im}\,T \cap V_1=0$ and $\mathrm{ker}\,T=V_0$ hold, which implies that the map is injective. By combining this with Lemma \[zero\], it is surjective.
Recall that if $a \in J_1(V)$, then $a(1)$ is a locally nilpotent derivation and $\exp(a(1))$ is a vertex algebra automorphism of $V$ [@MN Lemma 1.3]. In fact, we have the following:
\[J1\_exp\] If $a \in J_1(V)$, then $\exp(a(1)) \in \mathrm{Aut}^-\,V$.
\[J1\] Let $a \in CV$. Then $\mathrm{pr}_1(a) \in J_1(V)$ and $\exp(-\mathrm{pr}_1(a)(1))a \in V_{\geq 2} \cap CV$.
Set $a_n = \mathrm{pr}_n(a)$ for all $n \in \mathbb{N}$. Recall that $T$ is an operator of degree one and $a_n(0)$ is of degree $n-1$. Therefore, by $T =a(0)$, we have $a_n(0)=0$ for all $n \neq 2$. Thus, $a_1 \in J_1(V)$. According to Lemma \[J1\_exp\], $\exp(-a_1(1))$ is a vertex algebra automorphism. Hence, $\exp(-a_1(1))a \in CV$. It remains to prove that $\exp(-a_1(1))a \in V_{\geq 2}$. First, we prove that $\exp(-a_1(1))(a_0+a_1+\omega)=\omega$. Clearly, $\exp(-a_1(1))(a_0)=a_0$. Using $a(1)a=2a$ and $a_0 \in \mathbb{K}\bm{1}$, we have $2a_0=\mathrm{pr}_0(2a)=\mathrm{pr}_0(a(1)a)
=\mathrm{pr}_0(a_0(1)a_2+a_1(1)a_1+a_2(1)a_0+...)
=a_0(1)a_2+a_1(1)a_1+a_2(1)a_0=a_1(1)a_1 \label{J1_1}$. This gives $\exp(-a_1(1))a_1=a_1-a_1(1)a_1=a_1-2a_0$. By applying skew-symmetry, we have $a_1(1)\omega=L(0)a_1+TL(1)a_1+...=a_1$; thus, $\exp(-a_1(1))\omega=\omega-a_1(1)\omega+(1/2)a_1(1)^2\omega= \omega -a_1+a_0$. Hence, $\exp(-a_1(1))(a_0+a_1+\omega)=\omega$. According to Lemma \[zero\] and $(a-\omega)(0)=T-T=0$, we have $a-a_0-a_1-\omega \in \mathrm{Im}\,T$. Since $\exp(-a_1(1))$ is an automorphism, it preserves $\mathrm{Im}\,T$, which is a subset of $V_{\geq 2}$. Hence, $\exp(-\mathrm{pr}_1(a)(1))a=\exp(-\mathrm{pr}_1(a)(1))(a_0+a_1+\omega+a-a_0-a_1-\omega) \in \omega + \mathrm{Im}\,T \subset V_{\geq 2}$.
\[Aut\_m\] The map $J_1(V) \rightarrow \mathrm{Aut}^{-}\,V$, $a \mapsto \exp(a(1))$ is an isomorphism of groups, where $J_1(V)$ is considered as the additive group.
Let $a, b \in J_1(V)$. A similar computation as in the proof of Lemma \[J1\] shows that $\exp(a(1))\exp(b(1))\omega=\omega+a+b+(1/2)a(1)a+a(1)b+(1/2)b(1)b=\exp((a+b)(1))\omega$. The proof of Lemma \[auto\_unique\] also works for $\mathrm{Aut}^{-}\,V$. Hence, $\exp(a(1))\exp(b(1))=\exp((a+b)(1))$ holds. Since $\mathrm{pr}_1(\exp(a(1))\omega)=a$, the map is injective. Let $f \in \mathrm{Aut}^{-}\,V$. Since $\exp(-\mathrm{pr}_1(f(\omega))(1))f(\omega)=\omega$, we have $f=\exp(\mathrm{pr}_1(f(\omega))(1))$. Hence, the map is an isomorphism.
By combining Lemma \[J1\], Corollary \[rem\_cft\], and Theorem \[mor\], we have:
\[cft\] If $V_2 \cap CV_{CFT}=\{ \omega \}$, then all the elements in $CV_{CFT}$ are conjugate under $\mathrm{Aut}\,V$.
Let $a \in V_2 \cap CV$. According to Lemma \[zero\] and $a(0)=T=\omega(0)$, $a-\omega \in \mathrm{Im}\,T$. Thus, we have:
\[V2\] Let $a \in V_2 \cap CV$. Then, there exists $h \in V_1$ such that $a=\omega+Th$.
\[sslie\] If $V_1$ is $0$ or a semisimple Lie algebra, then $J_1(V)=0$ and $V_2 \cap CV_{CFT}=\{ \omega \}$.
Let $a \in V_2 \cap CV_{CFT}$. When $V_1=0$, the assertion immediately follows from Lemma \[V2\]. Hence, we may assume that $V_1$ is a semisimple Lie algebra. In this case, the center of the Lie algebra $V_1$ is $0$. Then, $J_1(V)=0$. By Lemma \[V2\], there exists $h \in V_1$ such that $a=\omega+Th$. Since $h(0)$ commutes with $L(0)$ and $a(1)=L(0)-h(0)$ is semisimple, the operator $h(0)$ is also semisimple. Since $(V, a)$ is of CFT type, the eigenvalues of $a(0)=L(0)-h(0)$ on $V_1$ must be positive integers. If $h \neq 0$, then $h(0)$ has both positive and negative eigenvalues, since $V_1$ is a semisimple Lie algebra. Thus, $h=0$, and $a= \omega$.
According to Lemma \[auto\_stable\], $\mathrm{Aut}\,V$ acts on $CV_{CFT}$. Furthermore, since $\mathrm{Aut}_\omega\,V$ preserves the grading, $\mathrm{Aut}_\omega\,V$ acts on $ V_2 \cap CV_{CFT}$. Let us denote the set of orbits by $CV_{CFT}/\mathrm{Aut}\,V$ and $(V_2 \cap CV_{CFT})/ \mathrm{Aut}_\omega\,V$. Then, we have a natural map $(V_2 \cap CV_{CFT})/ \mathrm{Aut}_\omega\,V \rightarrow CV_{CFT}/\mathrm{Aut}\,V$. The following lemma asserts that the map is surjective:
If $a \in V_{\geq 2} \cap CV_{CFT} $, then there exists $\psi \in \mathrm{Aut}\,V$ such that $\psi(\mathrm{pr}_2(a))=a$.
Set $\omega'=\mathrm{pr}_2(a)$. Let $b \in V_{\geq 2}$ such that $a= \omega' + Tb$ by Lemma \[zero\]. We denote by $b^{\omega'}_1$ the image of $b$ under the projection $\bigoplus_{k \geq 0} V^{\omega'}_{k} \to V^{\omega'}_1$. According to Lemma \[rem\_cft\], $\omega'$ is a conformal vector of CFT type. Thus, $a-\omega'=Tb \in \mathrm{Im}\,T \subset V^{\omega'}_{\geq 2}$. If $b^{\omega'}_1=0$, then $a-\omega' \in V^{\omega'}_{\geq 3}$, which implies that there exists $\psi \in \mathrm{Aut}\,V$ such that $\psi(\omega')=a$ by Theorem \[mor\]. Thus, it suffices to show that $b^{\omega'}_1=0$. Since $b \in V_{\geq 2}$, and $V_{\geq 2}= \bigoplus_{k \geq 0}
V_{\geq 2} \cap V^{\omega'}_{k}$, we have $b^{\omega'}_1 \in V_{\geq 2}$. By Lemma \[linear\], $b^{\omega'}_1(0)$ is semisimple on $V$. Thus, by combining this with $b^{\omega'}_1 \in V_{\geq 2}$, we obtain $b^{\omega'}_1(0)=0$. Let $h \in V_1$ such that $\omega'=\omega + Th$. Since $h=\omega(1)h=(\omega'(1)-Th(1))h=\omega'(1)h$, we have $h \in V^{\omega'}_1$. Since $b^{\omega'}_1, h \in V^{\omega'}_1$ and $b^{\omega'}_1(0)=0$, we have $0=b^{\omega'}_1(0)h=-h(0)b^{\omega'}_1$. Hence, $b^{\omega'}_1=\omega'(1)b^{\omega'}_1=(\omega+Th)(1)b^{\omega'}_1=\omega(1)b^{\omega'}_1$. Since $b^{\omega'}_1 \in V_{\geq 2}$, this implies that $b^{\omega'}_1=0$, as desired.
\[surj\_cft\] The natural map $(V_2 \cap CV_{CFT})/ \mathrm{Aut}_\omega\,V \rightarrow CV_{CFT}/\mathrm{Aut}\,V$ is surjective.
Shape of Conformal Vectors of strong CFT type
=============================================
In this section, we prove our first main result (Theorem \[main\]).
\[pr2s\] Let $a \in CV_{sCFT} \cap V_{\geq 2}$. Then, $\mathrm{pr}_2(a) \in CV_{sCFT}$.
Let $a \in CV_{sCFT} \cap V_{\geq 2}$. According to Corollary \[pr2\], $\mathrm{pr}_2(a) \in CV_{CFT}$. It remains to show that $\mathrm{pr}_2(a)(2)V_1^{\mathrm{pr}_2(a)}=0$. According to Proposition \[linear\], $V_1^{\mathrm{pr}_2(a)}=\bigoplus_{k \in \mathbb{N}} \mathrm{pr}_k(V_{\geq k} \cap V_1^a)$. Let $k \in \mathbb{N}$ and $v \in V_{\geq k} \cap V_1^a$. Since $a \in CV_{sCFT}$, $a(2)V_1^a=0$. According to Lemma \[bot\], $0=\mathrm{pr}_{k-1}(a(2)v)=\mathrm{pr}_2(a)(2)\mathrm{pr}_k(v).$ Thus, $\mathrm{pr}_2(a)(2)V_1^{\mathrm{pr}_2(a)}=0$, which proves the lemma.
The following lemma is a generalization of Proposition 4.8 in [@CKLW].
\[scft\] If $(V,\omega)$ is a simple VOA of strong CFT type, then $CV_{sCFT} \cap V_2 = \{ \omega \}$.
Let $a \in CV_{sCFT} \cap V_2$. According to Lemma \[V2\], we may assume that $a=\omega + Th$, where $h \in V_1$. Suppose that $h \neq 0$. Since the operator $a(1)$ is semisimple, and $L(0)$ commutes with $h(0)$, $h(0)$ is also semisimple. Since $V$ is a simple VOA of strong CFT type, according to \[Li\], the Lie algebra $V_1$ has a non-degenerate invariant bilinear form $( \; , \; )$, which coincides with the first product up to scalar factor. Let $k \in \mathbb{K}$ and $v \in V_1$ such that $h(0)v=kv$. Then, $0=(h(0)h,v)=(h,h(0)v)=k(h,v)$. If $k \neq 0$, then $(h,v)=0$. Since the bilinear form is non-degenerate, there exists a vector $b \in V_1$ such that $h(0)b=0$ and $(h,b) \neq 0$ (i.e., $h(1)b \neq 0$). Since $a(1)b=(L(0)-h(0))b=b$ and $a(2)b=(L(1)-2h(1))b=-2h(1)b \neq 0$, we have $b \in V_1^a$ and $a(2)V_1^a \neq 0$, which contradicts the assumption that $a$ is a conformal vector of strong CFT type. Therefore, $h=0$ and $CV_{sCFT} \cap V_2 = \{ \omega \}$.
At this point, our main result follows from Lemma \[pr2s\], Lemma \[scft\] , Lemma \[J1\] and Theorem \[mor\].
\[main\] If $(V, \omega)$ is a simple VOA of strong CFT type, then all the elements in $CV_{sCFT}$ are conjugate under $\mathrm{Aut}\,V$.
Structure of the Vertex Algebra Automorphism group
==================================================
In this section, we prove Theorem \[grp\] and Theorem \[classify\]. Recall that $\mathrm{Aut}^{0}\,V = \{f \in \mathrm{Aut}\, V \,|\, f(V_n)=V_n \text{ for all}\; n \in \mathbb{N} \; \}$. Clearly, $\mathrm{Aut}_{\omega}\,V \subset \mathrm{Aut}^{0}\,V$.
If $(V,\omega)$ is a simple VOA of strong CFT type, then $\mathrm{Aut}_{\omega}\,V = \mathrm{Aut}^{0}\,V$, for which see [@CKLW Corollary 4.11].
\[grp\] For a simple VOA $(V,\omega)$ of strong CFT type, $\mathrm{Aut}\,V = \mathrm{Aut}^{-}\,V \mathrm{Aut}^{0}V \mathrm{Aut}^{+}\,V.$ More precisely, for $f \in \mathrm{Aut}\,V$, there exist unique elements $g \in \mathrm{Aut}^{-}\,V$, $h \in \mathrm{Aut}^{0}\,V$ and $k \in \mathrm{Aut}^{+}\,V$ such that $f=ghk$ holds. In particular, if $J_1(V)=0$, then $\mathrm{Aut}\,V \cong \mathrm{Aut}^{+}\,V \rtimes \mathrm{Aut}^{0}\,V$.
Let $f \in \mathrm{Aut}\,V$. First, we show the existence of the decomposition. Set $a=f(\omega)$ and $b=\exp(-\mathrm{pr}_1(a)(1))a$. Since $b \in CV_{sCFT} \cap V_{\geq 2}$, we have $b \in CV^{+}$. Let $g=\psi_b^{-1} \circ \exp(-\mathrm{pr}_1(a)(1)) \circ f $. Since $\psi_b^{-1} \circ \exp(-\mathrm{pr}_1(a)(1)) \circ f(\omega) =\omega$, we have $g \in \mathrm{Aut}_{\omega}\,V$. It is easy to check that $\psi_b \circ g = g \circ \psi_{g^{-1}(b)}$. Hence, we obtain the decomposition.
Next, we show the uniqueness. Let $f=\exp(h(1)) \circ g \circ \psi_a=\exp(h'(1)) \circ g' \circ \psi_{a'}$, where $h,h' \in J_1(V)$, $g,g' \in \mathrm{Aut}_{\omega}V$ and $a,a' \in CV^{+}$. Computations similar to Lemma \[J1\] show that $\mathrm{pr}_1(f(\omega))= \mathrm{pr}_1(\exp(h(1)) \circ g \circ \psi_a(\omega))=h$ holds. Hence, $h=h'$. Since $\psi_a \in \mathrm{Aut}^{+}\,V$, we have $\mathrm{pr}_n \circ g \circ \psi_a|_{V_n}=g|_{V_n}$. This implies that $g=g'$. Finally, by $\psi_a(\omega)=a$, we have $a=a'$. Since $\mathrm{Aut}_{\omega}\,V$ preserves the degrees, $\mathrm{Aut}\,V \cong \mathrm{Aut}^{+}\,V \rtimes \mathrm{Aut}^{0}\,V$ if $J_1(V)=0$.
\[classify\] If $(V,\omega)$ is a simple VOA of strong CFT type, then the natural map $(V_2 \cap CV_{CFT})/ \mathrm{Aut}_\omega\,V \rightarrow CV_{CFT}/\mathrm{Aut}\,V$ is a bijection.
According to Corollary \[surj\_cft\], this map is surjective. Let $a, a' \in V_2 \cap CV_{CFT}$, satisfying $f(a)=a'$ for some $f \in \mathrm{Aut}\,V$. By Lemma \[V2\], we may assume that $a=\omega + Tb$ and $a'=\omega+Tb'$, where $b, b' \in V_1$. Let $g \in \mathrm{Aut}^{-}\,V$, $h \in \mathrm{Aut}^{0}\,V$, and $k \in \mathrm{Aut}^{+}\,V$, satisfying $f=ghk$. Since $\mathrm{pr}_1(f(a))=\mathrm{pr}_1(a')=0$, we have $k=1$. Since $a'=\mathrm{pr}_2(f(a))=\mathrm{pr}_2(gh(\omega+Tb))=\mathrm{pr}_2(g(\omega)+Tgh(b))=\omega+Th(b)$, we have $b'=h(b)$. Thus, $h(a)=a'$, which implies that the map $(V_2 \cap CV_{CFT})/ \mathrm{Aut}_\omega\,V \rightarrow CV_{CFT}/\mathrm{Aut}\,V$ is injective.
Vertex Operator Algebra with Positive-Definite Invariant Form
=============================================================
In this section, we examine VOAs over the real number field $\mathbb{R}$. Let $(V, \omega)$ be a simple VOA of strong CFT type over $\mathbb{R}$. We assume that the invariant bilinear form $( , )$ is positive-definite.
Set $Q_n=\{v \in V_n \;|\; L(1)v=0 \}$. Since $(V, \omega)$ is a simple VOA of strong CFT type, $V_n=Q_n \bigoplus L(-1)V_{n-1}$ holds.
The following lemma was proved in [@Mi (3.16)].
\[pos\_l\] For all $n \in \mathbb{N}$ and $v \in V_n$, $(\bm{1},v(2n-1)v)(-1)^n \geq 0$ and the equality holds if and only if $v=0$.
For the convenience of the reader we repeat the proof from \[Mi\], thus making our exposition self-contained.
First, let $v \in Q_n$. Since $(v,v)=(\bm{1},v(2n-1)v)(-1)^n$ and the invariant bilinear form is positive-definite, our assertion clearly holds. We apply induction to the degree $n$. The assertion is clear if $n=0$ and $n=1$, since $V_1=Q_1$. For $n \geq 2$, assume that the assertion is true for $n-1$. Let $v \in V_n$. We may assume that $v$ takes the form $a+Tb$, where $a \in Q_n$ and $b \in V_{n-1}$. Then, $(\bm{1},v(2n-1)v)(-1)^n =(\bm{1},a(2n-1)a+a(2n-1)Tb+Tb(2n-1)a+Tb(2n-1)Tb)(-1)^n. $ Since $a \in Q_n$, the first term on the right, viz., $(\bm{1},a(2n-1)a)(-1)^n$, is non-negative. By $(\bm{1},\mathrm{Im}\,T)=0$, $(\bm{1},(Tb)(2n-1)a)=(\bm{1},a(2n-1)Tb-T(a(2n)Tb)+...)=(\bm{1},a(2n-1)Tb)$ holds. Furthermore, $(\bm{1},a(2n-1)Tb)=(a,Tb)(-1)^n=0$ because $Q_n$ and $\mathrm{Im}\,T$ are orthogonal. Finally, we have $$\begin{aligned}
(\bm{1},Tb(2n-1)Tb)(-1)^n&=&-(2n-1)(\bm{1},[b(2n-2),T]b+T(b(2n-2)b))(-1)^n \nonumber \\
&=&(2n-1)(2n-2)(\bm{1},b(2n-3)b)(-1)^{n-1} \geq 0 \nonumber\end{aligned}$$ by the induction hypothesis. Hence, $(\bm{1},v(2n-1)v)(-1)^n \geq 0$. If the equality holds, then $(a,a)=0$ and $((2n-1)(2n-2)(\bm{1},b(2n-3)b)(-1)^{n-1})=0$ follow from the computations presented above. Since $n \geq 2$, again, according to the induction hypothesis, $a=0$ and $b=0$, which is the desired conclusion.
\[nil\] If $CV \nsubset V_0 \oplus V_1\oplus V_2$, then there exist $n \in \mathbb{N}$ and a non-zero vector $v \in V_n$ such that $v(2n-1)v=0$ and $n \geq 3$.
Let $a \in CV \backslash (V_0 \oplus V_1\oplus V_2)$. Set $a_k=\mathrm{pr}_k(a)$ and $n=\mathrm{max} \{\;k\;|\;a_k \neq 0 \;\}$. Since $a \notin V_0 \oplus V_1\oplus V_2$ , $n \geq 3$. We have $a(2n-1)a=0$, because $a$ is a conformal vector. Hence, $0=\mathrm{pr}_0(a(2n-1)a)=a_n(2n-1)a_n$.
By combining the above lemmas, we have $CV \subset V_0 \oplus V_1\oplus V_2$. Hence, to classify conformal vectors, it suffices to consider conformal vectors in $CV \cap V_2$ by Lemma \[J1\]. According to Lemma \[V2\], they are of the form $\omega + Th $, where $h \in V_1$ and $h(0)$ is semisimple on $V$. The following lemma shows that $h \in J_1(V)$.
If $h \in V_1$, $v \in V$ and $0 \neq k \in \mathbb{R}$ satisfy $h(0)v=kv$, then $v=0$.
Similarly to the proof of Lemma \[scft\], we have $k(v,v)=(h(0)v, v)=-(v,h(0)v)=-k(v,v)$. Hence, $v=0$.
By the above argument, we have the following:
\[pos\] Let $(V, \omega)$ be a simple VOA of strong CFT type over $\mathbb{R}$. Suppose that the invariant bilinear form is positive-definite. Then, $CV = \{ (1/2)a(1)a+a+\omega+Tb \;|\;a, b \in J_1(V)\}$ and $\mathrm{Aut}\,V=\mathrm{Aut}^{-}\,V \rtimes \mathrm{Aut}_{\omega}\,V$. In particular, if $J_1(V)=0$, then $CV = \{ \omega \}$ and $\mathrm{Aut}\,V=\mathrm{Aut}_{\omega}\,V$.
Matsuo and Nagatomo [@MN] have shown that the full vertex algebra automorphism group of the Heisenberg VOA of rank $1$ and level $1$ over $\mathbb{C}$ is isomorphic to $\mathbb{C} \rtimes {\mathbb{Z}}/ 2{\mathbb{Z}}$. The above theorem states that the full vertex algebra automorphism group of the Heisenberg VOA of rank $n$ and level $1$ over $\mathbb{R}$ is $\mathbb{R}^n \rtimes O(n,\mathbb{R})$, where $O(n,\mathbb{R})$ is the orthogonal group.
Let us make some remarks on the positive-definite invariant bilinear form. A unitary VOA is a VOA over $\mathbb{C}$ with a positive-definite invariant Hermitian form and an anti-involution $\sigma$ (see [@DL] for a precise definition). If $V$ is a unitary VOA with an anti-involution $\sigma$, then $V^\sigma=\{v \in V\, |\,
\sigma(v)=v \}$ is a VOA over $\mathbb{R}$ with a positive-definite invariant bilinear form, which gives a one-to-one correspondence between unitary VOAs and VOAs over $\mathbb{R}$ with a positive-definite invariant bilinear form (see for example [@CKLW Remark 5.4]).
Hence, we have:
\[uni\] Let $V$ be a simple unitary VOA of strong CFT type with an anti-involution $\sigma$. Then, $\mathrm{Aut}\,(V^\sigma) \cong \mathrm{Aut}^{-}\,V^\sigma \rtimes \mathrm{Aut}_{\omega}\,V^\sigma$. Furthermore, if $J_1(V)=0$, then $\mathrm{Aut}\,(V^\sigma) \cong \mathrm{Aut}_{\omega}\,V^\sigma$.
We remark that simple affine VOAs of level $k \in \mathbb{Z}_{>0}$, lattice VOAs and the moonshine module are unitary \[DL\]. They also satisfy the condition $J_1(V)=0$.
\[mon\] The moonshine module $V^\natural$ over $\mathbb{R}$ has only one conformal vector, and its full vertex algebra automorphism group is the Monster.
\[rational\] If $V$ is a rational $C_2$-cofinite simple VOA of strong CFT type, then $J_1(V)=0$. This follows immediately from [@Ma Theorem 3]. Hence, if $V$ is a rational $C_2$-cofinite simple unitary VOA of strong CFT type with an anti-involution $\sigma$, then the vertex algebra automorphism group of $V^\sigma$ coincides with the VOA automorphism group.
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to thank his research supervisor, Professor Atsushi Matsuo, for his insightful comments and advice. The author also wishes to express his gratitude to Hiroki Shimakura for carefully examining the manuscript and for his valuable comments. This work was supported by the Program for Leading Graduate Schools, MEXT, Japan.
[99]{}
R.E. Borcherds, Vertex algebras, Kac-Moody algebras, and the Monster, Proc. Nat’l. Acad. Sci. USA. 83 (1986), 3068-3071.
R.E. Borcherds, Monstrous moonshine and monstrous Lie superalgebras, Invent. Math. 109 (1992), 405-444.
S. Carpi, Y. Kawahigashi, R. Longo, and M. Weiner, From vertex operator algebras to conformal nets and back, Mem. Am. Math. Soc. 254 (2018), no.1213, vi+85 pp. arXiv:1503.01260.
C. Dong and R.L. Griess Jr., Automorphism groups and derivation algebras of finitely generated vertex operator algebras, Michigan Math. J. 50 (2002) 227-239.
C. Dong, and X. Lin, Unitary vertex operator algebras, J. Algebra, 397 (2014), 252-277.
C. Dong, H. Li, G. Mason, and P. Montague, The radical of a vertex operator algebra, in: Proc. of the Conference on the Monster and Lie Algebras at Ohio State University, May 1996, ed. by J. Ferrar and K. Harada, Walter de Gruyter, Berlin- New York, 1998, 17-25.
C. Dong and G. Mason, Rational vertex operator algebras and the effective central charge, Int. Math. Res. Not. (2004), 2989-3008.
C. Dong and K. Nagatomo, Automorphism groups and twisted modules for lattice vertex operator algebras, Contemp. Math. 248 (1999), 117-133
I. Frenkel, J. Lepowsky, and A. Meurman, Vertex Operator Algebras and the Monster, Academic Press, Boston 1988.
R.L. Griess Jr., The friendly giant, Invent. Math. 69 (1982), 1-102.
H. Li, Symmetric invariant bilinear forms on vertex operator algebras, J. Pure Appl. Algebra 96(1994), 279-297.
G. Mason, Lattice subalgebras of strongly regular vertex operator algebras, in Conformal Field Theory, Automorphic Forms and Related Topics, Contrib. Math. Comput. Sci. 8, Springer, Heidelberg, (2014). 31-53.
M. Miyamoto, A new construction of the moonshine vertex operator algebra over the real number field, Ann. of Math, 159 (2004), 535-596.
A. Matsuo and K. Nagatomo, A note on free bosonic vertex algebra and its conformal vectors, J. Algebra 212 (1999), 395-418.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper we improve traditional steepest descent methods for the direct minimization of the Gross-Pitaevskii (GP) energy with rotation at two levels. We first define a new inner product to equip the Sobolev space $H^1$ and derive the corresponding gradient. Secondly, for the treatment of the mass conservation constraint, we use a projection method that avoids more complicated approaches based on modified energy functionals or traditional normalization methods. The descent method with these two new ingredients is studied theoretically in a Hilbert space setting and we give a proof of the global existence and convergence in the asymptotic limit to a minimizer of the GP energy. The new method is implemented in both finite difference and finite element two-dimensional settings and used to compute various complex configurations with vortices of rotating Bose-Einstein condensates. The new Sobolev gradient method shows better numerical performances compared to classical $L^2$ or $H^1$ gradient methods, especially when high rotation rates are considered.'
author:
- 'Ionut Danaila and Parimah Kazemi[^1]'
title: 'A new Sobolev gradient method for direct minimization of the Gross–Pitaevskii energy with rotation'
---
Sobolev gradient, descent method, finite difference method, finite element method, Bose-Einstein condensate, vortices.
Introduction
============
First experimental realizations of Bose-Einstein condensates (BECs) in dilute alkali-metal gases [@anderson; @bradley; @davis] led to an explosion of mathematical and theoretical studies aimed at better understanding such systems. Recent efforts were devoted to documenting the superfluid nature of the condensate by providing evidence for the existence of quantized vortices when rotating the condensate. It was indeed experimentally observed [@madison; @matthews; @aboshaerr; @haljan] that instead of solid body rotation, the condensate rotates by forming vortices with quantized circulation. Initially, a few vortices are formed, and, as the rotation frequency increases, the vortices form an array similar to the Abrikosov lattice observed in type II superconductors. Since the rotating BEC is a highly controllable system with a simple theoretical description, it provides a perfect set-up for the theoretical study of macroscopic systems with quantized vortices.
In the zero-temperature limit, a dilute gaseous BEC is mathematically described by a macroscopic wave function derived in the framework of the Gross–Pitaevskii (GP) mean field theory. The spatial configuration of the wave function $\psi(\mathbf{x})$, with $\mathbf{x}=(x, y, z)^t$, is obtained by minimizing the GP energy in the rotating frame, $$E(\psi) = \int_{{{\mathds R}}^3} \frac{\hbar^2}{2m} | \nabla \psi |^2 + \tilde{V}_{trap}|\psi|^2 + \frac{\tilde{g}}{2}|\psi|^4 + i \hbar \psi^* \tilde{\mathbf{\Omega}}\cdot(\mathbf{x} \times \nabla)\psi,$$ subject to the normalization condition, $\int_{{{\mathds R}}^3} | \psi|^2 = N$, with $N$ the number of particles (atoms). In the previous expression, $\hbar$ is Planck’s constant, $m$ the atomic mass of the gas, $\tilde{\mathbf{\Omega}}$ the angular velocity vector, and $\tilde{V}_{trap}$ the magnetic trapping potential with trap frequencies ($\omega_x, \omega_y, \omega_z$). We denote by $\psi^*$ the complex conjugate of $\psi$. The interactions between atoms are described by $\tilde{g}=\frac{4 \pi \hbar^2 a_s}{m}$, with $a_s$ the $s$-wave scattering length. As in most of experimental settings, we consider in the following that $\tilde{\mathbf{\Omega}} = (0, 0, \tilde{\Omega})^t$ and that $\tilde{V}_{trap}$ has a lower bound and $\tilde{V}_{trap}(\mathbf{x}) \rightarrow \infty$, as $\mathbf{x} \rightarrow \infty$. Since from the previous assumption we can infer that $\psi(\mathbf{x}) \rightarrow 0$, as $\mathbf{x} \rightarrow \infty$, it suffices to work in a bounded domain ${\cal D} \subset {{\mathds R}}^3$ with homogeneous Dirichlet boundary conditions $\psi=0$ on $\partial D$.
In practice it is common to scale the energy so that the units become dimensionless. Using the scaling $\mathbf{r} = \mathbf{x}/d$, $u(\mathbf{r})=\psi(\mathbf{x}) d^{3/2}/\sqrt{N}$, $\Omega = \tilde{\Omega}/\omega_\perp$, with $d=\sqrt{\frac{\hbar}{m \omega_{\bot}}}$ the harmonic-oscillator length and $\omega_{\perp}=\min(\omega_x, \omega_y)$ the transverse trap frequency, we obtain the non-dimensional energy (per particle) functional: $$E(u) = \int_{\cal D} \frac{| \nabla u |^2}{2} + V_{trap}|u|^2 + \frac{g}{2}|u|^4 - \Omega i u^* (A^t \nabla) u,
\label{gp}$$ where $V_{trap}=\frac{1}{\hbar \omega_{\bot}} \tilde{V}_{trap}$, $g=\frac{4 \pi N a_s}{d}$, and $A=(y,-x, 0)$. The mass conservation constraint becomes: $$\int_{\cal D} |u|^2 = \| u \|^2 = 1,
\label{eq-psi-norm}$$ where we denote by $\| . \| = \| . \|_{L^2({\cal D}, {{\mathds C}})}$.
For given constants $\Omega, g$, and trapping potential function $V_{trap}$, the minimizer $u_g$ of the functional (\[gp\]) under the constraint (\[eq-psi-norm\]) is called the ground state of the condensate. Local minima of the energy functional with energies larger that $E(u_g)$ are called excited (or metastable) states of the condensate. For a detailed discussion of the derivation of the Gross-Pitaevskii energy and the physics of rotating Bose-Einstein condensates see, for example, [@fetter] and [@lieb].
The two key issues in numerically computing ground or excited states of BEC are [*(i)*]{} how to derive a numerical algorithm that starts from a chosen initial state and iteratively diminishes the energy of the solution to rapidly converge to a local minimum of the functional (\[gp\]), and [*(ii)*]{} how to take into account the mass constraint (\[eq-psi-norm\]). These two issues are obviously connected and have to be considered together in deriving efficient numerical algorithms. We present in this paper new approaches to address both issues and prove their superior numerical performance in the case of the energy minimization of the Gross-Pitaevskii energy with rotation.
Most of the numerical algorithms proposed in the literature use the so-called [*normalized gradient flow*]{} [@bao], that consists in two steps: the steepest descent method is applied to the unconstrained problem, $$\frac{\partial u}{\partial t} = - \frac{1}{2} \frac{\partial E(u)}{\partial u} = \frac{\nabla^2 u}{2} - V_{trap} u - g |u|^2u + i \Omega A^t \nabla u,
\label{eq-grad-flow}$$ to advance the solution from the discrete time level $t_n$ to $t_{n+1}$; the obtained predictor ${\tilde u}(r,t_{n+1})$ is then normalized in order to satisfy the unitary norm constraint and set the solution at $t_{n+1}$: $$u(\mathbf{r},t_{n+1}) \triangleq \frac{{\tilde u}(\mathbf{r},t_{n+1})}{\|{\tilde u}(\mathbf{r},t_{n+1})\|}.
\label{eq-steep-norm}$$
The gradient flow equation (\[eq-grad-flow\]) (or the related [*continuous gradient flow*]{} equation, see [@bao]) can be viewed as a complex heat equation and, consequently, solved by different classical time integration schemes (Runge-Kutta-Fehlberg [@garcia-ripoll], backward Euler [@bao; @du; @bao2008], second-order Strang time-splitting [@bao; @du], combined Runge-Kutta-Crank-Nicolson scheme [@danaila1; @danaila2; @danaila3], etc.), and different spatial discretization methods (Fourier spectral [@garcia-ripoll], finite elements [@du], finite differences [@bao; @danaila1; @danaila2; @danaila3], sine-spectral [@bao], Laguerre–Hermite pseudo-spectral [@bao2008], etc.).
It is interesting to note that in the descent method (\[eq-grad-flow\]), the right-hand side represents the $L^2$-gradient (or [*ordinary*]{} gradient) of the energy functional. An important improvement of the convergence rate of the descent method was obtained in [@garcia-ripoll] by replacing the ordinary gradient with the gradient defined on the Sobolev space $H^1({\cal D}, {{\mathds C}})$. The same Sobolev gradient method (see [@jwn] for various applications of this method) was recently used to minimize simpler Schr[ö]{}dinger type functionals in [@lookman]. Similar increase of the convergence rate over the ordinary gradient method was reported. A first new contribution of the present paper is to introduce a new definition of the inner product to equip the Sobolev space $H^1$ in the case of the GP energy with rotation. A proof of the existence of the asymptotic limit for the evolution equations associated with the Sobolev gradients in a Hilbert space setting is also given. When implemented in a finite difference or finite element settings, the new Sobolev gradient method shows better numerical performances compared to classical $L^2$ or $H^1$ gradient methods, especially when high rotation rates ($\Omega$) are considered.
The second important contribution of this work concerns the issue of the mass conservation constraint (\[eq-psi-norm\]). Instead of the classical (and very popular) normalization approach (\[eq-steep-norm\]), we suggest a projection method that preserves the norm of the initial state through the minimization procedure. The idea to project the (Sobolev) gradient into the [*tangent space*]{} associated to the constraint was already used to derive numerical algorithms for minimizing harmonic maps [@alouges1; @pierre], and, recently, to numerically find the smallest eigenvalue and corresponding eigenvectors of a Hermitian operator [@alouges]. Different algorithms based on the projected gradient were developed in these studies and successfully applied to different energy functionals: the Oseen-Frank energy for liquid crystals [@alouges1], the Dirichlet energy of harmonic maps [@pierre] and the Hartree-Fock energy for quantum chemical molecular systems [@alouges]. We derive here a projected Sobolev gradient method adapted to the Gross-Pitaevskii energy functional and provide an explicit expression of the projected gradient that allows to minimize trajectories when Hilbert spaces other than $L^2$ are considered. The new projection method proved very helpful in numerical implementations and allowed to avoid alternative methods to treat the mass constraint by adding to the energy functional a penalty term with a Lagrange multiplier (e. g. [@garcia-ripoll; @bao1]).
The organization of the paper is as follows. In section \[sec-gradients\] we introduce an alternate inner product on the Sobolev space $H^1$ and show that this inner product is equivalent to the traditional inner product on $H^1$. The corresponding new Sobolev gradient is also derived. We discuss in section \[sec-minim\] a constructive projection method for the mass constraint and give our existence and convergence result for the asymptotic limit of the evolution equation defined by the Sobolev gradients. In section \[sec-implement\] we give a discussion of the finite difference and finite element implementations in two-dimensions. The last section is devoted to numerical tests designed as benchmarks to compare performances of different Sobolev gradient methods. The effectiveness of the newly proposed Sobolev gradient method is proved by computing stationary states of rotating BEC that are physically relevant (high rotation and large interaction constants).
Gradient descent methods using several gradients {#sec-gradients}
================================================
In optimization problems that use a gradient descent or ascent technique, one usually has a choice of norms to use in the argument. If the norm has an associated inner product, then one can obtain a gradient with respect to this inner product (see [@jwn] for an explanation). In the example of the minimization problem of Schr[ö]{}dinger type functionals, the gradient represents the direction of change per unit time. Therefore, one wants to choose the gradient in the descent method so that the change in energy is maximal at each step. For the case of the Gross-Pitaevskii energy with rotation, we notice that the energy can be written as $$\label{gpcov}
E(u)= \int_{\cal D} \frac{| \nabla u + i \Omega A^t u|^2}{2} + V_{eff}|u|^2 +\frac{g}{2}|u|^4$$ where, the [*effective*]{} trapping potential is defined as: $$V_{eff}(r)= V_{trap}(r) - \frac{ \Omega^2 r^2}{2}.$$ This form of the energy suggests the definition of a new norm to equip the domain of the functional such that the functional is coercive with respect to this norm. This implies that if the size of the argument is large, then naturally the value of the functional will be large as well, making it suitable for rotating cases.
Inner products and norms {#subsec-norm}
------------------------
We define three inner products on $C^1({\cal D},{{\mathds C}})$ and study the completion of this space with respect to the norm arising from each of these inner products. Consider the inner products: $$\langle u , v \rangle_{L^2} = \int_{\cal D} \langle u , v \rangle,
\label{eq-innpL}$$ $$\langle u , v \rangle_{H} = \int_{\cal D} \langle u, v \rangle + \langle \nabla u , \nabla v \rangle,
\label{eq-innpH}$$ and $$\langle u , v \rangle_{H_A} = \int_{\cal D} \langle u , v \rangle + \langle \nabla_A u , \nabla_A v \rangle,
\label{eq-innpHA}$$ where $\nabla_A = \nabla + i \Omega A^t$, $\Omega$ is a fixed positive number. Here $\langle \cdot , \cdot \rangle$ denotes the complex inner product. Each of these inner products leads to a norm which we will denote by $\| \cdot \|_{L^2}, \| \cdot \|_{H},$ and $\| \cdot \|_{H_A}$. For $X=L^2, H, H_A$, consider the completion of $\{ u \in C^1({\cal D}, {{\mathds C}}): \|u\|_X < \infty \}$ with respect to each of the respective norms. In the first case, one obtains the Hilbert space $L^2=L^2({\cal D}, {{\mathds C}})$, in the second case $H^1=H^{1,2}({\cal D}, {{\mathds C}})$, and in the third case we call the resulting Hilbert space $H_A=H_A({\cal D}, {{\mathds C}})$ (see [@adams] for details on Sobolev spaces). Furthermore, the following calculation shows how the three norms are related. We first note that: $$\langle \nabla_A u , \nabla_A v \rangle =
\langle \nabla u , \nabla v \rangle + \Omega ^2 r^2\langle u , v \rangle +
i \Omega (\langle A^tu, \nabla v \rangle - \langle \nabla u , A^t v \rangle).
\label{gradAuv}$$ If $r_{\cal D}$ denotes the radius of $D$, one has $$\langle \nabla_A u , \nabla_A u \rangle =|\nabla u + i \Omega A^t u|^2 \leq 2( |\nabla u |^2 + r_{\cal D}^2 \Omega^2 |u|^2),$$ and, consequently, $$\| u \|_{H_A} \leq \int_{\cal D} (1+2 r_{\cal D}^2 \Omega^2)|u|^2 + 2 |\nabla u|^2 \leq c \|u\|_H^2$$ where $c=max(1+2 r_{\cal D}^2 \Omega^2, 2)$. Hence one has that the $H^1$ norm dominates the $H_A$ norm.
In the same time, using the identity $$\label{identity}
\int_{\cal D} \langle A^t u ,\nabla v \rangle = - \int_{\cal D} \langle \nabla u , A^t v \rangle,$$ we infer from that $$\label{HAidentity}
\int_{\cal D} \langle \nabla_A u , \nabla_A v \rangle =
\int_{\cal D} \langle \nabla u , \nabla v \rangle + \int_{\cal D} \Omega ^2 r^2\langle u , v \rangle -
2i \Omega \int_{\cal D} \langle \nabla u, A^t v \rangle.$$ and, consequently, $$\int_{\cal D} |\nabla u + i \Omega A^t u|^2 =
\int_{\cal D} |\nabla u|^2 + \Omega^2 r^2 |u|^2 - 2 \Omega \langle i \nabla u , A^t u \rangle.$$ Also, for $\epsilon > 0$, one has the inequality $$ab = \frac{a}{\epsilon} b \epsilon \leq \frac{1}{2}\left( \left(\frac{a}{\epsilon}\right)^2 + (b \epsilon)^2\right).$$ Now, using the above inequality and Cauchy-Schwartz one has that $$2|\langle i \nabla u , A^t u \rangle| \leq 2|\nabla u| |A^t u| \leq (\epsilon|\nabla u|)^2 + \frac{| A^t u|^2}{\epsilon^2}.$$ Thus $$- 2 \Omega \langle i \nabla u , A^t u \rangle \geq - \Omega((\epsilon|\nabla u|)^2 + \frac{| A^t u|^2}{\epsilon^2}) \geq -\Omega((\epsilon|\nabla u|)^2 + \frac{r_{\cal D}^2}{\epsilon^2}|u|^2).$$ From this one has that $$\begin{aligned}
\int_{\cal D} |\nabla u + i \Omega A^t u|^2 \geq \int_{\cal D} |\nabla u|^2 + \Omega^2 r^2 |u|^2 - \Omega((\epsilon|\nabla u|)^2 + \frac{r_{\cal D}^2}{\epsilon^2}|u|^2) = \\
\int_{\cal D} (1-\Omega \epsilon^2)|\nabla u|^2 + (\Omega^2r^2 - \Omega \frac{r_{\cal D}^2}{\epsilon^2})|u|^2 \geq \\
\int_{\cal D} (1-\Omega \epsilon^2)|\nabla u|^2 - \Omega \frac{r_{\cal D}^2}{\epsilon^2}|u|^2\end{aligned}$$ Now we choose $\epsilon$ so that $0 < 1 - \Omega \epsilon^2 < 1$ and let $k = 1+ \frac{\Omega}{\epsilon^2}r_{\cal D}^2 $. Since $k > 1$, we can write $$\begin{aligned}
\nonumber
k \int_{\cal D} |u|^2 + | \nabla_A u|^2 > \int_{\cal D} k|u|^2 + |\nabla_A u|^2 \geq\\
\int_{\cal D} |u|^2 + (1-\Omega \epsilon^2)| \nabla u|^2 > (1-\Omega \epsilon^2)\int_{\cal D} |u|^2 + |\nabla u|^2.\end{aligned}$$ and infer that the $H_A$ norm dominates the $H^1$ norm. Hence the two norms are equivalent. Furthermore, we have the following relationship between the three Hilbert spaces, $$H^{1,2}({\cal D} , {{\mathds C}}) = H_A({\cal D},{{\mathds C}}) \subset L^2({\cal D},{{\mathds C}}).$$ As sets $H^1$ and $H_A$ are equal. However, by using the equivalent norm induced on $H_A$, we will see that the numerical performance of the descent method is improved for the minimization of the GP energy with rotation.
Gradients {#subsec-grad}
---------
The next step in writing a descent method to directly minimize the energy, as given in equation or , is to obtain a gradient corresponding to each inner product. Taking the Fr[é]{}chet derivative of , one gets that $$\label{derivgp}
E'(u)h= \int_{\cal D} \Re \ (\langle \nabla u , \nabla h \rangle + \langle 2V_{trap} \ u + 2g |u|^2u - 2i \Omega A \nabla u , h\rangle)$$ or equivalently $$\label{derivgpcov}
E'(u)h= \int_{\cal D} \Re \ ( \langle \nabla_A u , \nabla_A h \rangle + \langle 2V_{eff} \ u + 2g |u|^2u, h \rangle).$$ Since $E'(u)$ is a continuous linear functional from $H^1$ to ${{\mathds R}}$ then for each $u \in H^1$, there exists a unique member of $H^1$ which we denote by $\nabla_H E(u)$ so that $$E'(u)h = \Re \langle \nabla_H E(u), h \rangle_H,$$ for all $h \in H^1$. We say that $\nabla_H E : H^1 \rightarrow H^1$ is a gradient for $E$ taken with respect to the $H^1$ inner product. Likewise $E'(u)$ is a continuous linear functional from $H_A$ to ${{\mathds R}}$, thus it has a representation like the one given above. We denote this gradient by $\nabla_{H_A} E : H_A \rightarrow H_A$ (see [@jwn] for a background on gradients obtained in this manner). Furthermore, we note from that for all $h \in C^{\infty}_c({\cal D}, {{\mathds C}})$ one has that $$\label{el}
E'(u)h= \Re \langle \nabla_X E, h \rangle_{X}= \int_{\cal D} \ \Re \langle -\nabla^2 u + 2V_{trap} u + 2g |u|^2u - 2 i \Omega A^t \nabla u, h \rangle.$$ When $X=L^2$, we directly obtain the expression of $\nabla_{L^2} E$, the $L^2$ (or [*ordinary*]{}) gradient of $E$, already recalled in . From a practical point of view, it is interesting to note that $H^1$ and $H_A$ gradients will be computed using different forms of : the corresponding strong formulation for the finite difference implementation (see also [@garcia-ripoll]) and the weak formulation for the finite element implementation (see also [@lookman]).
Constrained energy minimization {#sec-minim}
===============================
Projection method for the mass constraint {#subsec-mass}
-----------------------------------------
Before discussing the gradient descent method, we give a brief description of the projection used to deal with the mass constraint. In approximating stationary states, one could in principle use a normalized gradient flow in conjunction with a traditional Lagrange multiplier for the constraint [@garcia-ripoll; @bao1]. We adopt here a different approach and develop a projection method that will, in the continuous case, enforce the constraint for all time.
The method for enforcing the constraint is presented in [@jwn] for any general constraint and hence does not provide the needed expression for our case. For the unitary norm constraint, several projected gradient methods are developed in [@pierre; @alouges], based on the idea to directly compute the gradient in the [*tangent space*]{} to the unit sphere. In this work, in order to facilitate the numerical implementation, we first compute the gradient and then project it into the tangent space. For this purpose, it is very helpful to derive an explicit expression of the projected gradient that allows to preserve the unitary norm of the solution through the minimization procedure. It should be noted that explicit expressions of the projected gradient are given in [@alouges] for the ${{\mathds R}}^n$ gradient flow of the linear eigenvalue problem on the unit sphere and for the $L^2$ gradient flow of the Hartree-Fock nonlinear eigenvalue problem. We derive below an explicit expression of the projected gradient that allows to minimize trajectories when other Hilbert spaces than $L^2$ are considered.
Let $X = L^2$, $H^1,$ or $H_A$. As previously stated, for each $u \in X$, one can find a member of $X$, denoted by $\nabla_X E(u)$, so that $E'(u)h = \langle h , \nabla_X E(u) \rangle_X$. We called such an element of $X$ a gradient of $E$ at $u$. Consider $\beta : X \rightarrow {{\mathds R}}$, $$\label{beta}
\beta(u) = \int_{\cal D} |u|^2.$$ Since we want to minimize the energy $E(u)$ subject to the constraint $\beta(u)=1$, we obtain the [*tangent space*]{} for our problem: $$T_{u,X} = null ( \beta'(u)) = \{ w \in X: \langle u , w \rangle_{L^2} = 0 \}.$$ Note that $T_{u,X}$ is a closed linear subspace of $X$, and for each $u \in X$, there exists a unique orthogonal projection from $X$ onto $T_{u,X}$. We denote this projection by $P_{u,X}$. Note also that $P_{u,X}$ is a linear transformation with domain $X$ and range $T_{u,X}$. Thus, this transformation depends on the Hilbert space and $u \in X$.
Let $u_0 \in X$ so that $\beta(u_0)=1$ and write the descent method with the projected gradient: $$\label{sd}
z(0)=u_0 \text{ and } z'(t)=-P_{z(t),X} \nabla_X E(z(t)).$$ We can easily see that $\beta(z)$ is constant since $$(\beta (z))'(t) = \beta'(z(t)) z'(t) = - \beta'(z(t)) (P_{z(t),X} \nabla_X E(z(t)))= 0,$$ for all $t$, as $P_{z(t),X}$ is the projection of $X$ onto the null space of $\beta'(z(t))$. Thus $\beta (z)$ is constant and if $u=\lim_{t \rightarrow \infty} z(t)$, then $\beta(u) = \beta(u_0)$, and the norm of the initial state is preserved. In conclusion, by projecting the Sobolev gradient of $E$ at $z(t)$ into the null space of $\beta'(z(t))$ for each $t$, we get that $z(t)$ satisfies the mass constraint for all $t$ (see [@jwn] for a more detailed development on this topic).
For numerical implementation purposes, we give below an heuristic derivation of the explicit expression of the projection (see [@pkme] for a more rigorous demonstration). If, for the sake of simplicity, ${\cal G} = \nabla_X E(u)$ denotes the Sobolev gradient gradient of $E$ at $u$, the projected gradient is determined from the following two conditions: $$P_{u,X} {\cal G} \in T_{u,X},
\label{eq-projg-1}$$ $$\langle P_{u,X} {\cal G} , h \rangle_{X} = E'(u)h, \, \forall h \in T_{u,X}.
\label{eq-projg-2}$$ In order to satisfy , we choose the projected gradient of the form $P_{u,X} {\cal G} ={\cal G} - B v_X$, with $B\in {{\mathds R}}$ a constant and $v_X \in X$ such as $$\langle v_X , h \rangle_{X} = \langle u , h \rangle_{L^2}, \, \forall h \in X.
\label{eq-vX}$$ The constant $B$ is then obtained by imposing . The final expression that will be used for numerical implementation is: $$P_{u,X} {\cal G} = {\cal G} - \frac{\Re \langle u, {\cal G} \rangle_{L^2}}{\Re \langle u, v_X \rangle_{L^2}}\, v_X,
\label{eq-projG}$$ with $v_X$ computed from . Note that if $X=L^2$, $v_X=u$ and we recover the explicit expression of the projected gradient given in [@alouges1]. It is also important to note that, in regard to numerical consideration as well as obtaining global existence, uniqueness, and asymptotic convergence, we need that the map $u \rightarrow P_{u,X} \nabla_X E(u)$ be $C^1$ as a map from $X$ to $X$. Using the above expression for the projection, we present in the next section some convergence results.
Convergence results in an infinite dimensional Hilbert space {#subsec-theory}
------------------------------------------------------------
In this subsection we define the evolution equation we use in the Hilbert space setting and give our global existence and convergence result for the constrained minimization problem. Here we extend the results obtained in [@pkme] for the case of the Gross-Pitaevskii energy without rotation. In this work, as well as in [@pkme], we move away from the general theory of Sobolev gradients as presented in [@jwn], since the criteria for asymptotic convergence of the evolution equation for constrained minimization problems is not available in [@jwn].
The idea below the following analysis is to show that the GP energy functional with rotation has the same properties as the GP energy without rotation if the norm $\| \cdot \|_{H_A}$ is used. We thus can adapt the results obtained in [@pkme] to our case. We start by noting that, due to the mass conservation, one can add a multiple of $\int_{\cal D} |u|^2$ to the Gross-Pitaevskii energy and the resulting functional will have the same minimizers as the original functional. The idea is to obtain a functional that is uniformly and strictly convex. We remind the reader that for $X$ a Hilbert space, we say that $E: X \rightarrow {{\mathds R}}$ is uniformly and strictly convex if there exists $\epsilon > 0$ so that $E''(u)(h,h) \geq \epsilon |h|_X^2$ for all $h \in X$.
Indeed, let us consider the form of the energy functional, and suppose that there exists $1 > \delta > 0$ so that $V_{eff} > \delta$. We observe that $$\begin{aligned}
E''(u)(h,h) = \int_{\cal D} | \nabla_A h |^2 + 2V_{eff}|h|^2 + 2g(|u|^2 |h|^2 + 2(\Re \langle u, h \rangle))^2) \geq \\
\int_{\cal D} | \nabla_A h |^2 + 2V_{eff}|h|^2 \geq \delta \int_{\cal D} | \nabla_A h |^2 + |h|^2 = \delta \|h\|^2_{H_A},\end{aligned}$$ and infer that $E: H_A \rightarrow {{\mathds R}}$ is uniformly and strictly convex with the assumption that $V_{eff}$ is bounded away from zero. Due to the equivalence of norms, $E: H^1 \rightarrow {{\mathds R}}$ is also uniformly and strictly convex. Note that if $V_{eff}$ is not bounded away from zero, then one can obtain this property by adding a multiple of the constraint to the energy. This does not change the minimization problem as indicated by the following theorem.
\[shift\] Let $E$ be a $C^2$ function on a subspace $X$ contained in $L^2({\cal D})$. Let $$E_{\epsilon}(u) = E(u) + \epsilon \int_{\cal D} |u|^2.$$ Then for $\beta(u)= \int_{\cal D} |u|^2$ and $h \in null(\beta'(u))$, $E'(u)h=0$ iff $E_{\epsilon}'(u)h=0$.
Some other properties of the functional are required to obtain the asymptotic convergence of the evolution equation. In particular, we need the functional to be continuously twice Fr[é]{}chet differentiable and bounded from below. The latter two properties are standard and we therefore omit them. With these properties checked, we can give our global existence and convergence in the asymptotic limit. Using the space $H_A$, the proof of the following two theorems are identical to the ones given in [@pkme]. Thus we omit the proofs and refer the reader to this work.
\[thmsdexists\] Suppose $X$ is a Hilbert space and that $E: X \rightarrow {{\mathds R}}$ is continuously twice Fr[é]{}chet differentiable. Suppose also that $\beta:X \rightarrow \mathbb{R}$ is a given function such that, if $P_{u,X}$ denotes the orthogonal projection of $X$ onto the nullspace of $\beta'(u)$, then the map $u \rightarrow P_{u,X}$ is $C^1$. Then $z(t)$ given by is uniquely defined for all $t \geq 0$.
\[asconv\] Suppose the hypothesis of Theorem \[thmsdexists\] and that $z(t)$ is given by equation , with $\nabla_X E(u_0) \neq 0$. If $E: X \rightarrow {{\mathds R}}$ is uniformly and strictly convex, then $$\lim_{ t \rightarrow \infty} z(t) = u$$ exists. Furthermore, there exist two constants $m$ and $c$ so that $\|u - z(t)\|_X \leq m e^{-ct}$, and $E'(u)h=0$ for all $h \in null (\beta'(u))$.
From the above two theorems, if we have in mind that the functional $E$ defined in is continuously twice Fr[é]{}chet differentiable and uniformly and strictly convex when the domain is considered to be $H_A$ or $H$, we obtain the result that $E$ has a minimizer in $H_A$ and in $H^1$ that satisfies the constraint $\beta$. Furthermore, this minimizer is obtained as the limit of the trajectory we defined in . This convergence result is not only important on its own, but, as we shall see, plays an important role in the rate of convergence of our numerical simulations.
Numerical implementation {#sec-implement}
========================
In this section we explain in detail the setup for our simulations using the descent method with both finite differences and finite elements discretization in two space dimensions. Both implementations follow the general lines of the algorithm described below.
If $Y=L_G, H_G, H_{A_G}$ denotes the finite dimensional Hilbert spaces resulting after the discretization of the domain ${\cal D}$, the descent method takes the following discrete form:\
starting from $u_0 \in Y$, define a trajectory $z_n, n \geq 1$ as (forward Euler scheme): $$z_0=u_0, \quad z_{n+1} = z_n - \delta t_n \nabla_{z_n,Y} E_G(z_n),
\label{eq-algo-descent}$$ where $\nabla_{u,Y} E_G (u)$ denotes the gradient obtained with respect to each inner product and projected following . The [*time-step*]{} value $\delta t_n$ could be optimized when computed as the local minimum of the real valued function $$r \rightarrow E_G(z_n - r \nabla_{z_n,Y} E_G(z_n)).
\label{eq-algo-lines}$$ As convergence criterion, the algorithm stops when the relative change in energy $E_G$ is below of an imposed limit. [ We note that in the continuous steepest descent algorithm, the constraint was satisfied for all time $t$ and hence for the converged solution. In the discrete case, due to the first order discretization in time, it is easy to see from that the norm is conserved at time level $(n+1)$ up to an error of order $(\delta t_n)^2 \| \nabla_{z_n,Y} E_G(z_n) \|_{L^2}$. After each (or several) iteration(s), one could also normalize the solution, as in [@pierre] where a Sobolev descent method with step-size 1 is used. This results in an improvement in the accuracy to which the constraint is preserved. The main observation that we made was that even though we used a first order discretization in time, our projection method allowed to take larger time steps when compared to the method using the normalization alone.]{}
Finite differences
------------------
We discretize $\cal D$ into an $N$ by $N$ equally spaced ($\delta_x = \delta_y =\delta$) grid and let ${\cal D}_G$ be the set of all $K=N^2$ grid points. Let $X$ be the collection of all complex valued functions on $D_G$. For $f \in X$, $(D_1 f)(x,y)$ is the approximation to the partial derivative in the first independent variable at $(x,y)$ and $(D_2 f)(x,y)$ is the approximation to the partial derivative in the second independent variable at $(x,y)$. We have used a fourth-order centered finite-difference scheme to approximate the first partial derivatives. When compared to the classical second order scheme, this high-order approximation proved very helpful in computing complex configurations (vortex lattices within the condensate) with reasonably fine grids. Furthermore $Df = \binom{D_1f}{D_2f}$. For $(x,y)$ a grid point, we also define $D_{1,A}$ and $D_{2,A}$ by $$(D_{1,A} f)(x,y)= (D_1f) (x,y) + i \Omega y f(x,y)$$ and $$(D_{2,A} f)(x,y)= (D_2f) (x,y) - i \Omega x f(x,y).$$ We denote by $D_A = \binom{D_{1,A}}{D_{2,A}}$, the discretized form of the operator $\nabla_A $. The three inner products that equip $X$ are defined as: for $f,g \in X$, $$\langle f , g \rangle_{L^2} = \langle f , g \rangle,$$ $$\langle f , g \rangle_{H} = \langle f ,g \rangle + \langle D_1f , D_1g \rangle + \langle D_2f , D_2g \rangle,$$ and $$\langle f ,g \rangle_{H_A} = \langle f , g \rangle + \langle D_{1,A}f , D_{1,A}g \rangle+ \langle D_{2,A}f , D_{2,A}g \rangle,$$ where $\langle \cdot , \cdot \rangle$ denotes the complex ${{\mathds C}}^K$ inner product. Note that $\langle \cdot , \cdot \rangle_{L^2}$ is analogous to the $L^2({\cal D}, {{\mathds C}})$ inner product, $\langle \cdot, \cdot \rangle_H$ is analogous to the $H^{1,2}({\cal D}, {{\mathds C}})$ inner product, and $\langle \cdot , \cdot \rangle_{H_A}$ is analogous to the $H_A({\cal D},{{\mathds C}})$ inner product.
Since $D_1, \ D_2, \ D_{1,A}, \ D_{2,A}$ can be viewed as a linear transformation acting on $\mathbb{C}^K$, we think of each of these transformations as a $K \times K$ matrix. Let $D_M^*$ denote the conjugate transpose of the corresponding matrix. We note that we can write the $H$ and $H_A$ inner products as $$\label{gradH}
\langle f , g \rangle_H = \langle (I + D_1^*D_1 + D_2^*D_2) f , g \rangle_{L^2}.$$ and $$\label{gradHA}
\langle f , g \rangle_{H_A} = \langle (I + D_{1,A}^*D_{1,A} + D_{2,A}^*D_{2,A}) f , g \rangle_{L^2}.$$ The collection $X$ makes a finite dimensional Hilbert space with each of the above inner products. We denote the resulting Hilbert spaces by $L^2_G$, $H_G$, and $H_{A_G}$. Now, we discretize the energy functional as given in equations and . Here the subscript $G$ denotes that we are in the finite difference setting. $$\begin{aligned}
E_G(f)= \delta^2 \sum_{{\cal D}_G} \frac{1}{2} (|D_1 f|^2 + |D_2 f|^2) + V_{trap_G}|f|^2 + \frac{g}{2}|f|^4 - \Omega\, rot_G f,\end{aligned}$$ where for ${{\bm x}}\in {\cal D}_G$, $$rot_G f({{\bm x}}) = \Re \ (i f({{\bm x}})^*A({{\bm x}}) \binom{(D_1 f)({{\bm x}})}{(D_2 f))({{\bm x}})},$$ and $$A(x,y) = (y \ \ -x).$$ Equivalently, $$\begin{aligned}
E_G(f)= \delta^2 \sum_{ {\cal D}_G} \frac{1}{2} \left( |D_{1,A}f|^2 + |D_{2,A}|^2 \right) + V_{eff_G}|f|^2 + \frac{g}{2}|f|^4.\end{aligned}$$ If we take a derivative of $E_G$, we get that $$\begin{aligned}
\nonumber
E_G'(f)h= \delta^2 \Re \sum_{ {\cal D}_G} \ \langle D_1h , D_1f \rangle &+& \langle D_2h , D_2f \rangle \\
&+&
2\langle h , V_{trap_G} f + g |f |^2 f - \Omega i A^t (Df)\rangle.\end{aligned}$$ Observe that for each $f \in X$, $E_G'(f)$ is a continuous linear transformation on $X$ using any of the three norms we specified. Thus it has a representation with respect to each of the inner products we defined above. Using this representation, we will obtain a gradient. Since the $L^2$ inner product is proportional to the Euclidean inner product, the ordinary or Euclidean gradient (i.e. the list of partial derivatives of $E_G$ taken with respect to each of the $K$ independent variables) is easily derived if the real valued transformation $E_G'(f)$ is rewritten as: $$\label{gradh}
E_G'(f)h= \Re \langle h , \nabla_{L^2} E(f) \rangle _{L^2}.$$ We get that $$\begin{aligned}
\nabla_{L^2} E(f)= \delta^2( D_1^*D_1 f\ + \ D_2^*D_2 f\ + 2(V_{trap_G}f \ + g |f|^2 f\ -\Omega i A^t(Df))).\end{aligned}$$
We now derive the other two gradients, $\nabla_H E(f)$ and $\nabla_{H_A} E (f)$, with respect to the $H$ and $H_A$ inner products. From and we obtain that $$E_G'(f)h = \Re \langle h , \nabla_H E_G (f) \rangle_H = \Re \langle h , ( I+D^*D)\nabla_H E_G (f) \rangle_{L^2},$$ and $$E_G'(f)h = \Re \langle h , \nabla_{H_A} E_G (f) \rangle_{H_A} = \Re \langle h , (I+D_A^*D_A) \nabla_{H_A} E_G (f) \rangle_L.$$ By comparing these equations to , we finally get that $$\label{gH}
\nabla_H E_G (f) = (I+D^*D)^{-1} \nabla_{L^2} E(f),$$ and $$\label{gHA}
\nabla_{H_A} E_G (f) = ( I+D_A^*D_A)^{-1} \nabla_{L^2} E(f).$$
The discrete descent method using the above finite difference dscretization was implemented in Matlab. The Sobolev gradients are computed from and by solving linear systems at each time step using a preconditioned conjugate gradient method. Since this part is time consuming on fine grids, we used a linesearch algorithm to locally compute the time step from . This resulted in a significant reduction of the number of iterations needed to achieve convergence.
Finite elements
---------------
The finite-elements implementation uses the free software FreeFem++ [@freefem], which proposes a large variety of triangular finite elements (linear and quadratic Lagrangian elements, discontinuous $P1$, Raviart-Thomas elements, etc.) to solve partial differential equations (PDE) in two dimensions (2D). FreeFem++ is an integrated product with its own high level programming language with a syntax close to mathematical formulations.
It is therefore very easy to implement the variational formulations associated to the calculation of the three gradients, since the definitions of scalar products – use an integral form. Following the developments in section , and using as definition of the complex inner product $\langle u, v \rangle = u v^*$, the ordinary gradient is derived from and computed as the solution ${\cal G}=\nabla_{L^2} E$ of the problem with homogeneous Dirichlet boundary conditions: $$\begin{aligned}
\int_{\cal D} {\cal G} \, h &=& \mathrm{RHS}, \label{eq-FE-gl2}\\
\mathrm{RHS} &=& \int_{\cal D} \nabla u \nabla h + 2\left[ V_{trap}\, u + (g |u|^2 )u - i \Omega A^t\nabla u \right] h,
\label{eq-FE-rhs}\end{aligned}$$ where $h$ stands now for the real valued basis function of the finite element space. Following , the $H^1$ gradient is directly computed by solving the equation: $$\label{eq-FE-GH1}
\int_{\cal D} \nabla {\cal G} \, \nabla h + {\cal G} h = \mathrm{RHS}, \quad\mbox{where}\quad {\cal G} = \nabla_{H} E.$$ It is interesting to note that is directly derived from the weak formulation of , with the obvious advantage to obtain a simpler right-hand side , which is derived by integrating by parts the weak form of the $L^2$ gradient. Therefore, in order to solve , it is not necessary to explicitly compute the $L^2$ gradient (by solving ), as required for the finite-difference implementation.
Observing from that the $H_A$ scalar product could be expanded to obtain the equivalent definition: $$\label{eq-FE-HA}
<u, v> _{H_A} = \int_{\cal D} \langle \left[ 1 + \Omega^2 (y^2+x^2) \right] u , v \rangle + \langle \nabla u , \nabla v \rangle -2i\Omega \langle A^t\nabla u , v \rangle,$$ the $H_A$ gradient is directly computed as the solution ${\cal G} = \nabla_{H_A} E$ of the problem: $$\label{eq-FE-GHA}
\int_{\cal D} \left[ 1 + \Omega^2 (y^2+x^2) \right] {\cal G} h + \nabla {\cal G} \nabla h -2i\Omega (A^t\nabla {\cal G}) h = \mathrm{RHS}.$$ It is interesting to emphasize the fact that previous equations are solved in complex variables. The approach based on the separation of the real and imaginary part of the gradient used in [@lookman] is not possible when computing the $H_A$ gradient. The FreeFem scripts are written in an optimized form using the pre-computation and factorization of the complex matrices associated to linear systems given by (\[eq-FE-GH1\]) and (\[eq-FE-GHA\]). It is interesting to note that the same matrices are involved in the computation of $v_X$ from ; the projected gradient could be therefore optimized in the same way. The implementation uses P1 (piecewise linear) finite-elements, with a P4 representation of the nonlinear terms appearing in . A fifth order quadrature formula was used to compute two-dimensional integrals. The FreeFem scripts allow to switch to P2 (piecewise quadratic) finite elements by a simple change of the definition of the generic finite-elements space. Adaptive mesh refinement was used for simulations of rotating BEC with dense lattice of vortices.
Numerical experiments {#numerics}
=====================
We first use a test case with analytical manufactured solution to ascertain the convergence of the steepest descent method for each of the three gradients. Then, we use the numerical set-up to compute simple metastable states of rotating Bose-Einstein condensates with single or multiple vortices. The performances of the three methods are comparatively evaluated. Finally, the new $H_A$ gradient method is used to compute complex configurations relevant for real rotating condensates (Abrikosov vortex lattice and giant vortex).
Test case with manufactured solutions
-------------------------------------
This test case is used as benchmark for the evaluation of the descent method for each of the three gradients ($L^2, H, H_A$). We consider a non-linear problem close to the Gross-Pitaevskii equation: $$\label{eq-FF-testC-eq}
-\frac{1}{2} \nabla^2 u + C_{trap}\, u + g |u|^2 u -i\Omega (A^t\nabla)u = f,$$ corresponding to the minimization of the energy functional: $$\label{eq-FF-testR-en}
E(u,f) =
\int_{\cal D} \frac{1}{2}|\nabla u|^2
+ C_{trap} \,|u|^2+ {g}\,\frac{|u|^4}{2} -(f^* u + f u^*)
- \Omega \Re(iu^* A^t \nabla u).$$ For this energy functional, the $L^2$ gradient is expressed as in , with $V_{trap}=C_{trap}=const.$ and a supplementary term $-2\langle f, h\rangle$ to be added. It should be noted that this is a test case of minimization without constraint.
In order to test the implemented methods, we [*manufacture*]{} solutions of (\[eq-FF-testC-eq\]): we consider a given expression for $u$ and calculate the corresponding right-hand side $f(x,y)$. A simple way to construct such manufactured solutions is to consider solutions with azimuthal symmetry: $$\label{eq-FF-testC-sol}
u_f(x,y) = U(r) \, \exp(i m \theta),$$ where $(r,\theta)$ are cylindrical coordinates ($r=\sqrt{x^2+y^2}$). Since the Laplacian in cylindrical coordinates reads $$\nabla^2 = \frac{1}{r} \frac{{\partial }}{{\partial }r}\left(r \frac{{\partial }}{{\partial }r}\right) +
\frac{1}{r^2} \frac{{\partial }^2}{{\partial }\theta^2},$$ and the new term corresponding to the rotation becomes $$A^t \nabla u = y\frac{{\partial }u}{{\partial }x} - x \frac{{\partial }u}{{\partial }y} = - \frac{{\partial }u}{{\partial }\theta}.$$ we obtain that $$f = F(r) \, \exp(i m \theta),$$ with $$\label{eq-FF-testC-fr}
F(r) = -\frac{1}{2} \frac{1}{r} \frac{{\partial }}{{\partial }r}\left(r \frac{{\partial }U}{{\partial }r}\right)
+ \frac{1}{2} \frac{m^2}{r^2} U + C_{trap}\, U + g U^3 - m \Omega U.$$ We choose the domain ${\cal D}$ to be a circle of radius $R$ and $$U = r^2(R-r),
\label{eq-FF-testC-solU}$$ which satisfies the homogeneous boundary condition $u=0$ for $r=R$. For this choice, we obtain useful analytical formulas for $$F(r)= -\frac{1}{2}(4R-9r) + \frac{1}{2} m^2(R-r) + C_{trap}\, U + C_N U^3 - m \Omega U,$$ and energy $$\label{eq-FF-testC-enaU}
E(u,f) = 2\pi\left(-\frac{R^6}{20}-m^2 \frac{R^6}{120} -C_{trap} \frac{R^8}{168} -C_N \frac{3R^{14}}{20020} \right) + m \Omega \pi \frac{R^8}{84}.$$ The contour patterns for such solutions are displayed in Fig. \[fig-manuf\] for $m=1$ and $m=3$.
![Contour patterns of the manufactured solution corresponding to equations and . Azimuthal wave numbers $m=1$ and $m=3$.[]{data-label="fig-manuf"}](kazemi_fig1){width="0.7\columnwidth"}
The numerical application for manufactured solutions consider the following parameters: $$C_{trap}=20, \quad g=100, \quad R=1, \quad m=3, \quad \Omega=10.$$ For this case, the theoretical values for energy and angular momentum of the exact solution are: $E = -0.505553$ and $L_z = 0.1122$, respectively. The computation is considered as converged if the relative variation of the energy is less than $\varepsilon=10^{-8}$.
Tables \[tab-manuf-FD\] and \[tab-manuf-FE\] assess the convergence of the descent method by computing different norms of the difference between the exact and computed solutions. Performance of each gradient method are quantified by extracting the overall computing (CPU) time and the number $n$ of time steps necessary to achieve convergence.
All test cases considered $u_0=0$ as the initial guess for the descent method. Different initial conditions (e.g. $u_0$ computed as the solution of the corresponding linear problem) were tested with similar convergence results.
$N$ $n$ CPU E(u) $\|u-u_f\|_{\infty}$ $\|u-u_f\|_{L^2}$ $\|u-u_f\|_{H}$
-------- ---------- --------- ----------- ---------- ---------------------- ------------------- ----------------- --
$L$ $2^6$ 1137 35.41 -.4828 .0010 2.25e-4 .0270
$H$ $2^6$ 100 16.23 -.4828 2.81-4 1.66e-5 .0021
$H_A$ $2^6$ 38 6.39 -.4828 1.41e-4 3.85e-6 4.73e-4
$L$ $2^7$ 1960 308.13 -.4941 4.11e-4 2.84e-4 .0185
$H$ $2^7$ 40 36.34 -.4941 6.32e-5 4.81e-4 5.28e-4
$H_A$ $2^7$ 18 17.22 -.4941 2.05e-5 6.84e-7 6.26e-5
$L$ $2^8$ $>3000$ $>2.04e3$ -.4985
$H$ $2^8$ 30 154.47 -.4997 4.59-5 9.77e-6 .0013
$H_A$ $2^8$ 14 73.21 -.4997 1.56e-6 1.36e-6 1.504e-4
: Test case with manufactured solutions. Algorithm efficiency and convergence test for the finite difference implementation (variable time step computation).[]{data-label="tab-manuf-FD"}
$M$/Triangles $n$ CPU E(u) $\|u-u_f\|_{\infty}$ $\|u-u_f\|_{L^2}$ $\|u-u_f\|_{H}$ $\delta t$
-------- --------------- --------- --------- ---------- ---------------------- ------------------- ----------------- ------------ --
$L$ 100/1776 1176 85 -.4934 1.988e-3 1.001e-3 1.705e-2 8e-4
$H$ 100/1776 47 3.4 -.4934 1.883e-3 9.220e-4 1.668e-3 1
$H_A$ 100/1776 14 1 -.4934 1.880e-3 9.140e-4 1.665e-2 3
$L$ 200/7064 4292 1252 -.5025 7.492e-4 4.200e-4 7.401e-3 2e-4
$H$ 200/7064 47 13.8 -.5025 5.530e-4 2.232e-4 6.548e-3 1
$H_A$ 200/7064 14 4.1 -.5025 5.390e-4 2.119e-4 6.474e-3 3
$L$ 400/27604 $>8000$ $>9193$ -.5027 5e-5
$H$ 400/27604 47 54.2 -.5047 1.687e-4 6.8535e-5 3.954e-3 1
$H_A$ 400/27604 14 16.2 -.5047 1.549e-4 5.730e-5 3.791e-3 3
: Test case with manufactured solutions. Algorithm efficiency and convergence test for the finite element implementation (fixed time step computation). The triangular mesh is generated with $M$ points on the border of the domain.[]{data-label="tab-manuf-FE"}
The first obvious observation is that the descent method using the ordinary $L^2$ gradient has very slow convergence rate because of very small time steps imposed by the stability limit of the method. This was expected since this method is the equivalent to the explicit Euler integration scheme for the imaginary-time propagation equation. A similar result was reported in [@lookman] for simpler Schr[ö]{}dinger type energy functionals. Larger time steps are allowed in the $H^1$ and $H_A$ methods, since the Sobolev gradients represent a preconditioning of the ordinary gradient [@garcia-ripoll; @pierre; @alouges1].
For the descent methods using a constant time step $\delta t$ (finite element implementation), we compare the computations performed using the maximum value $(\delta t)_{max}$ allowed by the stability of each method. These values, displayed in Tab. \[tab-manuf-FE\], were obtained by successive tests: the value of $\delta t$ was increased by 20% for each new run, until the computation became unstable. It should be noted that we were not interested in a refined numerical evaluation of the stability limit of each method, since computations using a more precise estimation of $(\delta t)_{max}$ did not result in a significant variation of the CPU time. The same approach to compare methods using their maximum time step allowed by stability reasons will be applied to all subsequent computations in this section.
Tables \[tab-manuf-FD\] and \[tab-manuf-FE\] also allows to relate the computing cost to the complexity of each method. As already stated, the descent method using the $L^2$ gradient can be regarded as an explicit backward Euler scheme. It therefore has little complexity and the computing cost per iteration step ([*i.e.*]{} the ratio CPU/n) is very low. Sobolev gradients are computing by solving linear systems, which adds extra computational cost. For the finite-difference implementation, equations or are solved by a preconditioned conjugate-gradient method; since this part of the algorithm is time consuming, the CPU time per iteration step (CPU/n) is multiplied up to a factor of 8, when compared to the $L^2$ gradient method. The situation is different in the finite-element implementation. Since the weak formulation of the equation is used, the computation of all gradients needs to solve a linear system. In order to have an optimized numerical implementation that can switch between the three descent methods, the matrix of this system is stored and factorized before the time loop. As a consequence, even though the matrix of the system in is simpler (mass matrix) than in or , the ratio (CPU/n) is identical for the computation off all gradients.
In all numerical tests, the convergence of the $L^2$ gradient method needs a large number of time steps, and, consequently, much larger CPU times than the Sobolev gradient methods. Since the performances of the $L^2$ gradient method are very poor, it will not be used in the following numerical experiments. We shall now focus on the comparison between the $H$ and $H_A$ method. For this test case considering a large value of $\Omega$, the $H_A$ gradient allows for larger time steps and therefore the computational time is considerably reduced, by approximately a factor of 3. This suggests that the preconditioning of the gradient introduced by the new $H_A$ inner product is very effective for computing cases with high rotation frequencies $\Omega$ (it goes without saying that the $H$ and $H_A$ methods are equivalent for $\Omega \rightarrow 0$).
Simulations of rotating Bose-Einstein condensates
-------------------------------------------------
In computing stationary states of rotating Bose-Einstein condensates, the initial state $u_0$ in the descent method plays a crucial role. The algorithm usually starts from a wave function distribution derived from the Thomas-Fermi approximation. In the strong interaction regime (large values of $g$), it is reasonable to neglect the contribution of the kinetic energy and work with the simplified energy functional: $$E_{{{{TF}}}} (\rho) = \int_{\cal D} V_{trap}\rho + \frac{g}{2} |u|^4.$$ The minimizer of this energy corresponds to the Thomas-Fermi atomic density:
$$\rho_{{{{TF}}}}(r) = |u|^2 =\left(\frac{\mu - V_{trap}}{g}\right)_+,
\label{eq-rhoTF}$$
where $\mu$ is the chemical potential. Since $\mu$ is a Lagrange multiplier, imposing the mass constraint in yields a relation for $\mu$. After computing the value of $\mu$, the Thomas-Fermi radius of the condensate can be determined from ($\rho_{{{{TF}}}}(R_{TF})=0$). When a rotation $\Omega$ is applied, the Thomas-Fermi approximation stands with $V_{eff}$ replacing $V_{trap}$. The resulting radius $R_{{{{TF}}}}^\Omega$ is used to estimate the size of the domain $\cal D$ in simulations ($r_D > R_{{{{TF}}}}^\Omega$) .
We also mention that the converged final state is characterized by its energy $E(u)$ and angular momentum $L_z(u)$ which gives a measure of the rotation: $$L_z(u) = \int_{\cal D} \Re \left(i u^* (A^t \nabla) u \right).
\label{eq-LZ}$$
### Off-center vortex case: harmonic trapping potential and small $\Omega$
The second numerical experiment considers the classical harmonic trapping potential and an initial state computed from the Thomas-Fermi approximation plus a singly quantized vortex of center located at $(x_v, y_v)$. We use an ansatz for the vortex described in [@danaila1]. The parameters of the simulation are the following: $$g = 500, \quad V_{trap}=r^2/2, \quad \Omega = 0.4, \quad x_{v}=0.5, \quad y_{v}=0.$$ The Thomas-Fermi radius is for this case $ R_{{{TF}}}^{\Omega} = 5.246$ and the computational domain is circular of radius $R= 1.25 R_{{{TF}}}^\Omega=6.56$. The final converged state contains a single vortex centered at the origin (see Fig. \[fig-offcenter\]).
![Off-center vortex case. Initial state with an off-center vortex and final converged state with a centered vortex. Contours of atomic density $|u|^2$.[]{data-label="fig-offcenter"}](kazemi_fig2a "fig:"){width="0.40\columnwidth"} ![Off-center vortex case. Initial state with an off-center vortex and final converged state with a centered vortex. Contours of atomic density $|u|^2$.[]{data-label="fig-offcenter"}](kazemi_fig2b "fig:"){width="0.40\columnwidth"}
$N$ $n$ CPU E(u) $L_z(u)$
------- ------- ------ ---------- -------- ----------
$H$ $2^6$ 1313 169.51 8.3587 .9998
$H_A$ $2^6$ 1197 166.34 8.3587 .9998
$H$ $2^7$ 1184 866.88 8.3605 .9999
$H_A$ $2^7$ 1127 890.06 8.3605 .9999
$H$ $2^8$ 1274 4.9548e3 8.3606 .9999
$H_A$ $2^8$ 1244 4.7882e3 8.3606 .9999
: Off-center vortex case. Algorithm efficiency and characterization ($E(u), L_z(u)$) of the converged state state for the finite difference implementation (variable time step computation).[]{data-label="tab-exp2-FD"}
\
$M$/triangles $n$ CPU E(u) $L_z(u)$
------- --------------- ------ --------- -------- ----------
$H$ 100/1762 701 56.97 8.3819 .994598
$H_A$ 100/1762 703 57.47 8.3795 .994575
$H$ 200/7064 1667 537.85 8.3720 1.00042
$H_A$ 200/7064 1717 556.41 8.3694 1.00022
$H$ 400/27604 1788 2.335e3 8.3699 1.00052
$H_A$ 400/27604 1831 2.407e3 8.3673 1.00032
: Off-center vortex case. Algorithm efficiency and characterization ($E(u), L_z(u)$) of the converged state state for the finite element implementation. The time step is set to 0.1 for all computations.[]{data-label="tab-exp2-FE"}
The comparative results are presented in Tabs. \[tab-exp2-FD\] and \[tab-exp2-FE\]. It is important to note that the convergence test must be set to $\varepsilon=10^{-8}$ in order to obtain a final state with a vortex centered at the origin and $L_z=1$ (theoretical value reached for the finest meshes). A relaxed convergence criterion will result in a vortex that is not exactly centered since the convergence rate is very slow at the end of the simulation. As expected, the $H_1$ and $H_A$ perform similarly because of the low value of $\Omega$.
### Vortex array case: harmonic-plus-quartic trapping potential and large $\Omega$
The harmonic trapping potential physically sets an upper bound for the rotation frequency, since for $\Omega=1$ the centrifugal force balances the trapping force and the confinement of the condensate vanishes. To overcome this limitation, different forms of the trapping potential are currently experimentally and theoretically studied. We use in the third numerical experiment a combined harmonic-plus-quartic potential (see also [@kasamatsu1; @danaila2; @danaila3; @fetter]) with the following parameters $$g = 500, \quad V_{trap}=r^2/2+r^4/4, \quad \Omega = 2.$$ The Thomas-Fermi radius is for this case $R_{{{TF}}}^{\Omega} = 3.40$. The computational domain is circular of radius $R_{max}= 1.25 R_{{{TF}}}^{\Omega}$. The initial state contains a central vortex plus an array of 6 vortices equally distributed on the circle of radius $0.25 R_{max}$. All the vortices have a winding number $m=1$, except the first vortex that has $m=2$ (Fig. \[fig-array\]). Since vortices with winding number $m>1$ are not physically stable, the $m=2$ vortex will split into two singly quantized vortices. The final state contains therefore a central vortex an array of 7 vortices (Fig. \[fig-array\]). The convergence test is relaxed to $\varepsilon=10^{-6}$.
![Vortex array case. Initial state with 6 vortices and final converged state with an array of 7 vortices. Contours of atomic density $|u|^2$.[]{data-label="fig-array"}](kazemi_fig3a "fig:"){width="0.40\columnwidth"} ![Vortex array case. Initial state with 6 vortices and final converged state with an array of 7 vortices. Contours of atomic density $|u|^2$.[]{data-label="fig-array"}](kazemi_fig3b "fig:"){width="0.40\columnwidth"}
gradient $N$ $n$ CPU E(u) $L_z$
---------- ------- ----- ---------- --------- --------
$H$ $2^6$ 610 89.07 11.2679 6.4549
$H_A$ $2^6$ 459 79.76 11.2670 6.4576
$H$ $2^7$ 530 466.34 11.2971 6.4603
$H_A$ $2^7$ 442 447.92 11.2959 6.4603
$H$ $2^8$ 539 2.4760e3 11.2990 6.4605
$H_A$ $2^8$ 441 2.245e3 11.2977 6.4691
: Vortex array case. Algorithm efficiency and characterization ($E(u), L_z(u)$) of the converged state state for the finite difference implementation (variable time step computation).[]{data-label="tab-exp3-FD"}
\
gradient $M$/triangles $n$ CPU E(u) $L_z$
---------- --------------- ----- -------- --------- --------
$H$ 100/1762 507 42.28 12.0553 6.1297
$H_A$ 100/1762 330 27.70 12.1413 6.1654
$H$ 200/7064 418 138.53 11.5341 6.3920
$H_A$ 200/7064 270 90.11 11.6171 6.4135
$H$ 400/27604 420 550.10 11.4017 6.4641
$H_A$ 400/27604 262 346.87 11.4846 6.4840
: Vortex array case. Algorithm efficiency and characterization ($E(u), L_z(u)$) of the converged state for the finite element implementation. The maximum allowed time step is 0.1 for the $H$ gradient and 0.2 for the $H_A$ gradient.[]{data-label="tab-exp3-FE"}
Tables \[tab-exp3-FD\] and \[tab-exp3-FE\] show that the converged state is the same for both finite difference and finite element implementations. The $H_A$ method has better stability properties and allows a CPU time gain up to 36%. This gain was expected since $\Omega$ is large for this case.
### Giant vortex and Abrikosov vortex lattice
Finally, to show that the new method has the capability to handle more complicated cases, we produce the giant vortex using the $H_A$ gradient in conjunction with the new projection method proposed to enforce the mass constraint. We use the parameters of the previous numerical experiment (harmonic-plus-quartic trapping potential) and progressively increase $\Omega$ from 2 to 4. Each computation starts from an initial field representing the converged state previously obtained for a lower value of $\Omega$. The transition from a vortex lattice to the giant vortex is observed (Fig. \[fig-giant\]). The giant vortex is a hole in the condensate (the atomic density goes to zero inside) with multiple phase defects. This particular vortex structure, theoretically analyzed in numerous studies [@kasamatsu1; @danaila2; @danaila3; @fetter], was captured using both the finite elements and finite difference simulations.
![Giant vortex case. Converged states for $\Omega=2.5, 3, 4$ showing the formation of a hole in the condensate (giant vortex) for high rotation rates Contours of atomic density $|u|^2$.[]{data-label="fig-giant"}](kazemi_fig4){width="0.90\columnwidth"}
A last complex computational case is illustrated in Fig. \[fig-abrikosov\]. For a harmonic trapping potential and high rotation frequency ($\Omega=0.95$) an Abrikosov vortex lattice forms in the condensate. The difficulty in computing this case in the strong-interaction regime (large values of $g$) comes from the fact that the condensate becomes larger and the vortex lattice denser when the value of $g$ is increased. In order to increase convergence, each computation starts from an initial field representing the converged state obtained for a lower value of $g$. During the iterative process, new vortices nucleate at the boundaries and slowly move towards their final equilibrium locations. In computing such configurations, containing several hundreds of vortices, the adaptive mesh refinement capabilities of FreeFem proved very helpful in reducing the computational time and correctly capturing vortex positions.
![Abrikosov vortex lattice case. Converged states for $\Omega=0.95$ and increasing values of the interactions constant $g$. Finite elements computations using mesh adaptivity. Contours of atomic density $|u|^2$.[]{data-label="fig-abrikosov"}](kazemi_fig5){width="0.82\columnwidth"}
Summary
=======
The numerical study of a rotating Bose-Einstein condensate has been the subject of many numerical studies, both in two dimensions (2D) and three dimensions (3D). Since most of the studies [@du; @kasamatsu1; @kasamatsu2; @danaila1; @danaila2; @danaila3; @bao; @bao1] use the imaginary time propagation method (equivalent to the gradient flow model ), there are few studies using direct minimization by Sobolev gradient methods. Nevertheless, replacing the ordinary $L^2$ gradient in a descent method with the Sobolev $H^1$ gradient proved effective in minimizing the 3D Gross-Pitaevskii energy [@garcia1; @garcia-ripoll] or simpler Schr[ö]{}dinger type functionals [@lookman].
In this work we introduced a new inner product ($H_A$) to equip the domain of the GP energy functional with rotation and derived the corresponding gradient. We demonstrated that numerical performance is enhanced by replacing in the descent method the $L^2$ or $H^1$ gradients with the gradient obtained from the $H_A$ inner product. The gain in computational time proved very important when configurations with high rotation rates are computed. We also introduced a new projection method to enforce the mass constraint. This method avoids more complicated approaches using an energy functional with a penalty term, or the traditional normalization method that performs the descent over a path of functions with an imposed norm.
These two new tools allowed to implement robust descent methods using finite difference and finite element spatial discretization. Both numerical settings proved very efficient in computing various complex two-dimensional configurations of rotating Bose-Einstein condensates.
We finally emphasize the fact that the new gradient and projection method for the mass constraint have a more general interest and could be also used in conjunction with existing numerical schemes (such as sophisticated time stepping procedures) to study the energy minimization of Gross-Pitaevskii type functionals.
Acknowledgements {#acknowledgements .unnumbered}
================
P. Kazemi acknowledges support from CNRS and thanks Professor Sylvia Serfaty for hosting her at Laboratoire Jacques-Louis Lions. She also thanks Insitut f[ü]{}r Quantenphysik, Universit[ä]{}t Ulm for their generous hospitality during her visits. I. Danaila acknowledges helpful discussions with Professor F. Hecht in optimizing FreeFem scripts.
[99]{} J. R. Abo-Shaeer , C. Raman, J. M. Vogels, W. Ketterle, Observation of vortex lattices in Bose-Einstein condensates, Science, 292 (2001), pp. 476-479.
R. Adams, *Sobolev Spaces*, Academic, New York, 1975.
A. Aftalion and I. Danaila, Three-dimensional vortex conigurations in a rotating Bose Einstein condensate, Phys. Rev. A, 68 (2003), 023603.
A. Aftalion and I. Danaila, Giant vortices in combined harmonic and quartic traps, Phys. Rev. A, 69 (2004), 033608.
A. Aftalion and Q. Du, Vortices in a rotating Bose-Einstein condensate: critical angular velocities and energy diagrams in the Thomas-Fermi regime, Phys. Rev. A, 64 (2001), 063603.
F. Alouges, A new algorithm for computing liquid crystal stable configurations: the harmonic mapping case, SIAM J. Numer. Anal., 34 (1997), pp. 1708-1726.
F. Alouges and C. Audouze, Preconditioned gradient flows for nonlinear eigenvalue problems and applications to the Hartree-Fock functional, Numerical Methods for Partial Differential Equations, 25 (2008), pp. 380-400.
M. H. Anderson, J. P. Ensher, M. H. Matthews, C. E. Wieman, and E. A. Cornell, Observation of Bose-Einstein condensation in a dilute atomic vapor, Science, 269 (1995), pp. 198-201.
W. Bao and Q. Du, Computing the ground state solution of Bose–Einstein condensates by a normalized gradient flow, SIAM J. Sci. Comput. 25 , 5 (2004), pp. 1674-1697.
W. Bao and J. Shen, A generalized-Laguerre-Hermite pseudospectral method for computing symmetric and central vortex states in Bose-Einstein condensates, Journal of Computational Physics, 227 (2008),pp. 9778-9793.
W. Bao and W. Tang, Numerical solution of the Gross-Pitaevskii equation for Bose-Einstein condensation, Journal of Computational Physics, 187 (2003), pp. 230-254.
C. C. Bradley, A. Sacket, J. J. Tollett, R. G. Hulet, Evidence of Bose-Einstein condensation in an atomic gas with attractive interactions, Phys. Rev. Lett. , 75 (1995), pp. 1687-1690.
I. Danaila, Three-dimensional vortex structure of a fast rotating Bose-Einstein condensate with harmonic-plus-quartic confinement, Phys. Rev. A, 72 (2005), 013605.
K. B. Davis, M. O. Mewes, M. R. Andrews, N. J. Druten, D. S. Durfee, D. M. Kurn, and W. Ketterle, Bose-Einstein condensation in a gas of sodium atoms, Phys. Rev. Lett., 75 (1995), pp. 3969-3973.
A. L. Fetter, Rotating trapped Bose-Einstein condensates, Laser Physics 18, 1 (2008), pp. 1 - 11.
A. L. Fetter, B. Jackson, S. Stringari, Rapid rotation of a Bose-Einstein condensate in a harmonic plus quartic trap, Phys. Rev. A, 71 (2005) , 013605.
J. J. García-Ripoll and V. M. P[é]{}rez-García, Optimizing Schr[ö]{}dinger Functionals Using Sobolev gradients: application to quantum mechanics and nonlinear optics, SIAM J. Sci. Comput., 23 (2001), pp. 1315-1333.
J. J. García-Ripoll and V. M. P[é]{}rez-García, Vortex bending and tightly packed vortex lattices in Bose-Einstein condensates, Phys. Rev. A, 64 (2001), 053611.
P. C. Haljan , I. Coddington, P. Engels, E. A. Cornell, Driving Bose-Einstein condensate vorticity with a rotating normal cloud, Phys Rev Lett., 87 (2001), pp. 210403-210407.
F. Hecht, O. Pironneau, A. Le Hyaric and K. Ohtsuke, FreeFem++ (manual), www.freefem.org.
K. Kasamatsu, M. Tsubota, M. Ueda, Giant hole and Circular superflow in a fast rotating BEC, Phys. Rev. A, 66 (2002), 053606.
K. Kasamatsu, M. Tsubota, M. Ueda, Vortex lattice formation in a rotating Bose-einstein condensate, Phys. Rev. A, 65 (2002), 023603.
P. Kazemi and M. Eckart, *Minimizing the Gross-Pitaevskii energy functional with the Sobolev gradient – Analytical and numerical results*, to appear in the International Journal of Computational Methods, preprint available arXiv:0906.3206.
E. H. Lieb and R. Seiringer, Derivation of the Gross-Pitaevskii equation for rotating Bose gases, Comm. in Math. Physics, 264 , 2 (2006), pp. 505-537.
K. W. Madison, F. Chevy, W. Wohleben, J. Daribard, Vortex formation in a stirred Bose-Einstein condensate, Phys. Rev. Lett,. 84 (2000), pp. 806-809.
M. R. Matthews, B. P. Anderson, P. C. Hajlan, D. S. Hall, M. J. Holland, J. E. Williams, C. E. Weiman, E. A. Cornell, Watching a superfluid untwist itself: recurrence of rabi oscillations in a Bose-Einstein condensate, Phys. Rev. Lett., 83 (1999), pp. 3358-3361.
J. W. Neuberger, Sobolev Gradients and Differential Equations, Lecture notes in mathematics, 1670, Springer, Berlin / Heidelberg, 1997.
M. Pierre, Newton and conjugate gradient for harmonic maps from the disc into the sphere, ESAIM: Control, Optimisation and Calculus of Variations, 10 (2004), pp. 142-167.
N. Raza, S. Sial , S. S. Siddiqi , T. Lookman, Energy minimization related to the nonlinear Schr[ö]{}dinger equation, Journal of Computational Physics, 228 (2009), pp. 2572-2577.
[^1]: UPMC Univ Paris 06, UMR 7598, Laboratoire Jacques-Louis Lions, F-75005, Paris, France and CNRS, UMR 7598, Laboratoire Jacques-Louis Lions, F-75005, Paris, France
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We study numerically the behavior of qubit coupled to a quantum dissipative driven oscillator (resonator). Above a critical coupling strength the qubit rotations become synchronized with the oscillator phase. In the synchronized regime, at certain parameters, the qubit exhibits tunneling between two orientations with a macroscopic change of number of photons in the resonator. The life times in these metastable states can be enormously large. The synchronization leads to a drastic change of qubit radiation spectrum with appearance of narrow lines corresponding to recently observed single artificial-atom lasing \[O. Astafiev [*et al.*]{} Nature [**449**]{}, 588 (2007)\].'
author:
- 'O.V.Zhirov'
- 'D.L.Shepelyansky'
date: 'October 10, 2007'
title: Synchronization and bistability of qubit coupled to a driven dissipative oscillator
---
In physics there are not so many simple quantum problems which are exactly solvable [@morse]. Two of them are monochromatically driven two-level atom (spin-half or qubit) [@landau] and quantum oscillator (unitary or dissipative) [@perelomov; @weiss]. One atom weakly coupled to a field in a resonator is known as the Jaynes-Cummings model which is also integrable [@jaynes; @eberly; @scully]. At strong coupling the dynamics may become nontrivial with the emergence of classical [@zaslavsky; @milonni] and quantum chaos [@graham1] but in this case one should have many atoms which may absorb many photons. For one atom even with a strong coupling to a quantum photonic field the dynamics is still relatively simple due to a total energy balance [@graham].
This old problem of a two-level atom coupled to photons regained recently a significant interest due to appearance of long living superconducting qubits [@vion] which can be strongly coupled to a microwave resonator [@wallraff; @houck; @astafiev]. There are also other possibilities of superconducting qubit coupling to a quantum oscillator [@ilichev; @buisson]. The oscillator can be realized as a tank circuit tuned to the Rabi frequency [@ilichev] or as a current-biased dc SQUID [@buisson] allowing efficient energy exchange with a qubit. Possibilities of qubit coupling to a cooled nanomechanical resonator are actively discussed [@cleland; @rugar] and coupling between micro-mechanical cantilever and atomic spin found impressive experimental implementations (see [@berman] and Refs. therein). However, the most intriguing way seems to be the coupling with a microwave resonator where a lasing has been realized recently with 6 - 30 photons pumped into the resonator [@astafiev]. Contrary to recent interesting theoretical studies [@berman; @rodrigues; @shnirman] where pumping is applied to a spin or qubit we concentrate here on the case where a monochromatic pumping is applied to a dissipative oscillator (resonator). Such a dissipative oscillator can be also viewed as a semiclassical detector which performs monitoring of a qubit. This continuous type of measurements is now actively discussed for superconducting qubits and other solid-state devices [@korotkov]. The continuous measurement of a superconducting qubit is realized in [@ilichev].
=8.5cm
The Hamiltonian of our model reads $$\hat{H} = \hbar \omega_0 \hat{n}
- \hbar \Omega \sigma_x /2
+ g \hbar \omega_0 ( \hat{a} + \hat{a}^{\dag}) \sigma_z
+ f \cos{\omega t} \left( \hat{a} + \hat{a}^{\dag} \right)
\label{eq1}$$ where $g$ is a dimensionless coupling constant, the driving force amplitude and frequency are $f=\hbar \lambda \sqrt{n_p}$ and $\omega$, the oscillator frequency is $\omega_0$ and $\hbar \Omega$ is the qubit energy spacing. As in [@shnirman] we choose a qubit coupling via $\sigma_z$. We assume that the qubit life time is enormously long and that its dynamics is perturbed only by the coupling with the driven dissipative oscillator. The dissipation rate of oscillator is $\lambda$ and we assume the quality factor to be $Q=\omega_0/\lambda \sim 100$. The evolution of the whole system is described by the master equation for the density matrix $\hat{\rho}$ which has the standard form [@weiss]: $$\dot{\hat{\rho}} = - \frac{i}{\hbar} [\hat{H},\hat{\rho}] +
\lambda ( \hat{a} \hat{\rho} \hat{a}^{\dag}
- \frac{1}{2} \hat{a}^{\dag} \hat{\rho} \hat{a}
- \frac{1}{2} \hat{\rho} \hat{a}^{\dag} \hat{a})
\label{eq2}$$ The numerical simulations are done by direct solution of time evolution of $\hat{\rho}$ expanded in a finite basis of oscillator states $n$, by the state diffusion method [@stdif] and by the method of Quantum Trajectories (QT) [@percival]. We ensured that these methods give the same results but the majority of data is obtained with quantum trajectories which we found to be more suitable for massive simulations. In addition the QT have an advantage of providing a pictorial illustration of individual experimental runs. The numerical details are the same as in [@benenti; @zhirov] and we use here up to $n=70$ oscillator states which give good numerical convergence. Our results show that a coupling of two simple integrable models gives a nontrivial interesting behavior.
A typical example of QT is shown in Fig. \[fig1\]. It shows two main properties of the evolution: the oscillator spends a very long time at some average level $\langle n \rangle = n_-$ and then jumps to another significantly different value $n_+$. At the same time the polarization vector of qubit $\vec{\xi}$ defined as $\vec{\xi} = Tr(\hat{\rho} \vec{\sigma})$ also changes its orientation direction with a clear change of sign of $\xi_x$ from $\xi_x >0$ to $\xi_x<0$. The time averaged values of $\xi_{y,z}$ are zero but when they are taken at stroboscopic integer moments $\omega t/2\pi$ they also show transitions between two metastable states. The transition time is approximately $t_m \sim 1/\lambda$ being rather small compare to the life time in a metastable state. Inside such a state the degree of qubit polarization $\xi=| \vec{\xi} | $ is very close to unity showing that the qubit remains mainly in a pure state. The drops of $\xi$ appear only during transitions between metastable states. Special checks show that an inversion of $\xi_x$ by an additional pulse (e.g. from $\xi_x >0$ to $\xi_x <0$) produces a transition of oscillator to a corresponding state (from $n_-$ to $n_+$) after time $t_m \sim 1/ \lambda$. Thus we have here an interesting situation when a quantum flip of qubit produces a marcoscopic change of a state of detector (oscillator) which is continuously coupled to a qubit (we checked that even larger variation $n_{\pm} \sim n_p$ is possible by taking $n_p=40$). In addition to that inside a metastable state the coupling induces a [*synchronization*]{} of qubit rotation phase with the oscillator phase which in its turn is fixed by the phase of driving field. The synchronization is a universal phenomenon for classical dissipative systems [@pikovsky]. It is known that it also exists for dissipative quantum systems at small effective values of $\hbar$ [@zhirov]. However, here we have a new unusual case of qubit synchronization when a semiclassical system produces synchronization of a pure quantum two-level system.
=7.5cm
-0.0cm
The phenomenon of qubit synchronization is illustrated in a more clear way in Fig. \[fig2\]. The top panels taken at integer values $\omega t/2\pi$ show the existence of two fixed points in the phase space of oscillator (left) and qubit (right) coupled by quantum tunneling (the angles are determined as $\xi_x=\xi \cos \theta, \xi_y=\xi \sin \theta \sin \phi,
\xi_z = \xi \sin \theta \cos \phi$). A certain scattering of points in a spot of finite size should be attributed to quantum fluctuations. But the fact that on enormously long time (Fig. \[fig1\]) the spot size remains finite clearly implies that the oscillator phase $\varphi$ is locked with the driving phase $\omega t$ inducing the qubit synchronization with $\varphi$ and $\omega t$. The plot at $t$ values incommensurate with $2\pi/\omega$ (middle panels) shows that in time the oscillator performs circle rotations in $(p,x)$ plane with frequency $\omega$ while qubit polarization rotates around $x$-axis with the same frequency. Quantum tunneling gives transitions between two metastable states. The synchronization of qubit phase $\phi$ with oscillator phase $\varphi$ is clearly seen in bottom left panel where points form two lines corresponding to two metastable states. This synchronization disappears below a certain critical coupling $g_c$ where the points become scattered over the whole plane (panel bottom right). It is clear that quantum fluctuations destroy synchronization for $g < g_c$. Our data give $g_c \simeq 0.008$ for parameters of Fig. \[fig1\].
=4.2cm =4.2cm
-0.0cm
The variation of bistable states with coupling strength $g$ is shown in Fig. \[fig3\]. The difference between $n_+$ and $n_-$ grows with $g$. It is striking that $n_-$ may become close to zero. The direction of qubit polarization also changes in a smooth but nontrivial way. It is also important to note that according to our data the dispersion of oscillator wave function in metastable states is compatible with the dispersion of a coherent state with $n_\pm$. This corresponds to a wave packet collapse induced by dissipation (see [@benenti; @zhirov] and Refs. therein). The dependence of $n_{\pm}$ on driving frequency $\omega$ is shown in Fig. \[fig4\]. A symmetric double peak structure is evident: for $\omega > \omega_0$ the metastable state with $\xi_x <0$ has maximal $n$ value while for $\omega < \omega_0$ the state with maximal $n$ has $\xi_x >0$ (note color interchange). The peak width is approximately equal to the dissipation rate $\lambda$. With the increase of $g$ their form becomes asymmetric indicating importance of nonlinear effects. The splitting of peaks grows approximately linearly with $g$ (inset at Fig. \[fig4\]) and reminds the vacuum Rabi splitting effect [@eberly]. The shift $\Delta \omega_{\pm} $ explains two states $n_\pm$ of driven oscillator well described by $n_\pm = n_p \lambda^2/(4(\omega - \omega_0 -\Delta \omega_\pm)^2+\lambda^2)]$ (see dashed curves in Fig. \[fig3\] left traced with numerical values of $\Delta \omega_\pm $ from Fig. \[fig4\] inset). To estimate $\Delta \omega_\pm$ we note that the frequency of effective Rabi oscillations between quasi-degenerate levels is $\Omega_R \approx g \omega_0 \sqrt{n_\pm+1}$ [@jaynes; @scully] that gives $\Delta \omega_\pm \approx d \Omega_R/dn
\approx \pm g \omega_0/2\sqrt{n_\pm+1}$ in a good agreement with data of Fig. \[fig4\] for moderate $g$.
=7.5cm
-0.0cm
The properties of two metastable states are analyzed in Fig. \[fig5\]. The number of transitions $N_f$ between these states has a pronounced peak at $\Omega \approx 1.1 \omega_0$ that approximately corresponds to a resonance condition $\Omega - \omega \approx 2g\omega_0$. For $\Omega < 1.08 \omega_0$ there is an abrupt drop of $N_f$ and bistability becomes irregular disappearing for certain $\Omega$, but the synchronization still remains. Quite interestingly, the data show that for $\Omega > 1.1 \omega_0$ the life times of each state are rather different and enormously large, generally $\tau_- > \tau_+ \gg \omega_0/\lambda$.
=7.0cm
-0.0cm
The spectrum of qubit radiation $S(\nu)$ in presence of phase noise in $\phi$ is shown in Fig. \[fig6\]. It confirms the main features discussed above: for $\Omega/\omega_0=1.2$ the growth of driving power $n_p$ induces the synchronization of qubit with radiation suppression at qubit frequency $\Omega =1.2 \omega_0$ and appearance of narrow line with lasing at $\nu=\omega$. For $\Omega =\omega_0$ the radiation spectrum $S(\nu)$ at $n_p<1$ has two broad peaks at $\nu=\omega \pm g \omega_0$ corresponding to the vacuum Rabi splitting [@eberly] (narrow line from driving source at $\nu=\omega$ is also visible in this case). At strong driving $n_p > 1$ the synchronization takes place with appearance of one lasing line at $\nu = \omega$. For both values of $\Omega$ the transition to synchronization/lasing takes place at $n_p > n_{pl} \approx 2$. The spectrum $S(\nu)$ in Fig. \[fig6\] has close similarities with the spectrum observed recently in a single artificial-atom lasing [@astafiev] which appears at a similar threshold $n_{pl} \approx 1$. A shift related to splitting $\omega_\pm \approx g \omega_0/2\sqrt{n}$ is also seen experimentally. However, the vacuum Rabi splitting is not visible in Fig.3c of [@astafiev]. Exact comparison requires much more extended numerical simulations since in [@astafiev] $Q \approx 10^4$ while we have $Q \sim 100$.
=4.2cm =4.2cm
-0.0cm
In conclusion our numerical studies show a nontrivial behavior of a rather simple model given by Eqs. (\[eq1\],\[eq2\]). It is characterized by bistability and synchronization of qubit induced by its coupling to a quantum driven dissipative oscillator. As for the vacuum Rabi splitting [@eberly] it is important that the oscillator is quantum since the effect is absent for a classical dissipative oscillator with commuting $a, a^{\dag}$ (\[eq1\]). A better analytical description and understanding of the behavior discussed requires further studies. Especially interesting is the analysis of long life times $\tau_\pm$ in metastable states related to a macroscopic quantum tunneling.
Preparing the paper to submission we became aware of the preprint [@girvin] where a similar model is studied but synchronization is not discussed there.
This work is supported by the EC project EuroSQIP and RAS joint scientific program “Fundamental problems of nonlinear dynamics” (for OVZ).
[99]{}
P.M. Morse and H. Feshbach, [*Methods of Theoretical Physics*]{}, McGraw-Hill Sci. (1953). L.D. Landau and E.M. Lifshits, [*Quantum mechanics*]{}, Nauka, Moskva (1989). A. Perelomov, [*Generalized Coherent States and Their Applications*]{}, Springer, Berlin (1986). U. Weiss, [*Dissipative quantum mechanics*]{}. World Sci., Singapore (1999). E.T. Jaynes and F.W. Cummings, Proc. IEEE [**51**]{}, 89 (1963). J.J. Sanchez-Mondragon, N.B. Narozhny, and J.H. Eberly, Phys. Rev. Lett. [**51**]{}, 550 (1983). M.O. Scully and M.S. Zubairy, [*Quantum optics*]{}, Cambridge Univ. Press, Cambridge (1997). P.I. Belobrov, G.M. Zaslavsky, and G.K. Tartakovsky, Zh. Eksp. Teor. Fiz. [**71**]{}, 1799 (1976) \[Sov. Phys. JETP [**44**]{}, 945 (1976)\]. J.R. Ackerhalt, P.W. Milonni, and M.-L. Shin, Phys. Rep. [**128**]{}, 205 (1985). R. Graham and M. Höhnerbach, Phys. Rev. Lett. [**57**]{}, 1378 (1986). R. Graham and M. Höhnerbach, Z. Phys. B [**57**]{}, 233 (1984); M. Kuś, Phys. Rev. Lett. [**54**]{}, 1343 (1985). D. Vion [*et al.*]{}, Science [**296**]{}, 886 (2002). A. Wallraff [*et al.*]{}, Nature [**431**]{}, 162 (2004). A.A. Houck [*et al.*]{}, Nature [**449**]{}, 328 (2007). O. Astafiev [*et al.*]{}, Nature [**449**]{}, 588 (2007). E. Il’ichev [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 097906 (2003). J. Claudon [*et al.*]{}, Phys. Rev. Lett. [**93**]{}, 187003 (2004). A.N. Cleland and M.R. Geller, Phys. Rev. Lett. [**93**]{}, 070501 (2004). M. Poggio [*et al.*]{}, Phys. Rev. Lett. [**99**]{}, 017201 (2007). G.P. Berman [*al.*]{}, IEEE Trans. Nanotech. [**4**]{}, 14 (2005). D.A. Rodrigues, J. Imbers, and A.D. Armour, Phys. Rev. Lett. [**98**]{}, 067204 (2007). J. Hauss [*et al.*]{}, preprint arXiv:cond-mat/0701041v5 (2007). A.N. Korotkov, Phys. Rev. B [**60**]{}, 5737 (1999); [**ibid.**]{} [**67**]{}, 235408 (2003); A.N. Korotkov and D.V. Averin, [**ibid.**]{} [**64**]{}, 165310 (2001). N. Gisin and I.Percival, J.Phys. A [**25**]{}, 5677 (1992); [**ibid.**]{} [**26**]{}, 2233 (1993); [**ibid.**]{} [**26**]{}, 2245 (1993). T.A. Brun, I.C. Percival, and R. Schack, J. Phys. A [**29**]{}, 2077 (1996); T.A. Brun, Am. J. Phys. [**70**]{}, 719 (2002). G. Carlo, G. Benenti, and D.L. Shepelyansky, Phys. Rev. Lett. [**95**]{}, 164101 (2005). O.V. Zhirov and D.L. Shepelyansky, Eur. Phys. J. D [**38**]{}, 375 (2006). A. Pikovsky, M. Rosenblum, and J. Kurths, [*Synchronization: A Universal Concept in Nonlinear Sciences*]{}, Cambridge Univ. Press, Cambridge (2001). J. Gambetta [*et al.*]{}, preprint arXiv:0709.4264 (2007).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The set-theoretic axiom WISC states that for every set there is a *set* of surjections to it cofinal in *all* such surjections. By constructing an unbounded topos over the category of sets and using an extension of the internal logic of a topos due to Shulman, we show that WISC is independent of the rest of the axioms of the set theory given by a well-pointed topos. This also gives an example of a topos that is not a predicative topos as defined by van den Berg.'
author:
- '[^1]\'
title: The weak choice principle WISC may fail in the category of sets
---
Introduction
============
Well-known from algebra is the concept of a *projective object*: in a finitely complete category this is an object $P$ such that any epimorphism with codomain $P$ splits. The axiom of choice (AC) can be stated as saying that every set is projective in the category of sets. Various constructive set theories seek to weaken this, and in particular the axiom known as PAx (Presentation Axiom) [@Aczel_78] or CoSHEP (Category of Sets Has Enough Projectives) asks merely that every set $X$ has an epimorphism $P{\twoheadrightarrow}X$ where $P$ is a projective set. Many results that seem to rely on the axiom of choice, such as the existence of enough projectives in module categories, may be proved instead with PAx. As a link with a more well-known axiom, PAx imples the axiom of dependent choice.
There is, however, an even weaker option, here called WISC (to be explained momentarily). Consider the full subcategory $Surj/X {\hookrightarrow}{\bm{\mathrm{set}}}/X$ of surjections with codomain $X$, in some category ${\bm{\mathrm{set}}}$ of sets; clearly it is a large category. Then PAx implies the statement that $Surj/X$ has a *weakly initial object*, namely an object with a map to any other object, not necessarily unique (the axiom of choice says ${\mathrm{id}}_X\colon X\to X$ is weakly initial in $Surj/X$). Another way to think of the presentation axiom is that for every set $X$ there is a ‘cover’ $P {\twoheadrightarrow}X$ such that any surjection $Y{\twoheadrightarrow}P$ splits.
The axiom WISC (Weakly Initial Set of Covers), due to Toby Bartels and Mike Shulman, asks merely that the category $Surj/X$ has a weakly initial *set*, for every $X$. This is a set $I_X$ of objects (that is, of surjections to $X$) such that for any other object (surjection), there is a map from *some* object in $I_X$. To continue the geometric analogy, this is like asking that there is a set of covers of any $X$ such that each surjection $Y{\twoheadrightarrow}X$ splits locally over at least one cover in that set. An example implication of WISC is that the cohomology $H^1(X,G)$ defined by Blass in [@Blass_83] is indeed a set. The assertion that $H^1(X,G)$ is a proper class seems to be strictly weaker than $\neg$WISC, but to the author’s knowledge no models have yet been produced where this is the case.
The origin of the axiom WISC (see [@Roberts_12]) was somewhat geometric in flavour but the question naturally arises whether toposes, and in particular the category of sets, can fail to satisfy WISC. A priori, there is no particular reason why WISC should hold, so the burden is to supply an example where it fails. It goes without saying that neither AC nor PAx can hold in such an example.
The first result in this direction was from van den Berg (see [@vdBerg-Moerdijk_14][^2]) who proved that WISC implies the existence of a proper class of regular cardinals, and so WISC must fail in Gitik’s model of ZF [@Gitik_80]. This model is constructed assuming the existence of a proper class of certain large cardinals, and it has no regular cardinals bigger than $\aleph_0$. Working in parallel to the early development of the current paper, Karagila [@Karagila_14] gave a model of ZF in which there is a proper class of incomparable sets (sets with no injective resp. surjective functions between them) surjecting onto the ordinal $\omega$. This gave a large-cardinal-free proof that WISC was independent of the ZF axioms, answering a question raised by van den Berg.
The current paper started as an attempt to also give, via category-theoretic methods, a large-cardinals-free proof of the independence of WISC from ZF. Since the release of [@Karagila_14], this point is moot as far as independence from ZF goes. However, the proof in [@Karagila_14] relies on a symmetric submodel of a class-forcing model, which is rather heavy machinery. Thus this paper, while proving a slightly weaker result, does so with, in the opinion of the author, far less.
The approach we take is to consider the negation of WISC in the *internal logic* of a (boolean) topos. This allows us to interpret the theory of a well-pointed topos together with $\neg$WISC. However, since this internal version of WISC holds in any Grothendieck topos (assuming for example AC in the base topos of sets) [@vdBerg-Moerdijk_14], we necessarily consider a *non-bounded* topos over the base topos of sets (recall that boundedness of a topos is equivalent to it being a Grothendieck topos). In fact the topos we consider is a variant on the ‘faux topos’ mentioned in [@SGA4 IV 2.8] (wherein ‘topos’ meant what we now call a Grothendieck topos).
The reader familiar with such things may have already noticed that WISC or its negation is not the sort of sentence that can be written via the usual Kripke-Joyal semantics (see e.g. [@MacLane-Moerdijk §VI.6]) used for internal logic, as it contains unbounded quantifiers. As a result, we will be using an extension called the *stack semantics*, given by Shulman [@Shulman_10], that permits their use. The majority of the proof is independent of the details of the stack semantics, which are only used to translate WISC from a statement in a well-pointed topos to a general topos (in fact a locally connected topos, as this is the only case we will consider).
To summarise: starting from a well-pointed topos with natural number object we give a proper-class-sized group ${\mathcal{Z}}$ equipped with a certain topology, and consider the topos ${\mathcal{Z}}{\bm{\mathrm{set}}}$ of sets with a continuous action of this group. Of course, the preceeding sentence needs to be formalised appropriately, and we do this in terms of a base well-pointed topos and a large diagram of groups therein. We reduce the failure of WISC in the internal logic of ${\mathcal{Z}}{\bm{\mathrm{set}}}$ to simple group-theoretic statements. It should be pointed out that classical logic is used throughout, and all the toposes in this note are boolean.
Finally, the topos constructed as in the previous paragraph is not a *predicative topos* as defined in [@vandenBerg_12]. These are analogues of toposes that should capture predicative mathematics, as toposes capture the notion of intuitionistic mathematics. This apparent failure is understood and carefully discussed in *loc. cit.*; the example given in this paper is hopefully of use as a foil in the development of predicative toposes.
The author’s thanks go to Mike Shulman for helpful and patient discussions regarding the stack semantics. Thanks are also due to an anonymous referee who found an earlier version of this paper contained some critical errors.
WISC in the internal language
=============================
We use the following formulation of WISC, equivalent to the usual statement in a well-pointed topos and due to François Dorais [@Dorais_MO].
For every set $X$ there is a set $Y$ such that for every surjection $q\colon Z\to X$ there is a map $s\colon Y \to Z$ such that $q\circ s \colon Y\to X$ is a surjection.
The aim of this paper is to show that an internal version of $\neg$WISC is valid in the (non-well-pointed) topos constructed in section \[sec:construction\] below. The internal logic of a topos, in the generality required here, is given by the *stack semantics*. We refer to [@Shulman_10 section 7] for more details on the stack semantics, recalling purely what is necessary for the translation of WISC into the internal logic of a topos $S$ (Shulman takes weaker assumptions on $S$, but this extra generality is not needed here).
If $U$ is an object of $S$ we say that a formula of category theory $\phi$ with parameters in the category $S/U$ is a *formula over $U$*. We have[^3] the base change functor $p^*\colon S/U\to S/V$ for any map $p\colon V\to U$, and call the formula over $V$ given by replacing each parameter of $\phi$ by its image under $p^*$ the *pullback* of $\phi$ (denoted $p^*\phi$). Note that the language of category theory is taken to be two-sorted, so there are quantifiers for both objects and arrows separately. Here and later ${\twoheadrightarrow}$ denotes a map that is an epimorphism.
[(Shulman [@Shulman_10])]{}\[def:stack\_semantics\] Given the topos $S$, and a sentence $\phi$ over $U$, we define the relation $U\Vdash \phi$ recursively as follows
- $U\Vdash (f=g) \leftrightarrow f = g$
- $U\Vdash \top$ always
- $U\Vdash \bot \leftrightarrow U \simeq 0$
- $U\Vdash (\phi \wedge \psi) \leftrightarrow U\Vdash \phi$ and $U\Vdash \psi$
- $U\Vdash (\phi \vee \psi) \leftrightarrow U = V\cup W$, where $i\colon V{\hookrightarrow}U$ and $j\colon W{\hookrightarrow}U$ are subobjects such that $V\Vdash i^*\phi$ and $W\Vdash j^*\psi$
- $U\Vdash (\phi\Rightarrow \psi) \leftrightarrow$ for any $p\colon V\to U$ such that $V\Vdash p^*\phi$, also $V\Vdash p^*\psi$
- $U\Vdash \neg \phi \leftrightarrow U\Vdash (\phi\Rightarrow \bot)$
- $U\Vdash (\exists X)\phi(X) \leftrightarrow
\exists p\colon V\twoheadrightarrow U$ and $A\in \operatorname{Obj}(S/V)$ such that $V\Vdash p^*\phi(A)$
- $U\Vdash (\exists f\colon A\to B)\phi(f)\leftrightarrow
\exists p\colon V\twoheadrightarrow U$ and $g\colon p^*A \to p^*B \in \operatorname{Mor}(S/V)$ such that $V\Vdash p^*\phi(g)$
- $U\Vdash (\forall X)\phi(X) \leftrightarrow$ for any $p\colon V\to U$ and $A\in \operatorname{Obj}(S/V)$, $V\Vdash p^*\phi(A)$
- $U\Vdash (\forall f\colon A\to B)\phi(f)\leftrightarrow$ for any $p\colon V\to U$ and $j\colon p^*A \to p^*B \in \operatorname{Mor}(S/V)$, $V\Vdash p^*\phi(j)$
If $\phi$ is a formula over $1$ we say $\phi$ is *valid* if $1\Vdash \phi$.
Comparing with [@MacLane-Moerdijk §VI.6] one can recognise the Kripke-Joyal semantics as a fragment of the above, where attention is restricted to monomorphisms rather than arbitrary objects in slice categories, and all quantifiers are bounded.
Since our intended model will be built using not just an arbitrary topos, but a locally connected and cocomplete one, the following lemma will simplify working in the internal logic. The proof follows that of lemma 7.3 in [@Shulman_10]. We recall that a locally connected topos $E$ is a topos over ${\bm{\mathrm{set}}}$ with an additional left adjoint $\pi_0$ to the inverse image part of the global section functor, and an object $A$ is called *connected* if $\pi_0(A)=1$.
\[loc\_conn\_topos\_forall\] Let $E$ be a locally connected cocomplete topos. Then then if for any *connected* object $V$, arrow $p\colon V\to U$ and $A\in \operatorname{Obj}(S/V)$ we have $V\Vdash p^*\phi(A)$, then $U\Vdash (\forall X)\phi(X)$.
Here ‘locally connected cocomplete’ is relative to a base topos ${\bm{\mathrm{set}}}$ that is well-pointed (hence boolean) topos with natural number object (nno). We will refer to the objects of ${\bm{\mathrm{set}}}$ as ‘sets’, but without an implication that these arise from a particular collection of axioms. We will assume throughout that all toposes will come with an nno.
For a locally connected and cocomplete topos the statement of WISC translates, using definition \[def:stack\_semantics\] and applying lemma \[loc\_conn\_topos\_forall\], into the stack semantics as follows: $$\label{int_WISC}
\begin{aligned}
&\forall\ X\to U,\ U \text{ connected,} \\
&\exists\ V\stackrel{p}{{\twoheadrightarrow}} U,\ Y\to V,\\
&\forall\ W\stackrel{q}{\to} V,\ W \text{ connected,}\
Z\stackrel{g}{{\twoheadrightarrow}}W\times_U X, \\
&\exists\ T \stackrel{r}{{\twoheadrightarrow}} W,\ T\times_V Y \xrightarrow{({\mathrm{pr}}_1,l)} T\times_W Z, \\
&\text{the map}\ T\times_V Y \xrightarrow{({\mathrm{pr}}_1,l)} T\times_W Z \xrightarrow{r^*(g)} T\times_U X \text{ is an epi}.
\end{aligned}$$
Note also that “is an epi” is a proposition whose statement in the stack semantics is equivalent to the external statement (see discussion around example 7.10 of [@Shulman_10]). One does not need any knowledge of the stack semantics for the rest of this paper, and the uninitiated may choose to take (\[int\_WISC\]) as the *definition* of WISC in the internal language of a locally connected cocomplete topos, and ignore the stack semantics entirely.
We will give a boolean ${\bm{\mathrm{set}}}$-topos $E$ that is locally connected and cocomplete and in which the following statement, the negation of (\[int\_WISC\]), holds: $$\label{int_neg_WISC}
\begin{aligned}
&\exists\ X\to U,\ U \text{ connected,} \\
&\forall\ V\stackrel{p}{{\twoheadrightarrow}} U,\ Y\to V,\\
&\exists\ W\stackrel{q}{\to} V,\ W \text{ connected,}\
Z\stackrel{g}{{\twoheadrightarrow}}W\times_U X,\\
&\forall\ T \stackrel{r}{{\twoheadrightarrow}} W,\ T\times_V Y \xrightarrow{({\mathrm{pr}}_1,l)} T\times_W Z, \\
&\text{the map}\ T\times_V Y \xrightarrow{({\mathrm{pr}}_1,l)} T\times_W Z \xrightarrow{r^*(g)} T\times_U X \text{ is not epi}.
\end{aligned}$$
We denote the natural number object of $E$ by ${\mathbb{N}}_d$, which is given by the image of the nno ${\mathbb{N}}$ of ${\bm{\mathrm{set}}}$ under the inverse image part of the geometric morphism $E\to {\bm{\mathrm{set}}}$.
\[prop:implies\_neg\_WISC\] In a connected, locally connected cocomplete topos $E$ such that $\pi_0$ reflects epimorphisms, the statement $$\label{neg_WISC_simple}
\begin{aligned}
&\forall\ Y{\twoheadrightarrow}V,\ V \text{ connected,} \\
&\exists\ \Omega {\twoheadrightarrow}{\mathbb{N}}_d \text{ with } \pi_0(\Omega) \simeq \pi_0({\mathbb{N}}_d),\\
&\forall\ T {\twoheadrightarrow}V,\ T \text{ connected, } T\times_V Y \xrightarrow{l} \Omega, \\
& l \text{ is not epi}.
\end{aligned}$$ implies (\[int\_neg\_WISC\]), the negation of WISC in the internal language of $E$.
We give some facts about toposes that we will use in what follows. First, in a connected topos the terminal object is connected. Second, in a cocomplete topos one has infinitary extensivity, namely $A\times_B \coprod_{i\in I} C_i \simeq \coprod_{i\in I} A\times_B C_i$, and the initial object $0$ is *strict*: any map to it is an isomorphism. Third, since $\pi_0$ is a left adjoint, it preserves epimorphisms. Combined with the hypothesis on $\pi_0$ this means a map $f$ in $E$ is an epimorphism if and only if $\pi_0(f)$ is an epimorphism. Similarly $\pi_0$ preserves initial objects and the hypotheses imply it also reflects initial objects.
Now assume that (\[neg\_WISC\_simple\]) holds in $E$. In (\[int\_neg\_WISC\]) take $X\to U$ to be ${\mathbb{N}}_d \to 1$ (using $1$ is connected). Given an epimorphism $V{\twoheadrightarrow}1$, $V$ has a component as $\pi_0(V)\to 1$ is onto and $V = \coprod_{v\in\pi_0(V)} V_v$ (and $1$ is projective). Fix a component $V_0 {\hookrightarrow}V$.
Given any $Y\to V$, take $Y_0 = V_0 \times_V Y$ to get $Y_0 \to V_0$. If $Y_0$ is initial, then (\[int\_neg\_WISC\]) can be seen to hold by taking $W = V_0$ and $g={\mathrm{id}}$ since $T\times_V Y = T\times_{V_0} Y_0 = 0$ and as $r$ is an epi and $W$ is connected, $T\times {\mathbb{N}}_d$ is not initial.
Hence we can assume $Y_0$ is not initial, and hence has at least one component and so $Y_0 \to V_0$ is an epi. Fix some $\Omega {\twoheadrightarrow}{\mathbb{N}}_d$ inducing an isomorphism $\pi_0(\Omega) \simeq \pi_0({\mathbb{N}}_d)$ such that the rest of (\[neg\_WISC\_simple\]) holds. In (\[int\_neg\_WISC\]) take $q$ to be the inclusion $V_0 {\hookrightarrow}V$ (hence $W=V_0$, which is connected), and $Z = V_0 \times \Omega$ with the epimorphism $g$ the product of ${\mathrm{id}}_{V_0}$ and $\Omega {\twoheadrightarrow}{\mathbb{N}}_d$.
Now take any $T$ and pair of maps $T{\twoheadrightarrow}V_0$ and $T\times_V Y = T\times_{V_0} Y_0 \xrightarrow{({\mathrm{pr}}_1,l)} T\times_{V_0} Z = T \times \Omega$. We know that $T$ has a component by a similar argument to above, say $T_0 {\hookrightarrow}T$. Then $T_0 \to V_0$ is epi so (\[neg\_WISC\_simple\]) implies $T_0\times_{V_0} Y_0 = T_0 \times_V Y \to \Omega$ is not epi. This then implies $T_0 \times_V Y \to \Omega \to {\mathbb{N}}_d$ is not epi, since if it were, $\pi_0(T_0\times_V Y) \to \pi_0(\Omega) \xrightarrow{\sim} \pi_0({\mathbb{N}}_d)$ would be epi, implying $\pi_0(T_0\times_V Y) \to \pi_0(\Omega)$ and hence $T_0 \times_V Y \to \Omega$ was epi. Thus there is some component of ${\mathbb{N}}_d$ not in the image of this map, say indexed by $n\in {\mathbb{N}}$.
Then $T_0\times_V Y\to T_0\times {\mathbb{N}}_d$ is not epi, as the component of $T_0\times {\mathbb{N}}_d$ indexed by $n$ (isomorphic to $T_0$, which has $T_0 \to 1$ epi) is not in its image. It then follows that $T \times_V Y \to T\times {\mathbb{N}}_d$ is not epi, and so (\[int\_neg\_WISC\]) holds.
The construction {#sec:construction}
================
Given our base topos ${\bm{\mathrm{set}}}$, we can consider the category of objects in ${\bm{\mathrm{set}}}$ equipped with a linear order with no infinite descending chains, which we shall call ordinals, in analogy with material set theory. The usual Burali-Forti argument—which requires no Choice—tells us there is a large category $O$ with objects ordinals and arrows the order-preserving injections onto initial segments. This large category is a linear preorder and has no infinite strictly descending chains. That there are multiple representatives for a particular order type, that is, non-identical isomorphic ordinals, does not cause any problems. We also note that $O$ has small joins (defined up to isomorphism in $O$).
Given a topological group $G$, the category of sets with a continuous $G$ action forms a cocomplete boolean topos $G{\bm{\mathrm{set}}}$. In practice, one specifies a filter ${\mathcal{F}}$ of subgroups of $G$ and then those $G$-sets all of whose stabiliser groups belong to ${\mathcal{F}}$ are precisely those with a continuous action for the topology generated by ${\mathcal{F}}$.
For any group $G$, let ${\mathcal{C}}$ be a collection of finite-index subgroups closed under finite intersections. Then there is a filter ${\mathcal{F}}_{\mathcal{C}}$ with elements those subgroups $H \leq G$ containing a subgroup appearing in ${\mathcal{C}}$ (we say the filter is *generated* by ${\mathcal{C}}$). The category of continuous $G$-sets is then a full subcategory of the category of $G$-sets with finite orbits. The internal hom $Y^X$ is given by taking the set ${\bm{\mathrm{set}}}(X,Y)$ then retaining only those functions whose stabiliser under the $G$-action $f\mapsto g\cdot \left(f(g^{-1}\cdot -)\right)$ belongs to ${\mathcal{F}}_{\mathcal{C}}$. The subobject classifier is the two-element set with trivial $G$-action.
\[rem:nice\_cover\] Notice that every transitive $G$-set $X$ that is continuous with respect to the topology given by ${\mathcal{F}}_{\mathcal{C}}$ (all $G$-sets will be assumed continuous from now on) has an epimorphism from some $G/L$ where $L\in {\mathcal{C}}$. This is because any stabiliser $\operatorname{Stab}(x) \in {\mathcal{F}}_{\mathcal{C}}$, $x\in X$, is assumed to contain an element of ${\mathcal{C}}$.
\[eg:bounded\_depth\_subgroups\] For $\alpha$ an ordinal, let ${\mathbb{Z}}^\alpha$ be the set of functions $\alpha \to {\mathbb{Z}}$, considered as a group by pointwise addition. Consider functions $d\colon \alpha \to {\mathbb{N_+}}= \{1,2,3,\ldots\}$ such that $d(i) \not= 1$ for only finitely many $i\in \alpha$, which we shall call *local depth functions*. Such a function defines a subgroup $d{\mathbb{Z}}:= \prod_{i\in\alpha}d(i){\mathbb{Z}}\leq {\mathbb{Z}}^\alpha$ of finite index. The intersection of two such subgroups, given by $d_1$ and $d_2$, is given by the function $i\mapsto \operatorname{lcm}\{d_1(i)d_2(i)\}$. The subgroups belonging to the filter generated by this collection will be called *bounded depth subgroups*. From now on ${\mathbb{Z}}^\alpha$ will be regarded as having the topology generated by this filter.
If we are given a split open surjection $p\colon H \to G$ (with $p$ and its splitting continuous) there is a geometric morphism $(p^*\dashv p_*)\colon H{\bm{\mathrm{set}}}\to G{\bm{\mathrm{set}}}$ with $p^*$ fully faithful and possessing a left adjoint $p_! \dashv p^*$. Here $p^*$ sends a $G$-set to the same set with the $H$-action via $p$ and $p_!(X) = X/\ker(p)$ with the obvious $G$-action. The inverse image functor $p^*$ is in this case also a *logical* functor, meaning that it preserves the subobject classifier and internal hom, as well as finite limits. In the case that $G$ is the trivial group: $p^*$ is denoted $(-)_d$ and sends a set to the same set with the trivial action; $p_!$ is denoted $\pi_0$ and $\pi_0(X)$ is the set of orbits of the $H$-action.
For $\alpha {\hookrightarrow}\beta$ ordinals, there is a split open surjection ${\mathbb{Z}}^\beta \to {\mathbb{Z}}^\alpha$, projection being given by restriction of the domain, and the splitting given by extending a function by $0$. Note that a local depth function on $\alpha$ gives a local depth function on $\beta$ by extending it by $1$.
Now consider a functor ${\mathcal{G}}\colon O^{op}\to {\bm{\mathrm{Top}}}{\bm{\mathrm{Grp}}}_{sos}$, where ${\bm{\mathrm{Top}}}{\bm{\mathrm{Grp}}}_{sos}$ is the category of topological groups and split open surjections. Define the category ${\mathcal{G}}{\bm{\mathrm{set}}}$ with objects pairs $(\alpha,X)$ where $\alpha$ is an ordinal and $X$ is an object of ${\mathcal{G}}(\alpha){\bm{\mathrm{set}}}$, and arrows ${\mathcal{G}}{\bm{\mathrm{set}}}((\alpha,X),(\beta,Y)) = {\mathcal{G}}(\gamma)(X_\gamma,Y_\gamma)$ where $\gamma = \max\{\alpha,\beta\}$ and $X_\gamma,\ Y_\gamma$ are $X,Y$ considered as ${\mathcal{G}}(\gamma)$-sets via the inverse image functors as above. The hom-sets are defined without making any choices since $O$ is a linear preorder, and so $\gamma$ is either $\alpha$ or $\beta$ (and we can take $\gamma = \alpha$ if $\alpha\simeq\beta$). Composition is well defined due to the full faithfulness of the inverse image functors. The objects of ${\mathcal{G}}{\bm{\mathrm{set}}}$ will be referred to as ${\mathcal{G}}$-sets. Informally, this category is the colimit of the large diagram of inverse image functors.
The category ${\mathcal{G}}{\bm{\mathrm{set}}}$ is a connected, locally connected, atomic and cocomplete boolean ${\bm{\mathrm{set}}}$-topos. Moreover, $\pi_0$ reflects epimorphisms.
Let us first show that we have a topos. Finite limits exist because they can be calculated in any ${\mathcal{G}}(\alpha)$ where $\alpha$ is greater than all ordinals appearing in the objects in the diagram, and when the universal property is checked in ${\mathcal{G}}(\beta)$ for $\beta > \alpha$, the limit is preserved by the inverse image functor. Likewise the internal hom $(\alpha,X)^{(\beta,Y)}$ is defined as $X_\gamma^{Y_\gamma}$ in ${\mathcal{G}}(\gamma)$ ($\gamma = \max\{\alpha,\beta\}$) and its universal property is satisfied due to inverse image functors preserving internal homs. The subobject classifier $\mathbf{2}$ in ${\bm{\mathrm{set}}}$ is preserved by all inverse image functors ${\bm{\mathrm{set}}}\to {\mathcal{G}}(\alpha){\bm{\mathrm{set}}}$, so given any subobject in ${\mathcal{G}}{\bm{\mathrm{set}}}$ it has a classifying map to $\mathbf{2}$. Thus ${\mathcal{G}}{\bm{\mathrm{set}}}$ is a topos, and has a geometric morphism $((-)_d \dashv (-)^{\mathcal{G}})\colon {\mathcal{G}}{\bm{\mathrm{set}}}\to {\bm{\mathrm{set}}}$ as it is locally small ($(-)^{\mathcal{G}}:= {\mathcal{G}}{\bm{\mathrm{set}}}(1,-)$ is the global points functor). It is easy to check there is a functor $\pi_0$ sending a ${\mathcal{G}}(\alpha)$-set to its set of orbits and this is a left adjoint to $(-)_d$. Thus ${\mathcal{G}}{\bm{\mathrm{set}}}$ is locally connected. Since $(-)_d$ is fully faithful and logical ${\mathcal{G}}{\bm{\mathrm{set}}}$ is also connected and atomic respectively. Small colimits can be calculated in ${\mathcal{G}}(\alpha)$ where $\alpha$ is some small join of the ordinals appearing as the vertices of the diagram, and the universal property is verified since inverse image functors preserve all small colimits. Lastly, ${\mathcal{G}}{\bm{\mathrm{set}}}$ is boolean as $1 \to \mathbf{2} \leftarrow 1$ is a coproduct cocone, using the definition of colimits and the fact it is such in ${\bm{\mathrm{set}}}$.
To prove the last statement, suppose $X \to Y$ in ${\mathcal{G}}{\bm{\mathrm{set}}}$ (without loss of generality, take this in ${\mathcal{G}}(\alpha){\bm{\mathrm{set}}}$ for some $\alpha$) is such that $\pi_0$ induces an epimorphism of connected components. Then for each orbit of $Y$ there is an orbit of $X$ mapping to it, and equivariant maps between orbits are onto, so $X\to Y$ is onto as a map of sets and hence an epi.
The stack semantics in ${\mathcal{G}}{\bm{\mathrm{set}}}$ give a model of the structural set theory underlying ${\bm{\mathrm{set}}}$, minus any Choice that may hold in ${\bm{\mathrm{set}}}$ (see the discussion after lemma 7.13 in [@Shulman_10]). We will take a particular diagram of groups with the properties we need.
The diagram ${\mathcal{Z}}\colon \alpha \mapsto {\mathbb{Z}}^\alpha$, where ${\mathbb{Z}}^\alpha$ is regarding as having the topology given by the filter of bounded depth subgroups, gives rise to a connected, locally connected boolean topos ${\mathcal{Z}}{\bm{\mathrm{set}}}$ such that $\pi_0$ reflects epimorphisms.
If one is working in a setting that permits such reasoning, the proper class-sized group to which the introduction alludes is the colimit over the inclusions ${\mathcal{Z}}(\alpha) {\hookrightarrow}{\mathcal{Z}}(\beta)$ given by the splittings, for $\alpha {\hookrightarrow}\beta$. The rest of the paper will show that internal WISC fails in ${\mathcal{Z}}{\bm{\mathrm{set}}}$, and so WISC itself fails in the well-pointed topos given by the stack semantics of ${\mathcal{Z}}{\bm{\mathrm{set}}}$.
The failure of WISC {#sec:failure_of_WISC}
===================
We need some facts that hold in ${\mathcal{Z}}{\bm{\mathrm{set}}}$ regarding local depth functions. As a bit of notation, let us write ${\mathcal{Z}}/d{\mathbb{Z}}$ for the transitive ${\mathcal{Z}}$-set ${\mathbb{Z}}^\alpha/d{\mathbb{Z}}$ for $\alpha = \operatorname{dom}(d)$.
\[local\_depth\_bounds\] Let ${\mathcal{Z}}/d_1{\mathbb{Z}}\to {\mathcal{Z}}/d_2{\mathbb{Z}}$ be an equivariant map of ${\mathcal{Z}}$-sets. Then for every $i \in \alpha$ we have $d_2(i) \mid d_1(i)$.
The existence of the map implies $d_1{\mathbb{Z}}$ is conjugate to a subgroup of $d_2{\mathbb{Z}}$, but all groups here are abelian so it *is* a subgroup of $d_2{\mathbb{Z}}$. For the second statement, notice that the first statement implies $d_1(i){\mathbb{Z}}\leq d_2(i){\mathbb{Z}}\leq {\mathbb{Z}}$ for each $i\in \alpha$ and the result follows.
We also need to consider what taking pullbacks looks like from the point of view of local depth functions.
\[depth\_of\_fibred\_product\] Any orbit in $${\mathcal{Z}}/(d_1{\mathbb{Z}}\cap d_2{\mathbb{Z}}) \subset {\mathcal{Z}}/d_1{\mathbb{Z}}\times_{{\mathcal{Z}}/d_3{\mathbb{Z}}} {\mathcal{Z}}/d_2{\mathbb{Z}}$$ is isomorphic to a transitive ${\mathcal{Z}}$-set with local depth function $d$ given by $$d(i) = \operatorname{lcm}\{d_1(i),d_2(i)\},\quad \forall i \in \alpha$$ where $\alpha = \max\{\operatorname{dom}(d_1),\operatorname{dom}(d_2)\}$.
Notice that the fibred product as given is isomorphic to $$\prod_{i\in \alpha} {\mathbb{Z}}/d_1(i){\mathbb{Z}}\times_{{\mathbb{Z}}/d_3(i){\mathbb{Z}}} {\mathbb{Z}}/d_2(i){\mathbb{Z}}$$ where the ${\mathbb{Z}}^\alpha$ action is such that the $i^{th}$ coordinate—a copy of ${\mathbb{Z}}$—acts diagonally on the $i^{th}$ factor of the preceeding expression. The stabiliser of any $(n_i,n'_i)_{i\in \alpha}$ is then the product of the stabilisers of the ${\mathbb{Z}}$-action of the various ${\mathbb{Z}}/d_1(i){\mathbb{Z}}\times_{{\mathbb{Z}}/d_3(i){\mathbb{Z}}} {\mathbb{Z}}/d_2(i){\mathbb{Z}}$. We thus only need to consider the simpler problem of determining the stabilisers for a ${\mathbb{Z}}$-set ${\mathbb{Z}}/k{\mathbb{Z}}\times_{{\mathbb{Z}}/m{\mathbb{Z}}} {\mathbb{Z}}/l{\mathbb{Z}}$.
The stabiliser of $(0,0)$ is ${\mathbb{Z}}/(k{\mathbb{Z}}\cap l{\mathbb{Z}})$, from which the result follows by the description in example \[eg:bounded\_depth\_subgroups\] of the intersection of subgroups given by local depth functions. We only then need to consider the stabilisers of $(0,n)$ for $n\in {\mathbb{Z}}/l{\mathbb{Z}}$ as all others are equal to one of these by abelianness – but $\operatorname{Stab}(0,n)$ is again ${\mathbb{Z}}/(k{\mathbb{Z}}\cap l{\mathbb{Z}})$ using abelianness. The statement regarding local depth functions then follows.
We need a special collection of subgroups of ${\mathbb{Z}}^\alpha$ in the proof of theorem \[WISC\_fails\_in\_Zset\] below, namely those given by local depth functions $\delta[\alpha,n,i]\colon \alpha\to {\mathbb{N_+}}$ defined as $$\delta[\alpha,n,i](k) =
\begin{cases}
n & \text{if $k = i$;}\\
1 & \text{if $k \not= i$.}
\end{cases}$$ Note that the transitive ${\mathcal{Z}}$-set ${\mathcal{Z}}/\delta[\alpha,n,i]{\mathbb{Z}}$ has underlying set ${\mathbb{Z}}/n{\mathbb{Z}}$, and that $\Omega[\alpha,i] := \coprod_{n\in {\mathbb{N_+}}} {\mathcal{Z}}/\delta[\alpha,n,i]{\mathbb{Z}}$ is an object of ${\mathcal{Z}}{\bm{\mathrm{set}}}$ for any $\alpha\in O$ and $i\in\alpha$.
\[WISC\_fails\_in\_Zset\] The statement of WISC in the stack semantics in ${\mathcal{Z}}{\bm{\mathrm{set}}}$ fails.
In the notation of proposition \[prop:implies\_neg\_WISC\], taking transitive ${\mathcal{Z}}$-sets for connected objects, we need to show that for any $Y{\twoheadrightarrow}{\mathcal{Z}}/H$, there is an $\Omega$ such that for any $r\colon {\mathcal{Z}}/K \to {\mathcal{Z}}/H$, any $l\colon {\mathcal{Z}}/K\times_{{\mathcal{Z}}/H} Y \to \Omega$ is not an epimorphism.
Let us write $Y = \coprod_{y\in\pi_0(Y)} Y_y$, and note that this coproduct, like all colimits in ${\mathcal{Z}}{\bm{\mathrm{set}}}$ takes place in some ${\mathbb{Z}}^\alpha{\bm{\mathrm{set}}}$. In particular, by remark \[rem:nice\_cover\] each $Y_y$ has an epimorphism from some ${\mathcal{Z}}/d_y{\mathbb{Z}}$ for a local depth function $d_y \colon \alpha \to {\mathbb{N_+}}$. As a result $H \leq {\mathbb{Z}}^\alpha$, so fix some $d_H\colon \alpha \to {\mathbb{N_+}}$ to get an epimorphism ${\mathcal{Z}}/d_H{\mathbb{Z}}\to {\mathcal{Z}}/H$. Define $\Omega = \Omega[\alpha+1,\top_{\alpha+1}]$, where $\top_{\alpha+1}$ is the top element of the ordinal $\alpha+1$. Given ${\mathcal{Z}}/K \to {\mathcal{Z}}/H$, fix a local depth function $d_K\colon \beta \to {\mathbb{N_+}}$ such that $d_K{\mathbb{Z}}\leq K$ (without loss of generality, we can assume $\alpha \leq \beta$).
Since ${\mathcal{Z}}{\bm{\mathrm{set}}}$ is infinitary extensive, we have $${\mathcal{Z}}/K\times_{{\mathcal{Z}}/H} Y \simeq \coprod_{y\in\pi_0(Y)} {\mathcal{Z}}/K\times_{{\mathcal{Z}}/H} Y_y.$$ Any map $l\colon {\mathcal{Z}}/K\times_{{\mathcal{Z}}/H} Y \to \Omega$ is then given by a collection of maps $l_y \colon {\mathcal{Z}}/K\times_{{\mathcal{Z}}/H} Y_y \to \Omega$. We need to show that this collection of maps is not jointly surjective, and will do this by showing the image of $l_y$, for arbitrary $y$, must be contained in a strict subobject of $\Omega$ that is independent of $y$.
Given an epimorphism ${\mathcal{Z}}/d_y{\mathbb{Z}}\to Y_y$, consider, in ${\mathcal{Z}}/d_K{\mathbb{Z}}\times_{{\mathcal{Z}}/d_H{\mathbb{Z}}} {\mathcal{Z}}/d_y{\mathbb{Z}}$, an orbit ${\mathcal{Z}}/\delta_y{\mathbb{Z}}$ where $\delta_y(i) = \operatorname{lcm}\{d_K(i),d_y(i)\}$ for each $i\in \beta$, by lemma \[depth\_of\_fibred\_product\]. In particular, we have that $\delta_y(\top_{\alpha+1}) = d_K(\top_{\alpha+1}) =: N_0$ is independent of $y$.
Compose the inclusion ${\mathcal{Z}}/\delta_y{\mathbb{Z}}{\hookrightarrow}{\mathcal{Z}}/K \times_{{\mathcal{Z}}/H} Y_y$with $l_y$ to get a map $$l'_y\colon {\mathcal{Z}}/\delta_y{\mathbb{Z}}\to \Omega = \coprod_{n\in {\mathbb{N_+}}} {\mathcal{Z}}/\delta[\alpha,n,i]{\mathbb{Z}}.$$ Applying lemma \[local\_depth\_bounds\] to this map with $i = \top_{\alpha+1}$ we find that $n\mid N_0$ for any $n$ such that ${\mathcal{Z}}/\delta[\alpha,n,i]{\mathbb{Z}}\subset \operatorname{im}l'_y$. Thus the image of any $l_y$ and hence of $l$ is contained in $$\coprod_{n\mid N_0} {\mathcal{Z}}/\delta[\alpha,n,i]{\mathbb{Z}}\subsetneqq \Omega,$$ hence $l$ is not an epimorphism.
Recall that ETCS is a set theory defined by specifying the properties of the category of sets [@Lawvere_64], namely that it is a well-pointed topos (with nno) satisfying the axiom of choice. We can likewise specify a choiceless version, which is the theory of a well-pointed topos (with nno). Given a model ${\bm{\mathrm{set}}}$ of ETCS, we have constructed a well-pointed topos in which WISC is false. Thus we have our main result.
Assuming ETCS is consistent, so is the theory of a well-pointed topos with nno plus the negation of WISC.
Finally, we recall the definition from [@vandenBerg_12] of a predicative topos: this is a satisfying WISC (or, as called there, AMC).
The topos ${\mathcal{Z}}{\bm{\mathrm{set}}}$ is not a predicative topos.
[10]{}
*[SGA]{}4 – [T]{}héorie des topos et cohomologie étale des schémas. [T]{}ome 1: [T]{}héorie des topos*, Lecture Notes in Mathematics, Vol. 269, Springer-Verlag, Berlin, 1972. Séminaire de Géométrie Algébrique du Bois-Marie 1963–1964 (SGA 4), Dirigé par M. Artin, A. Grothendieck, et J. L. Verdier.
<span style="font-variant:small-caps;">Aczel, Peter</span>, ‘The type theoretic interpretation of constructive set theory’, in *Logic Colloquium ’77*, vol. 96 of *Stud. Logic Foundations Math.*, North-Holland, 1978, pp. 55–66.
<span style="font-variant:small-caps;">Blass, Andreas</span>, ‘Cohomology detects failures of the axiom of choice’, *Trans. Amer. Math. Soc.*, 279 (1983), 1, 257–269.
<span style="font-variant:small-caps;">Dorais(<http://mathoverflow.net/users/2000>), Fran[c c]{}ois G.</span>, ‘[O]{}n a weak choice principle’, MathOverflow, 2012. <http://mathoverflow.net/a/99934/> (version: 2012-06-18).
<span style="font-variant:small-caps;">Gitik, M.</span>, ‘All uncountable cardinals can be singular’, *Israel J. Math.*, 35 (1980), 1-2, 61–88.
<span style="font-variant:small-caps;">Karagila, A.</span>, ‘Embedding orders into cardinals with [$DC_\kappa$]{}’, *Fundamenta Mathematicae*, 226 (2014), 143–156. [arXiv:1212.4396]{}.
<span style="font-variant:small-caps;">Lawvere, F. William</span>, ‘An elementary theory of the category of sets (long version) with commentary’, *Repr. Theory Appl. Categ.*, (2005), 11, 1–35. Reprinted and expanded from Proc. Nat. Acad. Sci. U.S.A. [[**[5]{}**]{}2]{} (1964), With comments by the author and Colin McLarty.
<span style="font-variant:small-caps;">MacLane, S.</span>, and <span style="font-variant:small-caps;">I. Moerdijk</span>, *Sheaves in Geometry and Logic*, Springer-Verlag, 1992.
<span style="font-variant:small-caps;">Roberts, D. M.</span>, ‘Internal categories, anafunctors and localisation’, *Theory Appl. Categ.*, 26 (2012), 29, 788–829. [arXiv:1101.2363]{}.
<span style="font-variant:small-caps;">Shulman, Michael</span>, ‘Stack semantics and the comparison of material and structural set theories’, , 2010. [arXiv:1004.3802]{}.
<span style="font-variant:small-caps;">van den Berg, Benno</span>, ‘Predicative toposes’, , 2012. [arXiv:1207.0959]{}.
<span style="font-variant:small-caps;">van den Berg, Benno</span>, and <span style="font-variant:small-caps;">Ieke Moerdijk</span>, ‘The axiom of multiple choice and models for constructive set theory’, *Journal of Mathematical Logic*, 14 (2014), 1. [arXiv:1204.4045]{}.
[^1]: Supported by the Australian Research Council (grant number DP120100106). This paper will appear in the journal *Studia Logica*.
[^2]: In that paper, WISC is used in a guise of an equivalent axiom called AMC, the Axiom of Multiple Choice. To avoid confusion with other axioms with that name, this paper sticks with the term ‘’WISC’.
[^3]: Technically, this is only after choosing a splitting of the fibred category $S^\mathbf{2} \to S$, but in practice one only deals with a finite number of instances so this can be glossed over.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- |
Ties Behnke$^a$, Ralf Diener$^a$, Christoph Rosemann$^a$, Lea Steder$^a$[^1]\
DESY,\
Notkestrasse 85, 22607 Hamburg, Germany\
E-mail:
title: 'A Novel Self-supporting GEM-based Amplification Structure for a Time Projection Chamber at the ILC'
---
Introduction
============
Time Projection Chambers (TPCs) are discussed as tracking detectors for a number of projects in nuclear and particle physics. They are proposed because TPCs combine good spatial resolution with excellent, robust and efficient three-dimensional pattern recognition and the possibility to identify particles via their specific energy loss, $dE/dx$. Modern TPC systems are increasingly designed based on micro-pattern gas detectors (MPGD) for readout. These systems offer a number of potential advantages compared to the traditional wire-chamber based readout. MPGDs have amplification structures which are of a size ($\mathcal{O}$(100) ${\upmu}$m) similar to the resolution anticipated. Effects connected to the length scale in the amplification —as, for example, distortions introduced by drifting electrons transverse to the magnetic field— scale as the feature size and are therefore much reduced in MPGD based systems. Another advantage of MPGDs is that they suppress the backdrift of ions into the drift volume. This feature allows the construction of TPCs which can operate continuously. This is however —while essential for the application of MPGD in many instances— not further discussed in this paper. Here, a GEM (Gas Electron Multiplier) [@gem] based TPC system is described.
A typical GEM based readout system for a TPC consists of two or more GEM foils stacked on top of a readout pad plane. Electrons produced in the volume of the TPC are drifted towards the readout. Gas amplification takes place at each GEM. Per GEM an amplification of typically a factor of 100-200 is possible resulting in a total system amplification between $10^3$ and $10^4$. The performance of the system depends on the intrinsic amplification properties of the GEM, and on the mechanical properties like distances between GEMs, flatness of the GEM foils and stability of the overall system.
A technical problem for GEM systems is the mechanical support of a large scale amplification system on the end plate. The amplification system should cover the end plate without dead areas, should be mechanically stable and robust, and should minimize the amount of dead material introduced. The GEM foil should be supported stably, kept at a constant distance from the readout pad plane or its neighbor GEM, and not move under the influence of the electric or magnetic field. A number of different solutions to this problem have been published over the years [@heraB; @panda; @fair] including some which use also thin spacers between GEMs. However the solution proposed here is special insofar as there are no thicker spacers on the edges of the GEM system, as it is for example the case with the COMPASS GEM support system [@compass].
In this paper, a system is proposed to support the GEM foils by a ceramic grid. The system was originally developed for the TPC for the ILD (International Large Detector) experiment proposal at the ILC, but can be applied to other systems as well. The grid serves a dual function. When glued to the GEM a very stable, light-weight mechanical system is formed, with a very high bending strength, similar to the way modern sandwich materials are based on sheets of honeycomb material with thin skins on the top and the bottom. By properly adjusting the mesh size of the grid to the typical length scale by which GEM foils undulate, the GEM foils can be flattened without applying undue mechanical stress.
The design parameters and properties of the proposed system are evaluated based on GEM foils of $10 \times 10$ cm$^2$. In the following, the design of the system is described, and its performance is evaluated.
Mechanical Design of the Grid GEM {#sec:prod}
=================================
The central part of the proposed GEM system is a stack of GEM foils separated and supported by ceramic grids. The grid is made of an aluminum oxide ceramic (${{\mathrm{Al_2 O_3}}}$) [@ANC]. This material is very stiff, an excellent insulator and machinable by laser cutting. The support is designed as a grid with a mesh size of about . The bars of the grid are wide, which is the minimum width which currently is possible in this technology (see figure \[fig:grid\] for a technical drawing and list of specifications). A basic unit consists of a grid glued on both sides to a GEM foil producing a very stable and light-weight sandwich. Here, a system of ’GEM - two grids - GEM - two grids - GEM’ is used for a triple grid GEM stack, which is mounted on the readout plane. Such a stack provides wide transfer gaps. Several basic units can be combined in different ways and with additional or less grids as spacers, so that a very flexible system exists to build up a stack of GEMs on top of a readout plane, with different distances between the different units. The basic unit is produced by pulling the GEM foils on a flat vacuum table and then, using a proper mounting rig, positioning and glueing the GEM to the grid. A robot arm is used to meter the appropriate amounts of glue on the grid bars (a two component epoxy, Polybond EP 4619/3 [@PT]). Currently, only the outer bars of the grid are glued to the GEM. This is to avoid that glue leaks into the GEM holes in areas where the grid bars cover the active area of the GEM. In the future GEM foils will be specially produced which will provide narrow gluing surfaces without holes underneath the grid bars. It should be noted that the full mechanical strength will only be reached with such GEM foils.
(15,6.5) (0,0) (10,0.75) (0,6.3)[(a)]{} (9.5,6.3)[(b)]{} (9.5,0.9)[11]{} (9.5,1.4)[10]{} (9.5,1.9)[ 9]{} (9.5,2.4)[ 8]{} (9.5,2.9)[ 7]{} (9.5,3.4)[ 6]{} (9.5,3.9)[ 5]{} (9.5,4.4)[ 4]{} (9.5,4.9)[ 3]{} (9.5,5.4)[ 2]{}
A technical drawing of the grid is shown in figure \[fig:grid\](a). The relevant characteristics are summarized in table \[tab:grid\].
\[tab:grid\]
For the prototype system the grid covers about of the active area of the GEM foil. This is significantly less than in more conventional systems, where typically a stable frame external to the GEM takes the stress from stretching the GEM over the readout plane. Using naively the standard GEM provided mounting frames made from wide strips of glass fiber reinforced plastic and scaling it up, the support would take about of the total area.
The flatness of the GEM foils is another key parameter for their performance. The gain of a GEM depends on the potential applied across the GEM, and on the field on the top and at the bottom of the GEM. Changing the distance between the GEM and, e.g., the readout pad plane, will change locally the electric field and thus the gain. Therefore, for optimal performance a flat mounting is important. In addition, changes in the distance will create potentially areas with high fields which might be more sensitive to breakdown or instabilities. Studies done in simulation and validated with experimental measurements [@lea] have shown that fluctuations of the transfer distances between the GEMs —due to height deviations of the GEM foils of the order of — result in variations of the energy resolution of up to , which would still allow for a significant contribution to the particle identification via $dE/dx$ measurements in the TPC. The spatial resolution is to first order not affected by this height variations of the foils (see section \[sec:track\]).
To evaluate the feasibility and performance of the system, a triple GEM amplification structure was built. It is placed on a readout plane instrumented with pads with a pitch of , shown in figure \[fig:grid\](b). Adjacent rows of pads are staggered by half a pad pitch and the grid is aligned to the pads. For this photo, the grid has been placed directly onto the readout pad plane. About % of the pads are shadowed by the grid. A key objective therefore is to study the impact this shadowing has on the performance, overall and especially for those pads directly affected by the grid.
The assembled system was mechanically measured and the maximum deviation from a flat plane was found to be about . Its impact on the point resolution is sufficiently small to be neglected and expected to meet the envisaged particle identification capabilities via $dE/dx$ measurements.
The described system has been tested inside a prototype TPC, with and without magnetic field. It allows for stable operation using the same gas flow rate and high voltage parameters as for GEM stacks operated without a grid structure. No operational problems like charge-up or instabilities of the GEM foils were found over extended running periods of several weeks.
Experimental Setup to Study a Grid GEM TPC with Cosmic Muon Data {#sec:setup}
================================================================
The triple grid GEM stack composed of GEMs (double conical holes, hole pitch ${{\mathrm{\unit[140]{{\upmu}m}}}}$, inner/outer hole diameter) was operated in a prototype TPC with an active area of $\unit[27]{cm^2}$ and a drift length of at magnetic fields of up to . The GEM support structure was mounted on top of the readout plane described in section \[sec:prod\]. Ten complete pad rows were read out with charge sensitive preamplifiers followed by a flash ADC system [@daq]. As counting gas, a mixture of Argon and Methane was used. The drift field was set to in order to be —due to the velocity plateau for this gas at this value— fairly independent from small drift field variations. The GEMs were used with voltages across them ranging between and . Fields of were applied in the wide transfer gaps, while a field of was used in the wide induction region. With the current TPC prototype setup, an absolute gain measurement is not possible. Using the parametrization of a GEM based system described in [@sven], a gain of 10,000 has been estimated for the setup. Two scintillator counters —above and below the prototype— operated in coincidence were used to trigger on cosmic muons. The coordinate system used in the reconstruction is defined by the pad plane —x is pointing horizontally over the pad rows, while y follows the vertical columns— and the drift distance along the chamber axis, corresponding to the z axis.
The reconstruction of the recorded data is divided into three steps. First, the pad-wise charge deposition from different time bins (corresponding to z values) is recorded. Only hits which cross at the rising edge a threshold set at about twice the pedestal value are used. The hit reconstruction in time is terminated if the charge crosses a second, lower threshold. Since one hit is expected to extend across more than one pad, pads are then clustered row-wise into hits. A simple center of gravity method is used to determine the coordinates of this hit across the rows. The time of the reconstructed hit is calculated by looking at the distribution of the differences between bins which follow in time. The maximum of this distribution corresponding to the inflection point of the pulse in the rising edge, is used as an estimator for the time. Second, a track finding algorithm is applied, which assigns the hits to a track candidate. Finally, the tracks are fitted with an algorithm assuming a circular path. More details about the reconstruction algorithms can be found in [@matthias].
Grid Impact on Charge Measurement
=================================
Since parts of the grid shadow some pads, an impact on the reconstructed hit parameters for hits on these pads is expected. The impact of the grid on the charge has been studied by investigating the total amount of charge deposited throughout a long data taking period with about 61,000 triggers on different pads. The cosmic rays illuminate the sensitive area uniformly, such that deviations from an uniform charge distribution are likely to be caused by the grid structure.\
In figure \[fig:Qsumhhit\](a) the total recorded charge per pad integrated over the measurement run is shown. Each bin has been normalized to the total number of triggers recorded in this bin.
(15,6) (0.1,0.22) (9,0.15) (0.1,5.7)[(a)]{} (9,5.7)[(b)]{}
The structure of the grid is clearly visible through regions of reduced overall charge. Since the pads are highly elongated, the impact is expected to be different for vertical and horizontal bars. The vertical bars, going in the direction of increasing row numbers, cover on average close to of a pad in this region. For the horizontal bars, oriented parallel to the x axis, the bars cover only about .
The observed reduction in charge per pad is about for the vertical bars and for the horizontal bars, which is in rough agreement with the assumption that the charge reduction is to the first order proportional to the geometrical area shadowed by the grid.\
The results show a clear impact of the grid on the charge deposited on the pads. Most important is therefore the alignment of the vertical bars and the pad plane. By e.g. staggering the pads under the bar, it can be ensured that not a complete column of pads is shadowed by the grid. The horizontal bars should be aligned to the middle of a pad row in order to minimize their impact. For a large scale TPC, the design of the grid has to be adapted to the module layout to ensure that all pads are able to provide useful charge signals, which is desirable to preserve the good pattern recognition performance of the TPC. In summary, with a proper design of the system the impact of the grid on the charge determination can be minimized.
Grid Impact on Hit Reconstruction and Hit Efficiencies
======================================================
In this section, the impact of the grid on the number of pads contributing to a hit and on the single hit efficiency is presented. Hits are reconstructed by combining neighboring pads within a row, as described in section \[sec:setup\]. In figure \[fig:Qsumhhit\](b) and \[fig:vhit\](b), the average number of pads per hit is shown, studying the influence of horizontal (see figure \[fig:Qsumhhit\](b)) and vertical (see figure \[fig:vhit\](b)) grid bars. In the solid histogram in both figures the distribution for hits where the impact of the grid bar is negligible is shown, in figure \[fig:Qsumhhit\](b) hits that are located more than from a grid bar and in figure \[fig:vhit\](b) more than one row. In both cases, the average number of pads contributing to a hit is reduced close the the grid bars. As expected, the effect is more pronounced for vertical bars than for horizontal bars. The effect extends beyond the immediately affected pad.
(15,7.5) (2,0.5) (7.5,0) (0,7.2)[(a)]{} (7.5,7.2)[(b)]{}
For the horizontal bars, the number of pads per hit is reduced by , if the center of gravity of the hit is on top of the grid. For hits positioned two rows away no impact is seen. For the vertical bars, the number of pads per hit is reduced by . In both cases the error quoted is purely statistical. For hits with the reconstructed position 3 pads away, no effect is visible. Since the pad response function (PRF) has a width of for different drift distances, it can be stated that hits occurring more than two widths of the PRF —or one pad pitch— away from the grid structure are not influenced by the grid.
The changes on the number of pads contributing to a hit will have an impact on the hit reconstruction efficiency, and thus might ultimately have an impact on the efficiency with which tracks can be reconstructed.
In the following, the efficiency to reconstruct a hit is defined as the number of reconstructed hits relative to the expected number of hits at this position. The study was performed on a sample of about 42,000 single cosmic muon tracks. To determine the expected number of hits at a given position, tracks are searched for in the sensitive volume. The row for which the hit efficiency is investigated is excluded from the track finding and fitting. The expected hit position in the row under investigation is calculated from the parameters of the track. A hit reconstructed in this row is tagged as belonging to the track if it is located within one pad width of the expected hit position.
With the used setup, tracks could have at most 10 rows contributing. To ensure a sample of well defined tracks, all rows except the one under investigation are required to show a hit on the track.
(15,7.5) (0.1,0.15) (7.7,0) (0.1,7.2)[(a)]{} (7.7,7.2)[(b)]{}
The hit efficiency as a function of x for row seven (not influenced by a horizontal grid bar) is shown in figure \[fig:hiteff\](a). The binning is chosen such that each bin corresponds to half a pad pitch, since the row wise staggering corresponds exactly to this half pad pitch. A drop in the efficiency is clearly visible for the x regions affected by the vertical grid bars. At directly covered pads the efficiency drops to almost zero; at neighboring pads, the efficiency drops about . The effects of the horizontal bars are studied with a sample of tracks where hits are at least 3 pad rows away from any vertical grid. In figure \[fig:hiteff\](b) the hit efficiency is shown as a function of the row number, integrating over all x. The impact of the horizontal grids in row four and nine is negligible, as expected.
Impact of the Grid on the Track Reconstruction {#sec:track}
==============================================
Particle tracks are reconstructed from measured space points by fitting track parameters to these points. In this section, the impact of the grid structures on the point resolution and possible biases in the reconstruction of the points due to the grid are studied.
#### Reconstruction bias
The observed reduction of charge close to the grid can induce systematic shifts of the reconstructed hit position for hits close to a grid bar. The bias is studied in a sample of reconstructed tracks. The intercept between the track and the row position for the row investigated is calculated. The distance between this intercept and the measured hit position on this row is studied. For this study the hit investigated is included in the fit.
(15,7.5) (0.1,0) (7.7,0) (0.1,7.2)[(a)]{} (7.7,7.2)[(b)]{}
In figure \[fig:hitdist\](a), the impact of the horizontal grid is studied. The impact is very small, below and within the statistical uncertainty. In figure \[fig:hitdist\](b) the impact of the vertical bar is shown. For the vertical bar, hits close to the bar are observed to be systematically shifted. The distribution shows a significant tail. The shift in this particular test system was found to be as large as for hits reconstructed at the position of the grid. This is non negligible in size, but still small compared to the intrinsic resolution of the system. The absolute size of the effect will depend on many factors, including drift distance, angle, the details of the readout, etc..
#### Single Point Resolution
The single point resolution is of particular importance in tracking detectors. It determines the momentum resolution of the TPC and gives a handle to judge the performance.\
The single point resolution is calculated following the procedure outlined in [@dean]. To summarize briefly, for the hit under investigation the distance between this hit and the expected position of the track is calculated including and excluding the hit from the track fit. The width of the distribution of the geometrical mean of these two numbers has been shown to be an unbiased estimator of the single hit resolution.
\[tab:rescuts\]
The tracks are selected according to the cuts summarized in table \[tab:rescuts\]. Two samples of tracks are prepared which are sensitive to the impact of the horizontal and vertical bar, respectively. For the study of the horizontal bars, tracks with hits closer than three pad pitches to a vertical bar are excluded. The vertical bar is investigated by comparing tracks from the sample described above with those tracks that have been removed from this sample.
In figure \[fig:resallh\](a), the single point resolution for a sample of cosmic muons recorded with the TPC equipped with a grid GEM readout is shown. For comparison results from a similar run but recorded with a GEM based readout with conventionally framed GEMS are also shown. Both sets of data were taken at a magnetic field of and with identical GEM settings. In figure \[fig:resallh\](b) the same plots are shown, but this time only for tracks which do not contain hits within 3 columns of a vertical bar. These two sets of plots clearly demonstrate that the horizontal bars do not have a significant impact on the resolution, while hits close to a vertical bar are found to have a worse single point resolution.
(15,5.5) (0,0) (7.7,0) (0,5.2)[(a)]{} (7.7,5.2)[(b)]{}
The effect is studied in more detail by comparing a sample of tracks close to and crossing a vertical bar with a sample of tracks in a reference region. The single point resolution is calculated individually for tracks with hits close to (less than three pad pitches) and therefore affected by a vertical grid, and for tracks in the reference region far from both vertical grid bars. The results are shown in figure \[fig:reshv\]. The resolution deteriorates by about 20% compared to all other hits.
(15,5.5) (4,0.0)
Conclusion
==========
A novel scheme to mount GEM foils inside a TPC has been developed. This self-supporting structure is made of a ceramic grid glued on the GEM foils. The new support structure has been demonstrated to result in a stable, flat system which can be operated stably over extended periods of time.
In order to quantify the impact of the grid GEMs on the track reconstruction, cosmic muon tracks have been recorded in a magnetic field of . The results show that the main impact of the grid is a reduction in the reconstructed charge on all pads which are shadowed by the grid. The effect depends on the fraction of a pad covered by the grid. It has been shown that the effects can be controlled at a level below the intrinsic hit resolution.
The advantages of the developed grid support structure are the minimal amount of material, the achievable flatness without the need of stretching the foils, the almost edgeless module borders and the resulting possibility to cover large areas without significant gaps.
This work was supported by the Commission of the European Communities under the 6$^\mathrm{th}$ Framework Programme ’Structuring the European Research Area’, contract number RII3-026126. We would like to express our thanks to the DESY cryogenic operations team for the efficient operation of the 4T magnet facility.
[9]{}
F. Sauli, *GEM: A new concept for electron amplification in gas detectors*, Nucl. Instrum. Meth. A [**386**]{} (1997) 531.
T. Zeuner \[HERA-B Collaboration\], *The MSGC-GEM Inner Tracker for HERA-B* Nucl. Instrum. Meth. A [**446**]{} (2000) 324.
M. Vandenbroucke \[GEM-TPC Collaboration\], *A GEM-based TPC prototype for PANDA* IEEE Nucl. Sci. Symp. Conf. Rec. [**2009**]{} (2009) 1009.
F. Garcia, R. Turpeinen, R. Lauhakangas, E. Tuominen, M. Kalliokoski, R. Janik, P. Strmen and M. Pikna [*et al.*]{}, *Prototype development of a GEM-TPC for the Super-FRS of the FAIR facility* IEEE Nucl. Sci. Symp. Conf. Rec. [**2011**]{} (2011) 1788.
C. Altunbas et al. *Construction, test and commissioning of the triple-GEM tracking detectors for COMPASS.* Nucl. Instr. Meth. A [**490**]{} (2002) 177.
http://www.anceram.com/3\_keramik/produktbeschreibung\_en.php, *ANCeram Company website - technical ceramics*, 2010
http://www.polytec-pt.com, *Polytec PT company web site – in german*, 2010
L. Hallermann, *Analysis of GEM properties and development of a GEM support structure for the ILD Time Projection Chamber*, DESY-THESIS-2010-015.
M. Ball, N. Ghodbane, M. E. Janssen and P. Wienemann, *A DAQ system for linear collider TPC prototypes based on the ALEPH TPC electronics*, LC-DET-2004-013, arXiv:0407.120 \[physics\]
S. Lotze, *Ion Backdrift Minimisation in a GEM -Based TPC Readout*, PhD thesis, RWTH Aachen University.
M. E. Janssen, *Performance studies of a time projection chamber at the ILC and search for lepton flavour violation at HERA II*, DESY-THESIS-2008-011.
R. K. Carnegie, M. S. Dixit, J. Dubeau, D. Karlen, J. P. Martin, H. Mes and K. Sachs, *Resolution studies of cosmic ray tracks in a TPC with GEM readout* Nucl. Instrum. Meth. A [**538**]{} (2005) 372 \[physics/0402054\].
[^1]: Corresponding Author
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'N. La Palombara, P. Esposito, F. Pintore, L. Sidoli, S. Mereghetti, and A. Tiengo'
bibliography:
- 'biblio.bib'
date: 'Received / Accepted'
title: 'Spectral analysis of during its 2017 outburst and properties of the soft excess in X-ray binary pulsars [^1]'
---
Introduction\[intro\]
=====================
The X-ray pulsar was discovered with the *Rossi X-ray Timing Explorer* () in January 1998, during an observation in the direction of the Small Magellanic Cloud (SMC). The source (designated as XTE J0055-724) was detected at a flux level $\simeq 6 \times 10^{-11}$ in the energy range 2–10 keV, and displayed a periodic modulation with PS. = 59.0 $\pm$ 0.2 s [@Marshall+98]. A follow-up observation with led to the detection of the bright source SAX J0054.9-7226 within the 10 arcmin error circle of XTE J0055-724, at a flux level $\simeq 1.9 \times 10^{-11}$ between 2 and 10 keV. The measurement of a strong pulsed emission with PS. = 58.969 $\pm$ 0.001 s confirmed the source association with XTE J0055-724 [@Santangelo+98]. The uncertainty on the position was about 40 arcsec, thus including the and X-ray sources RX J0054.9-7226 (also known as 1WGA J0054.9-7226) and 2E 0053.2-7242. The latter was known to be variable on timescales from months to years and was proposed as a candidate High Mass X-ray Binary (HMXB) by @Bruhweiler+87 [source \#9], by @WangWu92 [source \#35], and by @White+94. The analysis of 13 archival PSPC and HRI observations revealed the former source three times between 1991 and 1996, at a luminosity level between $\simeq 8 \times 10^{34}$ and $4 \times10^{35}$ (in the energy range 0.1–2.4 keV). Moreover, in the 1991 observation a pulse period PS. = 59.072 $\pm$ 0.003 s was measured. This result confirmed the association of the and sources with RX J0054.9-7226 [@Israel+98], which can very likely be identified with 2E 0053.2-7242.
Between 1998 and 1999 detected four outbursts from , each with a duration of $\simeq$ 40 d. Based on the time spacing of these outburst, @Laycock+05 derived an orbital period = 123 $\pm$ 1 d. Later on, this result was confirmed by @Galache+08, who extracted the orbit ephemeris and derived a period = 122.10 $\pm$ 0.38 d from five source outbursts observed with between 2002 and 2004. During the periods of activity in 1998–1999 and in 2002–2004 the source showed a short-term spin-up and reached a 3–10 keV luminosity $\simeq 2 \times 10^{37}$ . Moreover, although during the $\sim$1100 days between the two groups of outbursts the source remained undetected, it experienced a long-term spin-up, with = -0.0227 $\pm$ 0.0006 s y$^{-1}$ [@Coe+10]. Finally, the source was detected again by between the end of 2008 and the beginning of 2011, although at a lower luminosity level ( = 10$^{36-37}$ ). By taking into account all the observations of this source over a time span of 13 years, @Klus+14 measured an average pulse period and source luminosity of 58.859 $\pm$ 0.005 s and (8.4$\pm$0.2)$\times 10^{36}$ , respectively. Throughout this time period experienced an average spin-up = -0.0206 $\pm$ 0.0005 s y$^{-1}$.
Three candidate counterparts were found within the 10 arcsec radius HRI error circle. All of them are early-type stars, but only one shows strong H$\alpha$ emission. Its radial velocity of 138 $\pm$ 27 km s$^{-1}$ is consistent with a location in the SMC. Moreover, it corresponds to object 810 of the catalogue of @MeyssonnierAzzopardi93, which includes the H$\alpha$ emission-line objects in the SMC. Therefore, this object was identified as the optical counterpart of [@Stevens+99] and, afterwards, it was classified as a B0e star [@Antoniou+09]. Moreover, it was extensively monitored with the *MAssive Compact Halo Objects* (MACHO) and *Optical Gravitational Lensing Experiment* (OGLE) projects. Based on the long-term MACHO and OGLE-II photometric data, @SchmidtkeCowley05 proposed an orbital period = 60.2 $\pm$ 0.8 d, subsequently corrected to 62.15 $\pm$ 0.04 d with the addition of the OGLE-III data [@Rajoelimanana+11]. The optical period is half that of the X-ray, but no significant X-ray flux was detected half a phase from the X-ray maximum [@Galache+08].
Since its discovery has been detected several times with the *Advanced Satellite for Cosmology and Astrophysics* () [@Yokogawa+00], the *X-ray Multi-Mirror* () mission [@Sasaki+03; @Haberl+08; @Sturm+13], and the telescope [@Laycock+10]. In their catalogue of the HMXBs in the SMC, @HaberlSturm16 reported that the 0.2–10 keV flux of has varied between $_{\rm min}$ = 7$\times 10^{-14}$ and $_{\rm max}$ = 6$\times 10^{-11}$ . Moreover, the catalogue of all the SMC observations performed with , , and , compiled by @Yang+17, shows that has been detected with at even lower luminosity levels, down to $\simeq 10^{34}$ . The measured pulse periods imply a long-term spin-up of the pulsar, with = -0.016 $\pm$ 0.004 s y$^{-1}$.
On March 30, 2017 (MJD 57842), XRT detected an outburst from [@Kennea+17]. The measured count rate in the Photon Counting mode exposure was 0.9 $\pm$ 0.2 c s$^{-1}$. A follow-up observation, performed again with on April 7 (MJD 57850), found that this source had brightened since the previous observation, up to an XRT count rate of 2.1 $\pm$ 0.1 c s$^{-1}$. The detected spectrum was well fitted with an absorbed power-law (PL) model, with a photon index $\Gamma$ = 0.96 $\pm$ 0.10. The measured flux value (in the energy range 0.5–10 keV) was = $8 \times 10^{-11}$ , corresponding to = $4 \times 10^{37}$ for an SMC distance of 62 kpc [@Graczyk+14]. Therefore, we triggered our Target-of-Opportunity (ToO) programme for the observation of transient binary pulsars in the SMC. Here, we report on the results obtained with the follow-up observation of .
Observation and data reduction {#data}
==============================
On April 14, 2017 (MJD 57857), two weeks after the beginning of the outburst, our target was observed with for a total exposure time of $\sim$ 14 ks. In Table \[observation\] we provide the set-up of the EPIC [@Struder+01] and MOS [@Turner+01] focal-plane cameras and of the Reflection Grating Spectrometer (RGS, @denHerder+01). All the collected events were processed with version 16 of the [*Science Analysis System*]{}[^2] (<span style="font-variant:small-caps;">sas</span>). We verified that the whole observation was free of soft-proton contamination, which could affect the data analysis. Therefore, we considered the full EPIC and RGS datasets for both timing and spectral analysis. The effective exposure times were $\simeq$ 10 ks for the camera and $\simeq$ 14 ks for the MOS and the RGS instruments; for the EPIC cameras they take into account the dead time of 29 % and 2.5 % for and MOS, respectively. In Table \[observation\] we summarize the observation.
In the case of the EPIC data, we selected mono- and bi-pixel events (with pattern between 0 and 4) for the camera and from 1- to 4-pixel events (with pattern between 0 and 12) for the MOS cameras. From the spatial point of view, the events were selected from circular regions around the source position. In all cameras the CCD edges or dark columns limited the extraction radius, which was 35 arcsec for the camera, 40 arcsec for the MOS1 camera, and 50 arcsec for the MOS2 camera. Although the source count rate (CR) was rather high (Table \[observation\]), we checked that the data of all the EPIC cameras were not affected by photon pile-up. To this end, we followed the same approach used by @LaPalombara+18: on the one hand, spectra with different pattern selections (only mono- or bi-pixel events), and on the other hand, spectra with or without the removal of the central part of the point spread function (PSF), where the possible pile-up is higher. In all cases the spectral analysis provided consistent results, thus showing that the pile-up was negligible. Therefore, for each camera the source events were selected from the whole circular region. The corresponding background events were selected from circular regions which were offset from the target position and free of sources. In the case it was on the same CCD of the source and, hence, it had a small radius of 35 arcsec. For the MOS cameras it was on a peripheral CCD and its radius was 200 and 250 arcsec for the MOS1 and MOS2, respectively.
------------ -------- -------------- ----------------- ------------------- ------------------- -----------------
Instrument Filter Mode Time Resolution Net Exposure Time Extraction Radius Net Count Rate
(ks) (arcsec) ()
Thin 1 Small Window 5.7 ms 10.0 35 9.49$\pm$0.03
MOS1 Thin 1 Small Window 0.3 s 13.8 40 2.64$\pm$0.01
MOS2 Thin 1 Small Window 0.3 s 14.0 50 3.02$\pm$0.01
RGS1 - Spectroscopy 4.8 s 14.7 - 0.274$\pm$0.003
RGS2 - Spectroscopy 9.6 s 14.6 - 0.090$\pm$0.002
------------ -------- -------------- ----------------- ------------------- ------------------- -----------------
Timing analysis {#timing}
===============
For the timing analysis of the EPIC events, we converted the arrival times to the solar system barycentre by using the <span style="font-variant:small-caps;">sas</span> tool <span style="font-variant:small-caps;">barycenter</span>. We defined three energy ranges (soft = 0.15–2 keV, hard = 2–12 keV, and total = 0.15–12 keV) and, for each range and each of the three cameras, we accumulated a light curve (with a time binning of 100 s). These curves were then corrected for the background and the extraction region by using the <span style="font-variant:small-caps;">sas</span> tool <span style="font-variant:small-caps;">epiclccorr</span>. In the total range the average CR was $\simeq$ 11.3, 3.6, and 3.5 cts s$^{-1}$, for the , MOS1, and MOS2 cameras, respectively. For each of the three energy ranges, we calculated the cumulative light curve as the sum of the light curves of the individual cameras. We show the three curves in Fig. \[lc\], together with the hardness ratio of the hard (H) to the soft (S) light curves (HR = H/S). In both the soft and hard energy ranges the average CR was $\simeq$ 9 cts s$^{-1}$. Although the source was variable over short timescales (with CR variations of up to $\sim$ 30 % between consecutive time bins), its flux does not increase or decrease over the observation timescale. Moreover, no correlation is found between the HR and the CR: the HR shows some bin-to-bin variability, but without any dependence on the CR or clear trend on long timescales.
We merged the event datasets of the three instruments for the measurement of the pulse period of the source in order to increase the count statistics. Then, by applying a standard phase-fitting technique, we obtained a best-fitting period of $P$ = 58.949(1) s. The three folded light curves, and the corresponding HR, are shown in Fig. \[flc2E\]. The pulse profile is characterized by similar properties at all energies: it shows a broad maximum around phase $\Phi \simeq$ 0.5, a small peak at phase $\Phi \simeq$ 0.9, and a minimum around phase $\Phi \simeq$ 0. However, the figure shows some energy dependence since in the hard range two distinct narrow peaks arise above the broad maximum, at $\Phi \simeq$ 0.45 and 0.65. The HR is rather correlated with the CR and shows two narrow peaks coincident with those of the hard range. The average pulsed fraction, defined as PF = (CR$_{\rm max}$ - CR$_{\rm min}$)/(2$\times$CR$_{\rm average}$), is high and depends on energy, since it is $\simeq$ 42 % for the soft range and $\simeq$ 64 % for the hard range.
In Fig. \[flc4E\] we show the folded light curve in four narrower energy bands (normalized to the average CR in each band). It proves that the pulse profile evolves from a single broad peak at the low-energy end to two narrow and very distinct peaks at the high-energy end. The flux variability increases with energy since the CR varies by a factor $\simeq$ 2 at E $<$ 1 keV, but increases by a factor $\simeq$ 3.5 for E $>$ 4.5 keV. The same happens for the PF, which increases from 30 % at E $<$ 1 keV up to above 60 % at E $>$ 4.5 keV.
EPIC spectroscopy {#EPIC}
=================
Since the light curve of shows that neither the flux nor the spectral properties of the source vary on the observation timescale, for each EPIC camera we accumulated its spectrum over the whole exposure. Each spectrum was extracted from the same region used for the light curve and was rebinned with a significance of at least 10 $\sigma$ for each energy bin. Then we used the <span style="font-variant:small-caps;">sas</span> tasks <span style="font-variant:small-caps;">rmfgen</span> and <span style="font-variant:small-caps;">arfgen</span> to generate the response matrices and ancillary files, respectively. We performed the spectral analysis in the energy range 0.2–12 keV, using version 12.9.1 of <span style="font-variant:small-caps;">xspec</span>, and calculated the spectral uncertainties and upper limits at the 90 % confidence level for one interesting parameter. Since is in the SMC, we assumed the value of 62 kpc [@Graczyk+14] for the source distance. We simultaneously fitted the spectra of the three EPIC cameras since their separate fits provided consistent results. Moreover, to account for possible uncertainties in instrumental responses, the relative normalizations of the three cameras were allowed to vary. For the spectral fitting we used the absorption model <span style="font-variant:small-caps;">tbnew</span> in <span style="font-variant:small-caps;">xspec</span>. To this end, we considered the interstellar abundances provided by @WilmsAllenMcCray00 and the photoelectric absorption cross sections of @Verner+96. We described the interstellar absorption with two different components: a Galactic component with solar abundances and column density fixed to $N_{\rm H}^{\rm GAL} = 6 \times 10^{20}$ cm$^{-2}$, the total Galactic absorption in the SMC direction [@DickeyLockman90], and an additional component ($N_{\rm H}^{\rm SMC+local}$), which accounts for the insterstellar absorption within the SMC and that local to the source, with free column density and abundances for elements heavier than helium set to 0.2, the canonical abundance value for the SMC [@RusselDopita92].
We found that no single-component model was able to describe the source spectrum successfully. For example, the fit with an absorbed PL model resulted in a best-fitting solution with /d.o.f. = 2.16/1214. In the medium panel of Fig. \[epic\_spectrum\] we report the corresponding data-model residuals, which show the presence of a significant soft excess (SE) below E $\simeq$ 1 keV. We described this feature with a black-body (BB) component and either a broad Gaussian emission line at $\simeq$ 0.95 keV or a model representing the spectral emission due to a collisionally ionized gas (<span style="font-variant:small-caps;">apec</span> in <span style="font-variant:small-caps;">xspec</span>). Moreover, in both cases we found two narrow emission lines at $\simeq$ 1.38 and 6.33 keV, which we modelled with two Gaussian emission lines of null width.
In Table \[epic\_fit\] we list the best-fitting parameters obtained for these spectral models. Two different possibilities were considered for the metal abundance of the APEC component. On the one hand, it was fixed at the estimated metallicity for the SMC [$Z = 0.2 Z_\odot$, @RusselDopita92], thus obtaining a fit quality equivalent to the value obtained with the BB+Gaussian model ( = 1.05). On the other hand, we left the metal abundance free to vary and obtained an almost equivalent fit quality ( = 1.06); in this case the best-fitting value of the abundance is 0.23$^{+0.31}_{-0.15}$, which is consistent with the expected value for the SMC. In all cases the best-fitting absorption value is $N_{\rm H}^{\rm SMC} = 1.2 \times 10^{21}$ cm$^{-2}$, which points out a significant absorption within the SMC or locally at the source.
The BB properties in the three models are similar. The radius of this component is $\simeq$ 110 km and it contributes $\simeq$ 3.5 % to the total flux. On the other hand, the contribution of the APEC model is $\simeq$ 1.5 %. The broad Gaussian emission line at $\simeq$ 0.95 keV of the first model can be attributed to either a mixture of several L$\alpha$ emission lines from Fe in various ionizations states (from [Fe xviii]{} to Fe [xx]{}) or a radiative recombination continuum (RRC) from O [viii]{} - Ne [ix]{}. The narrow Gaussian components at $\simeq$ 1.35 and 6.32 keV can be identified with Mg [xi]{} and neutral Fe K$\alpha$ emission lines, respectively. However, regarding the Fe line we note that its energy is slightly lower than the typical value of 6.4 keV, and moreover that it is characterized by a small instrinsic width.
[lccc]{} Model & <span style="font-variant:small-caps;">pl + bb + 3 Gaussian</span> & <span style="font-variant:small-caps;">pl + bb + apec + 2 Gaussian</span> & <span style="font-variant:small-caps;">pl + bb + apec + 2 Gaussian</span>\
parameter & & (fixed abundance) & (free abundance)\
$N_{\rm H}^{\rm SMC+local}$ (10$^{21}$ cm$^{-2}$) & 1.16$\pm$0.13 & 1.23$^{+0.09}_{-0.14}$ & 1.22$^{+0.21}_{-0.17}$\
$\Gamma$ & 0.78$^{+0.01}_{-0.02}$ & 0.76$\pm$0.01 & 0.76$\pm$0.02\
Flux$_{\rm PL}$ (0.2-12 keV, $\times 10^{-11}$ ) & 7.79$^{+0.06}_{-0.03}$ & 7.78$\pm$0.06 & 7.78$\pm$0.06\
$kT_{\rm BB}$ (eV) & 177$^{+11}_{-13}$ & 171$^{+11}_{-14}$ & 171$^{+14}_{-18}$\
$R_{\rm BB}$ (km) & 107$^{+18}_{-15}$ & 110$^{+25}_{-15}$ & 111$^{+26}_{-15}$\
Flux$_{\rm BB}$ (0.2-12 keV, $\times 10^{-12}$ ) & 3.0$\pm$0.3 & 2.8$\pm$0.3 & 2.8$^{+0.4}_{-0.5}$\
$kT_{\rm APEC}$ (keV) & - & 1.09$^{+0.16}_{-0.09}$ & 1.09$^{+0.17}_{-0.09}$\
Abundance (<span style="font-variant:small-caps;">apec</span>) & - & 0.2 (fixed) & 0.23$^{+0.31}_{-0.15}$\
$N_{\rm APEC}$ ($\times 10^{-3}$ cm$^{-5}$) & - & 1.0$^{+0.6}_{-0.4}$ & 0.9$^{+2.0}_{-0.8}$\
Flux$_{\rm APEC}$ (0.2-12 keV, $\times 10^{-13}$ ) & - & 9$^{+8}_{-2}$ & 8$^{+22}_{-6}$\
\
$E_{\rm line1}$ (keV) & 0.97$\pm$0.03 & - & -\
$\sigma_{\rm line1}$ (keV) & 0.09$^{+0.05}_{-0.04}$ & - & -\
Flux$_{\rm line1}$ ($\times 10^{-5}$ ) & 8$^{+5}_{-6}$ & - & -\
EW$_{\rm line1}$ (eV) & 20$^{+11}_{-9}$ & - & -\
$E_{\rm line2}$ (keV) & 1.38$^{+0.03}_{-0.02}$ & 1.38$^{+0.03}_{-0.02}$ & 1.38$^{+0.03}_{-0.02}$\
$\sigma_{\rm line2}$ (keV) & $<$ 0.04 & 0 (fixed) & 0 (fixed)\
Flux$_{\rm line2}$ ($\times 10^{-5}$ ) & 1.6$^{+0.7}_{-1.0}$ & 1.2$^{+0.8}_{-0.7}$ & 1.3$^{+0.7}_{-0.8}$\
EW$_{\rm line2}$ (eV) & 6$^{+4}_{-3}$ & 5$^{+4}_{-3}$ & 5$\pm$3\
$E_{\rm line3}$ (keV) & 6.32$^{+0.09}_{-0.27}$ & 6.33$^{+0.09}_{-0.15}$ & 6.33$^{+0.09}_{-0.27}$\
$\sigma_{\rm line3}$ (keV) & 0 (fixed) & 0 (fixed) & 0 (fixed)\
Flux$_{\rm line3}$ ($\times 10^{-5}$ ) & 1.3$\pm$0.7 & 1.2$\pm$0.7 & 1.2$^{+0.8}_{-0.7}$\
EW$_{\rm line3}$ (eV) & 18$\pm$11 & 18$\pm$11 & 18$\pm$9\
Flux$_{\rm BB}$/Flux$_{\rm TOT}$ (0.01-12 keV) & 3.8 % & 3.5 % & 3.6 %\
Flux$_{\rm APEC}$/Flux$_{\rm TOT}$ (0.01-12 keV) & - & 1.4 % & 1.3 %\
Unabsorbed flux (0.2-12 keV, $\times 10^{-11}$ ) & 8.12$^{+0.06}_{-0.07}$ & 8.16$^{+0.08}_{-0.07}$ & 8.15$^{+0.11}_{-0.08}$\
Luminosity (0.2-12 keV, $\times 10^{37}$ ) & 3.53$^{+0.02}_{-0.03}$ & 3.54$^{+0.04}_{-0.03}$ & 3.54$^{+0.05}_{-0.03}$\
/d.o.f. & 1.05/1204 & 1.05/1206 & 1.06/1205\
RGS Spectroscopy {#rgs}
================
We extracted both the first-order and the second-order spectra from the data of both RGS instruments. Then, we used the [sas]{} task [rgscombine]{} to combine the two spectra of the same order. However, we verified that the count statistics of the combined second-order spectrum was very limited compared to those of the first-order, thus providing no additional information. Therefore, we considered only the first-order spectrum. We rebinned it with a minimum of 30 counts per bin and analysed it with [xspec]{} in the energy range 0.4–2.1 keV.
We started the spectral analysis by fitting the spectrum with a simple absorbed PL model. For the interstellar absorption we adopted the same approach used for the EPIC spectra, considering both a fixed Galactic component $N_{\rm H}^{\rm GAL} = 6 \times 10^{20}$ cm$^{-2}$ with solar abundances and a free component $N_{\rm H}^{\rm SMC+local}$ with SMC abundances. In this way we obtained a best-fitting solution which left emission residuals at energies E $\simeq$ 0.49, 0.55, 0.65, 0.80, and 1.02 keV (Fig. \[RGS\_spectrum\], middle panel). They were described with Gaussian components, whose parameters are listed in Table \[RGS\_parameters\]. The intrinsic width of the lines at 0.65, 0.80, and 1.02 keV is well determined, while it is unconstrained for the two lines at 0.49 and 0.55 keV. Therefore, in the spectral fit of these two lines we fixed the width value at 0. In this way we obtained a good description of the RGS spectrum (Fig. \[RGS\_spectrum\], lower panel). The normalization of the lines at 0.49, 0.80, and 1.02 keV is significant at the 99 % confidence level, and then we can confirm their detection. On the other hand, for the other two lines at 0.55 and 0.65 keV the normalization is significant at the 90 % confidence level only, and hence their detection is only tentative. In Table \[RGS\_parameters\] we propose a possible identification of each line. The features at 0.65 and 1.02 keV can be attributed to the Ly$\alpha$ lines of the H-like O [viii]{} and Ne [x]{}, respectively. The lines at 0.49 and 0.55 keV are marginally consistent with N [vii]{} and O [vii]{} lines, respectively. On the other hand, the identification of the feature at 0.80 keV is more uncertain, although it can be due to various Fe ionization stages.
------------------- ----------------------- ------------ ---------------------- --------------------- -----------------------
Observed Ion Laboratory $\sigma$ Flux EW
Energy Energy
(eV) (eV) (eV) (10$^{-5}$ ) (eV)
490$^{+5}_{-1}$ N [vii]{} (?) 500.3 0 (fixed) 9.3$^{+6.8}_{-5.4}$ 9.7$^{+5.9}_{-8.4}$
550$^{+12}_{-31}$ O [vii]{} (?) 561.0 0 (fixed) 9.0$^{+7.9}_{-6.7}$ 12.8$^{+4.9}_{-11.5}$
653$^{+4}_{-5}$ O [viii]{} 653.5 3.0$^{+9.4}_{-3.0}$ 3.7$\pm$2.8 5.3$^{+3.8}_{-3.3}$
799$^{+4}_{-3}$ Fe [xvii-xviii]{} (?) 793.0 2.9$^{+8.4}_{-2.5}$ 3.5$^{+1.9}_{-2.1}$ 6.5$^{+3.7}_{-4.5}$
1017$^{+8}_{-9}$ Ne [x]{} 1022.0 11.0$^{+9.6}_{-4.4}$ 8.0$^{+3.3}_{-4.0}$ 20.3$^{+8.9}_{-8.5}$
------------------- ----------------------- ------------ ---------------------- --------------------- -----------------------
As a final test, we also tried to describe the RGS spectrum with the best-fitting PL+BB+APEC model used for the EPIC spectra (section \[EPIC\]). The results we found in this way are fully consistent with the previous values, since the continuum component of the RGS spectrum can also be described with the model used for the EPIC continuum. Moreover, in this case the residuals of the RGS spectrum are comparable to those obtained with a simple PL model (middle panel of Fig. \[RGS\_spectrum\]), thus proving that it is necessary to describe the observed emission lines with specific Gaussian components.
Phase-resolved spectral analysis {#resolved_spectral_analysis}
================================
The pulse profile of shows a clear energy dependence (Fig. \[flc2E\]), which suggests a spectral evolution with the spin phase. Therefore, we performed a phase-resolved spectral analysis by selecting two different spectra for each EPIC camera: spectrum A in the phase range $\Delta\Phi$ = 0.21–0.71 and spectrum B in the phase ranges $\Delta\Phi$ = 0–0.21, 0.71–1. In this way, we considered separately the parts of the folded light curve characterized by HR above or below 1, respectively.
We wanted to compare these two spectra with the best-fitting model used for the time-averaged spectrum in order to verify whether it can also provide an appropriate description for these two spectra. Moreover, in this case we wanted to assess if the best-fitting parameters have consistent or completely different values. To this end, the two spectra were independently fit using the PL+BB+APEC model (with the metal abundance fixed at 0.2) for the spectral continuum. In Table \[2spectra\] we list the results obtained in this way.
[cccc]{} Parameter & Spectrum A & Spectrum B\
$N_{\rm H}^{\rm SMC+local}$ (10$^{21}$ cm$^{-2}$) & 1.3$\pm$0.2 & 1.2$^{+0.2}_{-0.3}$\
$\Gamma$ & 0.78$^{+0.01}_{-0.02}$ & 0.72$\pm$0.03\
Flux$_{\rm PL}^{(a)}$ & 10.7$\pm$0.1 & 4.86$\pm$0.07\
$kT_{\rm BB}$ (eV) & 180$^{+20}_{-40}$ & 165$\pm$14\
$R_{\rm BB}$ (km) & 90$^{+50}_{-30}$ & 130$^{+30}_{-20}$\
Flux$_{\rm BB}^{(b)}$ & 1.8$\pm$0.5 & 3.4$^{+0.4}_{-0.3}$\
$kT_{\rm APEC}$ (keV) & 1.0$\pm$0.1 & 1.2$^{+0.3}_{-0.2}$\
$N_{\rm APEC}$ ($\times 10^{-3}$ cm$^{-5}$) & 1.4$^{+0.7}_{-0.6}$ & 1.0$^{+0.7}_{-0.6}$\
Flux$_{\rm APEC}^{(b)}$ & 1.2$^{+0.6}_{-0.7}$ & 0.9$^{+0.7}_{-0.5}$\
\
$E_{\rm line1}$ (keV) & 1.40$\pm$0.03 & 1.37$\pm$0.04\
$\sigma_{\rm line1}$ (keV) & 0 (fixed) & 0 (fixed)\
Flux$_{\rm line1}^{(c)}$ & 2.0$^{+1.2}_{-1.3}$ & 0.8$^{+0.9}_{-0.8}$\
EW$_{\rm line1}$ (eV) & 6$^{+4}_{-3}$ & 5$\pm$5\
$E_{\rm line2}$ (keV) & 6.3$^{+0.1}_{-0.5}$ & 6.3 (fixed)\
$\sigma_{\rm line2}$ (keV) & 0 (fixed) & 0 (fixed)\
Flux$_{\rm line2}^{(c)}$ & 1.3$^{+0.6}_{-1.2}$ & 1.3$^{+0.8}_{-0.9}$\
EW$_{\rm line2}$ (eV) & 16$^{+12}_{-14}$ & 29$\pm$15\
Flux$_{\rm BB}$/Flux$_{\rm TOT}$ (0.01–12 keV) & 1.7 % & 6.8 %\
Flux$_{\rm APEC}$/Flux$_{\rm TOT}$ (0.01–12 keV)& 1.5 % & 2.2 %\
Unabsorbed flux$^{(d)}$ & 1.10$\pm$0.01 & 0.53$\pm$0.01\
Luminosity$^{(e)}$ & 4.78$\pm$0.05 & 2.30$\pm$0.04\
/d.o.f. & 1.03/883 & 1.14/469\
$^{(a)}$ 0.2–12 keV, $\times 10^{-11}$\
$^{(b)}$ 0.2–12 keV, $\times 10^{-12}$\
$^{(c)}$ $\times 10^{-5}$\
$^{(d)}$ 0.2–12 keV, $\times 10^{-10}$\
$^{(e)}$ 0.2–12 keV, $\times 10^{37}$
We found that the SMC+local absorption is comparable in the two spectra and consistent with the best-fit value of the time-averaged spectrum. Therefore, the data show no evidence of absorption variability. The total source flux decreases by a factor of $\simeq$ 2 between spectrum A and B. This reduction is essentially due to the PL component, which in both cases contributes more than 90 % to the total flux. Its photon index is comparable to the value of the time-averaged spectrum in the case of spectrum A, and only slightly lower for spectrum B. Both the BB and the APEC components are significant at the 99 % confidence level in both spectra, but their behaviour is very different: on the one hand, the APEC flux is almost constant, so that its relative contribution to the total flux slightly increases in spectrum B; on the other hand, the BB flux doubles, so that its relative contribution to the total flux increases by a factor of $\simeq$ 4. We note that the BB temperature between the two phases is very similar and fully consistent with the time-averaged value. Therefore, the flux variation of this component is mainly due to the different size of its emission region. Instead the APEC temperature is slightly higher in spectrum B. The two emission lines detected in the time-averaged spectrum are much less evident in these two phase-resolved spectra. In spectrum A both lines are significant only at the 90 % confidence level, while in spectrum B the line at 1.38 keV remains undetected. Moreover, the energy of the Fe line is unconstrained in spectrum B. Therefore, in this case we had to fix it at the best-fit value of the time-averaged spectrum. In this way, we found that the Fe line is significant at the 90 % confidence level. This result is very probably due to the lower count statistics of these two spectra compared to the time-averaged spectrum.
We also performed a simultaneous fit of the two spectra in order to further constrain the variability of the continuum components. In this case we assumed a common value for and fixed the energy and width of the emission lines since the count statistics are not high enough to constrain them. For the fit we first considered two different solutions for the PL+BB+Gaussian model of the continuum: (1) a common value of the PL photon index and BB temperature, leaving both normalizations free to vary, and (2) a common BB component. We found that in the first case the null hypothesis probability (NHP) of the best-fit model is $\simeq$ 0.03, while in the second it reduces to $\simeq 9 \times 10^{-5}$, even if the photon index $\Gamma$ of the PL component is left free to vary between the two spectra. This means that a spectral fit with a constant BB component is statistically unacceptable and can be rejected. We repeated the same type of test also with the PL+BB+APEC model of the continuum, with three different solutions: (1) a common value of the PL photon index and of the BB and APEC temperatures and independent normalizations, (2) a common BB component, and (3) a common APEC component. For the best-fit model of solution (1) we found that NHP $\simeq$ 0.03. For solution (2) the spectral fit is significantly worse, since NHP $< 2 \times 10^{-4}$ even if both $\Gamma$ and the APEC temperature are left independent. In case (3), instead, the spectral fit is only slightly worse (NHP = 6 $\times 10^{-3}$); moreover, it becomes comparable to solution (1) if the two BB temperatures are independent. These results show that while there is no evidence of variability of the APEC component, a constant BB component is clearly rejected by the data.
In summary, the spectral variability observed between the two phase ranges can be attributed to the variation in the relative contribution of the continuum components since the BB flux increases and the PL flux decreases between phase ranges A and B, while the APEC remains almost constant. This behaviour explains the energy-dependent pulse shape shown in Fig. \[flc2E\].
/XRT and observations
======================
/XRT
----
The X-ray telescope (XRT) on board the satellite performed a dozen snapshot observations of between 2017 April 7 and May 1, with exposure times between a few hundred seconds and $\sim$ 3.3 ks each. In all cases the XRT instrument was operated in Windowed Timing (WT) mode. The data of all the observations were reduced following the standard procedures[^3], extracting source and background events from circular regions of radius 20 pixels. The event arrival times were reported to the solar system barycentre by using the tool <span style="font-variant:small-caps;">barycorr</span>. Because of the unknown orbital solution of SXP59.0, we did not correct the data for the orbital motion.
The /XRT lightcurve is shown in Fig. \[swift\_lc\]. The red symbol represents the /XRT count rate estimated from the source flux measured with (CR$_{\rm XRT} = f_{\rm XMM}$/(8.1$\times 10^{-11}$ ) cts s$^{-1}$). The /XRT followed the outburst decay which lasted at least 25 days. Based on the ephemerides provided by @Galache+08, we estimated the epoch of the centroid of the two maxima of the 122 d X-ray modulation which preceded and followed the time period covered by the observations: they are MJD 57801 $\pm$ 21 and 57923 $\pm$ 21, respectively. Therefore, the and observations were performed far from the expected epoch of flux peak. In each /XRT observation, we detected the source pulsation at $\simeq$ 58.9 s. Thanks to the tight temporal distance between the and /XRT observations, we were able to perform a phase-connection of the pulsations. Starting from the EPIC data, we phase-connected with a linear function the /XRT observation taken $\sim$1 day after. However, the linear function was not able to fully account for the phase variation and therefore we added a parabolic correction ($\dot{\nu}$), which significantly improved the fit. We then phase-connected all the other remaining observations until the orbital effects (not taken into account) started to be important and did not allow us to continue the analysis (Fig. \[phase\_connection\]). In particular, we were able to connect $\simeq$ 14 days and found this solution: PS. = 58.95392161(2) s, = -1.160(3)$\times 10^{-7}$ s s$^{-1}$, and $\chi^2_{\nu}=3.9$ for 48 dof. The reference epoch is MJD 57957.6212479.
{#ogle}
In order to put the observation of in the context of the evolution of its optical counterpart, we used the OGLE monitoring system of the X-ray variables (XROM[^4], @Udalski08) to look for the OGLE IV [@Udalski+15] optical light curve. In Fig. \[ogle\_lc\] we show the long-term OGLE IV light curve, which includes data from May 2010 to December 2017. It shows that starting from MJD 55720 (June 2011) the source brightness has continuously increased until the latest observations. The dashed vertical line at MJD 57857 marks the epoch of the observation. It was performed near the end of the OGLE IV monitoring campaign, but within a period which was not covered by any OGLE observation.
Discussion
==========
Since its discovery with in 1980, has been observed several times at very different luminosity levels. In the upper panel of Fig. \[overall\_lc\] we report the long-term evolution of its 0.2–10 keV luminosity, as estimated based on the observed flux level and assuming a source distance of 62 kpc. It shows that the source luminosity varies by more than three orders of magnitude between $\sim 10^{34}$ and $\sim 10^{37}$ . For the first two decades the source was detected only during rather bright states, with $10^{36}$ . In particular, at the beginning of 1998 (MJD 50835) the source luminosity increased of over 1.5 orders of magnitude in $\sim$ two months, up to $\simeq 3 \times 10^{37}$ . Afterwards, the high-throughput X-ray telescopes and detected it as well at significantly lower flux levels, $10^{35}$ . The observation performed in April 2017 caught at one of the highest flux levels ever detected, possibly comparable only to that observed with in 1998 (assuming that the latter does not include any contribution of other nearby sources). This ToO observation allowed us to perform a detailed timing and spectral analyses of this source down to $\simeq$ 0.2 keV.
The lower panel of Fig. \[overall\_lc\], where we give all the published period values obtained in previous observations, shows that the period varies in the range 58.8–59.1 s. Moreover, according to @Klus+14, all the pulse periods measured with between 1998 and 2011 are in the range $\simeq$ 58.3–59.3 s. This period variability cannot be due to the orbit of the neutron star (NS) around the companion star. If we assume a mass $M \sim 10 \msole$ for the Be star, the orbital period $P_{\rm orb} \simeq$ 122 d implies an average orbital velocity $v_{\rm orb} \simeq$ 90 km/s. The corresponding period variation is $\Delta P = P \times v_{\rm orb}/c \simeq$ 0.02 s, a value much smaller than the observed variation in period values. Therefore, is it very likely that the period variability is due to the pulsar spin-up and spin-down during the outburst and quiescence phases, respectively. The pulse period measured with in 2017 is well within the range of values previously reported. This implies that, although the source can have experienced various outburst and quiescence phases since 2011, its average spin period has not changed significantly. The phase-connection of the pulsations reported in Section \[xrt\] shows that the pulsar undergoes a very high spin-up during the outburst phases. Therefore, it is very likely that these spin-ups are balanced by a significant slow-down during the quiescent phases, when the pulsar is most probably in the propeller state.
In the last years three additional transient Be binary pulsars in the SMC have been observed at high spectral resolution during an outburst: [@Sidoli+15], [@LaPalombara+16], and [@LaPalombara+18]. Compared to , these sources are characterized by a shorter pulse period and, in the first two cases, by a higher outburst luminosity. However, a direct comparison of these binary pulsars with is useful, since they have similar spectral and timing properties. To this end, in Table \[transients\] we report the main properties of these four sources.
Parameter
--------------------------------------------- ------------------- ---------------- -------------------- ------------------------ ------------------- ------------------------ ------------------------
(0.2-12 keV, $\times 10^{37}$ ) 3.5
(days) 122.1
$P_{\rm spin}$ (s) 58.95
PF (%) 53
(10$^{20}$ cm$^{-2}$) 12$\pm$1
$kT_{\rm BB}$ (eV) 93$\pm$5 - 135$^{+14}_{-11}$ - 218$^{+13}_{-14}$ - 171$^{+11}_{-14}$
$R_{\rm BB}$ (km) 350$^{+80}_{-50}$ - 320$^{+125}_{-95}$ - 50$^{+6}_{-5}$ - 110$^{+25}_{-15}$
f$_{\rm BB}$/f$_{\rm PL}$ (%) 1.7 - 3.1 - 1.6 - 3.5
$kT_{\rm APEC}$ (keV) - 0.21$\pm$0.03 - 1.22$^{+0.07}_{-0.10}$ - 1.13$^{+0.10}_{-0.08}$ 1.09$^{+0.16}_{-0.09}$
$N_{\rm APEC}$ ($\times 10^{-3}$ cm$^{-5}$) - 25$^{+8}_{-6}$ - 5$\pm$1 - 4$\pm$1 1.0$^{+0.6}_{-0.4}$
f$_{\rm APEC}$/f$_{\rm PL}$ (%) - 7 - 1.8 - 4.5 1.4
N [vii]{} yes
O [vii]{} yes
O [viii]{} yes
Ne [ix]{} no
Ne [x]{} yes
Mg [xi]{} yes
Si [xiii]{} no
Si [xiv]{} no
E$_{\rm Fe K\alpha}$ (keV) 6.3
$d_{\rm BB}$ (km) 570
$R_{\rm m}$ (km) 1100
$^{(a)}$ For this source the APEC component is considered an alternative to the BB component.
The source pulsations of were detected, for the first time, even at E $<$ 1 keV. The pulse profile observed with is clearly energy dependent: while it shows a single broad maximum at low energies, two distinct narrow peaks arise at high energies. The pulsed fraction is $>$ 40 % at E $<$ 2 keV and increases further with the energy. Moreover, the spectral hardness increases with the source flux. Similar results were obtained during the previous observations of the source outbursts in 1991 with [@Israel+98] and in 1998 with [@Santangelo+98]. In both cases the pulse profile was asymmetric and the pulsed fraction was above 40 %. Moreover, data showed two distinct peaks above 1.5 keV. On the other hand, the source observations at low luminosity level performed with in 2002 [@Sasaki+03] and with in 2006 [@Laycock+10] revealed a single-peak sinusoidal profile, with an almost constant hardness ratio. Our results for are also similar to what we observed for , , and . All of them showed a double-peaked profile, with a high pulsed fraction in the case of and . Since the source was observed at a rather high luminosity level, it is very probable that the double-peaked profile is due to a fan-beam emission geometry.
The average EPIC spectrum of is dominated by a rather hard absorbed PL component, with photon index $\Gamma \simeq$ 0.8 and total absorption due not only to the Galactic absorption in the SMC direction but also to a local component in the SMC itself. Both results are in agreement with the results provided by the previous observations performed with [@Israel+98], [@Santangelo+98], [@Sasaki+03], and [@Laycock+10]. In all cases the best-fit photon index was in the range $\Gamma \sim$ 0.8–1 and the total absorption was $\sim$ (1–3)$\times10^{21}$ cm$^{-2}$, thus significantly higher than the Galactic value in the SMC direction ($N_{\rm H}^{\rm GAL} = 6 \times 10^{20}$ cm$^{-2}$, @DickeyLockman90). However, the spectral fit with this single-component model leaves large residuals at energies below $\sim$ 1 keV, which reveal the presence of a previously unknown soft excess above the main PL component. We fitted this soft excess with the sum of a soft (kT $\sim$ 0.2 keV) BB and either a broad Gaussian component at 0.95 keV or a hot (kT $\sim$ 1 keV) thermal plasma model (APEC). In both cases the sum of the two soft components contributes $\simeq$ 5 % to the total source luminosity. It is interesting to note that the properties of both thermal components are very similar to those of the corresponding components observed in , , and (Table \[transients\]). In particular, this is true in the case of , which was observed at a luminosity level almost equal to that of . However, we emphasize that, in the case of , both components are necessary to obtain an acceptable spectral fit, while either the BB or the APEC was required for the other sources. This is an important difference since it implies that the observed soft excess of is due to the coexistence of two different emission processes.
The observation of during its 2017 outburst allowed us to detect previously unknown narrow emission features in the source spectrum. The high-resolution RGS spectrum showed the presence of emission lines due to N, O, and Ne, and of some other emission features, which could be attributed to L-shell lines from Fe at various ionization levels. The EPIC spectrum revealed two emission lines at 1.38 and 6.32 keV, which we attributed to K$\alpha$ emission from Mg and neutral Fe, respectively (although they are characterized by a negligible intrinsic width and, in the case of the Fe line, the best-fit energy is slightly lower than the typical value of 6.4 keV). Similar emission features were already observed in the spectra of , , and during their outbursts (Table \[transients\]). This is particularly true in the case of , since it is the only other source which showed a K$\alpha$ line from neutral Fe (while the other two sources showed an emission line consistent with ionized Fe). From this point of view, it is interesting to note that both sources were observed at the same luminosity level.
The phase-resolved spectral analysis of the EPIC data allowed us to study the dependence of the pulse profile as a function of the energy and to investigate the spectral variability along the pulse. To this end, we analyzed separately the spectra corresponding to, respectively, the high hard state and the low soft state, which are chacterized by a flux difference of $\simeq$ 50 %. We found that, in both cases, all three continuum components of the time-averaged spectrum (PL, BB, and APEC) were required to obtain an acceptable fit, and that the PL always contributed more than 90 % to the total flux. Moreover, the PL photon index and the temperatures of the two thermal components were always consistent with those of the averaged spectrum. However, the behaviour of the single components was significantly different. While the APEC flux remained almost constant and the PL halved its flux, the BB doubled it. These findings were confermed by the simultaneous analysis of the two spectra which showed that a constant BB component is inconsistent with the data.
The observation of was performed during an outburst, when the source luminosity increased to $\simeq 3.5 \times 10^{37}$ . This luminosity level implies a high accretion rate onto the NS, which very likely occurs through an accretion disc. This scenario is supported by the high spin-up rate measured during the outburst. In this case, according to @Hickox+04, the observed SE can be due to the combination of two different types of processes: emission from photoionized or collisionally heated diffuse gas and reprocessing of hard X-rays from the NS by optically thick accreting material, most probably at the inner edge of the accretion disc. We have found that the SE of is due to the sum of a BB and a APEC component, where only the BB is clearly pulsating. If this component is due to reprocessing of the primary radiation, and $\Omega$ is the solid angle subtended by the reprocessing region, the relation between the reprocessed and the primary luminosity is given by $L_{\rm BB}$ = ($\Omega$/4$\pi$) $L_{\rm X}$. On the other hand, $L_{\rm BB} = \Omega d_{\rm BB}^2 \sigma T^4_{\rm BB}$, where $d_{\rm BB}$ is the distance of the reprocessing site from the source of the primary emission. This implies that $d_{\rm BB}^2 = L_{\rm X}$/($4\pi \sigma T^4_{\rm BB}$). Since for $L_{\rm X} \simeq 3.5 \times 10^{37}$ and $T_{\rm BB} \simeq$ 0.17 keV, the estimated distance of the BB component from the central NS is $d_{\rm BB} \simeq 570$ km. This value is the estimated radius of the inner edge of the accretion disc, which should be comparable to the magnetospheric radius of the accreting NS: $R_{\rm m} \sim 1.5 \times 10^8 m^{1/7} R_6^{10/7} L_{37}^{-2/7} B_{12}^{4/7}$ cm, where $m = M_{\rm NS}$/$\msole$, $R_6 = R_{\rm NS}$/(10$^6$ cm), $B_{12} = B_{\rm NS}$/(10$^{12}$ G), and $L_{37} = L_{\rm X}$/(10$^{37}$ ) . If we assume $m$ = 1.4, $R_6$ = 1, and $B_{12}$ = 1, since for $L_{37}$ = 3.5, we obtain $R_{\rm m} \simeq$ 1100 km. This value is comparable to $d_{\rm BB}$ within a factor $\simeq$ 2, thus supporting the hypothesis that the observed BB component arises from the reprocessing of the primary radiation at the inner edge of the accretion disc. Moreover, this value is equal to that obtained for and comparable to those obtained for and (Table \[transients\]). For the estimate of the magnetospheric radius we have assumed the same values of the NS parameters (mass, radius, and magnetic field) for all the sources. Therefore, the different values of $R_{\rm m}$ are only related to the different luminosity levels: it decreases as the luminosity increases. For all the binary pulsars listed in Table \[transients\], $d_{\rm BB}$ is comparable with $R_{\rm m}$ within a factor of 2–3. This difference can have various physical or geometrical causes, since the NS magnetic field can be different by $10^{12}$ G; moreover, our estimate of $R_{\rm m}$ can be affected by a tilted and/or warped accretion disc.
Regarding the non-pulsating APEC component, from its normalization of $\simeq 10^{-3}$ cm$^{-5}$ we can derive a value of $\simeq 5.5 \times 10^{59}$ cm$^{-3}$ for the emission measure $n^2 V$. Assuming a gas density $n < 10^{12}$ cm$^{-3}$, as we did in the case of the other sources listed in Table \[transients\], and a spherical emitting region for the optically thin plasma, we obtain a radius of $R \gsim 5 \times 10^{11}$ cm. Therefore, the size of this component is much larger than the inner radius of the accretion disc. We note that, assuming $v_{\rm w} \simeq$ 100 km s$^{-1}$ for the wind velocity of the Be star and $v_{\rm r}^2 = v_{\rm w}^2 + v_{\rm orb}^2$ for the relative velocity between the wind and the NS, from the relation $R_{\rm acc}$ = 2GM/$v_{\rm r}^2$ we obtain $R_{\rm acc} \simeq 2 \times 10^{12}$ cm for the accretion radius around the NS. This value is consistent with the estimated size of the APEC component. Therefore, it is possible that this component is related to the shock region around the NS caused by the accreted wind from the companion star.
The timing and spectral properties of the BB component support its identification with the reprocessed primary radiation from the inner edge of the accretion disc. As proposed in the case of [@LaPalombara+18], at each phase along the pulse period, the inner edge of the disc is swept by the beamed primary emission. Only a limited section of the disc surface is illuminated by the beam. Moreover, it is possible that the disc geometry allows us to see only a limited fraction of the disc edge. For these reasons, the reprocessed component can be detected only when this visible portion of the disc is hit by the primary beam. It is possible that the Fe line observed in the time-averaged spectrum also has the same origin, as we have demonstrated in the case of , but for we cannot prove it because we lack any evidence of its variability. On the other hand, the other lines observed in the EPIC and RGS spectra are probably due to photoionized matter in regions above the disc since they cannot be described with the BB+APEC model.
Conclusions
===========
The observation of reported in this paper completes our programme of ToO observations of X-ray binary pulsars in the SMC. Thanks to the low interstellar absorption towards the SMC direction and the high flux level reached by these sources during their outbursts, these observations allowed us to perform detailed spectral and timing analysis and to study the physical processes that occur in these sources.
In all cases we detected the observed source at $\gsim 4 \times 10^{37}$ and obtained high-resolution spectra with high count statistics. We obtained consistent results for these pulsars:
- the pulsed emission can be observed over the whole energy range between 0.2 and 12 keV, but the pulse profile is energy dependent and the pulsed fraction increases with energy;
- the continuum spectrum is dominated by a hard ($\Gamma < 1$) PL component, but shows a soft excess at E $\lsim$ 1 keV, which contributes a small percentage of the total flux and is mainly due to a variable BB component;
- the EPIC spectrum is characterized by an emission feature which can be identified with a neutral or ionized Fe K$\alpha$ transition line, while the RGS spectrum shows low-energy features due to ionized N, O, and Ne;
- the BB component which dominates the SE and the Fe K$\alpha$ emission line can be attributed to the reprocessing of the primary radiation by the optically thick material at the inner edge of the accretion disc, while the other emission features are most probably due to photoionized plasma above the accretion disc; only in the case of did we find evidence for a collisionally heated thermal plasma, probably due to the shocked wind accreted from the companion Be star.
We acknowledge the use of public data from the data archive. We acknowledge the financial contribution from the agreement ASI-INAF I/037/12/0. PE acknowledges funding in the framework of NWO Vidi award A.2320.0076 and in the framework of the project ULTraS, ASI-INAF contract N.2017-14-H.0.
[^1]: Based on observations obtained with XMM-Newton, an ESA science mission with instruments and contributions directly funded by ESA Member States and NASA
[^2]: https://xmm-tools.cosmos.esa.int/external/xmm\_user\_support/documentation/sas\_usg/USG/
[^3]: http://www.swift.ac.uk/analysis/xrt/index.php
[^4]: http://ogle.astrouw.edu.pl/ogle4/xrom/xrom.html
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract:
- 'In this paper, we develop a novel procedure for low-rank tensor regression, namely *mportance ketching ow-rank stimation for ensors* (ISLET). The central idea behind ISLET is *importance sketching*, i.e., carefully designed sketches based on both the responses and low-dimensional structure of the parameter of interest. We show that the proposed method is sharply minimax optimal in terms of the mean-squared error under low-rank Tucker assumptions and under randomized Gaussian ensemble design. In addition, if a tensor is low-rank with group sparsity, our procedure also achieves minimax optimality. Further, we show through numerical studies that ISLET achieves comparable or better mean-squared error performance to existing state-of-the-art methods whilst having substantial storage and run-time advantages including capabilities for parallel and distributed computing. In particular, our procedure performs reliable estimation with tensors of dimension $p = O(10^8)$ and is $1$ or $2$ orders of magnitude faster than baseline methods.'
- 'In this supplement, we provide additional notation, preliminaries, ISLET procedure for general order tensor estimations, more details on tuning parameter selection, and all proofs for the main results of the paper.'
author:
- 'Anru Zhang$^1$, Yuetian Luo$^1$, Garvesh Raskutti$^1$, and Ming Yuan$^2$'
title: 'ISLET: Fast and Optimal Low-rank Tensor Regression via Importance Sketching'
---
**Key words:** dimension reduction, high-order orthogonal iteration, minimax optimality, sketching, tensor regression.
Introduction {#sec:intro}
============
The past decades have seen a large body of work on tenors or multiway arrays [@kolda2009tensor; @sidiropoulos2017tensor; @cichocki2015tensor; @kroonenberg2008applied]. Tensors arise in numerous applications involving multiway data (e.g., brain imaging [@zhou2013tensor], hyperspectral imaging [@li2010tensor], recommender system design [@bi2018multilayer]). In addition, tensor methods have been applied to many problems in statistics and machine learning where the observations are not necessarily tensors, such as topic and latent variable models [@anandkumar2014tensor], additive index models [@balasubramanian2018tensor], high-order interaction pursuit [@hao2018sparse], among others. In many of these settings, the tensor of interest is *high-dimensional* in that the ambient dimension, i.e, the dimension of the target parameter is substantially larger than the sample size. However in practice, the tensor parameter often has intrinsic dimension-reduced structure, such as low-rankness and sparsity [@kolda2009tensor; @sun2016sparse; @udell2019big], which makes inference possible. How to exploit such structure for tensors poses new *statistical* and *computational challenges* [@raskutti2015convex].
From a statistical perspective, a key question is how many samples are required to learn the suitable dimension-reduced structure and what the optimal mean-squared error rates are. Prior work has developed various tensor-based methods with theoretical guarantees based on regularization approaches [@lee2010practical; @mu2014square; @raskutti2015convex; @tomioka2011statistical], the spectral method and projected gradient descent [@chen2016non], alternating gradient descent [@li2017parsimonious; @sun2017store; @zhou2013tensor], stochastic gradient descent [@ge2015escaping], and power iteration methods [@anandkumar2014tensor]. However a number of these methods are not statistically optimal. Furthermore, some of these methods rely on evaluation of a full gradient, which is typically costly in the high-dimensional setting. This leads to computational challenges including both the *storage* of tensors and *run-time* of the algorithm.
From a computational perspective, one approach to address both the storage and run-time challenge is *randomized sketching*. Sketching methods have been widely studied (see e.g. [@avron2016sharper; @avron2014subspace; @ban2019ptas; @boutsidis2017optimal; @clarkson2015input; @clarkson2017low; @dasarathy2015sketching; @diao2018sketching; @dobriban2018new; @haupt2017near; @mahoney2011randomized; @nelson2013osnap; @pagh2013compressed; @pham2013fast; @pilanci2015randomized; @raskutti2014statistical; @song2017low; @song2019relative; @sun2019low; @tropp2017practical; @wang2015fast; @woodruff2014sketching]). Many of these prior works on matrix or tensor sketching mainly focused on relative approximation error [@boutsidis2017optimal; @clarkson2017low; @nelson2013osnap; @raskutti2014statistical] after randomized sketching which either may not yield optimal mean-squared error rates under statistical settings [@raskutti2014statistical] or requires multiple sketching iterations [@pilanci2015randomized; @pilanci2016iterative].
In this article, we address both computational and statistical challenges by developing a novel sketching-based estimating procedure for tensor regression. The proposed procedure is provably fast and sharply minimax optimal in terms of mean-squared error under randomized Gaussian design. The central idea lies in constructing specifically designed structural sketches, namely *importance sketching*. In contrast with randomized sketching methods, importance sketching utilizes both the response and structure of the target tensor parameter and reduces the dimension of parameters (i.e., the number of columns) instead of samples (i.e., the number of rows), which leads to statistical optimality whilst maintaining the computational advantages of many randomized sketching methods. See more comparison between importance sketching in this work and sketching in prior literature in Section \[sec:literature-review\].
Problem Statement {#sec:problem-statement}
-----------------
Specifically, we focus on the following low-rank tensor regression model, $$\label{eq:model}
y_j = \langle {{\mathbfcal{X}}}_j, {{\mathbfcal{A}}}\rangle +\varepsilon_j,\quad j=1,\ldots, n,$$ where $y_j$ and $\varepsilon_j$ are responses and observation noise, respectively; $\{{{\mathbfcal{X}}}_j\}_{j=1}^n$ are tensor covariates with randomized design; ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1 \times\cdots\times p_d}$ is the order-$d$ tensor with parameters aligned in $d$ ways. Here $\langle \cdot, \cdot\rangle$ stands for the usual vectorized inner product. The goal is to recover ${{\mathbfcal{A}}}$ based on observations $\{y_j, {{\mathbfcal{X}}}_j\}_{j=1}^n$. In particular, when $d=2$, this becomes a low-rank matrix regression problem, which has been widely studied in recent years [@candes2011tight; @koltchinskii2011nuclear; @recht2010guaranteed]. The main focus of this paper is solving the underdetermined equation system, where the sample size $n$ is much smaller than the number of coefficients $\prod_{i=1}^d p_i$. This is because many applications belong to this regime. In particular, in the real data example to be discussed later, one MRI image is 121-by-145-by-121, which includes 2,122,945 parameters. Typically we can collect far less number of MRI images in practice.
The general regression model includes specific problem instances with different choices of design ${{\mathbfcal{X}}}$. Examples include matrix/tensor regression with general random or deterministic design [@chen2016non; @li2013tucker; @raskutti2015convex; @zhou2013tensor], matrix trace regression [@baldin2018optimal; @candes2011tight; @fan2017generalized; @fan2016shrinkage; @koltchinskii2011nuclear; @recht2010guaranteed], and matrix sparse recovery [@yu2018recovery]. Another example is *matrix/tensor recovery via rank-1 projections* [@cai2015rop; @chen2015exact; @hao2018sparse], which arise by setting ${{\mathbfcal{X}}}_j = {{\mathbf{u}}}_j \circ {{\mathbf{v}}}_j \circ {{\mathbf{w}}}_j$, where ${{\mathbf{u}}}_j, {{\mathbf{v}}}_j, {{\mathbf{w}}}_j$ are random vectors and “$\circ$" represents the outer product, which includes phase retrieval [@cai2016optimal; @candes2015phase] as a special case. The very popular matrix/tensor completion example [@candes2010power; @liu2013tensor; @montanari2016spectral; @xia2017polynomial; @xia2017statistically; @yuan2014tensor] arises by setting ${{\mathbfcal{X}}}_j = \left({{\mathbf{e}}}_{a_j} \circ {{\mathbf{e}}}_{b_j} \circ {{\mathbf{e}}}_{c_j}\right)$, where ${{\mathbf{e}}}_j$ is the $j$-th canonical vector and $\{a_j, b_j, c_j\}_{j=1}^n$ are randomly selected integers from $\{1,\ldots, p_1\}\times \{1,\ldots, p_2\}\times \{1,\ldots, p_3\}$. Specific applications of this low-rank tensor regression model includes neuroimaging analysis [@guhaniyogi2015bayesian; @li2017parsimonious; @zhou2013tensor], longitudinal relational data analysis [@hoff2015multilinear], 3D imaging processing [@guo2012tensor], etc.
For convenience of presentation, we specialize the discussions on order-3 tensors later, while the results can be extended to the general order-$d$ tensors. In the modern high-dimensional setting, a variety of matrix/tensor data satisfy intrinsic structural assumptions, such as low-rankness [@udell2019big] or sparsity [@zhou2013tensor], which makes the accurate estimation of ${{\mathbfcal{A}}}$ possible even if the sample size $n$ is smaller than the number of coefficients in the target tensor ${{\mathbfcal{A}}}$. We thus focus on the low Tucker rank $(r_1, r_2, r_3)$ tensor ${{\mathbfcal{A}}}$ with the following Tucker decomposition [@tucker1966some]: $$\label{eq:tucker-decomposition}
{{\mathbfcal{A}}}= \llbracket {{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket := {{\mathbfcal{S}}}\times_1 {{\mathbf{U}}}_1 \times_2 {{\mathbf{U}}}_2 \times_3 {{\mathbf{U}}}_3,$$ where ${{\mathbfcal{S}}}$ is a $r_1$-by-$r_2$-by-$r_3$ core tensor and ${{\mathbf{U}}}_k$ is a $p_k$-by-$r_k$ matrix with orthonormal columns for $k=1,2,3$. The rigorous definition of Tucker rank of a tensor and more discussions on tensor algebra are postponed to Section \[sec:notations\]. In addition, the canonical polyadic (CP) low-rank tensors have also been widely considered in recent literature [@hao2018sparse; @haupt2017near; @sun2017store; @zhou2013tensor]. Since any CP-rank-$r$ tensor ${{\mathbfcal{A}}}= \sum_{i=1}^r \lambda_i {{\mathbf{a}}}_i \circ {{\mathbf{b}}}_i \circ {{\mathbf{c}}}_i$ has the Tucker decomposition: ${{\mathbfcal{A}}}= \llbracket {{\mathbfcal{L}}}; {{\mathbf{A}}}, {{\mathbf{B}}}, {{\mathbf{C}}}\rrbracket$, where ${{\mathbfcal{L}}}$ is the $r$-by-$r$-by-$r$ diagonal tensor with diagonal entries $\lambda_1,\ldots, \lambda_r$, ${{\mathbf{A}}}= [{{\mathbf{a}}}_1, \ldots, {{\mathbf{a}}}_r]$, and likewise for ${{\mathbf{B}}}, {{\mathbf{C}}}$ [@kolda2009tensor], our results naturally adapt to low CP-rank tensor regression. Also, with a slight abuse of notation, we will refer to low-rank and low Tucker rank interchangeably throughout the paper.
Moreover, we also consider a sparse setting where there may exist a subset of modes, say $J_s \subseteq \{1,2, 3\}$, such that ${{\mathbfcal{A}}}$ is sparse along these modes, i.e. $$\label{eq:tucker-decomposition-sparse}
{{\mathbfcal{A}}}= \llbracket{{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket, \quad \|{{\mathbf{U}}}_k\|_0 = \sum_{i=1}^{p_k} 1_{\{({{\mathbf{U}}}_k)_{[i, :]}\neq 0\}} \leq s_k,\quad k \in J_s.$$
Our Contributions
-----------------
We make the following major contributions to low-rank tensor regression in this article. Firstly, we introduce the main algorithm – *mportance ketching ow-rank stimation for ensors* (ISLET). Our algorithm has three steps: (i) first we use the tensor technique high-order orthogonal iteration (HOOI) [@de2000best] or sparse tensor alternating thresholding - singular value decomposition (STAT-SVD) [@zhang2017optimal-statsvd] to determine the importance sketching directions. Here, HOOI and STAT-SVD are regular and sparse tensor low-rank decomposition methods respectively, whose explanations are postponed to the forthcoming Sections \[sec:regular-islet-procedure\] and \[sec:sparse-procedure\]; (ii) using the sketching directions from the first step, we perform importance sketching, then evaluate the dimension-reduced regression using the sketched tensors/matrices (to incorporate sparsity, we add a group-sparsity regularizer); (iii) we construct the final tensor estimator using the sketched components. Although the focus of this work is on low-rank tensor regression, we point out that our three-step procedure applies to general high-dimensional statistics problems with low-dimensional structure, provided that we can find a suitable projection operator in step (i), and inverse projection operator in step (iii).
One of the main advantages of ISLET is the scalability of the algorithm. The proposed procedure is computationally efficient due to the dimension reduction by importance sketchings. Most importantly, ISLET only require access to the full data twice, which significantly saves run time for large-scale settings when it is not possible to store all samples into the core memory. We also show that our algorithm can be naturally distributed across multiple machines that can significantly reduce computation time.
Secondly, we prove a deterministic oracle inequality for the ISLET procedure under the low-Tucker-rank assumption and general noise and design (Theorems \[th:upper\_bound\_general\] and \[th:upper\_bound\_sparse\_general\]). We additionally show that ISLET achieves the optimal mean-squared error (with the optimal constant for non-sparse ISLET) under randomized Gaussian design (Theorems \[th:upper\_bound\_regression\], \[th:lower-bound-regression\], \[th:upper\_bound\_sparse\_tensor\_regression\], and \[th:lower\_bound\_sparse\_tensor\_regression\]). The following informal statement summarizes two of the main results of the article:
\[ThmInformal\] Consider the regular tensor regression problem with Gaussian ensemble design, where ${{\mathbfcal{A}}}$ is Tucker rank-$(r_1,r_2,r_3)$, ${{\mathbfcal{X}}}_j$ has i.i.d. standard normal entries, $\varepsilon_j\overset{iid}{\sim}N(0, \sigma^2)$, and $\varepsilon_j, {{\mathbfcal{X}}}_j$ are independent.
- Under regularity conditions, ISLET achieves the following optimal rate of convergence with the matching constant, $$\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 = \left(1 + o(1)\right) \frac{m\sigma^2}{n},$$ where $m = r_1r_2r_3 + r_1(p_1-r_1) + r_2(p_2-r_2) + r_3(p_3-r_3)$ is exactly the degree of freedom of all Tucker rank-$(r_1, r_2, r_3)$ tensors in $\mathbb{R}^{p_1\times p_2\times p_3}$ and $\left\| \cdot \right\|_{{{\rm HS}}}$ is the Hilbert-Schmidt norm to be defined in Section \[sec:notations\].
- If in addition, holds with sparsity level $s_k$, then under regularity conditions, ISLET achieves the following optimal rate of convergence, $$\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \asymp \frac{m_s\sigma^2}{n},$$ where $m_s = r_1r_2r_3 + \sum_{k\in J_s} s_k \left(r_k + \log(p_k/s_k)\right) + \sum_{k\notin J_s} p_kr_k$ and “$\asymp$" denotes the asymptotic equivalence between two number series (see a more formal definition in Section \[sec:notations\]).
To the best of our knowledge, we are the first to develop the matching-constant optimal rate results for regular tensor regression under randomized Gaussian ensemble design, even for the low-rank matrix recovery case since it is not clear whether prior approaches (e.g. nuclear norm minimization) achieve sharp constants. We are also the first to develop the optimal rate results for tensor regression with sparsity condition (\[eq:tucker-decomposition-sparse\]).
Thirdly, proving the optimal mean-squared error bound presents a number of technical challenges and we introduce novel proof ideas to overcome these difficulties. In particular, one major difficulty lies in the analysis of reduced-dimensional regressions (see in the forthcoming Section \[sec:procedure\]) since we analyze sketched regression models. To this end, we introduce partial linear models for these reduced-dimensional regressions from which we develop estimation error upper bounds.
The final and most important computational contribution is to display through numerical studies the advantages of our ISLET algorithms. Compared to state-of-the-art tensor estimation algorithms including non-convex projected gradient descent (PGD) [@chen2016non], Tucker regression [@zhou2013tensor], and convex regularization [@tomioka2013convex], we show that our ISLET algorithm achieves comparable statistical performance with substantially faster computation. In particular, the runtime is 1-3 orders of magnitude faster than existing methods. In the most prominent example, our ISLET procedure can efficiently solve the ultrahigh-dimensional tensor regression with covariates of 7.68 terabytes. For the order-2 case, i.e., low-rank matrix regression, our simulation studies show that ISLET outperforms the classic nuclear norm minimization estimator. We also provide a real data application where we study the association between the attention-deficit/hyperactivity disorder disease and the high-dimensional MRI image tensors. We show that the proposed procedure provides significantly better prediction performance in much less time compared to state-of-the-art methods.
Related Literature {#sec:literature-review}
------------------
Our work is related to a broad range of literature varying from a number of communities including scientific computing, computer science, signal processing, applied mathematics, and statistics. Here we make an attempt to discuss existing results from these various communities however we do not claim that our literature survey is exhaustive.
Large-scale linear systems where the solution admits a low-rank tensor structure commonly arise after discretizing high-dimensional partial differential equations [@hofreither2018black; @hughes2005isogeometric; @lynch1964tensor] and various methods have been proposed. For example, [@bousse2017linear] developed algebraic and Gauss-Newton methods to solve the linear system with a CP low-rank tensor solution. [@ballani2013projection; @beylkin2005algorithms] proposed iterative projection methods to solve large-scale linear systems with Kronecker-product-type design matrices. [@georgieva2019greedy] introduced a greedy approach. [@kressner2016preconditioned; @kressner2010krylov] considered Riemannian optimization methods and tensor Krylov subspace methods, respectively. The readers are referred to [@grasedyck2013literature] for a recent survey. Different from these works, our proposed ISLET is a one-step procedure that only involves solving a simple least squares regression after performing dimension reduction on covariates by importance sketching (see Steps 1 and 2 in Section \[sec:regular-islet-procedure\]). Moreover, many prior works mainly focused on computational aspects of their proposed methods [@ballani2013projection; @bousse2018linear; @espig2012variational; @georgieva2019greedy; @grasedyck2013literature], while we show that ISLET is not only computationally efficient (see more discussion and comparison on computation complexity in Section \[sec:regular-islet-procedure\] Computation and Implementation part) but also has optimal theoretical guarantees in terms of mean square error under the statistical setting.
In addition, sketching methods play an important role in computation acceleration and has been widely considered in previous literature. For example, [@clarkson2017low; @meng2013low; @nelson2013osnap] provided accurate approximation algorithms based on sketching with novel embedding matrices, where the runtime is proportional to the number of the non-zero entries of the input matrix. Sketching methods have also been studied in robust $\ell_1$ low-rank matrix approximation [@markopoulos2014optimal; @markopoulos2017efficient; @meng2013cyclic; @song2017low; @zheng2012practical], general $\ell_p$ low-rank matrix approximation [@ban2019ptas; @chierichetti2017algorithms], low-rank tensor approximation [@song2019relative], etc. In the regression context, the sketching method has been considered for the least squares regression [@clarkson2017low; @diao2018sketching; @nelson2013osnap; @pilanci2016iterative; @raskutti2014statistical], $\ell_p$ regression [@clarkson2017low; @meng2013low; @nelson2013osnap], Kronecker product regression [@diao2018sketching], ridge regression [@avron2016sharper; @wang2017sketching], regularized kernel regression [@camoriano2016nytro; @zhang2014random], etc. Various types of random sketching matrices have been developed, including random sub-Gaussian [@pilanci2016iterative], random sampling [@drineas2012fast; @drineas2010effective], CountSketch [@charikar2002finding; @clarkson2015input], Sparse Johnson-Lindenstrauss transformation [@kane2014sparser], among many others. The readers are also referred to survey papers on sketching by Mahoney [@mahoney2011randomized] and Woodruff [@woodruff2014sketching]. The proposed method in this paper is different from these previous works in various aspects. First, many randomized sketching methods in the literature focused on relative approximation error [@mahoney2011randomized; @woodruff2014sketching] and the sketching matrices are constructed only based on covariates [@drineas2012fast; @drineas2010effective; @kane2014sparser; @pilanci2016iterative; @raskutti2014statistical]. In contrast, we explicitly construct “supervised" sketching matrices based on both the response $y_j$ and covariates ${{\mathbfcal{X}}}_j$ and obtain optimal bounds in mean square error under the statistical setting. Second, essentially speaking, our proposed importance sketching scheme reduces the number of columns (parameters) instead of the number of rows (samples) in the linear equation system. Third, different from the sketching on an overdetermined system of least squares [@clarkson2017low; @diao2018sketching; @nelson2013osnap; @pilanci2016iterative; @raskutti2014statistical], we mainly focus on the high-dimensional setting where the number of samples can be significantly smaller than the number of coefficients.
Organization
------------
In Section \[sec:notations\] we introduce important notation; then we present our ISLET procedure under non-sparse and sparse settings in Sections \[sec:regular-islet-procedure\] and \[sec:sparse-procedure\], respectively and illustrate the procedure from a sketching perspective in Section \[sec:sketching-perspective\]; in Section \[sec:oracle-inequality\] we provide general theoretical guarantees for our procedure which make no assumptions on the design or the noise distribution; in Section \[th:upper\_bound\_regression\] we specialize our bounds to tensor regression with low Tucker rank and assume the design is independent Gaussian; a simulation study showing the substantial computational benefits of our algorithm are provided in Section \[sec:numerical\]. Additional notation, discussion on general-order ISLET, simulation results, an application to attention deficit hyperactivity disorder (ADHD) MRI Imaging data analysis, and all technical proofs are provided in the supplementary materials [@ISLET-supplement].
Our Procedure: ISLET {#sec:procedure}
====================
In this section, we introduce the general procedure of importance sketching low-rank estimation for tensors (ISLET). Although for ease of presentation we will focus on order-3 tensors, the procedure for the general order-$d$ case can also be treated. Details of matrices and tensors greater than order 3 are provided in Section \[sec:general-order\] of the supplementary material [@ISLET-supplement].
Notation and Preliminaries {#sec:notations}
--------------------------
The following notation will be used throughout this article. Additional definitions can be found in Section \[sec:row-permutation-operator\] in the supplementary materials. Lowercase letters (e.g., $a, b$), lowercase boldface letters (e.g. ${{\mathbf{u}}}, {{\mathbf{v}}}$), uppercase boldface letters (e.g., ${{\mathbf{U}}}, {{\mathbf{V}}}$), and boldface calligraphic letters (e.g., ${{\mathbfcal{A}}}, {{\mathbfcal{X}}}$) are used to denote scalars, vectors, matrices, and order-3-or-higher tensors respectively. For simplicity, we denote ${{\mathbfcal{X}}}_j$ as the tensor indexed by $j$ in a sequence of tensors $\{{{\mathbfcal{X}}}_j\}$. For any two series of numbers, say $\{a_i\}$ and $\{b_i\}$, denote $a \asymp b$ if there exist uniform constants $c, C>0$ such that $ca_i \leq b_i\leq Ca_i, \forall i$ and $a = \Omega (b)$ if there exists uniform constant $c > 0$ such that $a_i \geq c b_i, \forall i$. We use bracket subscripts to denote sub-vectors, sub-matrices, and sub-tensors. For example, ${{\mathbf{v}}}_{[2:r]}$ is the vector with the $2$nd to $r$th entries of ${{\mathbf{v}}}$; ${{\mathbf{D}}}_{[i_1,i_2]}$ is the entry of ${{\mathbf{D}}}$ on the $i_1$-th row and $i_2$-th column; ${{\mathbf{D}}}_{[(r+1):p_1, :]}$ contains the $(r+1)$-th to the $p_1$-th rows of ${{\mathbf{D}}}$; ${{\mathbfcal{A}}}_{[1:s_1, 1:s_2, 1:s_3]}$ is the $s_1$-by-$s_2$-by-$s_3$ sub-tensor of ${{\mathbfcal{A}}}$ with index set $\{(i_1, i_2, i_3): 1 \leq i_1\leq s_1, 1\leq i_2 \leq s_2, 1\leq i_3\leq s_3\}$. For any vector ${{\mathbf{v}}}\in \mathbb{R}^{p_1}$, define its $\ell_q$ norm as $\|{{\mathbf{v}}}\|_q = \left(\sum_i |v_i|^q\right)^{1/q}$. For any matrix ${{\mathbf{D}}}\in \mathbb{R}^{p_1\times p_2}$, let $\sigma_k({{\mathbf{D}}})$ be the $k$-th singular value of ${{\mathbf{D}}}$. In particular, the least non-trivial singular value of ${{\mathbf{D}}}$, defined as $\sigma_{\min}({{\mathbf{D}}}) = \sigma_{p_1\wedge p_2}({{\mathbf{D}}})$, will be extensively used in later analysis. We also denote ${{\rm SVD}}_r({{\mathbf{D}}}) = [{{\mathbf{u}}}_1 ~ \cdots {{\mathbf{u}}}_r]$ and QR(${{\mathbf{D}}}$) as the subspace composed of the leading $r$ left singular vectors and the Q part of the QR orthogonalization of ${{\mathbf{D}}}$, respectively. The matrix Frobenius and spectral norms are defined as $ \|{{\mathbf{D}}}\|_F = \left(\sum_{i_1,i_2} {{\mathbf{D}}}_{[i_1,i_2]}^2\right)^{1/2} = (\sum_{i=1}^{p_1\wedge p_2}\sigma_i^2({{\mathbf{D}}}))^{1/2} \quad \text{and} \quad\|{{\mathbf{D}}}\| = \max_{{{\mathbf{u}}}\in \mathbb{R}^{p_2}}\|{{\mathbf{D}}}{{\mathbf{u}}}\|_2/\|{{\mathbf{u}}}\|_2 = \sigma_1({{\mathbf{D}}}).$ In addition, ${{\mathbf{I}}}_r$ represents the $r$-by-$r$ identity matrix. Let $\mathbb{O}_{p, r} = \{{{\mathbf{U}}}: {{\mathbf{U}}}^\top {{\mathbf{U}}}={{\mathbf{I}}}_r\}$ be the set of all $p$-by-$r$ matrices with orthonormal columns. For any ${{\mathbf{U}}}\in \mathbb{O}_{p, r}$, $P_{{{\mathbf{U}}}} = {{\mathbf{U}}}{{\mathbf{U}}}^\top$ represents the projection matrix onto the column space of ${{\mathbf{U}}}$; we also use ${{\mathbf{U}}}_\perp\in \mathbb{O}_{p, p-r}$ to represent the orthonormal complement of ${{\mathbf{U}}}$. For any event $A$, let ${\mathbb{P}}(A)$ be the probability that $A$ occurs.
For any matrix ${{\mathbf{D}}}\in \mathbb{R}^{p_1\times p_2}$ and order-$d$ tensor ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times \cdots\times p_d}$, let ${{\rm vec}}({{\mathbf{D}}})$ and ${{\rm vec}}({{\mathbfcal{A}}})$ be the vectorization of ${{\mathbf{D}}}$ and ${{\mathbfcal{A}}}$, respectively. The matricization $\mathcal{M}(\cdot)$ is the operation that unfolds or flattens the order-$d$ tensor ${{\mathbfcal{A}}}\in\mathbb{R}^{p_1\times \cdots \times p_d}$ into the matrix $\mathcal{M}_k({{\mathbfcal{A}}})\in \mathbb{R}^{p_k\times \prod_{j\neq k}p_j}$ for $k=1,\ldots, d$. Since the formal entry-wise definitions of matricization and vectorization is rather tedious, we leave them to Section \[sec:row-permutation-operator\] in the supplementary materials [@ISLET-supplement]. The Hilbert-Schmidt norm is defined as $\|{{\mathbfcal{A}}}\|_{{{\rm HS}}} = \left(\sum_{i_1,\ldots, i_d} {{\mathbfcal{A}}}_{[i_1,\ldots, i_d]}^2\right)^{1/2}.$ An order-$d$ tensor is rank-one if it can be written as the outer product of $d$ nonzero vectors. The CP-rank of any tensor ${{\mathbfcal{A}}}$ is defined as the minimal number $r$ such that ${{\mathbfcal{A}}}$ can be decomposed as ${{\mathbfcal{A}}}= \sum_{i=1}^r {{\mathbfcal{B}}}_i$ for rank-1 tensors ${{\mathbfcal{B}}}_i$. The Tucker rank (or multilinear rank) of a tensor ${{\mathbfcal{A}}}$ is defined as a $d$-tuple $(r_1, \ldots, r_d)$, where $r_k = \text{rank}(\mathcal{M}_k({{\mathbfcal{A}}}))$. The $k$-mode product of ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1 \times \ldots \times p_d}$ with a matrix ${{\mathbf{U}}}\in \mathbb{R}^{p_k\times r_k}$ is denoted by ${{\mathbfcal{A}}}\times_k {{\mathbf{U}}}$ and is of size $p_1 \times \cdots \times p_{k-1}\times r_k \times p_{k+1}\times \cdots \times p_d$, such that $$({{\mathbfcal{A}}}\times_k {{\mathbf{U}}})_{[i_1, \ldots, i_{k-1}, j, i_{k+1}, \ldots, i_d]} = \sum_{i_k=1}^{p_k} {{\mathbfcal{A}}}_{[i_1, i_2, \ldots, i_d]} {{\mathbf{U}}}_{[i_k, j]}.$$ For convenience of presentation, all mode indices $(\cdot)_k$ of an order-3 tensor are in the sense of modulo-3, e.g., $r_1=r_4$, $s_2 = s_5$, $p_0=p_3$, ${{\mathbfcal{X}}}\times_4 {{\mathbf{U}}}_4 = {{\mathbfcal{X}}}\times_1 {{\mathbf{U}}}_1$.
For any matrices ${{\mathbf{U}}}\in \mathbb{R}^{p_1\times p_2}$ and ${{\mathbf{V}}}\in \mathbb{R}^{m_1\times m_2}$, let $${{\mathbf{U}}}\otimes {{\mathbf{V}}}= \begin{bmatrix}
{{\mathbf{U}}}_{[1,1]}\cdot {{\mathbf{V}}}& \cdots & {{\mathbf{U}}}_{[1, p_2]}\cdot {{\mathbf{V}}}\\
\vdots & & \vdots\\
{{\mathbf{U}}}_{[p_1,1]}\cdot {{\mathbf{V}}}& \cdots & {{\mathbf{U}}}_{[p_1, p_2]}\cdot {{\mathbf{V}}}\\
\end{bmatrix}\in \mathbb{R}^{(p_1m_1)\times(p_2m_2)}$$ be the Kronecker product. Some intrinsic identities among Kronecker product, vectorization, and matricization, which will be used later in this paper, are summarized in Lemma \[lm:Kronecker-vectorization-matricization\] in the supplementary materials [@ISLET-supplement]. The readers can refer to [@kolda2009tensor] for a more comprehensive introduction to tensor algebra. Finally, we use $C, C_1, C_2, c$ and other variations to represent the large and small constants, whose actual value may vary from line to line.
Regular Low-rank Tensor Recovery {#sec:regular-islet-procedure}
--------------------------------
We first consider the tensor regression model , where ${{\mathbfcal{A}}}$ is low-rank without sparsity assumptions. The proposed algorithm of ISLET is divided into three steps and a pictorial illustration is provided in Figures \[fig:illu-1\] - \[fig:illu-3\] for readers’ better understanding. The pseudo-code is provided in Algorithm \[al:procedure\_regular\].
1. (Probing importance sketching directions) We first probe the importance sketching directions. When the covariates satisfy $\mathbb{E}{{\rm vec}}({{\mathbfcal{X}}}_j) {{\rm vec}}({{\mathbfcal{X}}}_j)^\top = {{\mathbf{I}}}_{p_1p_2p_3}$, we evaluate $$\label{eq:tilde-A}
\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j{{\mathbfcal{X}}}_j.$$ $\widetilde{{\mathbfcal{A}}}$ is essentially the covariance tensor between $y$ and ${{\mathbfcal{X}}}$. Since ${{\mathbfcal{A}}}=\llbracket {{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket$ has low Tucker rank, we perform the high-order orthogonal iterations (HOOI) on $\widetilde{{{\mathbfcal{A}}}}$ to obtain $\widetilde{{{\mathbf{U}}}}_k \in \mathbb{O}_{p_k, r_k}, k=1,2,3$ as initial estimates for ${{\mathbf{U}}}_k$. Here, HOOI is a classic method for tensor decomposition that can be traced back to Lathauwer, Moor, and Vandewalle [@de2000best]. The central idea of HOOI is the power iterated singular value thresholding. Then, the outcome of HOOI $\{\widetilde{{{\mathbf{U}}}}_k\}_{k=1}^3$ yield the following low-rank approximation for ${{\mathbfcal{A}}}$, $$\label{eq:bcA-approximation}
{{\mathbfcal{A}}}\approx \llbracket \widetilde{{{\mathbfcal{S}}}}; \widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3 \rrbracket,\quad \text{where}\quad \widetilde{{{\mathbfcal{S}}}} = \llbracket\widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket \in \mathbb{R}^{r_1\times r_2\times r_3}.$$ We further evaluate $$\begin{split}
\widetilde{{{\mathbf{V}}}}_k := {\rm QR}\left(\mathcal{M}_k^\top(\widetilde{{{\mathbfcal{S}}}})\right) \in \mathbb{O}_{r_{k+1}r_{k+2}, r_k}, \quad k=1,2,3.
\end{split}$$ $\{\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k\}_{k=1}^3$ obtained here are regarded as the *importance sketching directions*. As we will further illustrate in Section \[sec:oracle-regular\], the combinations of $\widetilde{{{\mathbf{U}}}}_k$ and $\widetilde{{{\mathbf{V}}}}_k$ provide approximations for singular subspaces of $\mathcal{M}_k({{\mathbfcal{A}}})$.
2. (Linear regression on sketched covariates) Next, we perform sketching to reduce the dimension of the original regression model . To be specific, we project the original high-dimensional covariates onto the dimension-reduced subspace “that is important in the covariance between $y$ and ${{\mathbfcal{X}}}$" and construct the following *importance sketching covariates*, $$\label{eq:importance-sketching-covariates}
\begin{split}
& \widetilde{{{\mathbf{X}}}} = \left[\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_3}\right] \in \mathbb{R}^{n\times m},\\
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times m_{{{\mathbfcal{B}}}}},\quad \left(\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\right)_{[i,:]} =
{{\rm vec}}\left({{\mathbfcal{X}}}_i \times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k} \in \mathbb{R}^{n\times m_{{{\mathbf{D}}}_k}}, \quad \left(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k}\right)_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{k\perp}^\top \mathcal{M}_k\left({{\mathbfcal{X}}}_{i} \times_{k+1} \widetilde{{{\mathbf{U}}}}_{k+1}^\top \times_{k+2} \widetilde{{{\mathbf{U}}}}_{k+2}^\top\right) \widetilde{{{\mathbf{V}}}}_k\right),
\end{split}$$ where $m_{{{\mathbfcal{B}}}}=r_1r_2r_3$, $m_{{{\mathbf{D}}}_k} = (p_k-r_k)r_k$, $k=1,2,3$, and $m = m_{{{\mathbfcal{B}}}} + m_{{{\mathbf{D}}}_1} + m_{{{\mathbf{D}}}_2} + m_{{{\mathbf{D}}}_3}$. Then, we evaluate the least-squares estimator of the sub-model with importance sketching covariates $\widetilde{{{\mathbf{X}}}}$, $$\label{eq:partial-regression}
\widehat{{{\boldsymbol{\gamma}}}} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^m}\left\|y - \widetilde{{{\mathbf{X}}}}{{\boldsymbol{\gamma}}}\right\|_2^2.$$ The dimension of sketching covariate regression is $m$, which is significantly smaller than the dimension of the original tensor regression model, $p_1p_2p_3$. Consequently, the computational cost can be significantly reduced.
3. (Assembling the final estimate) Then, $\widehat{{{\boldsymbol{\gamma}}}}$ is divided into four segments according to the block-wise structure of $\widetilde{{{\mathbf{X}}}} = [\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}, \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1}, \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2}, \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_3}]$, $$\label{eq:def_hat_Beta_hat_D}
\begin{split}
& {{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) = \widehat{{{\boldsymbol{\gamma}}}}_{[1:m_{{{\mathbfcal{B}}}}]},\\
& {{\rm vec}}(\widehat{{{\mathbf{D}}}}_1) = \widehat{{{\boldsymbol{\gamma}}}}_{[(m_{{{\mathbfcal{B}}}}+1): (m_{{{\mathbfcal{B}}}}+m_{{{\mathbf{D}}}_1})]},\\
& {{\rm vec}}(\widehat{{{\mathbf{D}}}}_2) = \widehat{{{\boldsymbol{\gamma}}}}_{[(m_{{{\mathbfcal{B}}}}+m_{{{\mathbf{D}}}_1}+1): (m_{{{\mathbfcal{B}}}}+m_{{{\mathbf{D}}}_1}+m_{{{\mathbf{D}}}_2})]},\\
& {{\rm vec}}(\widehat{{{\mathbf{D}}}}_3) = \widehat{{{\boldsymbol{\gamma}}}}_{[(m_{{{\mathbfcal{B}}}}+m_{{{\mathbf{D}}}_1}+m_{{{\mathbf{D}}}_2}+1):(m_{{{\mathbfcal{B}}}}+m_{{{\mathbf{D}}}_1}+m_{{{\mathbf{D}}}_2}+m_{{{\mathbf{D}}}_3})]}.
\end{split}$$ Finally, we construct the regression estimator $\widehat{{{\mathbfcal{A}}}}$ for the original problem using the regression estimator $\widehat{{{\boldsymbol{\gamma}}}}$ for the sub-model : let $\widehat{{{\mathbf{B}}}}_k = \mathcal{M}_k(\widehat{{{\mathbfcal{B}}}})$ and calculate $$\label{eq:hat_A_non-sparse}
\begin{split}
& \widehat{{{\mathbf{L}}}}_k = \left(\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k\right)\left(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\right)^{-1},\quad k=1,2,3, \quad \widehat{{{\mathbfcal{A}}}} = \left\llbracket\widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{L}}}}_1, \widehat{{{\mathbf{L}}}}_2, \widehat{{{\mathbf{L}}}}_3 \right\rrbracket.
\end{split}$$ More interpretation of is given in Section \[sec:oracle-regular\].
When $\mathbb{E}{{\rm vec}}({{\mathbfcal{X}}}){{\rm vec}}({{\mathbfcal{X}}})^\top \neq {{\mathbf{I}}}_{p_1p_2p_3}$, we could consider the following alternative ways to construct the initial estimate $\widetilde{{{\mathbfcal{A}}}}$. Firstly, in some cases we could do construction depending on the covariance structure of ${{\mathbfcal{X}}}$. For example, in the framework of tensor recovery via rank-one sketching (discussed in the introduction), we have ${{\mathbfcal{X}}}_j = {{\mathbf{u}}}_j\circ {{\mathbf{u}}}_j\circ {{\mathbf{u}}}_j$ and ${{\mathbf{u}}}_j \in \mathbb{R}^p$ has iid entry $N(0,1)$. By the high-order Stein’s identity [@janzamin2014score], one can show that $$\widetilde{{{\mathbfcal{A}}}} = \frac{1}{6}\left[\frac{1}{n}\sum_{j=1}^n y_j {{\mathbf{u}}}_j \circ {{\mathbf{u}}}_j \circ {{\mathbf{u}}}_j - \sum_{j=1}^p \left({{\mathbf{w}}}\circ {{\mathbf{e}}}_j \circ {{\mathbf{e}}}_j + {{\mathbf{e}}}_j \circ {{\mathbf{w}}}\circ {{\mathbf{e}}}_j + {{\mathbf{e}}}_j \circ {{\mathbf{e}}}_j \circ {{\mathbf{w}}}\right)\right],$$ is a proper initial unbiased estimator for ${{\mathbfcal{A}}}$ [@hao2018sparse Lemma 4]. Here, ${{\mathbf{w}}}= \frac{1}{n}\sum_{i=1}^n y_j {{\mathbf{u}}}_j$, ${{\mathbf{e}}}_j$ is the $j$th canonical basis in $\mathbb{R}^p$. Another commonly used setting in data analysis is the high-order Kronecker covariance structure: $\mathbb{E}({{\rm vec}}({{\mathbfcal{X}}}_j){{\rm vec}}({{\mathbfcal{X}}}_j)^\top) = {\boldsymbol{\Sigma}}_3\otimes {\boldsymbol{\Sigma}}_2\otimes {\boldsymbol{\Sigma}}_1$, where ${\boldsymbol{\Sigma}}_k \in \mathbb{R}^{p_k\times p_k}, k=1,2,3$ are covariance matrices along three modes, respectively [@he2014graphical; @lyu2019tensor; @manceur2013maximum; @pan2018covariate; @zhou2014gemini]. Under this assumption, we can first apply existing approaches to obtain estimators $\widehat{{\boldsymbol{\Sigma}}}_k$ for ${\boldsymbol{\Sigma}}_k$, then whiten the covariates by replacing ${{\mathbfcal{X}}}_j$ by $\llbracket {{\mathbfcal{X}}}_j; \widehat{{\boldsymbol{\Sigma}}}_1^{-1/2}, \widehat{{\boldsymbol{\Sigma}}}_2^{-1/2}, \widehat{{\boldsymbol{\Sigma}}}_3^{-1/2} \rrbracket$. After this pre-processing step, the other steps of ISLET still follow. Moreover, it still remains an open question how to perform initialization if ${{\mathbfcal{X}}}$ has the more general, unstructured, and unknown design.
In addition to high-order orthogonal iteration (HOOI), there are a variety of methods proposed in the literature to compute the low-rank tensor approximation, such as Newton-type optimization methods on manifolds [@elden2009newton; @ishteva2011best; @ishteva2009differential; @savas2010quasi], black box approximation [@bebendorf2011adaptive; @caiafa2010generalizing; @mahoney2008tensor; @oseledets2008tucker; @oseledets2010cross; @zhang2019cross], generalizations of Krylov subspace method [@goreinov2012wedderburn; @savas2013krylov], greedy approximation method [@georgieva2019greedy], among many others. Further, black box approximation methods [@bebendorf2011adaptive; @caiafa2010generalizing; @oseledets2008tucker; @oseledets2010cross; @zhang2019cross] can be applied even if the initial estimator $\widetilde{{{\mathbfcal{A}}}}$ does not fit into the core memory. When the tensor is further approximately CP low-rank, we can also apply the randomized compressing method [@sidiropoulos2012multi; @sidiropoulos2014parallel] or randomized block sampling [@vervliet2015randomized] to obtain the CP low-rank tensor approximation. Although the rest of our discussion will focus on the HOOI procedure for initialization, these alternative methods can also be applied to obtain an initialization for the ISLET algorithm.
\
.5cm
![Illustration for Step 3 of ISLET[]{data-label="fig:illu-3"}](illu-assemble2.png){width=".8\linewidth"}
[**Computation and implementation.**]{} We briefly discuss computational complexity and implementation aspects for the ISLET procedure here. It is noteworthy that ISLET accesses the sample only twice for constructing the covariance tensor (Step 1) and importance sketching covariates (Step 2), respectively. In large scale cases where it is difficult to store the whole dataset into random-access memory (RAM), this advantage can highly save the computational costs.
In addition, in the order-3 tensor case, when each mode shares the same dimension $p_k=p$ and rank $r_k=r$, the total number of observable values is $O(np^3)$ and the time complexity of ISLET is $O\left(np^3r + nr^6+Tp^4\right)$ where $T$ is the number of HOOI iterations. For general order-$d$ tensor regression, time complexity of ISLET is $O\left(np^dr + nr^{2d}+Tp^{d+1}\right)$. In contrast, the time complexity of the non-convex PGD [@chen2016non] is $O\left(T'(np^d+rp^{d+1})\right)$, where $T'$ is the number of iterations of gradient descent; [@bousse2018linear] introduced an optimization based method with time complexity $O(T'dnp^d r)$ where $T'$ is the number of iterations in Gauss-Newton method. We can see if $T'\geq r$, a typical situation in practice, ISLET is significantly faster than these previous methods.
It is worth pointing out that the computing time of ISLET is still high when the tensor parameter has a large order $d$. In fact, without any structural assumption on the design tensors ${{\mathbfcal{X}}}_j$, such a time cost may be unavoidable since reading in all data requires $O(np^d)$ operations. If there is extra structure on the design tensor, e.g., Kronecker product [@ballani2013projection; @hofreither2018black; @hughes2005isogeometric; @lynch1964tensor] and low separation rank [@beylkin2005algorithms; @georgieva2019greedy], the computing time can be significantly reduced by applying methods in this body of literature. Here, we mainly focus on the setting where ${{\mathbfcal{X}}}_j$ does not satisfy a clear structural assumption since in many real data applications, e.g., the neuroimaging data example studied in this and many other works [@allen2012regularized; @li2013tucker; @sun2017store; @zhou2013tensor], the design tensors ${{\mathbfcal{X}}}_j$ may not have a clear known structure.
Moreover, in the order-3 tensor case, instead of storing all $\{ {{\mathbfcal{X}}}_j \}_{j=1}^n$ in the memory which requires $O(np^3)$ RAM, ISLET only requires $O(p^3 + n(pr+r^3))$ RAM space if one chooses to access the samples from hard disks but not to store to RAM. This makes large-scale computing possible. We empirically investigate the computation cost by simulation studies in Section \[sec:numerical\].
The proposed ISLET procedure also allows convenient parallel computing. Suppose we distribute all $n$ samples across $B$ machines: $\{({{\mathbfcal{X}}}_{bi}, y_{bi})\}_{i=1}^{B_b}$, $b=1,\ldots, B$, where $B_b\approx n/B$. To evaluate the covariance tensor in Step 1, we can calculate $\widetilde{{{\mathbfcal{A}}}}_b = \sum_{i=1}^{B_i} y_{bi} {{\mathbfcal{X}}}_{bi}$ in each machine, then summarize them as $\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{b=1}^B\widetilde{{{\mathbfcal{A}}}}_b$; to construct sketching covariates and perform partial regression in Step 2, we calculate $$\label{eq:parallel-1}
{{\mathbf{y}}}_b = (y_{b1},\ldots, y_{bB_b})^\top \in \mathbb{R}^{B_b},$$ $$\label{eq:parallel-2}
\begin{split}
& \widetilde{{{\mathbf{X}}}}_{bi} = \left[\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}, bi} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1, bi} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2, bi} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_3, bi}\right] \in \mathbb{R}^{m},\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}, bi} =
{{\rm vec}}\left({{\mathbfcal{X}}}_{bi} \times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k,bi} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{k\perp}^\top \mathcal{M}_k\left({{\mathbfcal{X}}}_{bi} \times_{k+1} \widetilde{{{\mathbf{U}}}}_{k+1}^\top \times_{k+2} \widetilde{{{\mathbf{U}}}}_{k+2}^\top\right) \widetilde{{{\mathbf{V}}}}_k\right),
\end{split}$$ $$\label{eq:parallel-3}
\widetilde{{{\mathbf{G}}}}_b = \sum_{i=1}^{B_b} \widetilde{{{\mathbf{X}}}}_{bi}^\top\widetilde{{{\mathbf{X}}}}_{bi} ,\quad \widetilde{{{\mathbf{z}}}}_b = \sum_{i=1}^{B_b}\widetilde{{{\mathbf{X}}}}_{bi}^\top y_{bi}$$ in each machine. Then we combine the outcomes to $$\widehat{{{\boldsymbol{\gamma}}}} = \left(\sum_{b=1}^B \widetilde{{{\mathbf{G}}}}_b\right)^{-1} \left(\sum_{b=1}^B \widetilde{{{\mathbf{z}}}}_b\right).$$ The computational complexity can be reduced to $O\left(\frac{np^3r+nr^6}{B} + Tp^4\right)$ via the parallel scheme. In the large-scale simulation we present in this article, we implement this parallel scheme for speed-up.
To implement the proposed procedure, the input of Tucker rank are required as tuning parameters. When they are unknown in practice, we can perform cross-validation or an adaptive rank selection scheme. A more detailed description and numerical results are postponed to Section \[sec:tuning\] in the supplementary materials [@ISLET-supplement].
Sparse Low-rank Tensor Recovery {#sec:sparse-procedure}
-------------------------------
When the target tensor ${{\mathbfcal{A}}}$ is simultaneously low-rank and sparse, in the sense that holds for a subset $J_s\subseteq \{1,2,3\}$ known a priori, we introduce the following sparse ISLET procedure. The pseudo-code for sparse ISLET is summarized in Algorithm \[al:procedure\_sparse\].
1. (Probing sketching directions) When $\mathbb{E} {{\rm vec}}({{\mathbfcal{X}}}){{\rm vec}}({{\mathbfcal{X}}})^\top = {{\mathbf{I}}}_{p_1p_2p_3}$, we still evaluate the covariance tensor $\widetilde{{{\mathbfcal{A}}}}$ as Equation . Noting that ${{\mathbfcal{A}}}= \llbracket {{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket$ and $\{{{\mathbf{U}}}_k\}_{k\in J_s}$ are row-wise sparse, we apply the sparse tensor alternating thresholding SVD (STAT-SVD) [@zhang2017optimal-statsvd] on $\widetilde{{{\mathbfcal{A}}}}$ to obtain $\widetilde{{{\mathbf{U}}}}_k \in \mathbb{O}_{p_k, r_k}, k=1,2,3$ as initial estimates for ${{\mathbf{U}}}_k$. Here, STAT-SVD is a sparse tensor decomposition method proposed by [@zhang2017optimal-statsvd] with central ideas of the double projection & thresholding scheme and power iteration. Via STAT-SVD, we obtain the following sparse and low-rank approximation of ${{\mathbfcal{A}}}$, $${{\mathbfcal{A}}}\approx \llbracket\widetilde{{{\mathbfcal{S}}}}; \widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3\rrbracket, \quad \widetilde{{{\mathbf{U}}}}_k \in \mathbb{O}_{p_k, r_k}, \quad \widetilde{{{\mathbfcal{S}}}} = \llbracket \widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket \in \mathbb{R}^{r_1\times r_2\times r_3}.$$ We further evaluate $$\begin{split}
\widetilde{{{\mathbf{V}}}}_k = {\rm QR}\left(\mathcal{M}_k^\top(\widetilde{{{\mathbfcal{S}}}})\right) \in \mathbb{O}_{r_{k+1}r_{k+2}, r_k}.
\end{split}$$
2. (Group Lasso on sketched covariates) We perform sketching and construct the following importance sketching covariates based on $\{\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k\}_{k=1}^3$, $$\label{eq:tilde-X_B-tilde-X_E}
\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times (r_1r_2r_3)},\quad (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[i,:]} =
{{\rm vec}}\left({{\mathbfcal{X}}}_i \times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k} \in \mathbb{R}^{n\times p_kr_k}, \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})_{[i,:]} = {{\rm vec}}\left(\mathcal{M}_k\left({{\mathbfcal{X}}}_i\times_{k+1} \widetilde{{{\mathbf{U}}}}_{k+1}^\top\times_{k+2}\widetilde{{{\mathbf{U}}}}_{k+2}^\top\right)\widetilde{{{\mathbf{V}}}}_k\right).
\end{split}$$ Then we perform regression on sub-models with these reduced-dimensional covariates $\widetilde{{\mathbf{X}}}_{{{\mathbfcal{B}}}}$ and $\widetilde{{\mathbf{X}}}_{{{\mathbf{E}}}_k}$ respectively using least squares and group Lasso [@friedman2010note; @yuan2006model], $$\label{eq:regular-least-square}
\widehat{{{\mathbfcal{B}}}}\in \mathbb{R}^{r_1\times r_2\times r_3}, \quad {{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{r_1r_2r_3}} \|y - \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}{{\boldsymbol{\gamma}}}\|_2^2,\\$$ $$\label{eq:E_k-formula}
\widehat{{{\mathbf{E}}}}_k \in \mathbb{R}^{p_k\times r_k}, {{\rm vec}}(\widehat{{{\mathbf{E}}}}_k) = \left\{\begin{array}{ll}
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2, & \text{if } k \notin J_s;\\
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2 + \eta_k \sum_{j=1}^{p_k}\|{{\boldsymbol{\gamma}}}_{G_j^{k}}\|_2, & \text{if } k \in J_s.
\end{array}\right.$$ Here, $\{\eta_k\}_{k \in J_s}$ are the penalization level and $$\label{eq:partition-G}
G_j^k = \left\{j, j+p_k, \ldots, j+p_k(r_k-1)\right\}, \quad j=1,\ldots, p_k$$ form a partition of $\{1,\ldots, p_kr_k\}$ that is induced by the construction of $\widetilde{{\mathbf{X}}}_{{{\mathbf{E}}}_k}$ (details for why using group lasso can be found in Section \[sec:oracle-sparse\]).
3. (Constructing the final estimator) $\widehat{{\mathbfcal{A}}}$ can be constructed using the regression coefficients $\widehat{{{\mathbfcal{B}}}}$ and $\widehat{{\mathbf{E}}}_k$’s in the submodels and , $$\label{eq:hat_A-sparse}
\widehat{{{\mathbfcal{A}}}} = \left\llbracket \widehat{{{\mathbfcal{B}}}}, (\widehat{{{\mathbf{E}}}}_1(\widetilde{{{\mathbf{U}}}}_1^\top \widehat{{{\mathbf{E}}}}_1)^{-1}), (\widehat{{{\mathbf{E}}}}_2(\widetilde{{{\mathbf{U}}}}_2^\top \widehat{{{\mathbf{E}}}}_2)^{-1}), (\widehat{{{\mathbf{E}}}}_3(\widetilde{{{\mathbf{U}}}}_3^\top \widehat{{{\mathbf{E}}}}_3)^{-1})\right\rrbracket.$$ More interpretation of can be found in Section \[sec:oracle-sparse\].
Input: sample $\{y_j, {{\mathbfcal{X}}}_j\}_{j=1}^{n}$, Tucker rank ${\boldsymbol{r}}= (r_1, r_2, r_3)$. Calculate $\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j {{\mathbfcal{X}}}_j.$ Apply HOOI on $\widetilde{{{\mathbfcal{A}}}}$ and obtain initial estimates $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3$. Let $\widetilde{{{\mathbfcal{S}}}} = \llbracket\widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket$. Evaluate the sketching direction, $$\widetilde{{{\mathbf{V}}}}_k = {\rm QR}\left[\mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}})^\top\right], \quad k=1,2,3.$$ Construct $\widetilde{{{\mathbf{X}}}} = \left[\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1} ~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2} ~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_3}\right] \in \mathbb{R}^{n\times m}$, where $$\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times m_{{{\mathbfcal{B}}}}}, (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[i,:]} =
{{\rm vec}}\left({{\mathbfcal{X}}}_i \times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k} \in \mathbb{R}^{n\times m_{{{\mathbf{D}}}_k}}, (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k})_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{k\perp}^\top \mathcal{M}_k\left({{\mathbfcal{X}}}_{i} \times_{k+1}\widetilde{{{\mathbf{U}}}}_{k+1}^\top\times_{k+2} \widetilde{{{\mathbf{U}}}}_{k+2}^\top\right) \widetilde{{{\mathbf{V}}}}_k\right),
\end{split}$$ for $m_{{{\mathbfcal{B}}}} = r_1r_2r_3, m_{{{\mathbf{D}}}_k} = (p_k-r_k)r_k$, and $k=1,2,3$.
Solve $\widehat{{{\boldsymbol{\gamma}}}} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^m}\|y - \widetilde{{{\mathbf{X}}}}{{\boldsymbol{\gamma}}}\|_2^2$. Partition $\widehat{{{\boldsymbol{\gamma}}}}$ and assign each part to $\widehat{{{\mathbfcal{B}}}}, \widehat{{{\mathbf{D}}}}_1, \widehat{{{\mathbf{D}}}}_2, \widehat{{{\mathbf{D}}}}_3$, respectively, $$\begin{split}
& {{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) := \widehat{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}}= \widehat{{{\boldsymbol{\gamma}}}}_{[1:m_{{{\mathbfcal{B}}}}]},\\
& {{\rm vec}}(\widehat{{{\mathbf{D}}}}_k) := \widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{D}}}_k} = \widehat{{{\boldsymbol{\gamma}}}}_{\left[\left(m_{{{\mathbfcal{B}}}}+\sum_{k'=1}^{k-1} m_{{{\mathbf{D}}}_{k'}}+1\right): \left(m_{{{\mathbfcal{B}}}}+\sum_{k'=1}^{k}m_{{{\mathbf{D}}}_{k'}}\right)\right]},\quad k=1,2,3.\\
\end{split}$$ Let $\widehat{{{\mathbf{B}}}}_k = \mathcal{M}_k(\widehat{{{\mathbfcal{B}}}})$. Evaluate $$\widehat{{{\mathbfcal{A}}}} = \llbracket\widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{L}}}}_1, \widehat{{{\mathbf{L}}}}_2, \widehat{{{\mathbf{L}}}}_3\rrbracket, \quad \widehat{{{\mathbf{L}}}}_k = \left(\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k\right)\left(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\right)^{-1}, \quad k=1,2,3.$$
\[al:procedure\_regular\]
Input: sample $\{y_j, {{\mathbfcal{X}}}_j\}_{j=1}^{n}$, Tucker rank ${\boldsymbol{r}}= (r_1, r_2, r_3)$, sparsity index $J_s \subseteq \{1,2,3\}$. Evaluate $\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j {{\mathbfcal{X}}}_j.$ Apply STAT-SVD on $\widetilde{{{\mathbfcal{A}}}}$ with sparsity index $J_s$. Let the outcome be $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3$. Let $\widetilde{{{\mathbfcal{S}}}} = \llbracket\widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket$ and evaluate the probing direction, $$\widetilde{{{\mathbf{V}}}}_k = {\rm QR}\left[\mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}})^\top\right], \quad k=1,2,3.$$ Construct $$\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times (r_1r_2r_3)},\quad (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[i,:]} =
{{\rm vec}}({{\mathbfcal{X}}}_i \times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k} \in \mathbb{R}^{n\times (p_kr_k)}, \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})_{[i,:]} = {{\rm vec}}\left(\mathcal{M}_k\left({{\mathbfcal{X}}}_i\times_{k+1}\widetilde{{{\mathbf{U}}}}_{k+1}^\top\times_{k+2} \widetilde{{{\mathbf{U}}}}_{k+2}^\top\right)\widetilde{{{\mathbf{V}}}}_k\right).
\end{split}$$ Solve $$\begin{split}
\widehat{{{\mathbfcal{B}}}}\in \mathbb{R}^{r_1r_2r_3},& \quad {{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{r_1r_2r_3}} \|y - \widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}}{{\boldsymbol{\gamma}}}\|_2^2;
\end{split}$$ $$\begin{split}
\widehat{{{\mathbf{E}}}}_k \in \mathbb{R}^{p_k\times r_k}, {{\rm vec}}(\widehat{{{\mathbf{E}}}}_k) =\left\{\begin{array}{ll}
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2 + \lambda_k \sum_{j=1}^{p_k}\|{{\boldsymbol{\gamma}}}_{G_j^{k}}\|_2, & k \in J_s;\\
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2, & k\notin J_s.
\end{array} \right.
\end{split}$$ Evaluate $$\widehat{{{\mathbfcal{A}}}} = \left\llbracket\widehat{{{\mathbfcal{B}}}}; (\widehat{{{\mathbf{E}}}}_1(\widetilde{{{\mathbf{U}}}}_1^\top \widehat{{{\mathbf{E}}}}_1)^{-1}), (\widehat{{{\mathbf{E}}}}_2(\widetilde{{{\mathbf{U}}}}_2^\top \widehat{{{\mathbf{E}}}}_2)^{-1}), (\widehat{{{\mathbf{E}}}}_3(\widetilde{{{\mathbf{U}}}}_3^\top \widehat{{{\mathbf{E}}}}_3)^{-1})\right\rrbracket.$$
\[al:procedure\_sparse\]
A Sketching Perspective of ISLET {#sec:sketching-perspective}
--------------------------------
While one of the main focuses of this article is on low-rank tensor regression, from a sketching perspective, ISLET can be seen as a special case of a more general algorithm that broadly applies to high-dimensional statistical problems with dimension-reduced structure. In fact the three steps of the ISLET procedure are completely general and are summarized informally here:
1. (Probing projection directions) For the tensor regression problem, we use the HOOI [@de2000best] or STAT-SVD [@zhang2017optimal-statsvd] approach for finding the informative low-rank sub-spaces we project/sketch along. More generally if we let $\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j {{\mathbfcal{X}}}_j$ where ${{\mathbfcal{X}}}_j$ has ambient dimension $p$, we can define a general projection operator (with a slight abuse of notation) $\mathcal{P}_m(.) : \mathbb{R}^p \rightarrow \mathbb{R}^m$ indexed by low dimension $m$ and let $\mathcal{S}(\widetilde{{{\mathbfcal{A}}}})$ be the $m$-dimensional subspace of $\mathbb{R}^p$ determined by performing $\mathcal{P}_m(\widetilde{{{\mathbfcal{A}}}})$.
2. (Estimation in subspaces) The second step involves first projecting the data ${{\mathbfcal{X}}}$ on to the subspace $\mathcal{S}(\widetilde{{{\mathbfcal{A}}}})$, specifically $\widetilde{{{\mathbf{X}}}} = \mathcal{P}_{\mathcal{S}(\widetilde{{{\mathbfcal{A}}}})}({{\mathbfcal{X}}}) \in \mathbb{R}^{n \times m}$. Then perform regression or other procedures of choice using the sketched data $\widetilde{{{\mathbf{X}}}}$ to determine the dimension-reduced parameter $\widehat{{{\boldsymbol{\gamma}}}} \in \mathbb{R}^m$.
3. (Embedding to high-dimensional space) Finally, we need to project the estimator back to the high-dimensional space $\mathbb{R}^{p}$ by applying an equivalent to the inverse of the projection operator $\mathcal{P}^{-1}_{\mathcal{S}(\widetilde{{{\mathbfcal{A}}}})} : \mathbb{R}^m \rightarrow \mathbb{R}^p$. For low-rank tensor regression we require the formula .
The description above illustrates that the idea of ISLET is applicable to more general high-dimensional problems with dimension-reduced structure. In fact, the well-regarded *sure independence screening* in high-dimensional sparse linear regression [@fan2008sure; @xue2011sure] can be seen as a special case of this idea. To be specific, consider the high-dimensional linear regression model, $$y_i = X_{[i,:]} {\boldsymbol{\beta}}+ \varepsilon_i,\quad i=1,\ldots, n,$$ where ${\boldsymbol{\beta}}$ is the $m$-sparse vector of interests and $y_i\in \mathbb{R}$ and $X_{[i,:]}^\top\in \mathbb{R}^{p}$ are observable response and covariate. Then the $m$-dimensional subspace $\mathcal{S}(\widetilde{{\boldsymbol{\beta}}})$ in Step 1 can be the co-ordinates corresponding to the $m$ largest entries of $\widetilde{{\boldsymbol{\beta}}} = \sum_{i=1}^n X_{[i,:]}^\top y_i$; Step 2 corresponds to the dimension reduced least squares in sure independence screening; the inverse operator in Step 3 is simply filling in $0$’s in the co-ordinates that do not correspond to $\mathcal{S}(\widetilde{{\boldsymbol{\beta}}})$. In addition, this idea applies more broadly to problems such as matrix and tensor completion. One of the novel contributions of this article is finding suitable projection and inverse operators for low-rank tensors.
We can also contrast this approach with prior approaches that involve randomized sketching [@dobriban2018new; @pilanci2015randomized; @raskutti2014statistical]. These prior approaches showed that the randomized sketching may lose data substantially, increases the variance, and yield sub-optimal result for many statistical problems. There are two key differences with how we exploit sketching in our context: (1) we sketch along the parameter directions of ${{\mathbfcal{X}}}$, reducing the data from $\mathbb{R}^{n \times p}$ to $\mathbb{R}^{n \times m}$; whereas approaches in [@dobriban2018new; @pilanci2015randomized; @raskutti2014statistical] sketch along the sample directions, reducing the data from $\mathbb{R}^{n \times p}$ to $\mathbb{R}^{m \times p}$, which reduces the effective sample size from $n$ to $m$; (2) secondly and most importantly rather than using the randomized sketching that is *unsupervised* without the response $y$, our importance sketching is *supervised* that is obtained using both the response $y$ and covariates ${{\mathbfcal{X}}}$. Then we sketch along the subspace $\mathcal{S}(\widetilde{{{\mathbfcal{A}}}})$ which contains information on the low-dimensional structure of the parameter $\mathcal{A}$. This is why our general procedure has both desirable statistical and computational properties.
Oracle Inequalities {#sec:oracle-inequality}
===================
In this section, we provide general oracle inequalities without focusing on specific design, which provides a general guideline for the theoretical analyses of our ISLET procedure. We first introduce a quantification of the errors in sketching directions obtained in the first step of ISLET. Let ${{\mathbf{V}}}_k\in \mathbb{O}_{r_{k+1}r_{k+2}, r_k}$ be the right singular subspace of $\mathcal{M}_k({{\mathbfcal{S}}})$, where ${{\mathbfcal{S}}}$ is the core tensor in the Tucker decomposition of ${{\mathbfcal{A}}}$: ${{\mathbfcal{A}}}= \llbracket {{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket$. By Lemma \[lm:Kronecker-vectorization-matricization\] in the supplementary material [@ISLET-supplement], $$\label{eq:def_W}
\begin{split}
{{\mathbf{W}}}_1 := ({{\mathbf{U}}}_{3}\otimes &{{\mathbf{U}}}_{2}){{\mathbf{V}}}_1 \in \mathbb{O}_{p_{2}p_{3}, r_1}, ~~ {{\mathbf{W}}}_2 := ({{\mathbf{U}}}_{3}\otimes {{\mathbf{U}}}_{1}){{\mathbf{V}}}_2 \in \mathbb{O}_{p_{1}p_{3}, r_2},\\
& \text{and}\quad {{\mathbf{W}}}_3 := ({{\mathbf{U}}}_{2}\otimes {{\mathbf{U}}}_{1}){{\mathbf{V}}}_3 \in \mathbb{O}_{p_{1}p_{2}, r_3}
\end{split}$$ are the right singular subspaces of $\mathcal{M}_1({{\mathbfcal{A}}}), \mathcal{M}_2({{\mathbfcal{A}}})$, and $\mathcal{M}_3({{\mathbfcal{A}}})$, respectively. Recall that we initially estimate ${{\mathbf{U}}}_k$ and ${{\mathbf{V}}}_k$ by $\widetilde{{{\mathbf{U}}}}_k$ and $\widetilde{{{\mathbf{V}}}}_k$, respectively in Step 1 of ISLET. Define $$\widetilde{{{\mathbf{W}}}}_1 = (\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2}) \widetilde{{{\mathbf{V}}}}_1,\quad \widetilde{{{\mathbf{W}}}}_2 = (\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{1}) \widetilde{{{\mathbf{V}}}}_2, \quad \text{and}\quad \widetilde{{{\mathbf{W}}}}_3 = (\widetilde{{{\mathbf{U}}}}_{2}\otimes \widetilde{{{\mathbf{U}}}}_{1}) \widetilde{{{\mathbf{V}}}}_3$$ in parallel to . Intuitively speaking, $\{\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{W}}}}_k\}_{k=1}^3$ can be seen as the initial sample approximations for $\{{{\mathbf{U}}}_k, {{\mathbf{W}}}_k\}_{k=1}^3$. Therefore, we quantify the *sketching direction error* by $$\label{eq:theta-sin-theta-distance}
\begin{split}
\theta := \max_{k=1,2,3}\left\{\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\|, \|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\|\right\}.
\end{split}$$ Next, we provide the oracle inequality via $\theta$ for ISLET under regular and sparse settings, respectively in the next two subsections.
Regular Tensor Regression and Oracle Inequality {#sec:oracle-regular}
------------------------------------------------
In order to study the theoretical properties of the proposed procedure, we need to introduce another representation of the original model . Decompose the vectorized parameter ${{\mathbfcal{A}}}$ as follows, $$\label{eq:decomposition-A}
\begin{split}
{{\rm vec}}({{\mathbfcal{A}}}) = & P_{\widetilde{{{\mathbf{U}}}}}{{\rm vec}}({{\mathbfcal{A}}}) + P_{\widetilde{{{\mathbf{U}}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}})\\
= & P_{\widetilde{{{\mathbf{U}}}}_3\otimes\widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1} {{\rm vec}}({{\mathbfcal{A}}}) + P_{\mathcal{R}_1(\widetilde{{{\mathbf{W}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_{1\perp})} {{\rm vec}}({{\mathbfcal{A}}}) + P_{\mathcal{R}_2(\widetilde{{{\mathbf{W}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{2\perp})} {{\rm vec}}({{\mathbfcal{A}}})\\
& + P_{\mathcal{R}_3(\widetilde{{{\mathbf{W}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_{3\perp})} {{\rm vec}}({{\mathbfcal{A}}}) + P_{\widetilde{{{\mathbf{U}}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}})\\
= & (\widetilde{{{\mathbf{U}}}}_3\otimes\widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1) {{\rm vec}}(\widetilde{{{\mathbfcal{B}}}}) + \mathcal{R}_1(\widetilde{{{\mathbf{W}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_{1\perp}) {{\rm vec}}(\widetilde{{\mathbf{D}}}_1) + \mathcal{R}_2(\widetilde{{{\mathbf{W}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{2\perp}) {{\rm vec}}(\widetilde{{\mathbf{D}}}_2) \\
& + \mathcal{R}_3(\widetilde{{{\mathbf{W}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_{3\perp}) {{\rm vec}}(\widetilde{{\mathbf{D}}}_3) + P_{\widetilde{{{\mathbf{U}}}}_\perp} {{\rm vec}}({{\mathbfcal{A}}}).
\end{split}$$ (See the proof of Theorem \[th:upper\_bound\_general\] for a detailed derivation of ). Here, $$\begin{split}
\qquad \widetilde{{{\mathbf{U}}}} = \left[\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1 ~~ \mathcal{R}_1(\widetilde{{{\mathbf{W}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_{1\perp}) ~~ \mathcal{R}_2\left( \widetilde{{{\mathbf{W}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{2\perp}\right) ~~ \mathcal{R}_3\left( \widetilde{{{\mathbf{W}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_{3\perp}\right)\right], \end{split}$$ $$\begin{split}
\widetilde{{{\mathbfcal{B}}}} := & \left\llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket \in \mathbb{R}^{r_1r_2r_3} \quad \text{and}\quad \widetilde{{{\mathbf{D}}}}_k := \widetilde{{{\mathbf{U}}}}_{k\perp}^\top \mathcal{M}_k({{\mathbfcal{A}}}) \widetilde{{{\mathbf{W}}}}_k \in \mathbb{R}^{(p_k-r_k)\times r_k}
\end{split}$$ are the singular subspace of the “Cross structure" and the low-dimensional projections of ${{\mathbfcal{A}}}$ onto the “body" and “arms" formed by sketching directions $\{\widetilde{{{\mathbf{U}}}}_k,\widetilde{{{\mathbf{V}}}}_k\}_{k=1}^3$, respectively (See Figure \[fig:illustration\_U\] for an illustration of $\widetilde{{\mathbf{U}}}$, $\widetilde{{\mathbfcal{B}}}$, and $\widetilde{{\mathbf{V}}}_k$). Due to different alignments, the $i$-th row of $\{{{\mathbf{W}}}_k\otimes {{\mathbf{U}}}_{k\perp}\}_{k=1}^3$ does not necessarily correspond to the $i$-th entry of ${{\rm vec}}({{\mathbfcal{A}}})$ for all $1\leq i \leq p_1p_2p_3$. We thus permute the rows of $\{\widetilde{{\mathbf{W}}}_k\otimes \widetilde{{\mathbf{U}}}_{k\perp}\}_{k=1}^3$ to match each row of $\mathcal{R}_k(\widetilde{{\mathbf{W}}}_k\otimes \widetilde{{{\mathbf{U}}}}_{k\perp})$ to the corresponding entry in ${{\rm vec}}({{\mathbfcal{A}}})$. The formal definition of the row-wise permutation operator $\mathcal{R}_k$ is rather clunky and postponed to Section \[sec:row-permutation-operator\] in the supplementary materials. Intuitively speaking, $P_{\widetilde{{{\mathbf{U}}}}}{{\rm vec}}({{\mathbfcal{A}}})$ represents the projection of ${{\mathbfcal{A}}}$ onto to the Cross structure and $P_{\widetilde{{{\mathbf{U}}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}})$ can be seen as a residual. If the estimates $\{\widetilde{{\mathbf{U}}}_k, \widetilde{{\mathbf{W}}}_k\}_{k=1}^3$ are close enough to $\{{{\mathbf{U}}}_k, {{\mathbf{W}}}_k\}_{k=1}^3$, i.e., $\theta$ defined in is small, we expect that the residual $P_{\widetilde{{\mathbf{U}}}_{\perp}}{{\rm vec}}({{\mathbfcal{A}}})$ has small amplitude.
![Illustration of Decomposition . Here, we assume $\widetilde{{{\mathbf{U}}}}_k^\top = [{{\mathbf{I}}}_{r_k} ~ \boldsymbol{0}_{r_k \times (p_k-r_k)}]$, $k=1,2,3$, for a better visualization. The gray, green, blue, and red cubes represent the subspaces of $\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes\widetilde{{{\mathbf{U}}}}_1$, $\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{1\perp}$, $\widetilde{{{\mathbf{U}}}}_3\otimes\widetilde{{{\mathbf{U}}}}_{2\perp}\otimes\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_{3\perp}\otimes\widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1$. The gray cube also corresponds to the projected parameters $\widetilde{{{\mathbfcal{B}}}}$; matricizations of green, blue and red cubes correspond to the projected parameters $\widetilde{{\mathbf{U}}}_{1\perp}^\top \mathcal{M}_1({{\mathbfcal{A}}})(\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_2)$, $\widetilde{{\mathbf{U}}}_{2\perp}^\top \mathcal{M}_2({{\mathbfcal{A}}})(\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_1)$, and $\widetilde{{\mathbf{U}}}_{3\perp}^\top \mathcal{M}_3({{\mathbfcal{A}}})(\widetilde{{\mathbf{U}}}_2\otimes \widetilde{{\mathbf{U}}}_1)$, respectively. The three plains in the right panel correspond to the subspace of $\widetilde{{\mathbf{V}}}_1$, $\widetilde{{\mathbf{V}}}_2$, and $\widetilde{{\mathbf{V}}}_3$, respectively.[]{data-label="fig:illustration_U"}](cross){height="4cm" width=".6\linewidth"}
Based on , we can re-write the original regression model into the following partial regression model, $$\label{eq:low-dimensional-regression}
\begin{split}
y_j = & (\widetilde{{\mathbf{X}}}_{{{\mathbfcal{B}}}})_{[j,:]} {{\rm vec}}(\widetilde{{\mathbfcal{B}}}) + \sum_{k=1}^3 (\widetilde{{\mathbf{X}}}_{{{\mathbf{D}}}_k})_{[j,:]}{{\rm vec}}(\widetilde{{\mathbf{D}}}_k) + {{\rm vec}}({{\mathbfcal{X}}}_j)^\top P_{\widetilde{{\mathbf{U}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}}) + \varepsilon_j\\
= & \widetilde{{\mathbf{X}}}_{[j,:]} \widetilde{{{\boldsymbol{\gamma}}}} + \widetilde{\varepsilon}_j, \quad j=1,\ldots, n.
\end{split}$$ (See the proof of Theorem \[th:upper\_bound\_general\] for a detailed derivation of .) Here,
- $\widetilde{\varepsilon}_j = {{\rm vec}}({{\mathbfcal{X}}}_j)^\top P_{\widetilde{{{\mathbf{U}}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}}) + \varepsilon_j$ is the oracle noise; $\widetilde{{\boldsymbol{\varepsilon}}} = (\widetilde{\varepsilon}_1,\ldots, \widetilde{\varepsilon}_{n})^\top$;
- $\widetilde{{\mathbf{X}}}_{{{\mathbfcal{B}}}}, \widetilde{{\mathbf{X}}}_{{{\mathbf{D}}}_k}$ are sketching covariates introduced in Equation ;
- $\widetilde{{{\boldsymbol{\gamma}}}} = \left[{{\rm vec}}(\widetilde{{\mathbfcal{B}}})^\top, {{\rm vec}}(\widetilde{{\mathbf{D}}}_1)^\top, {{\rm vec}}(\widetilde{{\mathbf{D}}}_2)^\top, {{\rm vec}}(\widetilde{{\mathbf{D}}}_3)^\top\right]^\top = \widetilde{{\mathbf{U}}}^\top {{\rm vec}}({{\mathbfcal{A}}}) \in \mathbb{R}^m$ is the dimension-reduced parameter.
reveals the essence of the least squares estimator in the ISLET procedure – the outcomes of and , i.e., $\widehat{{\mathbfcal{B}}}$ and $\widehat{{\mathbf{D}}}_k$, are sample-based estimates of $\widetilde{{\mathbfcal{B}}}$ and $\widetilde{{\mathbf{D}}}_k$. Finally, based on the detailed algebraic calculation in Step 3 and the proof of Theorem \[th:upper\_bound\_general\], $$\label{eq:bcA-identity}
\begin{split}
& {{\mathbfcal{A}}}= \left\llbracket \widetilde{{{\mathbfcal{B}}}}; \widetilde{{\mathbf{L}}}_1, \widetilde{{\mathbf{L}}}_2, \widetilde{{\mathbf{L}}}_3\right\rrbracket, \quad \widetilde{{\mathbf{L}}}_k = \left(\widetilde{{\mathbf{U}}}_k\widetilde{{\mathbf{B}}}_k \widetilde{{\mathbf{V}}}_k + \widetilde{{\mathbf{U}}}_{k\perp}\widetilde{{\mathbf{D}}}_k\right)\left(\widetilde{{\mathbf{B}}}_k\widetilde{{\mathbf{V}}}_k\right)^{-1}.
\end{split}$$ is essentially a higher-order version of the Schur complement formula (also see [@cai2016structured]). Finally, we apply the plug-in estimator to obtain the final estimator $\widehat{{\mathbfcal{A}}}$ (Equation in Step 3 of the ISLET procedure).
Based on previous discussions, it can be seen that the estimation error of the original tensor regression is driven by the error of the least squares estimator $\widehat{{\boldsymbol{\gamma}}}$, i.e., $\|(\widetilde{{\mathbf{X}}}^\top\widetilde{{\mathbf{X}}})^{-1}\widetilde{{\mathbf{X}}}^\top\widetilde{\boldsymbol{\varepsilon}}\|_2^2$. We have the following oracle inequality for the proposed ISLET procedure.
\[th:upper\_bound\_general\] Suppose ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times p_2 \times p_3}$ has Tucker rank-$(r_1, r_2, r_3)$ tensor and $\widehat{{{\mathbfcal{A}}}}$ is the outcome of Algorithm \[al:procedure\_regular\]. Assume the sketching directions $\{\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k\}_{k=1}^3$ satisfy $\theta<1/2$ (see for the definition of $\theta$) and $\left\|\widehat{{{\mathbf{D}}}}_k(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\| \leq \rho$. We don’t impose other specific assumptions on ${{\mathbfcal{X}}}_i$ and $\varepsilon_i$. Then, we have $$\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \leq (1 + C(\theta+\rho))\left\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top\widetilde{{\boldsymbol{\varepsilon}}}\right\|_2^2$$ for uniform constant $C>0$ that does not rely on any other parameters.
See Appendix \[sec:proof\_upper\_bound\_general\] for a complete proof. In particular, the proof contains three major steps. After introducing a number of notation, we first transform the original regression model to the partial regression model , then rewrite the upper bound $\|(\widetilde{{\mathbf{X}}}^\top\widetilde{{\mathbf{X}}})^{-1}\widetilde{{\mathbf{X}}}^\top\widetilde{{\boldsymbol{\varepsilon}}}\|_2^2$ to $\|\widehat{{{\mathbfcal{B}}}} - \widetilde{{\mathbfcal{B}}}\|_{{{\rm HS}}}^2 + \sum_{k=1}^3 \|\widehat{{{\mathbf{D}}}}_k-\widetilde{{\mathbf{D}}}_k\|_F^2$. Next, we introduce a factorization of ${{\mathbfcal{A}}}$ in parallel with the one of $\widehat{{\mathbfcal{A}}}$, based on which the loss $\|\widehat{{\mathbfcal{A}}}- {{\mathbfcal{A}}}\|_{{{\rm HS}}}$ is decomposed into eight terms. Finally, we introduce a novel deterministic error bound for the “Cross scheme" (Lemma \[lm:FGH\] in the supplementary material [@ISLET-supplement]; also see [@zhang2019cross]), carefully analyze each term in the decomposition of $\|\widehat{{\mathbfcal{A}}}- {{\mathbfcal{A}}}\|_{{{\rm HS}}}$, and finalize the proof.
Theorem \[th:upper\_bound\_general\] shows that once the sketching directions $\widetilde{{{\mathbf{U}}}}$ and $\widetilde{{{\mathbf{V}}}}$ are reasonably accurate, the estimation error for $\widehat{{{\mathbfcal{A}}}}$ will be close to the error of partial linear regression in Equation . This bound is general and deterministic, which can be used as a key step in more specific settings of low-rank tensor regression.
Sparse Tensor Regression and Oracle Inequality {#sec:oracle-sparse}
----------------------------------------------
Next, we study the oracle performance of the proposed procedure for sparse tensor regression, where ${{\mathbfcal{A}}}$ further satisfies the sparsity constraint . As in the previous section, we decompose the vectorized parameter as $$\label{eq:decomposition-A-sparse1}
\begin{split}
{{\rm vec}}({{\mathbfcal{A}}}) = & P_{\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_2\otimes \widetilde{{\mathbf{U}}}_1}{{\rm vec}}({{\mathbfcal{A}}}) + P_{(\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_2\otimes \widetilde{{\mathbf{U}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\\
= & (\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_2 \otimes\widetilde{{\mathbf{U}}}_3) {{\rm vec}}(\widetilde{{\mathbfcal{B}}}) + P_{(\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_2\otimes \widetilde{{\mathbf{U}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}});
\end{split}$$ $$\label{eq:decomposition-A-sparse2}
\begin{split}
{{\rm vec}}({{\mathbfcal{A}}}) = & P_{\mathcal{R}_k(\widetilde{{\mathbf{W}}}_k\otimes {{\mathbf{I}}}_{p_k})}{{\rm vec}}({{\mathbfcal{A}}}) + P_{\mathcal{R}_k(\widetilde{{\mathbf{W}}}_k\otimes {{\mathbf{I}}}_{p_k})_\perp}{{\rm vec}}({{\mathbfcal{A}}})\\
= & \mathcal{R}_k(\widetilde{{\mathbf{W}}}_k\otimes {{\mathbf{I}}}_{p_k}) {{\rm vec}}(\widetilde{{\mathbf{E}}}_k) + P_{\mathcal{R}_k(\widetilde{{\mathbf{W}}}_k\otimes {{\mathbf{I}}}_{p_k})_\perp}{{\rm vec}}({{\mathbfcal{A}}}), \quad k=1,2,3.
\end{split}$$ Here, $$\label{eq:tilde-B-tilde-E}
\begin{split}
& \widetilde{{{\mathbfcal{B}}}} := \llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket \in \mathbb{R}^{r_1r_2r_3};\\
& \widetilde{{{\mathbf{E}}}}_k := \mathcal{M}_k\left({{\mathbfcal{A}}}\times_{(k+1)}\widetilde{{{\mathbf{U}}}}_{k+1}^\top\times_{(k+2)}\widetilde{{{\mathbf{U}}}}_{k+2}^\top\right)\widetilde{{{\mathbf{V}}}}_{k}\in \mathbb{R}^{p_k\times r_k},\quad k=1,2,3,
\end{split}$$ are the low-dimensional projections of ${{\mathbfcal{A}}}$ onto the importance sketching directions. Since $\{{{\mathbf{U}}}_k, {{\mathbf{W}}}_k\}$ are the left and right singular subspaces of $\mathcal{M}_k({{\mathbfcal{A}}})$, we can show $P_{({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})$ and $P_{\mathcal{R}_k({{\mathbf{W}}}_k\otimes {{\mathbf{I}}}_{p_k})_\perp}{{\rm vec}}({{\mathbfcal{A}}})$ are zeros. Thus if the estimates $\{\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{W}}}}_k\}_{k=1}^3$ are sufficiently accurate, i.e., $\theta$ defined in Eq. is small, we expect that the residuals $P_{(\widetilde{{\mathbf{U}}}_3\otimes \widetilde{{\mathbf{U}}}_2\otimes \widetilde{{\mathbf{U}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})$ and $P_{\mathcal{R}_k(\widetilde{{\mathbf{W}}}_k\otimes {{\mathbf{I}}}_{p_k})_\perp}{{\rm vec}}({{\mathbfcal{A}}})$ have small amplitudes. Then, based on a more detailed calculation in the proof of Theorem \[th:upper\_bound\_sparse\_general\], the model of sparse and low-rank tensor regression $y_j = \langle {{\mathbfcal{X}}}_j, {{\mathbfcal{A}}}\rangle + \varepsilon_j$ can be rewritten as the following partial linear regression, $$\label{eq:partial-linear-model-B}
y_j = (\widetilde{{\mathbf{X}}}_{{\mathbfcal{B}}})_{[j,:]}{{\rm vec}}(\widetilde{{\mathbf{B}}}) + (\widetilde{\boldsymbol{\varepsilon}}_{{\mathbfcal{B}}})_j,$$ $$\label{eq:partial-linear-model-Ek}
y_j = (\widetilde{{\mathbf{X}}}_{{{\mathbf{E}}}_k})_{[j,:]}{{\rm vec}}(\widetilde{{\mathbf{E}}}_k) + (\widetilde{\boldsymbol{\varepsilon}}_{{{\mathbf{E}}}_k})_j, \quad k=1,2,3.$$ Here, $\widetilde{{\mathbf{X}}}_{{{\mathbfcal{B}}}}$ and $\widetilde{{\mathbf{X}}}_{{{\mathbf{E}}}_k}$ are the covariates defined in Equation and\
$\widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}= ((\widetilde{\varepsilon}_{{\mathbfcal{B}}})_1,\ldots, (\widetilde{\varepsilon}_{{\mathbfcal{B}}})_{n})^\top$, $\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k} = ((\widetilde{\varepsilon}_{{{\mathbf{E}}}_k})_1,\ldots, (\widetilde{\varepsilon}_{{{\mathbf{E}}}_k})_{n})^\top$ are oracle noises defined as $$\label{eq:epsilon_B-epsilon_E}
\begin{split}
& (\widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}})_j = \left\langle {{\rm vec}}({{\mathbfcal{X}}}_j), P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)_{\perp}}{{\rm vec}}({{\mathbfcal{A}}})\right\rangle + \varepsilon_j\\
\text{and}\quad & (\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k})_j = \left\langle {{\rm vec}}({{\mathbfcal{X}}}_{j}), P_{\left(\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes{{\mathbf{I}}}_{p_k})\right)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\rangle + \varepsilon_j.
\end{split}$$ Therefore, the Step 2 of sparse ISLET can be interpreted as the estimation of $\widetilde{{\mathbfcal{B}}}$ and $\widetilde{{\mathbf{E}}}_k$.
We apply regular least squares to estimate $\widetilde{{\mathbfcal{B}}}$ and $\widetilde{{\mathbf{E}}}_k$ for $k\notin J_s$. For any sparse mode $k\in J_s$, $\widetilde{{{\mathbf{E}}}}_k$ are group sparse due to the definition and the assumption that ${{\mathbf{U}}}_k$ are row-wise sparse. Specifically, $\widetilde{{\mathbf{E}}}_k$ satisfies $$\label{eq:tilde-E-group-sparse}
\begin{split}
\left\|{{\rm vec}}(\widetilde{{{\mathbf{E}}}}_k)\right\|_{0, 2} := \sum_{i=1}^{p_k} 1_{\left\{\left({{\rm vec}}(\widetilde{{\mathbf{E}}}_k)\right)_{G_i^k} \neq 0\right\}}\leq s_k,
\end{split}$$ where $$G_i^k = \left\{i, i+p_k, \ldots, i+p_k(r_k-1)\right\}, \quad i=1,\ldots, p_k, \quad \forall k\in J_s,$$ is a partition of $\{1,\ldots, p_kr_k\}$ (see the proof for Theorem \[th:upper\_bound\_sparse\_general\] for a more detailed argument for ). By detailed calculations in Step 3 of the proof for Theorem \[th:upper\_bound\_general\], one can verify that $${{\mathbfcal{A}}}= \left\llbracket \widetilde{{{\mathbfcal{B}}}}, (\widetilde{{{\mathbf{E}}}}_1(\widetilde{{{\mathbf{U}}}}_1^\top \widetilde{{{\mathbf{E}}}}_1)^{-1}), (\widetilde{{{\mathbf{E}}}}_2(\widetilde{{{\mathbf{U}}}}_2^\top \widetilde{{{\mathbf{E}}}}_2)^{-1}), (\widetilde{{{\mathbf{E}}}}_3(\widetilde{{{\mathbf{U}}}}_3^\top \widetilde{{{\mathbf{E}}}}_3)^{-1})\right\rrbracket.$$ Then the finally sparse ISLET estimator $\widehat{{\mathbfcal{A}}}$ in can be seen as the plug-in estimator.
To ensure that the group Lasso estimator in provides a stable estimation for the proposed procedure, we introduce the following group restricted isometry condition, which can also be seen as an extension of restricted isometry property (RIP), a commonly used condition in compressed sensing and high-dimensional linear regression literature [@candes2005decoding].
\[con:group-RIP\] We say a matrix ${{\mathbf{X}}}\in \mathbb{R}^{n\times p}$ satisfies the group restricted isometry property (GRIP) with respect to partition $G_1,\ldots, G_m \subseteq \{1,\ldots, p\}$, if there exists $\delta>0$ such that $$n(1-\delta)\|{{\mathbf{v}}}\|_2^2 \leq \|{{\mathbf{X}}}{{\mathbf{v}}}\|_2^2 \leq n(1+\delta)\|{{\mathbf{v}}}\|_2^2$$ for all group-wise sparse vector $v$ satisfying $\sum_{k=1}^m 1_{\{{{\mathbf{v}}}_{G_k} \neq 0\}}\leq s$.
We still use $\theta$ defined in Eq. to characterize the sketching direction errors. The following oracle inequality holds for sparse tensor regression with importance sketching.
\[th:upper\_bound\_sparse\_general\] Consider the sparse low-rank tensor regression . Suppose $\theta<1/2$, the importance sketching covariates $\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}$ and $\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}$ ($k\notin J_s$) are non-singular. For any $k\in J_s$, $\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}$ satisfies group restricted isometry property (Condition \[con:group-RIP\]) with respect to partition $G_1^k,\ldots, G_{p_k}^k$ in and $\delta < 1/3$. We apply the proposed Algorithm \[al:procedure\_sparse\] with group Lasso penalty $$\eta_k = C_1\max_{i=1,\ldots, p_k}\left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k, [:, G_i^k]})^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2$$ for $k\in J_s$ and some constant $C_1 \geq 3$. We also assume $\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top\widehat{{{\mathbf{E}}}}_k(\widetilde{{{\mathbf{U}}}}_k^\top \widehat{{{\mathbf{E}}}}_k)^{-1}\| \leq \rho$. Then, $$\label{ineq:oracle-error-bound}
\begin{split}
& \left\|\widehat{{{\mathbfcal{A}}}}-{{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \leq (1 +C_2s(\theta+ \rho)) \Bigg( \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbf{B}}}\right\|_2^2 \\
& + \sum_{k \notin J_s} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2 + C_3\sum_{k\in J_s} s_k \cdot\max_{i = 1,\ldots, p_k} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k, [:, G_i^k]})^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n\right\|_2^2\Bigg).
\end{split}$$
See Appendix \[sec:proof\_upper\_bound\_sparse\_general\].
In the oracle error bound , $\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbf{B}}}\|_2^2$,\
$ \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbf{B}}}\right\|_2^2$, and $s_k \max_{i = 1,\ldots, p_k} \|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k, [:, G_i^k]})^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n\|_2^2$ correspond to the estimation errors of $\widehat{{{\mathbfcal{B}}}}$, $\widehat{{{\mathbf{E}}}}_k$ of the non-sparse mode, and $\widehat{{{\mathbf{E}}}}_k$ of sparse mode, respectively. When the group restricted isometry property (Condition \[con:group-RIP\]) is replaced by group restricted eigenvalue condition (see, e.g., [@lounici2011oracle]), a similar result to Theorem \[th:upper\_bound\_sparse\_general\] can be derived.
Fast Low-rank Tensor Regression via ISLET {#sec:regression}
=========================================
We further study the low-rank tensor regression with Gaussian ensemble design, i.e., ${{\mathbfcal{X}}}_i$ has i.i.d. standard normal entries. This has been considered a benchmark setting for low-rank tensor/matrix recovery literature [@candes2011tight; @chen2016non]. For convenience, we denote ${\boldsymbol{p}}= (p_1,p_2,p_3), {\boldsymbol{r}}= (r_1,r_2,r_3)$, $p = \max\{p_1, p_2, p_3\}$, and $r = \max\{r_1, r_2, r_3\}$. We discuss the regular low-rank and sparse low-rank tensor regression in the next two subsections, respectively.
Regular Low-rank Tensor Regression with ISLET
---------------------------------------------
We have the following theoretical guarantee for ISLET under Gaussian ensemble design.
\[th:upper\_bound\_regression\] Consider the tensor regression model , where ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times p_2 \times p_3}$ is Tucker rank-$(r_1, r_2, r_3)$, ${{\mathbfcal{X}}}_i$ has i.i.d. standard normal entries, and $\varepsilon\overset{iid}{\sim}N(0, \sigma^2)$. Denote $\widetilde{\sigma}^2 = \|{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2 + \sigma^2$, $\lambda_0 = \min_k \lambda_k, \lambda_k = \sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{A}}}))$, $\kappa = \max_k \|\mathcal{M}_k({{\mathbfcal{A}}})\|/\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{A}}}))$, and $m = r_1r_2r_3 + \sum_{k=1}^3(p_k-r_k)r_k$. If $n_1\wedge n_2 \geq \frac{C\widetilde{\sigma}^2(p^{3/2} + \kappa pr)}{\lambda_0^2}$, then the sample-splitting ISLET estimator (see the forthcoming Remark \[rm:sample-splitting\]) satisfies $$\left\|\widehat{{{\mathbfcal{A}}}}-{{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2\leq \frac{m}{n_2}\left(\sigma^2 + \frac{C_1\widetilde{\sigma}^4mp}{n_1^2\lambda_0^2}\right)\left(1 + C_2\sqrt{\frac{\log p}{m}} + C_3\sqrt{\frac{m\widetilde{\sigma}^2}{(n_1\wedge n_2)\lambda_0^2}}\right)$$ with probability at least $1 - p^{-C_4}$.
See Section \[sec:proof\_upper\_bound\_regression\] for details. Specifically, we first derive the estimation error upper bounds for sketching directions $\widetilde{{{\mathbf{U}}}}_k$ via the deterministic error bound of HOOI [@zhang2019HOOI]. Then we apply concentration inequalities to obtain upper bounds for $\left\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top\widetilde{\boldsymbol{\varepsilon}}\right\|_2^2$ and $\|\widehat{{\mathbf{D}}}_k(\widehat{{\mathbf{B}}}_k\widetilde{{\mathbf{V}}}_k)^{-1}\|$ for $k=1,2,3$. Finally, the oracle inequality of Theorem \[th:upper\_bound\_general\] leads to the desired upper bound.
In Theorem \[th:upper\_bound\_regression\], we show that as long as the sample size $n=\Omega(p^{3/2}r + pr^2)$, ISLET achieves consistent estimation under regularity conditions. This sample complexity outperforms many computationally feasible algorithms in previous literature, e.g., $n= \Omega(p^2r\textrm{polylog}(p))$ in projected gradient descent [@chen2016non], sum of nuclear norm minimization [@tomioka2011statistical], and square norm minimization [@mu2014square]. To the best of our knowledge, ISLET is the first computationally efficient algorithm that achieves this sample complexity result.
On the other hand, [@mu2014square] showed that the direct nonconvex Tucker rank minimization, a computationally infeasible method, can do exact recovery with $O(pr + r^3)$ linear measurements in the noiseless setting. [@bousse2018linear] showed that if tensor parameter ${{\mathbfcal{A}}}$ is CP rank-$r$, the linear system $y_j = \langle {{\mathbfcal{A}}}, {{\mathbfcal{X}}}_j\rangle, j=1,\ldots,n$ has a unique solution with probability one if one has $O(pr)$ measurements. It remains an open question whether the sample complexity of $n = \Omega(p^{3/2}r + pr^2)$ is necessary for all computationally efficient procedures.
\[rm:sample-splitting\] The direct analysis for the proposed ISLET in Algorithm \[al:procedure\_regular\] is technically involved, among which one major difficulty is the dependency between the sketching directions $\widetilde{{\mathbf{U}}}_k$ obtained in Step 1 and the regression noise $\widetilde{{\boldsymbol{\varepsilon}}}$ in Step 2. To overcome this difficulty, we choose to analyze a modified procedure with the sample splitting scheme: we randomly split all $n$ samples into two sets with cardinality $n_1$ and $n_2$, respectively. Then we use the first set of $n_1$ samples to construct the covariance tensor $\widetilde{{{\mathbfcal{A}}}}$ (Step 1) and use the second set of $n_2$ samples to evaluate the importance sketching covariates (Step 2). As illustrated by numerical studies in Section \[sec:numerical\], such a scheme is mainly for technical purposes and is not necessary in practice. Simulations suggest that it is preferable to use all samples $\{y_i, {{\mathbfcal{X}}}_i\}_{i=1}^n$ for both constructing the initial estimate $\widetilde{{{\mathbfcal{A}}}}$ and performing linear regression on sketching covariates.
We further consider the statistical limits for low-rank tensor regression with Gaussian ensemble. Consider the following class of general low-rank tensors, $$\mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}} = \left\{{{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times p_2\times p_3}: \text{Tucker rank}({{\mathbfcal{A}}}) \leq (r_1,r_2,r_3) \right\}.$$ The following minimax lower bound holds for all low-rank tensors in $\mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}$.
\[th:lower-bound-regression\] If $n > m+1$, the following non-asymptotic lower bound in estimation error hold, $$\label{ineq:hat-A-regression-lower}
\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}}\mathbb{E} \left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \frac{m}{n-m-1}\cdot \sigma^2.$$ If $n \leq m + 1$, $$\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}}\mathbb{E} \left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 = +\infty.$$
See Appendix \[sec:proof\_lower-bound-regression\].
Combining Theorems \[th:upper\_bound\_regression\] and \[th:lower-bound-regression\], we can see that as long as the sample size satisfies $\frac{m\widetilde\sigma^2}{n_1\lambda_0^2} = o(1)$, $\frac{m(p_1+p_2+p_3)\widetilde{\sigma}^4}{n_1n_2\lambda_0^2}=o(\sigma^2)$, and $n_2 = (1+o(1))n$, the statistical loss of the proposed method is sharp with matching constant to the lower bound.
If the order of tensor reduces to two, the tensor regression becomes the well-regarded *low-rank matrix recovery* in literature [@candes2011tight; @recht2010guaranteed]: $$y_i = \left\langle{{\mathbf{X}}}_i, {{\mathbf{A}}}\right\rangle + \varepsilon_i, \quad i=1,\ldots, n.$$ Here, ${{\mathbf{A}}}\in \mathbb{R}^{p_1\times p_2}$ is the unknown rank-$r$ target matrix, $\{{{\mathbf{X}}}_i\}_{i=1}^n$ are design matrices, and $\varepsilon_i \sim N(0, \sigma^2)$ are noises. The low-rank matrix recovery, including its instances such as phase retrieval [@candes2015phase], has been widely considered in recent literature. Various methods, such as nuclear norm minimization [@candes2010matrix; @recht2010guaranteed], projected gradient descent [@toh2010accelerated], singular value thresholding [@cai2010singular], Procrustes flow [@tu2016low], etc, have been introduced and both the theoretical and computational performances have been extensively studied. By similar proof of Theorem \[th:upper\_bound\_regression\], the following upper bound for matrix ISLET estimator $\widehat{{{\mathbf{A}}}}$ (Algorithm \[al:procedure\_regular\_matrix\] in the supplementary material [@ISLET-supplement]) $$\left\|\widehat{{{\mathbf{A}}}} - {{\mathbf{A}}}\right\|_F^2 \leq \frac{m}{n_2}\left(\sigma^2 + \frac{C_1\widetilde{\sigma}^4mp}{n_1^2\lambda_0^2}\right)\left(1 + C_2\sqrt{\frac{\log p}{m}} + C_3\sqrt{\frac{m\widetilde{\sigma}^2}{(n_1\wedge n_2)\lambda_0^2}}\right)$$ can be established with high probability. Here, $m=(p_1+p_2-r)r$, $\lambda_0 = \sigma_r({{\mathbf{A}}}), \widetilde{\sigma}^2 = \|{{\mathbf{A}}}\|_F^2+\sigma^2$. The lower bound similarly to Theorem \[th:lower-bound-regression\] also holds.
Sparse Tensor Regression with Importance Sketching {#sec:sparse-regression}
--------------------------------------------------
We further consider the simultaneously sparse and low-rank tensor regression with Gaussian ensemble design. We have the following theoretical guarantee for sparse ISLET. Due to the same reason as for regular ISLET (see Remark \[rm:sample-splitting\]), the sample splitting scheme is introduced in our technical analysis.
\[th:upper\_bound\_sparse\_tensor\_regression\] Consider the tensor regression model , where ${{\mathbfcal{A}}}$ is simultaneously low-rank and sparse , ${{\mathbfcal{X}}}_i$ has i.i.d. standard Gaussian entries, and $\varepsilon_i\overset{iid}{\sim}N(0, \sigma^2)$. Denote $\lambda_0 = \min_k \sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{A}}}))$, $s_k = p_k$ if $k\notin J_s$, $m_s = r_1r_2r_3 + \sum_{k\in J_s} s_k(r_k + \log p_k) + \sum_{k \notin J_s} p_kr_k$, and $\kappa = \max_k \|\mathcal{M}_k({{\mathbfcal{A}}})\|/\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{A}}}))$. We apply the proposed Algorithm \[al:procedure\_sparse\] with sample splitting scheme (see Remark \[rm:sample-splitting\]) and group Lasso penalty $\eta_k = C_0 \widetilde{\sigma}\sqrt{n_2(r_k + \log(p_k))}$. If $\log(p_1)\asymp \log(p_2)\asymp \log(p_3)\asymp\log (p)$, $$n_1 \geq \frac{C_1\kappa^2\widetilde{\sigma}^2}{\lambda_0^2}\left(s_1s_2s_3\log(p) + \sum_{k=1}^3(s_k^2r_k^2 + r_{k+1}^2r_{k+2}^2)\right),\quad n_2 \geq \frac{C_2 m_s\kappa^2\widetilde{\sigma}^2}{\lambda_0^2},$$ the output $\widehat{{{\mathbfcal{A}}}}$ of sparse ISLET satisfies $$\label{ineq:sparse-tensor-regression}
\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \leq \frac{C_3 m_s}{n_2}\left(\sigma^2 + \frac{C_4 m_s\kappa^2\widetilde{\sigma}^2}{n_1}\right)$$ with probability at least $1 - p^{-C}$.
See Appendix \[sec:proof\_upper\_bound\_sparse\_tensor\_regression\].
We further consider the following class of simultaneously sparse and low-rank tensors, $$\mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}, {\boldsymbol{s}}} = \left\{{{\mathbfcal{A}}}= \llbracket {{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket: {{\mathbf{U}}}_k\in \mathbb{O}_{p_k, r_k}, \|{{\mathbf{U}}}_k\|_{0,2} \leq s_k, k\in J_s \right\}.$$ The following minimax lower bound of the estimation risk holds in this class.
\[th:lower\_bound\_sparse\_tensor\_regression\] There exists constant $C>0$ such that whenever $m_s\geq C$, the following lower bound holds for any arbitrary estimator $\widehat{{{\mathbfcal{A}}}}$ based on $\{{{\mathbfcal{X}}}_i, y_i\}_{i=1}^n$, $$\inf_{{{\mathbfcal{A}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}, {\boldsymbol{s}}}}\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \frac{cm_s}{n}\sigma^2.$$
See Appendix \[sec:proof-lower\_bound\_sparse\_tensor\_regression\].
Combining Theorems \[th:upper\_bound\_sparse\_tensor\_regression\] and \[th:lower\_bound\_sparse\_tensor\_regression\], we can see the proposed procedure achieves optimal rate of convergence if $\frac{m_s\|{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2}{n_1\sigma^2} = O(1)$ and $n_2\asymp n$.
Numerical Analysis {#sec:numerical}
==================
In this section, we conduct a simulation study to investigate the numerical performance of ISLET. In each study, we construct sensing tensors $\boldsymbol{\mathcal{X}}_j\in \mathbb{R}^{p\times p \times p}$ with independent standard normal entries. In the non-sparse settings, using the Tucker decomposition we generate the core tensor ${{\mathbfcal{S}}}\in \mathbb{R}^{r\times r\times r}$ and ${{\mathbf{E}}}_k \in \mathbb{R}_{p, r}$ with i.i.d. Gaussian entries, the coefficient tensor ${{\mathbfcal{A}}}= \llbracket \boldsymbol{\mathcal{S}}; {{\mathbf{E}}}_1; {{\mathbf{E}}}_2; {{\mathbf{E}}}_3\rrbracket$; in the sparse settings, we construct ${{\mathbfcal{S}}}$ and ${{\mathbfcal{A}}}$ in the same way and generate ${{\mathbf{E}}}_k$ as $$\quad ({{\mathbf{E}}}_k)_{[i, :]}
= \left\{\begin{array}{ll}
(\bar{{{\mathbf{E}}}}_k)_{[j,:]}, & i\in \Omega_k, \text{ and $i$ is the $j$-th element of $\Omega_k$};\\
0, & i \notin \Omega_k,
\end{array}\right.$$ where $\Omega_k$ is a uniform random subset of $\{1,\ldots, p\}$ with cardinality $s_k$ and $\bar{{{\mathbf{E}}}}_k$ has $s_k$-by-$r$ i.i.d. Gaussian entries. Finally, let the response $y_j = \langle {{\mathbfcal{X}}}_j, {{\mathbfcal{A}}}\rangle + \varepsilon_j, j = 1, 2, \ldots, n$, where $\varepsilon_j \overset{iid}\sim N(0, \sigma^2)$. We report both the average root mean squared error (RMSE) $\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}||_{{{\rm HS}}}/ ||{{\mathbfcal{A}}}\|_{{{\rm HS}}}$ and the run time for each setting. Unless otherwise noted, the reported results are based on the average of 100 repeats and on a computer with Intel Xeon E5-2680 2.50GHz CPU. Additional simulation results of tuning-free ISLET and approximate low-rank tensor regression are collected in Sections \[sec:tuning\] and \[sec:additional-simu\] in the supplementary material [@ISLET-supplement].
Since we proposed to evaluate sketching directions and dimension-reduced regression (Steps 1 and 2 of Algorithm \[al:procedure\_regular\]) both using the complete sample, but introduced a sample splitting scheme (Remark \[rm:sample-splitting\]) to prove Theorems \[th:upper\_bound\_regression\] and \[th:upper\_bound\_sparse\_tensor\_regression\], we investigate how the sample splitting scheme affects the numerical performance of ISLET in this simulation setting. Let $n$ vary from 1000 to 4000, $p = 10$, $r = 3, 5$, $\sigma = 5$. In addition to the original ISLET without splitting, we also implement sample-splitting ISLET, where a random $n_1 \approx \{\frac{3}{10}n, \frac{4}{10}n, \frac{5}{10}n\}$ samples are allocated for importance direction estimation (Step 1 of ISLET) and $n-n_1$ are allocated for dimension-reduced regression (Step 2 of ISLET). The results plotted in Figure \[fig: 1\] clearly show that the no-sample-splitting scheme yields much smaller estimation error than all sample-splitting approaches. Although the sample splitting scheme brings advantages for our theoretical analyses for ISLET, it is not necessary in practice. Therefore, we will only perform ISLET without sample splitting for the rest of the simulation studies.
We also compare the performance of non-sparse ISLET with a number of contemporary methods, including non-convex projected gradient descent (non-convex PGD) [@chen2016non], Tucker low-rank regression via alternating gradient descent (Tucker regression)[^1] [@li2013tucker; @zhou2013tensor], and convex regularization low-rank tensor recovery (convex regularization)[^2] [@liu2013tensor; @raskutti2015convex; @tomioka2011statistical]. We implement all four methods for $p=10$, but only the ISLET and non-convex projected PGD for $p=50$, as the time cost of Tucker regression and convex regularization are beyond our computational limit if $p=50$. Results for $p=10$ and $p=50$ are respectively plotted in Panels (a)(b) and Panels (c)(d) of Fig. \[fig: 4\]. Plots in Fig. \[fig: 4\] (a) and (c) show that the RMSEs of ISLET, tucker tensor regression and non-convex PGD are close, and all of them are slightly better than the convex regularization method; Figure \[fig: 4\] (b) and (d) further indicate that ISLET is much faster than other methods – the advantage significantly increases as $n$ and $p$ grow. In particular, ISLET is about 10 times faster than non-convex PGD when $p = 50, n = 12000$. In summary, the proposed ISLET achieves similar statistical performance within in a significantly shorter time period comparing to the other state-or-the-art methods.
Next, we investigate the performance of ISLET when $p$ and $n$ substantially grow. Let $p = 100, 150, 200$, $r = 3, 5$, $n\in [8000, 20000]$. The results in RMSE and run time are shown in Fig. \[fig: 9\] (a), (b), (c), and (d), respectively. We can see that the estimation error significantly decays as the sample size $n$ grows, the dimension $p$ decreases, or the Tucker rank $r$ decreases.
We further fix $r = 2, n=30000$ and let $p$ grow to 400. Now the space cost for storing $\{{{\mathbfcal{X}}}_i\}_{i=1}^n$ reaches $400^3\times 30000\times 4\text{bytes} = 7.68$ terabytes, which is far beyond the volume of most personal computing devices. Since each sample is used only twice in ISLET, we perform this experiment in a parallel way. To be specific, in each machine $b = 1,\ldots, 40$, we store the random seed, draw pseudo random tensor ${{\mathbfcal{X}}}_{bi}$, evaluate $y_{bi}$ and $\widetilde{{{\mathbfcal{A}}}}_b$ by the procedure in Section \[sec:regular-islet-procedure\], and clean up the memory of ${{\mathbfcal{X}}}_{bi}$. After synchronizing the outcomes and obtaining the importance sketching directions, for each machine $b=1,\ldots, 40$, we generate pseudo-random covariates ${{\mathbfcal{X}}}_{bi}$ again using the stored random seeds, evaluate $\widetilde{{{\mathbf{G}}}}_b$ and $\widetilde{{{\mathbf{X}}}}_{bi}$ by -, and clean up the memory of ${{\mathbfcal{X}}}_{bi}$ again. The rest of the procedure follows from Section \[sec:regular-islet-procedure\] and the original ISLET in Algorithm \[al:procedure\_regular\]. The average RMSE and run time for five repeats are shown in Figure \[fig: 14\]. We clearly see that ISLET yields good statistical performance within a reasonable amount of time, while the other contemporary methods can hardly do so in such a ultrahigh-dimensional setting.
In addition, we explore the numerical performance of ISLET for simultaneously sparse and low-rank tensor regression. To perform sparse ISLET (Algorithm \[al:procedure\_sparse\]), we apply the *gglasso* package[^3] [@yang2015fast] for group Lasso and penalty level selection. Let $n$ vary from 1500 to 4000, $p = 20, 25, 30$, $r = 3, 5$, $\sigma = 5$, $s = s_1 = s_2 = s_3 = 8$. The result is shown in Fig. \[fig: 10\]. Similar to the non-sparse ISLET, as sample size $n$ increases or Tucker rank $r$ decreases, the average estimation errors decrease.
We also compare sparse ISLET with slice-sparse non-convex PGD proposed by [@chen2016non]. Let $n \in [5000, 12000]$, $p = 50$, $r = 3, 5$, $\sigma = 5$, $s_1 = s_2 = s_3 = 15$. From Fig. \[fig: 12\], we can see that ISLET yields much smaller estimation error with significantly shorter time than non-convex PGD – the difference between two algorithms becomes more significant as $n$ grows.
Finally, if the tensor is of order $2$, tensor regression becomes the classic *low-rank matrix recovery* problem [@candes2011tight; @recht2010guaranteed]. Among existing approaches for low-rank matrix recovery, the nuclear norm minimization (NNM) has been proposed and extensively studied in recent literature. We compare the numerical performance of matrix ISLET (see Algorithm \[al:procedure\_regular\_matrix\] in Section \[sec:general-order\] for implementation details) and NNM that aims to solve [^4] $$\sum_{i=1}^n (y_i - \langle {{\mathbf{X}}}_i, {{\mathbf{A}}}\rangle )^2 + \lambda ||{{\mathbf{A}}}||_\ast,$$ where $\|{{\mathbf{A}}}\|_\ast = \sum_i \sigma_i({{\mathbf{A}}})$ is the matrix nuclear norm. We consider two specific settings: (1) $p_1 = p_2 = 50$, $r = 2$, $\sigma = 10$, $n \in [2000, 16000]$; (2) $p_1 = p_2 = 100, r = 4, \sigma = 10, n \in [2000, 28000]$. From Figure \[fig: 5\], we find that ISLET has similar, or sometimes even better performance than NNM in estimation error. On the other hand, the run time of ISLET is negligibly small compared to NNM.
Discussion {#sec:discussion}
==========
In this article, we develop a general importance sketching algorithm for high-dimensional low-rank tensor regression. In particular, to sufficiently reduce the dimension of the higher-order structure, we propose a fast algorithm named *mportance ketching ow-rank stimation for ensors* (ISLET). The proposed algorithm includes three major steps: we first apply tensor decomposition approaches, such as HOOI and STAT-SVD, to obtain importance sketching directions; then we perform regression using the sketched tensor/matrices (in the sparse case, we add group-sparsity regularizers); finally we assemble the final estimator. We establish deterministic oracle inequalities for the proposed procedure under general design and noise distributions. We also prove that ISLET achieves optimal mean-squared error rate under Gaussian ensemble design – regular ISLET can further achieves the optimal constant for mean-squared error. As illustrated in simulation studies, the proposed procedure is computationally efficient comparing to contemporary methods. Although the presentation mainly focuses on order-3 tensors here, the method and theory for the general order-$d$ tensors can be elaborated similarly.
It is also noteworthy that the storage cost for Tucker decomposition in the proposed procedure grows exponentially with the order $d$. Thus, if the target tensor has a large order, it is more desirable to consider other low-rank approximation methods than Tucker, such as the CP decomposition [@bousse2017linear; @bousse2018linear], Hierarchical Tucker (HT) decomposition [@ballani2013projection; @grasedyck2010hierarchical; @hackbusch2009new], and Tensor Train (TT) decomposition [@oseledets2011tensor; @oseledets2009breaking], etc. The ISLET framework can be adapted to these structures as long as there are two key components: there exists a sketching approach for dimension reduction and a computational inversion step for embedding the low-dimensional estimate back to the high-dimensional space (also see Section \[sec:sketching-perspective\]). Whether these components hold for the previously described methods remains an interesting open question.
In addition to low-rank tensor regression, the idea of ISLET can be applied to various other high-dimensional problems. First, *high-order interaction pursuit* is an important topic in high-dimensional statistics that aims at the interaction among three or more variables in the regression setting. This problem can be transformed to the tensor estimation based on a number of rank-1 projections by the argument in [@hao2018sparse]. Similarly to analysis on tensor regression in this paper, the idea of ISLET can be used to develop an optimal and efficient procedure for high-order interaction pursuit with provable advantages over other baseline methods.
In addition, *matrix/tensor completion* has attracted significant attention in the recent literature [@candes2010power; @liu2013tensor; @xia2017polynomial; @xia2017statistically; @yuan2014tensor]. The central task of matrix/tensor completion is to complete the low-rank matrix/tensor based on a limited number of observable entries. Since each observable entry in matrix/tensor completion can be seen as a special rank-one projection of the original matrix/tensor, the idea behind ISLET can be used to achieve a more efficient algorithm in matrix/tensor completion with theoretical guarantees. It will be an interesting future topic to further investigate the performance of ISLET on other high-dimensional problems.
[100]{}
Genevera I Allen. Regularized tensor factorizations and higher-order principal components analysis. , 2012.
Animashree Anandkumar, Rong Ge, Daniel Hsu, Sham M Kakade, and Matus Telgarsky. Tensor decompositions for learning latent variable models. , 15(1):2773–2832, 2014.
Haim Avron, Kenneth L Clarkson, and David P Woodruff. Sharper bounds for regression and low-rank approximation with regularization. , 6, 2016.
Haim Avron, Huy Nguyen, and David Woodruff. Subspace embeddings for the polynomial kernel. In [*Advances in Neural Information Processing Systems*]{}, pages 2258–2266, 2014.
Krishnakumar Balasubramanian, Jianqing Fan, and Zhuoran Yang. Tensor methods for additive index models under discordance and heterogeneity. , 2018.
Nicolai Baldin and Quentin Berthet. Optimal link prediction with matrix logistic regression. , 2018.
Jonas Ballani and Lars Grasedyck. A projection method to solve linear systems in tensor format. , 20(1):27–43, 2013.
Frank Ban, Vijay Bhattiprolu, Karl Bringmann, Pavel Kolev, Euiwoong Lee, and David P Woodruff. A ptas for $\ell_p$-low rank approximation. In [*Proceedings of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms*]{}, pages 747–766. SIAM, 2019.
Mario Bebendorf. Adaptive cross approximation of multivariate functions. , 34(2):149–179, 2011.
Gregory Beylkin and Martin J Mohlenkamp. Algorithms for numerical analysis in high dimensions. , 26(6):2133–2159, 2005.
Xuan Bi, Annie Qu, and Xiaotong Shen. Multilayer tensor factorization with applications to recommender systems. , 46(6B):3308–3333, 2018.
M Bouss[é]{}, I Domanov, and L De Lathauwer. Linear systems with a multilinear singular value decomposition constrained solution. , 2017.
Martijn Bouss[é]{}, Nico Vervliet, Ignat Domanov, Otto Debals, and Lieven De Lathauwer. Linear systems with a canonical polyadic decomposition constrained solution: Algorithms and applications. , 25(6):e2190, 2018.
Christos Boutsidis and David P Woodruff. Optimal cur matrix decompositions. , 46(2):543–589, 2017.
Jian-Feng Cai, Emmanuel J Cand[è]{}s, and Zuowei Shen. A singular value thresholding algorithm for matrix completion. , 20(4):1956–1982, 2010.
T Tony Cai, Xiaodong Li, and Zongming Ma. Optimal rates of convergence for noisy sparse phase retrieval via thresholded wirtinger flow. , 44(5):2221–2251, 2016.
T Tony Cai and Anru Zhang. Sparse representation of a polytope and recovery of sparse signals and low-rank matrices. , 60(1):122–132, 2014.
T Tony Cai and Anru Zhang. : Matrix recovery via rank-one projections. , 43(1):102–138, 2015.
T Tony Cai and Anru Zhang. Rate-optimal perturbation bounds for singular subspaces with applications to high-dimensional statistics. , 46(1):60–89, 2018.
Tianxi Cai, T. Tony Cai, and Anru Zhang. Structured matrix completion with applications to genomic data integration. , 111(514):621–633, 2016.
Cesar F Caiafa and Andrzej Cichocki. Generalizing the column–row matrix decomposition to multi-way arrays. , 433(3):557–573, 2010.
Raffaello Camoriano, Tom[á]{}s Angles, Alessandro Rudi, and Lorenzo Rosasco. Nytro: When subsampling meets early stopping. In [*Artificial Intelligence and Statistics*]{}, pages 1403–1411, 2016.
Emmanuel J Candes, Xiaodong Li, and Mahdi Soltanolkotabi. Phase retrieval via wirtinger flow: Theory and algorithms. , 61(4):1985–2007, 2015.
Emmanuel J Candes and Yaniv Plan. Matrix completion with noise. , 98(6):925–936, 2010.
Emmanuel J Candes and Yaniv Plan. Tight oracle inequalities for low-rank matrix recovery from a minimal number of noisy random measurements. , 57(4):2342–2359, 2011.
Emmanuel J Candes and Terence Tao. Decoding by linear programming. , 51(12):4203–4215, 2005.
Emmanuel J Cand[è]{}s and Terence Tao. The power of convex relaxation: Near-optimal matrix completion. , 56(5):2053–2080, 2010.
Moses Charikar, Kevin Chen, and Martin Farach-Colton. Finding frequent items in data streams. In [*International Colloquium on Automata, Languages, and Programming*]{}, pages 693–703. Springer, 2002.
Han Chen, Garvesh Raskutti, and Ming Yuan. Non-convex projected gradient descent for generalized low-rank tensor regression. , 2016.
Yuxin Chen, Yuejie Chi, and Andrea J Goldsmith. Exact and stable covariance estimation from quadratic sampling via convex programming. , 61(7):4034–4059, 2015.
Flavio Chierichetti, Sreenivas Gollapudi, Ravi Kumar, Silvio Lattanzi, Rina Panigrahy, and David P Woodruff. Algorithms for $\ell_p$ low-rank approximation. In [*Proceedings of the 34th International Conference on Machine Learning-Volume 70*]{}, pages 806–814. JMLR. org, 2017.
Andrzej Cichocki, Danilo Mandic, Lieven De Lathauwer, Guoxu Zhou, Qibin Zhao, Cesar Caiafa, and Huy Anh Phan. Tensor decompositions for signal processing applications: From two-way to multiway component analysis. , 32(2):145–163, 2015.
Kenneth L Clarkson and David P Woodruff. Input sparsity and hardness for robust subspace approximation. In [*2015 IEEE 56th Annual Symposium on Foundations of Computer Science*]{}, pages 310–329. IEEE, 2015.
Kenneth L Clarkson and David P Woodruff. Low-rank approximation and regression in input sparsity time. , 63(6):54, 2017.
Gautam Dasarathy, Parikshit Shah, Badri Narayan Bhaskar, and Robert D Nowak. Sketching sparse matrices, covariances, and graphs via tensor products. , 61(3):1373–1388, 2015.
Lieven De Lathauwer, Bart De Moor, and Joos Vandewalle. On the best rank-1 and rank-(r 1, r 2,..., rn) approximation of higher-order tensors. , 21(4):1324–1342, 2000.
Huaian Diao, Zhao Song, Wen Sun, and David Woodruff. Sketching for kronecker product regression and p-splines. In [*International Conference on Artificial Intelligence and Statistics*]{}, pages 1299–1308, 2018.
Edgar Dobriban and Sifan Liu. A new theory for sketching in linear regression. , 2018.
Petros Drineas, Malik Magdon-Ismail, Michael W Mahoney, and David P Woodruff. Fast approximation of matrix coherence and statistical leverage. , 13(Dec):3475–3506, 2012.
Petros Drineas and Michael W Mahoney. Effective resistances, statistical leverage, and applications to linear equation solving. , 2010.
Lars Eld[é]{}n and Berkant Savas. A newton–grassmann method for computing the best multilinear rank-(r\_1, r\_2, r\_3) approximation of a tensor. , 31(2):248–271, 2009.
Mike Espig, Wolfgang Hackbusch, Thorsten Rohwedder, and Reinhold Schneider. Variational calculus with sums of elementary tensors of fixed rank. , 122(3):469–488, 2012.
Jianqing Fan, Wenyan Gong, and Ziwei Zhu. Generalized high-dimensional trace regression via nuclear norm regularization. , 2017.
Jianqing Fan and Jinchi Lv. Sure independence screening for ultrahigh dimensional feature space. , 70(5):849–911, 2008.
Jianqing Fan, Weichen Wang, and Ziwei Zhu. A shrinkage principle for heavy-tailed data: High-dimensional robust low-rank matrix recovery. , 2016.
Jerome Friedman, Trevor Hastie, and Robert Tibshirani. A note on the group lasso and a sparse group lasso. , 2010.
Rong Ge, Furong Huang, Chi Jin, and Yang Yuan. Escaping from saddle points—online stochastic gradient for tensor decomposition. In [*Conference on Learning Theory*]{}, pages 797–842, 2015.
Irina Georgieva and Clemens Hofreither. Greedy low-rank approximation in tucker format of solutions of tensor linear systems. , 358:206–220, 2019.
SA Goreinov, Ivan V Oseledets, and Dmitry V Savostyanov. Wedderburn rank reduction and krylov subspace method for tensor approximation. part 1: Tucker case. , 34(1):A1–A27, 2012.
Lars Grasedyck. Hierarchical singular value decomposition of tensors. , 31(4):2029–2054, 2010.
Lars Grasedyck, Daniel Kressner, and Christine Tobler. A literature survey of low-rank tensor approximation techniques. , 36(1):53–78, 2013.
Rajarshi Guhaniyogi, Shaan Qamar, and David B Dunson. Bayesian tensor regression. , 2015.
Weiwei Guo, Irene Kotsia, and Ioannis Patras. Tensor learning for regression. , 21(2):816–827, 2012.
Wolfgang Hackbusch and Stefan K[ü]{}hn. A new scheme for the tensor representation. , 15(5):706–722, 2009.
Botao Hao, Anru Zhang, and Guang Cheng. Sparse and low-rank tensor estimation via cubic sketchings. , 2018.
Jarvis Haupt, Xingguo Li, and David P Woodruff. Near optimal sketching of low-rank tensor regression. , 2017.
Shiyuan He, Jianxin Yin, Hongzhe Li, and Xing Wang. Graphical model selection and estimation for high dimensional tensor data. , 128:165–185, 2014.
Peter D Hoff. Multilinear tensor regression for longitudinal relational data. , 9(3):1169, 2015.
Clemens Hofreither. A black-box low-rank approximation algorithm for fast matrix assembly in isogeometric analysis. , 333:311–330, 2018.
Thomas JR Hughes, John A Cottrell, and Yuri Bazilevs. Isogeometric analysis: Cad, finite elements, nurbs, exact geometry and mesh refinement. , 194(39-41):4135–4195, 2005.
Mariya Ishteva, P-A Absil, Sabine Van Huffel, and Lieven De Lathauwer. Best low multilinear rank approximation of higher-order tensors, based on the riemannian trust-region scheme. , 32(1):115–135, 2011.
Mariya Ishteva, Lieven De Lathauwer, P-A Absil, and Sabine Van Huffel. Differential-geometric newton method for the best rank-(r 1, r 2, r 3) approximation of tensors. , 51(2):179–194, 2009.
Majid Janzamin, Hanie Sedghi, and Anima Anandkumar. Score function features for discriminative learning: Matrix and tensor framework. , 2014.
Daniel M Kane and Jelani Nelson. Sparser johnson-lindenstrauss transforms. , 61(1):4, 2014.
Tamara G Kolda and Brett W Bader. Tensor decompositions and applications. , 51(3):455–500, 2009.
Tamara Gibson Kolda. , volume 2. United States. Department of Energy, 2006.
Vladimir Koltchinskii. A remark on low rank matrix recovery and noncommutative bernstein type inequalities. In [*From Probability to Statistics and Back: High-Dimensional Models and Processes–A Festschrift in Honor of Jon A. Wellner*]{}, pages 213–226. Institute of Mathematical Statistics, 2013.
Vladimir Koltchinskii, Karim Lounici, and Alexandre B Tsybakov. Nuclear-norm penalization and optimal rates for noisy low-rank matrix completion. , 39(5):2302–2329, 2011.
Daniel Kressner, Michael Steinlechner, and Bart Vandereycken. Preconditioned low-rank riemannian optimization for linear systems with tensor product structure. , 38(4):A2018–A2044, 2016.
Daniel Kressner and Christine Tobler. Krylov subspace methods for linear systems with tensor product structure. , 31(4):1688–1714, 2010.
Pieter M Kroonenberg. , volume 702. John Wiley & Sons, 2008.
Beatrice Laurent and Pascal Massart. Adaptive estimation of a quadratic functional by model selection. , pages 1302–1338, 2000.
Jason D Lee, Ben Recht, Nathan Srebro, Joel Tropp, and Ruslan R Salakhutdinov. Practical large-scale optimization for max-norm regularization. In [*Advances in Neural Information Processing Systems*]{}, pages 1297–1305, 2010.
Erich L Lehmann and George Casella. . Springer Science & Business Media, 2006.
Lexin Li and Xin Zhang. Parsimonious tensor response regression. , pages 1–16, 2017.
Nan Li and Baoxin Li. Tensor completion for on-board compression of hyperspectral images. In [*2010 IEEE International Conference on Image Processing*]{}, pages 517–520. IEEE, 2010.
Xiaoshan Li, Da Xu, Hua Zhou, and Lexin Li. Tucker tensor regression and neuroimaging analysis. , pages 1–26, 2018.
Ji Liu, Przemyslaw Musialski, Peter Wonka, and Jieping Ye. Tensor completion for estimating missing values in visual data. , 35(1):208–220, 2013.
Karim Lounici, Massimiliano Pontil, Sara Van De Geer, and Alexandre B Tsybakov. Oracle inequalities and optimal inference under group sparsity. , 39(4):2164–2204, 2011.
RE Lynch, JOHN R Rice, and DONALD H Thomas. Tensor product analysis of partial difference equations. , 70(3):378–384, 1964.
Xiang Lyu, Will Wei Sun, Zhaoran Wang, Han Liu, Jian Yang, and Guang Cheng. Tensor graphical model: Non-convex optimization and statistical inference. , 2019.
Michael W Mahoney. Randomized algorithms for matrices and data. , 3(2):123–224, 2011.
Michael W Mahoney, Mauro Maggioni, and Petros Drineas. Tensor-cur decompositions for tensor-based data. , 30(3):957–987, 2008.
Ameur M Manceur and Pierre Dutilleul. Maximum likelihood estimation for the tensor normal distribution: Algorithm, minimum sample size, and empirical bias and dispersion. , 239:37–49, 2013.
Panos P Markopoulos, George N Karystinos, and Dimitris A Pados. Optimal algorithms for $ l\_ ${$1$}$ $-subspace signal processing. , 62(19):5046–5058, 2014.
Panos P Markopoulos, Sandipan Kundu, Shubham Chamadia, and Dimitris A Pados. Efficient l1-norm principal-component analysis via bit flipping. , 65(16):4252–4264, 2017.
Pascal Massart. . Springer, 2007.
Deyu Meng, Zongben Xu, Lei Zhang, and Ji Zhao. A cyclic weighted median method for l1 low-rank matrix factorization with missing entries. In [*Twenty-Seventh AAAI Conference on Artificial Intelligence*]{}, 2013.
Xiangrui Meng and Michael W Mahoney. Low-distortion subspace embeddings in input-sparsity time and applications to robust linear regression. In [*Proceedings of the forty-fifth annual ACM symposium on Theory of computing*]{}, pages 91–100. ACM, 2013.
Andrea Montanari and Nike Sun. Spectral algorithms for tensor completion. , 2016.
Cun Mu, Bo Huang, John Wright, and Donald Goldfarb. Square deal: Lower bounds and improved relaxations for tensor recovery. In [*ICML*]{}, pages 73–81, 2014.
Jelani Nelson and Huy L Nguy[ê]{}n. Osnap: Faster numerical linear algebra algorithms via sparser subspace embeddings. In [*2013 IEEE 54th Annual Symposium on Foundations of Computer Science*]{}, pages 117–126. IEEE, 2013.
Ivan V Oseledets. Tensor-train decomposition. , 33(5):2295–2317, 2011.
Ivan V Oseledets, DV Savostianov, and Eugene E Tyrtyshnikov. Tucker dimensionality reduction of three-dimensional arrays in linear time. , 30(3):939–956, 2008.
Ivan V Oseledets, Dmitry V Savostyanov, and Eugene E Tyrtyshnikov. Cross approximation in tensor electron density computations. , 17(6):935–952, 2010.
Ivan V Oseledets and Eugene E Tyrtyshnikov. Breaking the curse of dimensionality, or how to use svd in many dimensions. , 31(5):3744–3759, 2009.
Rasmus Pagh. Compressed matrix multiplication. , 5(3):9, 2013.
Yuqing Pan, Qing Mai, and Xin Zhang. Covariate-adjusted tensor classification in high dimensions. , pages 1–15, 2018.
Ninh Pham and Rasmus Pagh. Fast and scalable polynomial kernels via explicit feature maps. In [*Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining*]{}, pages 239–247. ACM, 2013.
Mert Pilanci and Martin J Wainwright. Randomized sketches of convex programs with sharp guarantees. , 61(9):5096–5115, 2015.
Mert Pilanci and Martin J Wainwright. Iterative hessian sketch: Fast and accurate solution approximation for constrained least-squares. , 17(1):1842–1879, 2016.
Garvesh Raskutti and Michael Mahoney. A statistical perspective on randomized sketching for ordinary least-squares. , 2014.
Garvesh Raskutti, Ming Yuan, and Han Chen. Convex regularization for high-dimensional multi-response tensor regression. , 2015.
Benjamin Recht, Maryam Fazel, and Pablo A Parrilo. Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization. , 52(3):471–501, 2010.
Berkant Savas and Lars Eld[é]{}n. Krylov-type methods for tensor computations i. , 438(2):891–918, 2013.
Berkant Savas and Lek-Heng Lim. Quasi-newton methods on grassmannians and multilinear approximations of tensors. , 32(6):3352–3393, 2010.
Nicholas D Sidiropoulos, Lieven De Lathauwer, Xiao Fu, Kejun Huang, Evangelos E Papalexakis, and Christos Faloutsos. Tensor decomposition for signal processing and machine learning. , 65(13):3551–3582, 2017.
Nicholas D Sidiropoulos and Anastasios Kyrillidis. Multi-way compressed sensing for sparse low-rank tensors. , 19(11):757–760, 2012.
Nicholas D Sidiropoulos, Evangelos E Papalexakis, and Christos Faloutsos. Parallel randomly compressed cubes: A scalable distributed architecture for big tensor decomposition. , 31(5):57–70, 2014.
Zhao Song, David P Woodruff, and Peilin Zhong. Low rank approximation with entrywise l 1-norm error. In [*Proceedings of the 49th Annual ACM SIGACT Symposium on Theory of Computing*]{}, pages 688–701. ACM, 2017.
Zhao Song, David P Woodruff, and Peilin Zhong. Relative error tensor low rank approximation. In [*Proceedings of the Thirtieth Annual ACM-SIAM Symposium on Discrete Algorithms*]{}, pages 2772–2789. Society for Industrial and Applied Mathematics, 2019.
Will Wei Sun and Lexin Li. Sparse low-rank tensor response regression. , 2016.
Will Wei Sun and Lexin Li. Store: sparse tensor response regression and neuroimaging analysis. , 18(1):4908–4944, 2017.
Yiming Sun, Yang Guo, Charlene Luo, Joel Tropp, and Madeleine Udell. Low-rank tucker approximation of a tensor from streaming data. , 2019.
Kim-Chuan Toh and Sangwoon Yun. An accelerated proximal gradient algorithm for nuclear norm regularized linear least squares problems. , 6(615-640):15, 2010.
Ryota Tomioka and Taiji Suzuki. Convex tensor decomposition via structured schatten norm regularization. In [*Advances in neural information processing systems*]{}, pages 1331–1339, 2013.
Ryota Tomioka, Taiji Suzuki, Kohei Hayashi, and Hisashi Kashima. Statistical performance of convex tensor decomposition. In [*Advances in Neural Information Processing Systems*]{}, pages 972–980, 2011.
Joel A Tropp, Alp Yurtsever, Madeleine Udell, and Volkan Cevher. Practical sketching algorithms for low-rank matrix approximation. , 38(4):1454–1485, 2017.
Stephen Tu, Ross Boczar, Max Simchowitz, Mahdi Soltanolkotabi, and Ben Recht. Low-rank solutions of linear matrix equations via procrustes flow. In [*International Conference on Machine Learning*]{}, pages 964–973, 2016.
Ledyard R Tucker. Some mathematical notes on three-mode factor analysis. , 31(3):279–311, 1966.
Madeleine Udell and Alex Townsend. Why are big data matrices approximately low rank? , 1(1):144–160, 2019.
Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. , 2010.
Nico Vervliet and Lieven De Lathauwer. A randomized block sampling approach to canonical polyadic decomposition of large-scale tensors. , 10(2):284–295, 2015.
Jialei Wang, Jason D Lee, Mehrdad Mahdavi, Mladen Kolar, Nathan Srebro, et al. Sketching meets random projection in the dual: A provable recovery algorithm for big and high-dimensional data. , 11(2):4896–4944, 2017.
Yining Wang, Hsiao-Yu Tung, Alexander J Smola, and Anima Anandkumar. Fast and guaranteed tensor decomposition via sketching. In [*Advances in Neural Information Processing Systems*]{}, pages 991–999, 2015.
David P Woodruff. Sketching as a tool for numerical linear algebra. , 10(1–2):1–157, 2014.
Dong Xia and Ming Yuan. On polynomial time methods for exact low rank tensor completion. , 2017.
Dong Xia, Ming Yuan, and Cun-Hui Zhang. Statistically optimal and computationally efficient low rank tensor completion from noisy entries. , 2017.
Lingzhou Xue and Hui Zou. Sure independence screening and compressed random sensing. , pages 371–380, 2011.
Dan Yang, Zongming Ma, and Andreas Buja. A sparse singular value decomposition method for high-dimensional data. , 23(4):923–942, 2014.
Yi Yang and Hui Zou. A fast unified algorithm for solving group-lasso penalize learning problems. , 25(6):1129–1141, 2015.
Ming Yu, Zhaoran Wang, Varun Gupta, and Mladen Kolar. Recovery of simultaneous low rank and two-way sparse coefficient matrices, a nonconvex approach. , 2018.
Ming Yuan and Yi Lin. Model selection and estimation in regression with grouped variables. , 68(1):49–67, 2006.
Ming Yuan and Cun-Hui Zhang. On tensor completion via nuclear norm minimization. , pages 1–38, 2014.
Anru Zhang. Cross: Efficient low-rank tensor completion. , 47(2):936–964, 2019.
Anru Zhang and Rungang Han. Optimal sparse singular value decomposition for high-dimensional high-order data. , page to appear, 2018.
Anru Zhang, Yuetian Luo, Garvesh Raskutti, and Ming Yuan. Supplement to “[ISLET]{}: Fast and optimal low-rank tensor regression via importance sketching", 2018.
Anru Zhang, Yuetian Luo, Garvesh Raskutti, and Ming Yuan. A sharp blockwise tensor perturbation bound for higher-order orthogonal iteration. , 2019.
Anru Zhang and Dong Xia. Tensor [SVD]{}: Statistical and computational limits. , 64(11):7311–7338, 2018.
Lijun Zhang, Mehrdad Mahdavi, Rong Jin, Tianbao Yang, and Shenghuo Zhu. Random projections for classification: A recovery approach. , 60(11):7300–7316, 2014.
Yinqiang Zheng, Guangcan Liu, Shigeki Sugimoto, Shuicheng Yan, and Masatoshi Okutomi. Practical low-rank matrix approximation under robust l 1-norm. In [*2012 IEEE Conference on Computer Vision and Pattern Recognition*]{}, pages 1410–1417. IEEE, 2012.
Hua Zhou. Matlab tensorreg toolbox version 1.0, 2017. Available online at https://hua-zhou.github.io/TensorReg/.
Hua Zhou, Lexin Li, and Hongtu Zhu. Tensor regression with applications in neuroimaging data analysis. , 108(502):540–552, 2013.
Shuheng Zhou. Gemini: Graph estimation with matrix variate normal instances. , 42(2):532–562, 2014.
[Supplement to “ISLET: Fast and Optimal Low-rank Tensor]{}
[Anru Zhang, Yuetian Luo, Garvesh Raskutti, and Ming Yuan]{}
Additional Notation and Preliminaries {#sec:row-permutation-operator}
=====================================
To conveniently specify the dimensions of tensors, for an order-$d$ tensor ${{\mathbfcal{A}}}$ with dimensions $p_1\times \cdots \times p_d$, we denote $p_{-k} = p_1\cdots p_d/p_k$ for $k=1,\ldots, d$. Then the mode-$k$ matricization of ${{\mathbfcal{A}}}$, denoted as $\mathcal{M}_k({{\mathbfcal{A}}})$, has dimension $p_k\times p_{-k}$. For any matrix ${{\mathbf{D}}}\in \mathbb{R}^{p_1\times p_2}$ and order-$d$ tensor ${{\mathbfcal{A}}}$, we formally define the vectorization as $$\begin{split}
& {{\rm vec}}({{\mathbf{D}}}) \in \mathbb{R}^{(p_1p_2)}, \quad {{\rm vec}}({{\mathbf{D}}})_{[i_1 + (i_2-1)p_1]} = {{\mathbf{D}}}_{[i_1,i_2]}; \\
& {{\rm vec}}({{\mathbfcal{A}}})\in \mathbb{R}^{(p_1\cdots p_d)}, \quad {{\rm vec}}({{\mathbfcal{A}}})_{[i_1+p_1(i_2-1)+\cdots + (i_d-1)p_1\cdots p_d]} = {{\mathbfcal{A}}}_{[i_1, \ldots, i_d]}.
\end{split}$$ For any tensor ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times \cdots \times p_d}$, the Mode-$k$ matricization is formally defined as $$\mathcal{M}_k({{\mathbfcal{A}}}) \in \mathbb{R}^{p_k\times p_{-k}}, \quad {{\mathbfcal{A}}}_{[i_1,\ldots, i_d]} = \left(\mathcal{M}_k({{\mathbfcal{A}}})\right)_{\left[i_k, j\right]}, \quad j = 1 + \sum_{\substack{l=1\\l\neq k}}^d\left\{(i_l-1)\prod_{\substack{m=1\\m\neq k}}^{l-1}p_m\right\}$$ for any $1\leq i_l \leq p_l, l=1,\ldots, d$. Also see [@kolda2009tensor Section 2.4] for more discussions on tensor matricizations.
In order to better illustrate the proposed procedure, we have introduced a row-permutation operator $\mathcal{R}_k$ that matches the index of ${{\mathbf{W}}}_k\otimes {{\mathbf{V}}}_k$ to ${{\rm vec}}({{\mathbfcal{A}}})$. In particular if ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times p_2\times p_3}, {{\mathbf{W}}}_k \in \mathbb{R}^{p_{-k}\times r_k}, {{\mathbf{V}}}_k\in \mathbb{R}^{p_k\times r_k}$, $\mathcal{R}_k$ is defined as follows: $$\begin{split}
& \left(\mathcal{R}_1\left({{\mathbf{W}}}_1 \otimes {{\mathbf{V}}}_1\right)\right)_{[i_1 + (i_2-1)p_1+(i_3-1)p_1p_2, :]} = \left({{\mathbf{W}}}_1\otimes {{\mathbf{V}}}_1\right)_{[i_1 + (i_2-1)p_1+(i_3-1)p_1p_2,:]},\\
& \left(\mathcal{R}_2\left({{\mathbf{W}}}_2 \otimes {{\mathbf{V}}}_2\right)\right)_{[i_1 + (i_2-1)p_1+(i_3-1)p_1p_2, :]} = \left({{\mathbf{W}}}_2\otimes {{\mathbf{V}}}_2\right)_{[i_2+(i_1-1)p_2+(i_3-1)p_2p_1, :]},\\
& \left(\mathcal{R}_3\left({{\mathbf{W}}}_3 \otimes {{\mathbf{V}}}_3\right)\right)_{[i_1 + (i_2-1)p_1+(i_3-1)p_1p_2, :]} = \left({{\mathbf{W}}}_3\otimes {{\mathbf{V}}}_3\right)_{[i_3+(i_1-1)p_3+(i_2-1)p_1p_3, :]}
\end{split}$$ for $1\leq i_1\leq p_1, 1\leq i_2\leq p_2, 1\leq i_3\leq p_3$.
ADHD MRI Imaging Data Analysis {#sec:real_data}
==============================
In this section, we display the value of our method on predicting attention deficit hyperactivity disorder (ADHD) with magnetic resonance imaging (MRI) dataset provided by Neuro Bureau[^5]. The dataset involves 973 subjects, where each subject is associated with a $121$-by-$145$-by-$121$ MRI image and several demographic variables. After removing the missing values, we obtain 930 samples, among which 356 and 574 are diagnosed and control subjects, respectively.
We aim to do prediction based on the association between the diagnosis label $y_i$ of $i^{th}$ observation and its covariates with MRI imaging ${{\mathbfcal{X}}}_i$, demographic variables age $x^1_{i}$, gender $x^2_{i}$, and handedness $x^3_{i}$. To better cope the job of predicting binary response $y_i$ and incorporate the demographic information in addition to tensor image covariates, we apply importance sketching, the central idea of ISLET, for dimension reduction. The 5-fold cross-validation is applied to examine the prediction power. Specifically for $l=1,\ldots, 50$, we randomly partition all 930 subjects into 5 uniform subsets $\{\Omega_j^{(l)}\}_{j=1,\ldots, 5} \subseteq \{1,\ldots, 930\}$. For $j=1,\ldots, 5$, we assign one fold $\Omega_j^{(l)}$ and the other four folds $\Omega_{-j}^{(l)} = \cup_{j'\neq j}\Omega_{j'}$ as the testing and training sets, respectively. We apply Step 1 of sparse ISLET (described in Section \[sec:sparse-procedure\]) on $\{y_i, {{\mathbfcal{X}}}_i\}_{i\in \Omega_{(-j)}^{(l)}}$ to obtain $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3$ and construct the importance sketching covariates $\widetilde{{{\mathbf{x}}}}_i = {{\rm vec}}({{\mathbfcal{X}}}_i\times_1\widetilde{U}_1^\top,\times_1\widetilde{U}_1^\top,\times_1\widetilde{U}_1^\top)$, perform logistic regression for $y_{i}$ versus the combined covariates $\left[\widetilde{{{\mathbf{x}}}}_i, x^{1}_{i}, x^{2}_{i}, x^{3}_{i}\right]$, $i\in \Omega_{-j}^{(l)}$ and possible $\ell_1$ regularizer to get the estimates. Then we use estimates and $\left[\widetilde{{{\mathbf{x}}}}_i, x_{1}^{i}, x_{2}^{i}, x_{3}^{i}\right], i\in \Omega_{j}^{(l)}$ to predict the labels of samples in the testing set $\Omega_{j}^{(l)}$. For comparison, we also perform Tucker regression and Tucker regression with regularizer proposed by [@li2013tucker; @zhou2013tensor] under the same setting. Since it is computationally intensive to perform full Tucker regression on complete tensor covariates of dimension $121\times 145 \times 121$, we follow the procedure described in [@li2013tucker; @zhou2013tensor] and apply the discrete cosine transformation to downsize the MRI data to $12\times 14\times 12$ using the code available at the authors’ website [@zhou2017matlab]. For all methods, we input Tucker rank $(r,r,r)$ for $r = 3,4,5$ and other regularization tuning parameters selected via cross validation. We repeat experiments for $l=1,\ldots, 50, j=1,\ldots, 5$ and take average to ensure stable estimations of the prediction accuracy for both procedures.
The average prediction accuracy with standard deviation in the parenthesis and runtime for both methods are shown in Table \[tab: ADHD realdata\]. We can see the importance sketching method performs significantly better than Tucker regression in both the prediction accuracy and runtime for all different Tucker rank choices. Particularly for the importance sketching, adding $\ell_1$ regularizer provides more accurate prediction but costs more time. In addition, compared to the downsizing method by [@zhou2013tensor; @li2013tucker] that deterministically relies on external information, our importance sketching is fully data-driven. We can also see downsizing the tensor covariates to 3-by-3-by-3 by importance sketching provides more prediction power than downsizing to 12-by-14-by-12 by deterministic methods. This reveals the runtime advantage and immediately demonstrates the advantage of the proposed method over other state-of-the-art approaches.
------------ --- -------------- ------------------ -------------- ---------------
IS IS Tucker Reg. Tucker Reg.
+ regularizer + regularizer
Prediction 3 0.684(0.010) **0.686**(0.009) 0.624(0.014) 0.647(0.009)
Accuracy 4 0.673(0.009) **0.682**(0.008) 0.609(0.014) 0.648(0.007)
5 0.653(0.009) **0.674**(0.007) 0.591(0.015) 0.644(0.007)
Runtime 3 **0.008** 0.392 14.291 3.03
Unit: 4 **0.024** 1.003 22.088 5.761
seconds 5 **0.064** 3.339 33.392 13.710
------------ --- -------------- ------------------ -------------- ---------------
: Importance sketching (IS) vs. Tucker regression in prediction accuracy and runtime[]{data-label="tab: ADHD realdata"}
ISLET for General Order Tensor Estimation {#sec:general-order}
=========================================
For completeness, we provide the ISLET procedure for general order-$d$ low-rank tensor estimation in this section. The procedure for $d\geq 3$ is provided in Algorithms \[al:procedure\_regular\_general\_order\] and the one for $d = 2$ (i.e., the low-rank matrix estimation) is provided in Algorithm \[al:procedure\_regular\_matrix\]. The sparse versions for $d\geq 3$ and $d=2$ are provided in Algorithms \[al:procedure\_sparse\_general\_order\] and \[al:procedure\_sparse\_matrix\], respectively.
Input: $y_1,\ldots, y_n\in \mathbb{R}, {{\mathbfcal{X}}}_1,\ldots, {{\mathbfcal{X}}}_n \in \mathbb{R}^{p_1\times \cdots \times p_d}$, rank ${\boldsymbol{r}}= (r_1,\ldots, r_d)$. Evaluate $\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j {{\mathbfcal{X}}}_j.$ Apply order-$d$ HOOI on $\widetilde{{{\mathbfcal{A}}}}$ to obtain initial estimates $\widetilde{{{\mathbf{U}}}}_k, k =1,\ldots, d$. Let $\widetilde{{{\mathbfcal{S}}}} = \llbracket\widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \ldots, \widetilde{{{\mathbf{U}}}}_d^\top\rrbracket$. Evaluate the sketching directions, $$\widetilde{{{\mathbf{V}}}}_k = {\rm QR}\left[\mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}})^\top\right], \quad k=1,\ldots, d.$$ Construct $\widetilde{{{\mathbf{X}}}} = \left[\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1} ~ \cdots ~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_d}\right] \in \mathbb{R}^{n\times m}$, where $$\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times m_{{{\mathbfcal{B}}}}},\quad (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[i,:]} =
{{\rm vec}}\left({{\mathbfcal{X}}}_i \times_{l=1}^d \widetilde{{{\mathbf{U}}}}_l^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k} \in \mathbb{R}^{n\times m_{{{\mathbf{D}}}_k}}, \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k})_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{k\perp}^\top \mathcal{M}_k\left({{\mathbfcal{X}}}_{i} \times_{\substack{l=1\\l\neq k}}^d\widetilde{{{\mathbf{U}}}}_{l}^\top\right) \widetilde{{{\mathbf{V}}}}_k\right)
\end{split}$$ for $m_{{{\mathbfcal{B}}}} = r_1\cdots r_d, m_{{{\mathbf{D}}}_k} = (p_k-r_k)r_k$, $k=1,\ldots, d$, and $m = m_{{{\mathbfcal{B}}}} + m_{{{\mathbf{D}}}_1}+\cdots +m_{{{\mathbf{D}}}_d}$.
Solve $\widehat{{{\boldsymbol{\gamma}}}} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^m}\|y - \widetilde{{{\mathbf{X}}}}{{\boldsymbol{\gamma}}}\|_2^2$. Partition $\widehat{{{\boldsymbol{\gamma}}}}$ to $\widehat{{{\mathbfcal{B}}}}, \widehat{{{\mathbf{D}}}}_1, \ldots, \widehat{{{\mathbf{D}}}}_d$, $$\begin{split}
& {{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) := \widehat{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}}= \widehat{{{\boldsymbol{\gamma}}}}_{[1:m_{{{\mathbfcal{B}}}}]},\\
& {{\rm vec}}(\widehat{{{\mathbf{D}}}}_k) := \widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{D}}}_k} = \widehat{{{\boldsymbol{\gamma}}}}_{\left[\left(m_{{{\mathbfcal{B}}}}+\sum_{k'=1}^{k-1} m_{{{\mathbf{D}}}_{k'}} +1\right): \left(m_{{{\mathbfcal{B}}}}+\sum_{k'=1}^k m_{{{\mathbf{D}}}_{k'}}\right)\right]}, \quad k=1,\ldots, d.
\end{split}$$ Let $\widehat{{{\mathbf{B}}}}_k = \mathcal{M}_k(\widehat{{{\mathbfcal{B}}}})$, evaluate $$\widehat{{{\mathbfcal{A}}}} = \llbracket\widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{L}}}}_1, \ldots, \widehat{{{\mathbf{L}}}}_d\rrbracket, \quad \widehat{{{\mathbf{L}}}}_k = \left(\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k\right)\left(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\right)^{-1}, \quad k=1,\ldots, d.$$
\[al:procedure\_regular\_general\_order\]
Input: $y_1,\ldots, y_n\in \mathbb{R}, {{\mathbf{X}}}_1,\ldots, {{\mathbf{X}}}_n \in \mathbb{R}^{p_1\times p_2}$, rank $r$. Evaluate $\widetilde{{{\mathbf{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j {{\mathbf{X}}}_j.$ and let $\widetilde{{{\mathbf{U}}}}_1 = {{\rm SVD}}_r(\widetilde{{{\mathbf{A}}}}), \widetilde{{{\mathbf{U}}}}_2 = {{\rm SVD}}_r(\widetilde{{{\mathbf{A}}}}^\top)$. Construct $\widetilde{{{\mathbf{X}}}} = \left[\widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}}~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1} \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2}\right] \in \mathbb{R}^{n\times r(p_1+p_2-r)}$, where $$\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}}\in \mathbb{R}^{n\times r^2},\quad (\widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}})_{[i,:]} =
{{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{X}}}_i \widetilde{{{\mathbf{U}}}}_2\right),\\
\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k} \in \mathbb{R}^{n\times (p_k-r)r}, &\quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1})_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{X}}}_{i} \widetilde{{{\mathbf{U}}}}_2\right),\quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2})_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{X}}}_{i}^\top \widetilde{{{\mathbf{U}}}}_1\right).
\end{split}$$
Solve $\widehat{{{\boldsymbol{\gamma}}}} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^m}\|y - \widetilde{{{\mathbf{X}}}}{{\boldsymbol{\gamma}}}\|_2^2$. Partition $\widehat{{{\boldsymbol{\gamma}}}}$ and assign to $\widehat{{{\mathbf{B}}}}, \widehat{{{\mathbf{D}}}}_1, \widehat{{{\mathbf{D}}}}_2$, $$\begin{split}
{{\rm vec}}(\widehat{{{\mathbf{B}}}}) := \widehat{{{\boldsymbol{\gamma}}}}_{[1:r^2]},\quad
{{\rm vec}}(\widehat{{{\mathbf{D}}}}_1) := \widehat{{{\boldsymbol{\gamma}}}}_{\left[\left(r^2+1\right): rp_1\right]}, \quad {{\rm vec}}(\widehat{{{\mathbf{D}}}}_2) := \widehat{{{\boldsymbol{\gamma}}}}_{\left[\left(rp_1 + 1\right): \left(r(p_1+p_2-r)\right)\right]}.
\end{split}$$ Evaluate $$\widehat{{{\mathbfcal{A}}}} = \widehat{{{\mathbf{L}}}}_1 \widehat{{{\mathbf{B}}}}\widehat{{{\mathbf{L}}}}_2^\top, \quad \widehat{{{\mathbf{L}}}}_1 = \left(\widetilde{{{\mathbf{U}}}}_1\widehat{{{\mathbf{B}}}} + \widetilde{{{\mathbf{U}}}}_{1\perp}\widehat{{{\mathbf{D}}}}_1\right)\widehat{{{\mathbf{B}}}}^{-1}, \quad \widehat{{{\mathbf{L}}}}_2 = \left(\widetilde{{{\mathbf{U}}}}_2\widehat{{{\mathbf{B}}}}^\top + \widetilde{{{\mathbf{U}}}}_{2\perp}\widehat{{{\mathbf{D}}}}_2\right)\left(\widehat{{{\mathbf{B}}}}^\top\right)^{-1}.$$
\[al:procedure\_regular\_matrix\]
Input: $y_1,\ldots, y_n \in\mathbb{R}$, ${{\mathbfcal{X}}}_1,\ldots, {{\mathbfcal{X}}}_n\in \mathbb{R}^{p_1\times \cdots \times p_d}$, rank ${\boldsymbol{r}}= (r_1, r_2, \ldots, r_d)$, sparsity index $J_s$. Evaluate $\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n}\sum_{j=1}^{n} y_j {{\mathbfcal{X}}}_j.$ Apply STAT-SVD on $\widetilde{{{\mathbfcal{A}}}}$ with sparsity index $J_s$. Let the outcome be $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3, \ldots ,\widetilde{{{\mathbf{U}}}}_d $. Let $\widetilde{{{\mathbfcal{S}}}} = \llbracket\widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \ldots, \widetilde{{{\mathbf{U}}}}_d^\top\rrbracket$ and evaluate the probing directions $\widetilde{{{\mathbf{V}}}}_k = {\rm QR}\left[\mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}})^\top\right], k=1,\ldots, d.$ Construct $$\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times (r_1\cdots r_d)},\quad (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[i,:]} =
{{\rm vec}}\left({{\mathbfcal{X}}}_i \times_{l=1}^d \widetilde{{{\mathbf{U}}}}_l^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k} \in \mathbb{R}^{n\times (p_kr_k)}, \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})_{[i,:]} = {{\rm vec}}\left(\mathcal{M}_k\left({{\mathbfcal{X}}}_i\times_{\substack{l=1\\l\neq k}}^{d}\widetilde{{{\mathbf{U}}}}_{l}^\top\right)\widetilde{{{\mathbf{V}}}}_k\right), \quad k=1,\ldots, d.
\end{split}$$ Solve $$\begin{split}
& \widehat{{{\mathbfcal{B}}}}\in \mathbb{R}^{r_1\cdots r_d}, \quad {{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{r_1\cdots r_d}} \|y - \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}{{\boldsymbol{\gamma}}}\|_2^2;\\
\widehat{{{\mathbf{E}}}}_k \in \mathbb{R}^{p_k\times r_k}, \quad & {{\rm vec}}(\widehat{{{\mathbf{E}}}}_k) =\left\{\begin{array}{ll}
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{p_kr_k}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2 + \lambda_k \sum_{j=1}^{p_k}\|{{\boldsymbol{\gamma}}}_{G_j^{k}}\|_2, & k \in J_s;\\
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{p_kr_k}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2, & k\notin J_s.
\end{array} \right.
\end{split}$$ Evaluate $$\widehat{{{\mathbfcal{A}}}} = \llbracket\widehat{{{\mathbfcal{B}}}}; (\widehat{{{\mathbf{E}}}}_1(\widetilde{{{\mathbf{U}}}}_1^\top \widehat{{{\mathbf{E}}}}_1)^{-1}), \ldots, (\widehat{{{\mathbf{E}}}}_d(\widetilde{{{\mathbf{U}}}}_d^\top \widehat{{{\mathbf{E}}}}_d)^{-1})\rrbracket$$
\[al:procedure\_sparse\_general\_order\]
Input: $y_1,\ldots, y_n \in\mathbb{R}$, ${{\mathbf{X}}}_1,\ldots, {{\mathbf{X}}}_n\in \mathbb{R}^{p_1 \times p_2}$, rank $r$, sparsity index $J_s \subseteq \{1,2\}$. Evaluate $\widetilde{{{\mathbf{A}}}} = \frac{1}{n_1}\sum_{j=1}^{n} y_j {{\mathbf{X}}}_j.$ Apply sparse matrix SVD (the Two-Way Iterative Thresholding in [@yang2014sparse] or the order-2 version of STAT-SVD in [@zhang2017optimal-statsvd]) on $\widetilde{{{\mathbf{A}}}}$ with sparsity index $J_s$. Let the estimated left and right subspaces be $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2$. Construct $$\begin{split}
& \widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}}\in \mathbb{R}^{n\times (r^2)},\quad (\widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}})_{[i,:]} =
{{\rm vec}}(\widetilde{{{\mathbf{U}}}}_1^\top{{\mathbf{X}}}_i \widetilde{{{\mathbf{U}}}}_2),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k} \in \mathbb{R}^{n\times (p_k r)}, \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_1})_{[i,:]} = {{\rm vec}}\left({{\mathbf{X}}}_i\widetilde{{{\mathbf{U}}}}_{2}\right), \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_2})_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{X}}}_i\right).
\end{split}$$ Solve $ \widehat{{{\mathbf{B}}}}\in \mathbb{R}^{r\times r}, {{\rm vec}}(\widehat{{{\mathbf{B}}}}) = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{r^2}} \|y - \widetilde{{{\mathbf{X}}}}_{{\mathbf{B}}}{{\boldsymbol{\gamma}}}\|_2^2$; $$\begin{split}
\widehat{{{\mathbf{E}}}}_k \in \mathbb{R}^{p_k\times r_k}, \quad {{\rm vec}}(\widehat{{{\mathbf{E}}}}_k) =\left\{\begin{array}{ll}
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{p_k r}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2 + \lambda_k \sum_{j=1}^{p_k}\|{{\boldsymbol{\gamma}}}_{G_j^{k}}\|_2, & k \in J_s;\\
{\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{p_k r}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\|_2^2, & k\notin J_s.
\end{array} \right.
\end{split}$$ Evaluate $$\widehat{{{\mathbf{A}}}} =\widehat{{{\mathbf{E}}}}_1(\widetilde{{{\mathbf{U}}}}_1^\top \widehat{{{\mathbf{E}}}}_1)^{-1}\widehat{{{\mathbf{B}}}}(\widetilde{{{\mathbf{U}}}}_2^\top \widehat{{{\mathbf{E}}}}_2)^{-\top}\widehat{{{\mathbf{E}}}}_2^\top.$$
\[al:procedure\_sparse\_matrix\]
More Details on Tuning Parameter Selection {#sec:tuning}
==========================================
The implementation of ISLET requires the rank ${\boldsymbol{r}}$ as inputs. When ${\boldsymbol{r}}$ is unknown in practice, we propose a two-stage-scheme for adaptive low-rank tensor regression. First, we input a conservatively large value of ${\boldsymbol{r}}_{ini}$ into ISLET to obtain $\widehat{{{\mathbfcal{B}}}}, \widehat{{{\mathbf{D}}}}_k$ (regular case) or $\widehat{{{\mathbfcal{B}}}}, \widehat{{{\mathbf{E}}}}_k$ (sparse case), based on which we estimate the rank $\widehat{{\boldsymbol{r}}}$ by the “Cross scheme" introduced recently by [@zhang2019cross]. Then, we run ISLET again with $\widehat{{\boldsymbol{r}}}$ to obtain the final estimates. The pseudo-codes for regular and sparse order-$d$ tensor regression are provided in Algorithms \[al:procedure\_regular\_general\_order\_unknown\_r\] and \[al:procedure\_sparse\_general\_order\_unknown\_r\], respectively.
Input: $y_1,\ldots, y_n\in \mathbb{R}, {{\mathbfcal{X}}}_1,\ldots, {{\mathbfcal{X}}}_n \in \mathbb{R}^{p_1\times \cdots \times p_d}$, rank ${\boldsymbol{r}}_{ini} = (r_{1,ini},\ldots, r_{d,ini})$. Apply Algorithms \[al:procedure\_regular\], \[al:procedure\_regular\_general\_order\], \[al:procedure\_regular\_matrix\] with rank ${\boldsymbol{r}}_{ini}$ to obtain $\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k$, $\widehat{{{\mathbfcal{B}}}}$, and $\widehat{{{\mathbf{D}}}}_k$ for $k=1,\ldots, d$. Denote $\widehat{{\mathbf{B}}}_k = \mathcal{M}_k(\widehat{{\mathbfcal{B}}})$. Evaluate ${{\mathbf{U}}}_k^{(B)}$ and ${{\mathbf{V}}}_k^{(A)}$ via SVDs. Then rotate, $${{\mathbf{U}}}_k^{(B)} \in \mathbb{O}_{r_{k,ini}}, \text{ as the left singular vectors of }\widehat{{{\mathbf{B}}}}_k,$$ $${{\mathbf{V}}}_k^{(A)} \in \mathbb{O}_{r_{k,ini}}, \text{ as the right singular vectors of }\left(\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k\right);$$ $${{\mathbf{A}}}_k = \left(\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k\right){{\mathbf{V}}}_k^{(A)}\in\mathbb{R}^{p_k\times r_{k, ini}},$$ $$\boldsymbol{J}_k = ({{\mathbf{U}}}_k^{(B)})^\top \cdot \left(\widehat{{\mathbf{B}}}_k\widetilde{{\mathbf{V}}}_k\right) \cdot {{\mathbf{V}}}_k^{(A)} \in \mathbb{R}^{r_{k,ini}\times r_{k, ini}}.$$ $\boldsymbol{J}_{k, [1:s, 1:s]}$ is not singular and $\|{{\mathbf{A}}}_{k, [:, 1:s]} \boldsymbol{J}_{k, [1:s, 1:s]}^{-1}\| \leq 3$ [**then**]{} $\widehat r_k = s$; [**break**]{} from the loop; $\widehat{r}_k$ is still unassigned [**then**]{} $\widehat{r}_k = 0$. Apply Algorithm \[al:procedure\_regular\] again with rank $\widehat{{\boldsymbol{r}}} = (\widehat{r}_1,\ldots, \widehat{r}_d)$. Let the final output be $\widehat{{{\mathbfcal{A}}}}$.
\[al:procedure\_regular\_general\_order\_unknown\_r\]
Input: $y_1,\ldots, y_n\in \mathbb{R}, {{\mathbfcal{X}}}_1,\ldots, {{\mathbfcal{X}}}_n \in \mathbb{R}^{p_1\times \cdots \times p_d}$, rank ${\boldsymbol{r}}_{ini}$, sparsity index $J_s$. Apply Algorithms \[al:procedure\_sparse\], \[al:procedure\_sparse\_general\_order\], or \[al:procedure\_sparse\_matrix\] with rank ${\boldsymbol{r}}_{ini}$ to obtain $\widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k$, $\widehat{{{\mathbfcal{B}}}}$, and $\widehat{{{\mathbf{E}}}}_k$ for $k=1,\ldots, d$. Denote $\widehat{{\mathbf{B}}}_k = \mathcal{M}_k(\widehat{{\mathbfcal{B}}})$. Evaluate ${{\mathbf{U}}}_k^{(B)}$ and ${{\mathbf{V}}}_k^{(A)}$ via SVDs, then rotate, $${{\mathbf{U}}}_k^{(B)} \in \mathbb{O}_{r_{k,ini}}, \text{ as the left singular vectors of }\widehat{{{\mathbf{B}}}}_k,$$ $${{\mathbf{V}}}_k^{(A)} \in \mathbb{O}_{r_{k,ini}}, \text{ as the right singular vectors of }\widehat{{{\mathbf{E}}}}_k;$$ $${{\mathbf{A}}}_k = \widehat{{{\mathbf{E}}}}_k{{\mathbf{V}}}_k^{(A)}\in\mathbb{R}^{p_k\times r_{k, ini}},\quad \boldsymbol{J}_k = ({{\mathbf{U}}}_k^{(B)})^\top \cdot \left(\widehat{{\mathbf{B}}}_k\widetilde{{\mathbf{V}}}_k\right) \cdot {{\mathbf{V}}}_k^{(A)} \in \mathbb{R}^{r_{k,ini}\times r_{k, ini}}.$$ $\boldsymbol{J}_{k, [1:s, 1:s]}$ is not singular and $\|{{\mathbf{A}}}_{k, [:, 1:s]} \boldsymbol{J}_{k, [1:s, 1:s]}^{-1}\| \leq 3$ [**then**]{} $\widehat r_k = s$; [**break**]{} from the loop; $\widehat{r}_k$ is still unassigned [**then**]{} $\widehat{r}_k = 0$. Apply Algorithm \[al:procedure\_sparse\] again with rank $\widehat{{\boldsymbol{r}}} = (\widehat{r}_1,\ldots, \widehat{r}_d)$. Let the final output be $\widehat{{{\mathbfcal{A}}}}$.
\[al:procedure\_sparse\_general\_order\_unknown\_r\]
Next, we perform simulation studies to verify the proposed rank selection scheme in both the regular and sparse cases. In particular, let $p = 20, 30$, ${\boldsymbol{r}}_{ini} = \lfloor {\boldsymbol{p}}/3 \rfloor$, $n \in[2000, 5000]$, $\sigma = 5$, $s= 12$, and the actual rank $r = 3, 5$. We randomly generate the regular and sparse regression settings as described in Section \[sec:numerical\], then perform Algorithms \[al:procedure\_regular\_general\_order\_unknown\_r\] and \[al:procedure\_sparse\_general\_order\_unknown\_r\]. The average estimation error results are plots in Figures \[fig:unknown-rank-regular\] and \[fig:unknown-rank-sparse\] respectively for the regular and sparse cases. We can see from both cases that the estimation errors with known rank are close to the one without known rank and the difference decreases when the sample size gets larger.
Simulation Study on Approximate Low-rank Tensor Regression {#sec:additional-simu}
==========================================================
We provide simulation results on the performance of ISLET when the parameter ${{\mathbfcal{A}}}$ is approximately low rank. Specifically, we first simulate the exact low Tucker rank tensor ${{\mathbfcal{A}}}_0$ in the same way as the one in previous settings and simulate ${{\mathbfcal{Z}}}$ as the perturbation tensor with i.i.d. standard normal entries. Then we set ${{\mathbfcal{A}}}= {{\mathbfcal{A}}}+ \frac{\tau \|{{\mathbfcal{A}}}\|_F {{\mathbfcal{Z}}}}{p^3}$. The response $y_j$ and covariate ${{\mathbfcal{X}}}_j$ are generated the same to previous settings. Let $\sigma = 5, p = 20, n = [2000,8000], s_1 = s_2 = s_3 = 12, \tau = 0,0.1,0.3,0.5$. $\tau$ here characterizes how close ${{\mathbfcal{A}}}$ is to the exact low-rank tensor – ${{\mathbfcal{A}}}$ is exact low rank if $\tau = 0$. We apply ISLET in both the regular and sparse regimes with the tuning parameter selection scheme described in Algorithms \[al:procedure\_regular\_general\_order\_unknown\_r\] and \[al:procedure\_sparse\_general\_order\_unknown\_r\] The results are collected in the Figure \[fig:approx-low-rank\]. We can see that the estimation error decreases as $\tau$ decreases or $n$ increases; generally speaking, ISLET achieve good performance under both the regular and sparse regime when the true parameter ${{\mathbfcal{A}}}$ is only approximately low rank.
Proofs {#sec:proofs}
======
We collect all proofs of the main technical results in this section.
Proof of Theorem \[th:upper\_bound\_general\] {#sec:proof_upper_bound_general}
---------------------------------------------
This theorem aims to develop a deterministic error bound for $\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\|_{{{\rm HS}}}^2$ in terms of the sketching direction error $\theta$, $\rho$, and error term $\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}\|_2^2$. Since the proof is long and technically challenging, we divide the whole argument into six steps for a better presentation. In Step 1, we introduce the notation to be used throughout the proof. In Step 2, we transform the original high-dimensional low-rank tensor regression model to dimension-reduced one . We also rewrite the key quantities in the upper bound $\|(\widetilde{{\mathbf{X}}}^\top\widetilde{{\mathbf{X}}})^{-1}\widetilde{{\mathbf{X}}}^\top\widetilde{{\boldsymbol{\varepsilon}}}\|_2^2$ to $\|\widehat{{{\mathbfcal{B}}}} - \widetilde{{\mathbfcal{B}}}\|_{{{\rm HS}}}^2 + \sum_{k=1}^3 \|\widehat{{{\mathbf{D}}}}_k-\widetilde{{\mathbf{D}}}_k\|_F^2$. In step 3, we introduce the factorization for ${{\mathbfcal{A}}}$ and $\widehat{{\mathbfcal{A}}}$. Based on this factorization and the property of orthogonal projection, in step 4, we decompose the loss $\|\widehat{{\mathbfcal{A}}}- {{\mathbfcal{A}}}\|_{{{\rm HS}}}$ into eight terms. In step 5, we bound some intermediate error terms in terms of $\theta$ and $\rho$ using properties of the spectral norm and least singular value. In the last Step 6, we finish the proof by bounding each of the eight terms in Step 4 using the results in Step 2, 5, and Lemma \[lm:FGH\].
1. For simplicity, we denote $${{\mathbf{x}}}_j = {{\rm vec}}({{\mathbfcal{X}}}_j)\in \mathbb{R}^{p_1 p_2 p_3}, \quad {{\mathbf{X}}}_{jk} = \mathcal{M}_k({{\mathbfcal{X}}}_j)\in \mathbb{R}^{p_k \times (p_{k+1}p_{k+2})},$$ $${{\mathbf{a}}}= {{\rm vec}}({{\mathbfcal{A}}})\in \mathbb{R}^{p_1p_2p_3},\quad {{\mathbf{A}}}_k = \mathcal{M}_k({{\mathbfcal{A}}}) \in \mathbb{R}^{p_k\times(p_{k+1}p_{k+2})}$$ as the vectorized and matricized tensor covariates and parameter. (Note that ${{\mathbf{X}}}_{jk}$ is a matrix rather than the $(j,k)$-th entry of ${{\mathbf{X}}}$. Instead, we use ${{\mathbf{X}}}_{[j,k]}$ to denote the specific $(i,j)$-th entry of the matrix ${{\mathbf{X}}}$ in our notation system.) All mode indices $(\cdot)_k$ are in module-3, e.g., $p_{4} = p_1$, ${{\mathbf{A}}}_4 = {{\mathbf{A}}}_1$, ${{\mathbf{X}}}_{j5} = {{\mathbf{X}}}_{j2}$, etc. Recall $${{\mathbf{W}}}_1 = ({{{\mathbf{U}}}}_3\otimes {{{\mathbf{U}}}}_2){{\mathbf{V}}}_1, \quad{{\mathbf{W}}}_2 = ({{{\mathbf{U}}}}_3\otimes {{{\mathbf{U}}}}_1){{\mathbf{V}}}_2, \quad {{\mathbf{W}}}_3 = ({{{\mathbf{U}}}}_2\otimes {{{\mathbf{U}}}}_1){{\mathbf{V}}}_3,$$ $$\widetilde{{\mathbf{W}}}_1 = (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2)\widetilde{{\mathbf{V}}}_1, \quad \widetilde{{\mathbf{W}}}_2 = (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_1)\widetilde{{\mathbf{V}}}_2, \quad \widetilde{{{\mathbf{W}}}}_3 = (\widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)\widetilde{{\mathbf{V}}}_3.$$ Define $$\label{eq:def-tilde-B-D_k}
\begin{split}
\widetilde{{{\mathbfcal{B}}}} = & \left\llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket = \left\llbracket {{\mathbfcal{S}}}\times_1 {{\mathbf{U}}}_1\times_2 {{\mathbf{U}}}_2\times_3 {{\mathbf{U}}}_3; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top , \widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket \in \mathbb{R}^{r_1\times r_2\times r_3};\\
\widetilde{{{\mathbf{D}}}}_1 = & \widetilde{{{\mathbf{U}}}}_{1\perp}^\top \mathcal{M}_1({{\mathbfcal{A}}}\times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3) \widetilde{{{\mathbf{V}}}}_1 \overset{\text{Lemma \ref{lm:Kronecker-vectorization-matricization}}}{=} \widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1 \widetilde{{{\mathbf{W}}}}_1 \in \mathbb{R}^{(p_1-r_1)\times r_1},\\
\widetilde{{{\mathbf{D}}}}_2 = & \widetilde{{{\mathbf{U}}}}_{2\perp}^\top \mathcal{M}_2({{\mathbfcal{A}}}\times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_3 \widetilde{{{\mathbf{U}}}}_3) \widetilde{{{\mathbf{V}}}}_2 = \widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{A}}}_2 \widetilde{{{\mathbf{W}}}}_2 \in \mathbb{R}^{(p_2-r_2)\times r_2},\\
\widetilde{{{\mathbf{D}}}}_3 = & \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \mathcal{M}_3({{\mathbfcal{A}}}\times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2) \widetilde{{{\mathbf{V}}}}_3 = \widetilde{{{\mathbf{U}}}}_{3\perp}^\top {{\mathbf{A}}}_3 \widetilde{{{\mathbf{W}}}}_3 \in \mathbb{R}^{(p_3-r_3)\times r_3}.
\end{split}$$ Intuitively speaking, $\widetilde{{\mathbf{B}}}$ is the parameter core tensor lying in the singular subspaces $\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1$ and $\widetilde{{{\mathbf{D}}}}_1, \widetilde{{{\mathbf{D}}}}_2, \widetilde{{{\mathbf{D}}}}_3$ are the parameter matrices corresponding to the arm-minus-body part lying in the singular subspace of $\mathcal{R}_1\left(\widetilde{{{\mathbf{W}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_{1\perp}\right)$, $\mathcal{R}_2\left(\widetilde{{{\mathbf{W}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{2\perp}\right)$, $\mathcal{R}_3\left(\widetilde{{{\mathbf{W}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_{3\perp}\right)$.
2. In this step, we introduce an important decomposition for $y_j$ and the error term $\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{\mathbf{X}}})^{-1}\widetilde{{\mathbf{X}}}^\top \widetilde{\boldsymbol{\varepsilon}}\|_2^2$. In correspondence to $\widehat{{\boldsymbol{\gamma}}}$ , we construct $\widetilde{{{\boldsymbol{\gamma}}}}$ as $$\label{eq:gamma_tilde}
\widetilde{{{\boldsymbol{\gamma}}}} = \left({{\rm vec}}(\widetilde{{{\mathbfcal{B}}}})^\top, {{\rm vec}}(\widetilde{{{\mathbf{D}}}}_1)^\top, {{\rm vec}}(\widetilde{{{\mathbf{D}}}}_2)^\top, {{\rm vec}}(\widetilde{{{\mathbf{D}}}}_3)^\top\right)^\top \in \mathbb{R}^{m}.$$
Then for $j=1,\ldots, n$, the response $y_j$ can be decomposed as $$\label{ineq:y_j^(2)-decompose}
\begin{split}
y_j = & \langle {{\mathbfcal{X}}}_j, {{\mathbfcal{A}}}\rangle +\varepsilon_j = \left\langle {{\mathbf{x}}}_j, {{\mathbf{a}}}\right\rangle +\varepsilon_j\\
= & \left\langle {{\mathbf{x}}}_j, P_{\widetilde{{{\mathbf{U}}}}} {{\mathbf{a}}}\right\rangle +\varepsilon_j + \left\langle {{\mathbf{x}}}_j, P_{\widetilde{{{\mathbf{U}}}}_{\perp}} {{\mathbf{a}}}\right\rangle \\
= & \left\langle {{\mathbf{x}}}_j, P_{\widetilde{{{\mathbf{U}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_3} {{\mathbf{a}}}\right\rangle + \sum_{k=1}^3\left\langle {{\mathbf{x}}}_j, P_{\mathcal{R}_k\left(\widetilde{{{\mathbf{U}}}}_{k\perp}\otimes \widetilde{{{\mathbf{W}}}}_k\right)} {{\mathbf{a}}}\right\rangle + \widetilde{\varepsilon}_j\\
\overset{\eqref{eq:def-tilde-B-D_k}}{=} & \left\langle (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)^\top {{\mathbf{x}}}_j, ~ (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)^\top {{\mathbf{a}}}\right\rangle\\
& + \sum_{k=1}^3 \left\langle \widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{X}}}_{jk}\widetilde{{{\mathbf{W}}}}_k, \widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_k \right\rangle + \widetilde{\varepsilon}_j \\
\overset{\eqref{eq:def-tilde-B-D_k}}{=} & (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[j, :]} {{\rm vec}}(\widetilde{{{\mathbfcal{B}}}}) + \sum_{k=1}^3(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k})_{[j,:]} {{\rm vec}}(\widetilde{{{\mathbf{D}}}}_k) + \widetilde{\varepsilon}_j = \widetilde{{{\mathbf{X}}}}_{[j,:]}\cdot \widetilde{{{\boldsymbol{\gamma}}}} + \widetilde{\varepsilon}_j.
\end{split}$$ Given the definitions of $\widehat{{{\mathbf{D}}}}_k$, $\widehat{{{\mathbfcal{B}}}}$ and $\widehat{{{\boldsymbol{\gamma}}}}$ and the fact that $\widetilde{{{\mathbf{X}}}}$ is non-singular, $\widehat{{\boldsymbol{\gamma}}}$ can be rewritten into the following vectorized form, $$\begin{split}
\widehat{{{\boldsymbol{\gamma}}}} = & {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{m}} \sum_{i=1}^{n}\left(y_i - \widetilde{{{\mathbf{X}}}}_{[i, :]}{{\boldsymbol{\gamma}}}\right)^2 = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{m}}\left\|y - \widetilde{{{\mathbf{X}}}}{{\boldsymbol{\gamma}}}\right\|_2^2\\
= & \left(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}}\right)^{-1} \widetilde{{{\mathbf{X}}}}^\top y = \left(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}}\right)^{-1} \widetilde{{{\mathbf{X}}}}^\top \left(\widetilde{{{\mathbf{X}}}}\widetilde{{{\boldsymbol{\gamma}}}} + \widetilde{{\boldsymbol{\varepsilon}}}\right)\\
= & \widetilde{{{\boldsymbol{\gamma}}}} + \left(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}}\right)^{-1} \widetilde{{{\mathbf{X}}}}^\top\widetilde{{\boldsymbol{\varepsilon}}}.
\end{split}$$ where $m = r_1 r_2 r_3 + \sum_{k=1}^3 (p_k -r_k)r_k $. Thus, by the definition of $\widetilde{{{\boldsymbol{\gamma}}}}$ , $\widehat{{{\boldsymbol{\gamma}}}}$ , $\widehat{{{\mathbfcal{B}}}}$ and $\widehat{{\mathbf{D}}}_k$ , we have $$\label{th:hat_B-B}
\begin{split}
& \|\widehat{{{\mathbfcal{B}}}} - \widetilde{{{\mathbfcal{B}}}}\|_{{{\rm HS}}}^2 + \sum_{k=1}^3\|\widehat{{{\mathbf{D}}}}_k - \widetilde{{{\mathbf{D}}}}_k\|_F^2 = \left\|\widehat{{{\boldsymbol{\gamma}}}} - \widetilde{{{\boldsymbol{\gamma}}}}\right\|_2^2 = \left\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}\right\|_2^2 := \kappa^2.
\end{split}$$
3. In this step, we introduce the factorization for ${{\mathbfcal{A}}}$ . Since the left and right singular subspaces of ${{\mathbf{A}}}_k$ are ${{\mathbf{U}}}_k$ and ${{\mathbf{W}}}_k$, respectively, $$\label{ineq:sigma_min-tildeUAW}
\begin{split}
& \sigma_{r_k} \left(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_k\right) = \sigma_{r_k} \left(\widetilde{{{\mathbf{U}}}}_k^\top P_{{{\mathbf{U}}}_k} {{\mathbf{A}}}_k P_{{{\mathbf{W}}}_k}\widetilde{{{\mathbf{W}}}}_k\right) = \sigma_{r_k} \left((\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{U}}}_k) {{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k {{\mathbf{W}}}_k ({{\mathbf{W}}}_k^\top\widetilde{{{\mathbf{W}}}}_k)\right) \\
\geq & \sigma_{\min}(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{U}}}_k) \cdot \sigma_{\min}({{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k{{\mathbf{W}}}_k) \cdot \sigma_{\min}({{\mathbf{W}}}_k^\top\widetilde{{{\mathbf{W}}}}_k)\\
= & \sqrt{1 - \|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\|^2} \cdot \sigma_{r_k}({{\mathbf{A}}}_k) \cdot \sqrt{1 - \|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\|^2}\\
\geq & \sigma_{r_k}({{\mathbf{A}}}_k)(1-\theta^2)>0.
\end{split}$$ Here, the last but one equality is due to the property of $\sin\Theta$ distance (c.f., Lemma 1 in [@cai2018rate]). Thus, ${{\rm rank}}(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_k) = r_k$, which is a full rank matrix. Thus, $$\label{eq:factor-bcA}
\begin{split}
& {{\mathbfcal{A}}}= \left\llbracket {{\mathbfcal{B}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3 \right \rrbracket\\
= & \left\llbracket \llbracket {{\mathbfcal{B}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3 \rrbracket; {{\mathbf{U}}}_1 (\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{U}}}_1)^{-1}\widetilde{{{\mathbf{U}}}}_1^\top , {{\mathbf{U}}}_2 (\widetilde{{{\mathbf{U}}}}_2^\top {{\mathbf{U}}}_2)^{-1}\widetilde{{{\mathbf{U}}}}_2^\top, {{\mathbf{U}}}_3 (\widetilde{{{\mathbf{U}}}}_3^\top {{\mathbf{U}}}_3)^{-1}\widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket\\
= & \left\llbracket {{\mathbfcal{A}}}; {{\mathbf{U}}}_1 (\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{U}}}_1)^{-1}\widetilde{{{\mathbf{U}}}}_1^\top , {{\mathbf{U}}}_2 (\widetilde{{{\mathbf{U}}}}_2^\top {{\mathbf{U}}}_2)^{-1}\widetilde{{{\mathbf{U}}}}_2^\top, {{\mathbf{U}}}_3 (\widetilde{{{\mathbf{U}}}}_3^\top {{\mathbf{U}}}_3)^{-1}\widetilde{{{\mathbf{U}}}}_3^\top \right \rrbracket\\
= & \left\llbracket {{\mathbfcal{A}}}; {{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_1(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_1)^{-1}\widetilde{{{\mathbf{U}}}}_1^\top, {{\mathbf{A}}}_2\widetilde{{{\mathbf{W}}}}_2(\widetilde{{{\mathbf{U}}}}_2^\top {{\mathbf{A}}}_2\widetilde{{{\mathbf{W}}}}_2)^{-1}\widetilde{{{\mathbf{U}}}}_2^\top , {{\mathbf{A}}}_3\widetilde{{{\mathbf{W}}}}_3(\widetilde{{{\mathbf{U}}}}_3^\top {{\mathbf{A}}}_3\widetilde{{{\mathbf{W}}}}_3)^{-1}\widetilde{{{\mathbf{U}}}}_3^\top \right \rrbracket\\
\end{split}$$ The fourth equality is because the left singular space and right singular space of ${{\mathbf{A}}}_k$ is ${{\mathbf{U}}}_k$ and ${{\mathbf{W}}}_k$.
Recall $$\widehat{{{\mathbfcal{A}}}} = \left\llbracket \widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{L}}}}_1, \widehat{{{\mathbf{L}}}}_2, \widehat{{{\mathbf{L}}}}_3 \right\rrbracket, \quad \widehat{{{\mathbf{L}}}}_k = (\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k)(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1},\quad k=1,2,3.$$ Denote $\widetilde{{{\mathbf{B}}}}_k = \mathcal{M}_k(\widetilde{{{\mathbfcal{B}}}})$, $\widehat{{{\mathbf{B}}}}_k = \mathcal{M}_k(\widehat{{{\mathbfcal{B}}}})$. In parallel to the definition of $\widehat{{{\mathbf{L}}}}_k$, we define $$\label{eq:tilde-L_k}
\begin{split}
\widetilde{{{\mathbf{L}}}}_1 = & (\widetilde{{{\mathbf{U}}}}_1\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1 + \widetilde{{{\mathbf{U}}}}_{1\perp}\widetilde{{{\mathbf{D}}}}_1)(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \\
= & \left(\widetilde{{{\mathbf{U}}}}_1\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1 (\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})\widetilde{{{\mathbf{V}}}}_1 + \widetilde{{{\mathbf{U}}}}_{1\perp}\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2)\widetilde{{{\mathbf{V}}}}_1\right)\\
& ~~ \cdot \left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2)\widetilde{{{\mathbf{V}}}}_1\right)^{-1}\\
= & {{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_1\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1 \widetilde{{{\mathbf{W}}}}_1\right)^{-1}.
\end{split}$$ Similarly, $$\begin{split}
\widetilde{{{\mathbf{L}}}}_2 = & (\widetilde{{{\mathbf{U}}}}_2\widetilde{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2 + \widetilde{{{\mathbf{U}}}}_{2\perp}\widetilde{{{\mathbf{D}}}}_2)(\widetilde{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1} = {{\mathbf{A}}}_2\widetilde{{{\mathbf{W}}}}_2\left(\widetilde{{{\mathbf{U}}}}_2^\top {{\mathbf{A}}}_2 \widetilde{{{\mathbf{W}}}}_2\right)^{-1},\\
\widetilde{{{\mathbf{L}}}}_3 = & (\widetilde{{{\mathbf{U}}}}_3\widetilde{{{\mathbf{B}}}}_3\widetilde{{{\mathbf{V}}}}_3 + \widetilde{{{\mathbf{U}}}}_{3\perp}\widetilde{{{\mathbf{D}}}}_3)(\widetilde{{{\mathbf{B}}}}_3\widetilde{{{\mathbf{V}}}}_3)^{-1} = {{\mathbf{A}}}_3\widetilde{{{\mathbf{W}}}}_3\left(\widetilde{{{\mathbf{U}}}}_3^\top {{\mathbf{A}}}_3 \widetilde{{{\mathbf{W}}}}_3\right)^{-1}.
\end{split}$$ Thus, in addition to $\widehat{{\mathbfcal{A}}}= \llbracket\widehat{{\mathbfcal{B}}}; \widehat{{\mathbf{L}}}_1, \widehat{{\mathbf{L}}}_2, \widehat{{\mathbf{L}}}_3\rrbracket$, we have $$\label{eq: tensorA-indentity}
{{\mathbfcal{A}}}= \llbracket\widetilde{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{L}}}}_1, \widetilde{{{\mathbf{L}}}}_2, \widetilde{{{\mathbf{L}}}}_3 \rrbracket$$
4. Next, we analyze the estimation error of $\widehat{{{\mathbfcal{A}}}}$. First, the error bound of $\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}$ can be decomposed into eight parts, $$\label{eq:error-decomposition}
\begin{split}
& \|\widehat{{{\mathbfcal{A}}}}-{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2 = \left\|\llbracket\widehat{{{\mathbfcal{A}}}}-{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1} + P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}, P_{\widetilde{{{\mathbf{U}}}}_2} + P_{\widetilde{{{\mathbf{U}}}}_{2\perp}}, P_{\widetilde{{{\mathbf{U}}}}_3} + P_{\widetilde{{{\mathbf{U}}}}_{3\perp}} \rrbracket\right\|_{{{\rm HS}}}^2\\
= & \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket\right\|_{{{\rm HS}}}^2 + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket\right\|_{{{\rm HS}}}^2 \\
& + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket\right\|_{{{\rm HS}}}^2 + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \rrbracket\right\|_{{{\rm HS}}}^2 \\
& + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \rrbracket\right\|_{{{\rm HS}}}^2 + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \rrbracket\right\|_{{{\rm HS}}}^2\\
& + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket\right\|_{{{\rm HS}}}^2 + \left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \rrbracket\right\|_{{{\rm HS}}}^2.
\end{split}$$ Here we used the fact that $P_{\widetilde{{{\mathbf{U}}}}_1}$ and $P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}$ are orthogonal complementary. We aim to apply Lemma \[lm:FGH\] to analyze each term above in the next two steps.
5. Before giving the upper bounds for each term of , we denote $$\begin{split}
\lambda_k = & \max\left\{\left\|\widehat{{{\mathbf{D}}}}_k(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\|, \left\|\widetilde{{{\mathbf{D}}}}_k(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\|\right\},\\
\pi_k = & \|(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\widetilde{{{\mathbf{B}}}}_k\|,\quad k=1, 2, 3
\end{split}$$ and aim to provide upper bounds for $\lambda_k, \pi_k$ in this step. By definition of $\widetilde{{\mathbf{B}}}_k$ and the fact that the right singular vector of ${{\mathbf{A}}}_k$ is ${{\mathbf{W}}}_k$, $$\label{ineq:pi_k}
\begin{split}
\pi_1 = & \left\|(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}\widetilde{{{\mathbf{B}}}}_1\right\| = \left\|(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1 \widetilde{{{\mathbf{W}}}}_1)^{-1} \widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1(\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})\right\|\\
\leq & \left\|\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1 \widetilde{{{\mathbf{W}}}}_1\right)^{-1}\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1\right\| = \left\|\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1{{\mathbf{W}}}_1^\top \widetilde{{{\mathbf{W}}}}_1\right)^{-1}\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{A}}}_1{{\mathbf{W}}}_1\right\|\\
\leq & \left\|({{\mathbf{W}}}_1^\top \widetilde{{{\mathbf{W}}}}_1)^{-1}\right\| = \sigma_{\min}^{-1}(\widetilde{{{\mathbf{W}}}}_1^\top {{\mathbf{W}}}_1) = \left(1-\|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\|^2\right)^{-1/2} \\
\leq & \frac{1}{(1-\theta^2)^{1/2}}.
\end{split}$$ Similarly, the same upper bounds also applies to $\pi_2$ and $\pi_3$.
Based on definitions of $\widetilde{{{\mathbf{D}}}}_k$ and $\widetilde{{{\mathbf{B}}}}_k$ and the fact that the left singular subspace of ${{\mathbf{A}}}_k$ is ${{\mathbf{U}}}_k$, we have $$\label{ineq:tilde-D-tilde-B-tilde-V}
\begin{split}
& \|\widetilde{{{\mathbf{D}}}}_k(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\|^2+1 = \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k)^{-1}\right\|^2 + 1\\
= & \left\|\begin{bmatrix}
{{\mathbf{I}}}_{r_k} \\
\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k)^{-1}
\end{bmatrix}\right\|^2 = \left\|\begin{bmatrix}
\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_k(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_k)^{-1}\\
\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k)^{-1}
\end{bmatrix}\right\|^2 \\
= & \left\|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k \left(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k\right)^{-1}\right\|^2 = \left\|{{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k \left(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{U}}}_k {{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k\right)^{-1}\right\|^2 \\
= & \left\|{{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k \left({{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k\right)^{-1} \left(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{U}}}_k\right)^{-1}\right\|^2\\
= & \left\|\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{U}}}_1\right)^{-1}\right\|^2 = \sigma_{\min}^{-2}\left(\widetilde{{{\mathbf{U}}}}_1^\top {{\mathbf{U}}}_1\right) = \left(1 - \|\sin\Theta(\widetilde{{{\mathbf{U}}}}_1, {{\mathbf{U}}}_1)\|^2\right)^{-1} \leq \frac{1}{1-\theta^2},
\end{split}$$ which implies $$\|\widetilde{{\mathbf{D}}}_k (\widetilde{{\mathbf{B}}}_k\widetilde{{\mathbf{V}}}_k)^{-1}\| \leq \sqrt{\frac{1}{1-\theta^2} - 1} = \sqrt{\frac{\theta^2}{1-\theta^2}}.$$ By the assumption of the theorem that $\|\widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}\| \leq \rho$ and $\theta\leq 1/2$, we have $$\label{ineq:lambda_k}
\lambda_k \leq \max\left\{\rho, \frac{\theta}{\sqrt{1-\theta^2}}\right\} \leq \rho + \frac{2}{\sqrt{3}}\theta, \quad k=1,2,3.$$
6. Now we are ready to give upper bounds for all terms in .
- First, by definition of $\widehat{{{\mathbfcal{B}}}}$, $\widehat{{{\mathbfcal{A}}}}$ , $$\label{eq:hat-A-project}
\begin{split}
& \llbracket \widehat{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket = \left\llbracket \llbracket \widehat{{\mathbfcal{B}}}; \widehat{{{\mathbf{L}}}}_1, \widehat{{{\mathbf{L}}}}_2, \widehat{{{\mathbf{L}}}}_3\rrbracket; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket\\
= & \left\llbracket \widehat{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{U}}}}_1^\top \widehat{{\mathbf{L}}}_1, \widetilde{{{\mathbf{U}}}}_2^\top \widehat{{\mathbf{L}}}_2, \widetilde{{{\mathbf{U}}}}_3^\top \widehat{{\mathbf{L}}}_3 \right\rrbracket.
\end{split}$$ Here, $$\widetilde{{\mathbf{U}}}_k^\top\widehat{{\mathbf{L}}}_k = \widetilde{{{\mathbf{U}}}}_k^\top \left((\widetilde{{{\mathbf{U}}}}_k\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k + \widetilde{{{\mathbf{U}}}}_{k\perp}\widehat{{{\mathbf{D}}}}_k)(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right) = (\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k) (\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1} = {{\mathbf{I}}}_{r_k}.$$ Similarly, we have $\widetilde{{\mathbf{U}}}_k^\top\widetilde{{\mathbf{L}}}_k = {{\mathbf{I}}}_{r_k}$.
Thus, $\llbracket \widehat{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket = \widehat{{\mathbfcal{B}}}$. By definition of $\widetilde{{{\mathbfcal{B}}}}$ , we have $$\label{ineq:thm1-term1}
\begin{split}
\left\|\llbracket (\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket\right\|_{{{\rm HS}}}^2 = & \left\|\llbracket \widehat{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket - \llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket \right\|_{{{\rm HS}}}^2 \\
= & \|\widehat{{{\mathbfcal{B}}}} - \widetilde{{{\mathbfcal{B}}}}\|_{{{\rm HS}}}^2.
\end{split}$$
- Note that $$\label{ineq:thm1-term1.5}
\begin{split}
& \left\|\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \rrbracket\right\|_{{{\rm HS}}}^2 \\
\overset{ \eqref{eq: tensorA-indentity},\eqref{eq:hat-A-project}}{=} & \left\|\llbracket \widehat{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{U}}}}_{1\perp}^\top \widehat{{\mathbf{L}}}_1, \widetilde{{{\mathbf{U}}}}_2^\top\widehat{{{\mathbf{L}}}}_2, \widetilde{{{\mathbf{U}}}}_3^\top\widehat{{{\mathbf{L}}}}_3 \rrbracket - \llbracket\widetilde{{\mathbfcal{B}}}; \widetilde{{\mathbf{U}}}_{1\perp}^\top\widetilde{{{\mathbf{L}}}}_1, \widetilde{{\mathbf{U}}}_{2}^\top\widetilde{{{\mathbf{L}}}}_2, \widetilde{{\mathbf{U}}}_{3}^\top\widetilde{{{\mathbf{L}}}}_3\rrbracket\right\|_{{{\rm HS}}}^2\\
\overset{\eqref{eq:hat_A_non-sparse}\eqref{eq:tilde-L_k}}{=} & \left\|\llbracket \widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, {{\mathbf{I}}}, {{\mathbf{I}}}\rrbracket - \llbracket \widetilde{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{D}}}}_1(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, {{\mathbf{I}}}, {{\mathbf{I}}}\rrbracket\right\|_{{{\rm HS}}}^2\\
\overset{\text{Lemma \ref{lm:Kronecker-vectorization-matricization}}}{=} & \left\|\widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}\widehat{{{\mathbf{B}}}}_1 - \widetilde{{{\mathbf{D}}}}_1(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}\widetilde{{{\mathbf{B}}}}_1 \right\|_F^2\\
\end{split}$$ By the first part of Lemma \[lm:FGH\], $$\begin{split}
& \left\|\widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}\widehat{{{\mathbf{B}}}}_1 - \widetilde{{{\mathbf{D}}}}_1(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}\widetilde{{{\mathbf{B}}}}_1 \right\|_F^2\\
\leq & \left(\pi_1\|\widehat{{{\mathbf{D}}}}_1-\widetilde{{{\mathbf{D}}}}_1\|_F + \lambda_1\|\widehat{{{\mathbf{B}}}}_1- \widetilde{{{\mathbf{B}}}}_1\|_F + \pi_1\lambda_1 \|\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1 - \widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1\|_F\right)^2\\
\overset{\eqref{ineq:pi_k}\eqref{ineq:lambda_k}}{\leq} & \left(\frac{1}{\sqrt{1-\theta^2}}\|\widehat{{{\mathbf{D}}}}_1-\widetilde{{{\mathbf{D}}}}_1\|_F + (\rho+\frac{2}{\sqrt{3}}\theta)\kappa + (\rho+\frac{2}{\sqrt{3}}\theta)\frac{1}{\sqrt{1-\theta^2}}\kappa\right)^2\\
\leq & \frac{1}{1-\theta^2}\|\widehat{{{\mathbf{D}}}}_1 - \widetilde{{{\mathbf{D}}}}_1\|_F^2 + C_1(\rho+\theta)\|\widehat{{{\mathbf{D}}}}_1-\widetilde{{{\mathbf{D}}}}_1\|_F\kappa + C_2(\rho+\theta)^2\kappa^2\\
\leq & \|\widehat{{\mathbf{D}}}_1 - \widetilde{{\mathbf{D}}}_1\|_F^2 + 2\theta^2\|\widehat{{\mathbf{D}}}_1-\widetilde{{\mathbf{D}}}_1\|_F^2 + C_1(\rho+\theta)\|\widehat{{\mathbf{D}}}_1 - \widetilde{{\mathbf{D}}}_1\|_F\kappa + C_2(\rho+\theta)^2\kappa^2\\
\leq & \|\widehat{{{\mathbf{D}}}}_1 - \widetilde{{{\mathbf{D}}}}_1\|_F^2 + C(\rho+\theta)\kappa^2.
\end{split}$$ Here, the last inequality is due to the fact that $\|\widehat{{\mathbf{D}}}_1 - \widetilde{{\mathbf{D}}}_1\|_F\leq \kappa$. Therefore, $$\label{ineq:thm1-term2}
\begin{split}
& \left\|\left\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_{2}^\top, \widetilde{{{\mathbf{U}}}}_3\right\rrbracket\right\|_{{{\rm HS}}}^2 \leq \|\widehat{{\mathbf{D}}}_1 - \widetilde{{\mathbf{D}}}_1\|_F^2 + C(\rho+\theta)\kappa^2;\\
\text{similarly} \quad & \left\|\left\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1}^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_3\right\rrbracket\right\|_{{{\rm HS}}}^2 \leq \|\widehat{{{\mathbf{D}}}}_2 - \widetilde{{{\mathbf{D}}}}_2\|_F^2 + C(\rho+\theta)\kappa^2,\\
& \left\|\left\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1}^\top, \widetilde{{{\mathbf{U}}}}_{2}^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}\right\rrbracket\right\|_{{{\rm HS}}}^2 \leq \|\widehat{{{\mathbf{D}}}}_3-\widetilde{{{\mathbf{D}}}}_3\|_F^2 + C(\rho+\theta)\kappa^2.
\end{split}$$
- By similar argument as , we have $$\begin{split}
& \left\|\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_3\rrbracket\right\|_F^2 \\
= & \left\|\llbracket \widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \widehat{{{\mathbf{D}}}}_2(\widehat{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1}, {{\mathbf{I}}}\rrbracket - \llbracket \widetilde{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{D}}}}_1(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \widetilde{{{\mathbf{D}}}}_2(\widetilde{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1}, {{\mathbf{I}}}\rrbracket\right\|_F^2\\
\end{split}$$ By the second part of Lemma \[lm:FGH\], $$\begin{split}
& \left\|\llbracket \widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \widehat{{{\mathbf{D}}}}_2(\widehat{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1}, {{\mathbf{I}}}\rrbracket - \llbracket \widetilde{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{D}}}}_1(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \widetilde{{{\mathbf{D}}}}_2(\widetilde{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1}, {{\mathbf{I}}}\rrbracket\right\|_F^2\\
\leq & \left(\lambda_1\lambda_2\|\widehat{{{\mathbfcal{B}}}}-\widetilde{{{\mathbfcal{B}}}}\|_F + \sum_{k=1,2}\pi_k \lambda_1\lambda_2/\lambda_k \|\widehat{{{\mathbf{D}}}}_k-\widetilde{{{\mathbf{D}}}}_k\|_F + \sum_{k=1,2}\pi_k\lambda_1\lambda_2 \|\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k - \widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\|_F\right)^2\\
\overset{\eqref{th:hat_B-B}}{\leq} & (\lambda_1\lambda_2 + \pi_1\lambda_2+\pi_2\lambda_1+\pi_1\lambda_1\lambda_2 + \pi_2\lambda_1\lambda_2)^2\kappa^2 \overset{\eqref{ineq:pi_k}}{\leq} C(\rho+\theta)^2\kappa^2.
\end{split}$$ Therefore, $$\label{ineq:thm1-term3}
\begin{split}
& \left\|\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket\right\|_F^2 \leq C(\rho+\theta)^2\kappa^2;\\
\text{similarly,}\quad & \left\|\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}^\top\rrbracket\right\|_F^2 \leq C(\rho+\theta)^2\kappa^2,\\
& \left\|\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}^\top\rrbracket\right\|_F^2 \leq C(\rho+\theta)^2\kappa^2.
\end{split}$$
- By the second part of Lemma \[lm:FGH\], $$\label{ineq:thm1-term4}
\begin{split}
& \left\|\llbracket(\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}); \widetilde{{{\mathbf{U}}}}_{1\perp}^\top, \widetilde{{{\mathbf{U}}}}_{2\perp}^\top, \widetilde{{{\mathbf{U}}}}_{3\perp}\rrbracket\right\|_F^2 \\
= & \Big\|\llbracket \widehat{{{\mathbfcal{B}}}}; \widehat{{{\mathbf{D}}}}_1(\widehat{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \widehat{{{\mathbf{D}}}}_2(\widehat{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1}, \widehat{{{\mathbf{D}}}}_3(\widehat{{{\mathbf{B}}}}_3\widetilde{{{\mathbf{V}}}}_3)^{-1}\rrbracket\\
& - \llbracket \widetilde{{{\mathbfcal{B}}}}; \widetilde{{{\mathbf{D}}}}_1(\widetilde{{{\mathbf{B}}}}_1\widetilde{{{\mathbf{V}}}}_1)^{-1}, \widetilde{{{\mathbf{D}}}}_2(\widetilde{{{\mathbf{B}}}}_2\widetilde{{{\mathbf{V}}}}_2)^{-1}, \widetilde{{{\mathbf{D}}}}_3(\widetilde{{{\mathbf{B}}}}_3\widetilde{{{\mathbf{V}}}}_3)^{-1}\rrbracket\Big\|_F^2\\
\leq & \Big(\lambda_1\lambda_2\lambda_3\|\widehat{{{\mathbfcal{B}}}}-\widetilde{{{\mathbfcal{B}}}}\|_F + \sum_{k=1,2,3}\pi_k \lambda_1\lambda_2\lambda_3/\lambda_k \|\widehat{{{\mathbf{D}}}}_k-\widetilde{{{\mathbf{D}}}}_k\|_F \\
& \quad + \sum_{k=1,2,3}\pi_k\lambda_1\lambda_2\lambda_3 \|\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k - \widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\|_F\Big)^2\\
\overset{{\eqref{th:hat_B-B}}\eqref{ineq:pi_k}}{\leq} & C(\rho+\theta)^4\kappa^2.
\end{split}$$
Combining , , , and , we finally have $$\begin{split}
& \left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \leq \|\widehat{{{\mathbfcal{B}}}}-{{\mathbfcal{B}}}\|_F^2 + \sum_{k=1} \|\widehat{{{\mathbf{D}}}}_k-\widetilde{{{\mathbf{D}}}}_k\|_F^2 + C(\rho+\theta)\kappa^2 = (1+C(\rho+\theta))\kappa^2.
\end{split}$$
In summary, we have finished the proof of this theorem.$\square$
Proof of Theorem \[th:upper\_bound\_sparse\_general\] {#sec:proof_upper_bound_sparse_general}
-----------------------------------------------------
This theorem gives a deterministic error bound of $\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\|_{{{\rm HS}}}^2$ in terms of $\theta, \rho$ and $\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbf{B}}}\|_2^2$, $\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2^2$, $\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k, [:, G_i^k]})^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n\|_2^2$ for the sparse ISLET estimator $\widehat{{{\mathbfcal{A}}}}$ in the sparse low-rank tensor regression model. To prove this theorem, we first rewrite the original high-dimensional regression model to four dimension-reduced ones , . Then we derive error bounds for the least square estimator or group Lasso estimator in terms of $\|\widehat{{\mathbfcal{B}}}- {{\mathbfcal{B}}}\|_{{{\rm HS}}}^2$ or $\|\widehat{{\mathbf{E}}}_k-\widetilde{{\mathbf{E}}}_k\|_F^2$ for each of these dimension-reduced regression models. The rest of the proof aims to assemble the upper bound for $\|\widehat{{\mathbfcal{A}}}- {{\mathbfcal{A}}}\|_{{{\rm HS}}}^2$, which essentially follows from Steps 3-6 in the proof of Theorem \[th:upper\_bound\_general\].
Denote $${{\mathbf{A}}}_k = \mathcal{M}_k({{\mathbfcal{A}}}), \quad {{\mathbf{a}}}= {{\rm vec}}({{\mathbfcal{A}}}), \quad {{\mathbf{X}}}_{jk} = \mathcal{M}_k({{\mathbfcal{X}}}_j), \quad {{\mathbf{x}}}_j={{\rm vec}}({{\mathbfcal{X}}}_j), \quad 1\leq j \leq n,\quad k=1,2,3;$$ $$\label{eq:B-E}
\begin{split}
& \widetilde{{{\mathbfcal{B}}}} = \llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket;\\
& \widetilde{{{\mathbf{E}}}}_k = \mathcal{M}_k({{\mathbfcal{A}}}\times_{k+1}\widetilde{{{\mathbf{U}}}}_{k+1}^\top\times_{k+2}\widetilde{{{\mathbf{U}}}}_{k+2}^\top)\widetilde{{{\mathbf{V}}}}_{k} = {{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_k \in \mathbb{R}^{p_k\times r_k},\quad k=1,2,3;
\end{split}$$ $$\label{eq:gamma_B-gamma_E}
\widetilde{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}}= {{\rm vec}}(\widetilde{{{\mathbfcal{B}}}}) \in \mathbb{R}^{p_1p_2p_3}, \quad \widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k} = {{\rm vec}}(\widetilde{{{\mathbf{E}}}}_k)\in \mathbb{R}^{p_kr_k},\quad k=1,2,3.$$ Then similarly as the argument in the proof of Theorem \[th:upper\_bound\_general\], we can write down the following partial regression formulas that relate $y_j$ and $({{\mathbfcal{X}}}_j, {{\mathbfcal{A}}})$, $$\label{eq:B-regression-model}
\begin{split}
y_j = & \langle {{\mathbfcal{X}}}_j, {{\mathbfcal{A}}}\rangle +\varepsilon_j = \langle {{\mathbf{x}}}_j, {{\mathbf{a}}}\rangle +\varepsilon_j \\
= & \left\langle {{\mathbf{x}}}_j, P_{\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2 \otimes \widetilde{{{\mathbf{U}}}}_1}{{\mathbf{a}}}\right\rangle + \varepsilon_j + \langle {{\mathbf{x}}}_j, P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)_{\perp}} {{\mathbf{a}}}\rangle \\
= & \left\langle (\widetilde{{{\mathbf{U}}}}_3\otimes\widetilde{{{\mathbf{U}}}}_2\otimes\widetilde{{{\mathbf{U}}}}_1)^\top {{\mathbf{x}}}_j, (\widetilde{{{\mathbf{U}}}}_3\otimes\widetilde{{{\mathbf{U}}}}_2\otimes\widetilde{{{\mathbf{U}}}}_1)^\top {{\mathbf{a}}}\right\rangle + (\widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}})_j\\
\overset{\eqref{eq:B-E}\eqref{eq:gamma_B-gamma_E}}{=} & (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{[j, :]} \widetilde{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}}+ (\widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}})_j,
\end{split}$$ $$\label{eq:E_k-regression-model}
\begin{split}
y_j = & \langle {{\mathbfcal{X}}}_j, {{\mathbfcal{A}}}\rangle +\varepsilon_j \\
= & \left\langle {{\mathbfcal{X}}}_j, P_{\mathcal{R}_k\left(\widetilde{{{\mathbf{W}}}}_k\otimes {{\mathbf{I}}}_{p_k}\right)}[{{\mathbfcal{A}}}] \right\rangle + \varepsilon_j + \left\langle {{\mathbfcal{X}}}_j, P_{\left(\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes {{\mathbf{I}}}_{p_k})\right)_\perp}[{{\mathbfcal{A}}}] \right\rangle \\
= & \left\langle{{\mathbf{X}}}_{jk}\widetilde{{{\mathbf{W}}}}_k, ~ {{\mathbf{A}}}_{k}\widetilde{{{\mathbf{W}}}}_k \right\rangle + (\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k})_j\\
\overset{\eqref{eq:B-E}\eqref{eq:gamma_B-gamma_E}}{=} & (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})_{[j, :]} \widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k} + (\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k})_j
\end{split}$$ for $j=1,\ldots, n$ and $k=1,2,3$. We discuss the estimation errors of $\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}$ ($k\in J_s$), $\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}$ ($k\notin J_s$), and $\widehat{{{\mathbfcal{B}}}}$ separately as below.
- For any $k\in J_s$, due to the definition that $$\widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}={{\rm vec}}(\widetilde{{\mathbf{E}}}_k), \quad \widetilde{{\mathbf{E}}}_k = {{\mathbf{A}}}_k\widetilde{{\mathbf{W}}}_k,$$ and the left singular vectors of ${{\mathbf{A}}}_k$ is ${{\mathbf{U}}}_k$ that satisfying $\|{{\mathbf{U}}}_k\|_0 = \sum_{i=1}^{p_k} 1_{\{({{\mathbf{U}}}_k)_{[i, :]}\neq 0\}} \leq s_k$, $\widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}$ is correspondingly group-wise sparse. More specifically, let $G_k^i = \{i, i+p_k, \ldots, i+p_k(r_k-1)\}$ with $i=1,\ldots, p_k$ be a partition of $\{1,\ldots, p_kr_k\}$. Then $$\label{eq:group-wise-sparsity}
\widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}^i := (\widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k})_{G_k^i} \in \mathbb{R}^{r_k},\quad \sum_{i=1}^{p_k} 1_{\{\widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}^i \neq 0\}} \leq s_k.$$ Accordingly, $\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}\in \mathbb{R}^{n_2\times (p_kr_k)}$ are with grouped covariates with respect to $\{G_k^1,\ldots, G_k^{p_k}\}$: $$\label{eq:tilde-X-E-k}
\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i = (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})_{[:, G_k^i]} \in \mathbb{R}^{n\times r_k}, \quad i=1,\ldots, p_k.$$ Recall $\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}$ is the group Lasso estimator, $$\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{(p_kr_k)}} \|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k} {{\boldsymbol{\gamma}}}\|_2^2 + \eta_k \sum_{i=1}^{p_k} \|{{\boldsymbol{\gamma}}}_{G_k^i}\|_2.$$ By the group-wise sparsity structure , the partial linear regression model , the assumption that $\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}\in \mathbb{R}^{n_2\times (p_k r_k)}$ satisfies GRIP assumption with $\delta<1/4$, and $\eta_k = C\max_{1\leq i\leq p_k} \|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i)^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2$ for constant $C\geq 3$, Lemma \[lm:group-oracle-RIP\] yields $$\label{ineq:hat-D_k-D_k}
\|\widehat{{{\mathbf{E}}}}_k - \widetilde{{{\mathbf{E}}}}_k\|_F = \|\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k} - \widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}\|_2 \leq \frac{C\sqrt{s_k} \eta_k}{n} \leq C\sqrt{s_k}\max_{1\leq i\leq p_k} \|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i)^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n\|_2,\quad \forall k\in J_s.$$
- For $k\notin J_s$, recall $\widehat{{{\mathbf{E}}}}_k$ is evaluated via the least square estimator, $${{\rm vec}}(\widehat{{{\mathbf{E}}}}_k) = \widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k},\quad \widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{(p_kr_k)}} \left\|y - \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}{{\boldsymbol{\gamma}}}\right\|_2^2.$$ By linear regression model and the definition of the least square estimator, $$\label{eq:hat_Ek-Ek-identity}
\|\widehat{{{\mathbf{E}}}}_k - \widetilde{{{\mathbf{E}}}}_k\|_F = \|\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k} - \widetilde{{{\boldsymbol{\gamma}}}}_{{{\mathbf{E}}}_k}\|_2 = \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2.$$
- In addition, recall $${{\rm vec}}(\widehat{{{\mathbfcal{B}}}}) = \widehat{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}},\quad \widehat{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}}= {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{r_1r_2r_3}} \|y - \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}{{\boldsymbol{\gamma}}}\|_2^2.$$ By linear regression model and the definition of the least square estimator $\widehat{{{\boldsymbol{\gamma}}}}_{{{\mathbfcal{B}}}}$, $$\label{eq:hat_B-B-identity}
\|\widehat{{{\mathbfcal{B}}}} - {{\mathbfcal{B}}}\|_{{{\rm HS}}}^2 = \|\widehat{{{\boldsymbol{\gamma}}}}_{{\mathbfcal{B}}}- {{\boldsymbol{\gamma}}}_{{\mathbfcal{B}}}\|_2^2 = \|(\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\|_2^2.$$
Given $\theta = \max\{\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\|, \|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\|\} \leq 1/2$, similarly as the proof of Theorem \[th:upper\_bound\_general\], one can show $\widetilde{{{\mathbf{U}}}}_k^\top \widetilde{{{\mathbf{E}}}}_k$ is non-singular. Therefore, $$\begin{split}
\|\widehat{{\mathbfcal{B}}}- {{\mathbfcal{B}}}\|_{{{\rm HS}}}^2 + \sum_{k=1}^3 \|\widehat{{\mathbf{E}}}_k-\widetilde{{\mathbf{E}}}_k\|_F^2 \leq & \left\|(\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\right\|_2^2 + C\sum_{k\in J_s} s_k\max_{1\leq i\leq p_k} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i)^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n_2\right\|_2^2\\
& + \sum_{k \notin J_s} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2.
\end{split}$$ The rest of the proof directly follows from Steps 3 - 6 in Theorem \[th:upper\_bound\_sparse\_general\].$\square$
Proof of Theorem \[th:upper\_bound\_regression\] {#sec:proof_upper_bound_regression}
------------------------------------------------
The goal of Theorem \[th:upper\_bound\_regression\] is to give a probabilistic error bound for regular tensor regression via ISLET. The high level idea is to first derive the error bound for importance sketching regression by a perturbation bound of the HOOI outcome (Theorem 1 in [@zhang2019HOOI]), and then apply the oracle inequality in Theorem \[th:upper\_bound\_general\] to obtain the final estimation error rate. For a better presentation, we divide the long proof into six steps. First in Step 1, we bound the initialization error of $\widetilde{{{\mathbf{U}}}}_k^{(0)}$ using perturbation theory [@cai2018rate] and concentration inequality (Lemmas \[lm:concatenation-singular-value\] and \[lm:concentration-Gaussian-ensemble\]). Then in Step 2, we aim to apply Theorem 1 in [@zhang2019HOOI] to get an error bound for the importance sketching directions $\widetilde{{{\mathbf{U}}}}_k$. The central goal of Step 3 is to prove an error bound for $\theta$. In Steps 4, we move on to the second batch of sample and derive error bounds for a few intermediate terms. In step 5, we evaluate key quantities $\rho$ and $\left\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}\right\|_2^2$ in the context of Theorem \[th:upper\_bound\_general\]. Finally, we plug in all quantities to Theorem \[th:upper\_bound\_general\] and finish the proof.
We begin the proof by introducing some notations. Throughout the proof, the mode indices $(\cdot)_k$ are presented in modulo 3: e.g., ${{\mathbf{U}}}_4 = {{\mathbf{U}}}_1$, ${{\mathbf{V}}}_5 = {{\mathbf{V}}}_2$. For convenience, we denote $$\widetilde{\sigma}^2 = \|{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2 + \sigma^2, \quad {{\mathbf{A}}}_k = \mathcal{M}_k({{\mathbfcal{A}}}),\quad \widetilde{{{\mathbf{A}}}}_k = \mathcal{M}_k(\widetilde{{{\mathbfcal{A}}}}), \quad {{\mathbf{X}}}_{ik} = \mathcal{M}_k({{\mathbfcal{X}}}_i)$$ for $k=1,2,3$. $p=\max\{p_1,p_2,p_3\}$, $r = \max\{r_1, r_2, r_3\}$. To avoid repeating similar notations consecutively, throughout the proof of this theorem we slightly abuse the notation and denote $${{\mathbf{U}}}_{k+2}\otimes {{\mathbf{U}}}_{k+1} = \left\{\begin{array}{ll}
{{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2, & k=1;\\
{{\mathbf{U}}}_3 \otimes {{\mathbf{U}}}_1, & k=2;\\
{{\mathbf{U}}}_2 \otimes {{\mathbf{U}}}_1, & k=3
\end{array}\right.$$ without ambiguity. Other related notations, e.g., $({{\mathbf{U}}}_{k+2 \perp}{{\mathbf{V}}}) \otimes {{\mathbf{U}}}_{k+1}$, are defined in a similar fashion.
The rest of the proof for Theorem \[th:upper\_bound\_regression\] is divided into 6 steps.
1. We first develop the error bound for $\widetilde{{{\mathbf{U}}}}_1^{(0)}$, $\widetilde{{{\mathbf{U}}}}_2^{(0)}$, and $\widetilde{{{\mathbf{U}}}}_3^{(0)}$. Particularly, we aim to show that $$\label{ineq:U_k^{(0)}-upper-bound}
{\mathbb{P}}\left(\left\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k^{(0)}, {{\mathbf{U}}}_k)\right\| \leq \left(\frac{C\widetilde{\sigma}\sqrt{p_k/n_1}}{\lambda_k} + \frac{\widetilde{\sigma}^2 \sqrt{p_1p_2p_3}/n_1}{\lambda_k^2}\right)\wedge 1, k=1,2,3\right) \geq 1 - p^{-C}.$$ We only focus on $\widetilde{{{\mathbf{U}}}}_1^{(0)}$ as the conclusions for $\widetilde{{{\mathbf{U}}}}_2^{(0)}$ and $\widetilde{{{\mathbf{U}}}}_3^{(0)}$ similarly follow. Recall the baseline unbiased estimator $$\widetilde{{{\mathbfcal{A}}}} = \frac{1}{n_1} \sum_{i=1}^{n_1} y_i^{(1)} {{\mathbfcal{X}}}_i^{(1)} =\frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbfcal{X}}}_i^{(1)}, {{\mathbfcal{A}}}\rangle + \varepsilon_i^{(1)}\right){{\mathbfcal{X}}}_i^{(1)} \in \mathbb{R}^{p_1\times p_2\times p_3}.$$ Since the left and right singular subspaces of ${{\mathbf{A}}}_1$ are ${{\mathbf{U}}}_1$ and ${{\mathbf{W}}}_1$, respectively, we further have $\widetilde{{{\mathbf{A}}}}_1\in \mathbb{R}^{p_1\times (p_2p_3)}$ and $$\begin{split}
\widetilde{{{\mathbf{A}}}}_1 = & \mathcal{M}_1\left(\widetilde{{{\mathbfcal{A}}}}\right) = \frac{1}{n_1} \sum_{i=1}^n y_i^{(1)} {{\mathbf{X}}}_{i1}^{(1)} = \frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{X}}}_{i1}^{(1)}, {{\mathbf{A}}}_1 \rangle +\varepsilon_i^{(1)}\right) {{\mathbf{X}}}_{i1}^{(1)} \\
= & \frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{X}}}_{i1}^{(1)}, P_{{{\mathbf{U}}}_1}{{\mathbf{A}}}_1 P_{{{\mathbf{W}}}_1}\rangle +\varepsilon_i^{(1)}\right){{\mathbf{X}}}_{i1}^{(1)}\\
= & \frac{1}{n_1}\sum_{i=1}^{n_1} \left({{\rm tr}}\left(({{\mathbf{X}}}_{i1}^{(1)})^\top {{\mathbf{U}}}_1{{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1 {{\mathbf{W}}}_1^\top\right) +\varepsilon_i^{(1)}\right){{\mathbf{X}}}_{i1}^{(1)}\\
= & \frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i^{(1)}\right){{\mathbf{X}}}_{i1}^{(1)}.
\end{split}$$ Since $\widetilde{{{\mathbf{U}}}}^{(0)}_1 = {{\rm SVD}}_{r_1}(\widetilde{{{\mathbf{A}}}}_1)$, the one-sided perturbation bound [@cai2018rate Proposition 1] yields $$\label{ineq:th2-initialization-perturbation}
\begin{split}
\left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_1^{(0)}, {{\mathbf{U}}}_1\right)\right\| \leq \frac{\sigma_{r_1}({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1)\|{{\mathbf{U}}}_{1\perp}^\top \widetilde{{{\mathbf{A}}}}_1 P_{({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1)^\top}\|}{\sigma_{r_1}^2({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1) - \sigma_{r_1+1}^2(\widetilde{{{\mathbf{A}}}}_1)}\wedge 1
\end{split}$$ To proceed, we analyze $\sigma_{\min}^2\left({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1\right)$, $\sigma_{r_1+1}(\widetilde{{\mathbf{A}}}_1)$, and $\|{{\mathbf{U}}}_{1\perp}^\top \widetilde{{{\mathbf{A}}}}_1 P_{({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1)^\top}\|$, respectively.
- $$\begin{split}
& \sigma_{\min}^2\left({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1\right) \overset{\text{Lemma \ref{lm:concatenation-singular-value}}}{\geq} \sigma_{\min}^2\left({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1 {{\mathbf{W}}}_1\right) + \sigma_{\min}^2\left({{\mathbf{U}}}_1^\top \widetilde{{{\mathbf{A}}}}_1 ({{\mathbf{W}}}_1)_{\perp}\right)\\
= & \sigma_{\min}^2\left(\frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i^{(1)}\right){{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}{{\mathbf{W}}}_1\right)\\
& + \sigma_{\min}^2\left(\frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i^{(1)}\right){{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{W}}}_1)_{\perp}\right).
\end{split}$$
By Lemma \[lm:concentration-Gaussian-ensemble\], ${{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1{{\mathbf{W}}}_1\in \mathbb{R}^{r_1\times r_1}$, and $n_1 \geq Cp^{3/2}r_1$, we have $$\begin{split}
& \sigma_{\min}\left(\frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i^{(1)}\right){{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}{{\mathbf{W}}}_1\right) \\
\geq & \sigma_{\min}({{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1) - \left\|\frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i^{(1)}\right){{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1 - {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\right\|\\
\overset{\text{Lemma \ref{lm:concentration-Gaussian-ensemble}}}{\geq} & \sigma_{r_1}({{\mathbf{A}}}_1) - C\sqrt{\frac{\log p }{n_1} \left(2r_1\|{{\mathbf{A}}}_1\|_F^2 + \sigma^2\right)} \geq (1-c)\sigma_{r_1}({{\mathbf{A}}}_1)
\end{split}$$ with probability at least $1 - p^{-c}$. When ${{\mathbf{X}}}_{i1}^{(1)}$ has i.i.d. Gaussian entries and ${{\mathbf{W}}}_1$ is fixed orthogonal matrix, ${{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{W}}}_1)_{\perp}\in \mathbb{R}^{r_1\times (p_{-1}-r_1)}$ and $\left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i\right)\in \mathbb{R}$ are independently Gaussian distributed and $$\left\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\right\rangle + \varepsilon_i^{(1)}\sim N(0, \widetilde{\sigma}^2).$$ By Lemma \[lm:concentration-independent\], $$\begin{split}
& \sigma_{\min}^2\left(\frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle + \varepsilon_i^{(1)}\right){{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}{{\mathbf{W}}}_{1\perp}\right) \\
\geq & \widetilde{\sigma}^2 \cdot \frac{n_1-C_1\sqrt{n_1\log p}}{n_1^2}\cdot \left(\sqrt{p_{-1}-r_1} - \sqrt{r_1} - C_2\sqrt{\log p}\right)^2\\
\geq & \frac{\widetilde{\sigma}^2}{n_1} \cdot \left(1 - C_1\sqrt{\frac{\log p}{n_1}}\right)\cdot \left(p_{-1} - C_3\sqrt{p_{-1}r_1}-C_2\sqrt{p_{-1}\log p}\right)\\
\geq & \frac{\widetilde{\sigma}^2}{n_1} \left(p_{-1} - C_4\sqrt{p_{-1}r_1} - C_5\sqrt{p_{-1}\log p}\right)
\end{split}$$ with probability at least $1-p^{-c}$. To sum up, $$\label{ineq:thm3-1}
\sigma_{\min}^2\left({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1\right) \geq (1-c)\sigma_{r_1}^2({{\mathbf{A}}}_1) + \frac{\widetilde{\sigma}^2}{n_1} \cdot \left(p_{-1} - C_1 \sqrt{p_{-1}r_1}-C_2 \sqrt{p_{-1}\log p}\right)$$ with probability at least $1 - p^{-c}$.
- Next, we consider $\sigma_{r_1+1}(\widetilde{{\mathbf{A}}}_1)$, note that $$\begin{split}
\sigma_{r_1+1}(\widetilde{{\mathbf{A}}}_1) = & \min_{{{\rm rank}}(M)\leq r_1} \left\|\widetilde{{\mathbf{A}}}_1 - {{\mathbf{M}}}\right\| \leq \left\|\widetilde{{{\mathbf{A}}}}_1 - P_{{{\mathbf{U}}}_1}\widetilde{{\mathbf{A}}}_1\right\| \leq \|{{\mathbf{U}}}_{1\perp}^\top \widetilde{{\mathbf{A}}}_1\|\\
= & \left\|\frac{1}{n_1} \sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 \rangle +\varepsilon_i^{(1)}\right){{\mathbf{U}}}_{1\perp}^\top {{\mathbf{X}}}_{i1}^{(1)}\right\|.
\end{split}$$ Since $$\left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}{{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 {{\mathbf{W}}}_1\rangle +\varepsilon_i^{(1)}\right) \sim N\left(0, \widetilde{\sigma}^2\right),$$ which is also independent of ${{\mathbf{U}}}_{1\perp}^\top {{\mathbf{X}}}_{i1}^{(1)}$. Thus, $$\label{ineq:thm3-2}
\begin{split}
\sigma_{r_1+1}^2(\widetilde{{\mathbf{A}}}_1) = & \left\|\frac{1}{n_1} \sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 \rangle +\varepsilon_i^{(1)}\right){{\mathbf{U}}}_{1\perp}^\top {{\mathbf{X}}}_{i1}^{(1)}\right\|^2 \\
\leq & \widetilde{\sigma}^2\cdot \frac{n_1+C(\sqrt{n_1\log p} + \log p)}{n_1^2}\cdot \left(\sqrt{p_1 -r_1} + \sqrt{p_{-1}} + C\sqrt{\log p}\right)^2\\
\leq & \frac{\widetilde\sigma^2}{n_1}\left(1 + C\sqrt{\frac{\log p}{n_1}}\right)\left(p_{-1}+C\sqrt{p_{-1}p_1} + C\sqrt{p_{-1}\log p} + Cp_1 + C\log p\right)\\
\leq & \frac{\widetilde{\sigma}^2}{n_1} \cdot \left(p_{-1} + C\sqrt{p_{-1}p_1} + C\sqrt{p_{-1}\log p} + Cp_1 + C\log p\right)
\end{split}$$ with probability at least $1 - p^{-c}$.
- Then we consider $\left\|{{\mathbf{U}}}_{1\perp}^\top \widetilde{{\mathbf{A}}}_1 P_{({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)^\top}\right\|$. Note that $${{\mathbf{U}}}_{1\perp}^\top \widetilde{{\mathbf{A}}}_1 P_{({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)^\top} = \frac{1}{n_1}\sum_{i=1}^{n_1} \left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}{{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1{{\mathbf{W}}}_1 \rangle +\varepsilon_i^{(1)}\right) {{\mathbf{U}}}_{1\perp}^\top {{\mathbf{X}}}_{i1}^{(1)} P_{({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)^\top},$$ Here, $\left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} {{\mathbf{W}}}_1, {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1{{\mathbf{W}}}_1\rangle +\varepsilon_i\right)\sim N(0, \widetilde{\sigma}^2)$; by independence, conditioning on fixed value of ${{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}$, ${{\mathbf{U}}}_{1\perp}^\top {{\mathbf{X}}}_{i1}^{(1)}$ is still standard normal, and then $${{\mathbf{U}}}_{1\perp}^\top {{\mathbf{X}}}_{i1}^{(1)} P_{({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)^\top}\Big| {{\mathbf{U}}}_1^\top{{\mathbf{X}}}_{i1}^{(1)}$$ is a $(p_1-r_1)$-by-$r_1$ i.i.d. standard Gaussian matrix. By Lemma \[lm:concentration-independent\], we have $$\label{ineq:thm3-3}
\begin{split}
\left\|{{\mathbf{U}}}_{1\perp}^\top \widetilde{{\mathbf{A}}}_1 P_{({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)^\top}\right\| \leq & \widetilde{\sigma} \sqrt{\frac{n_1+C_1\sqrt{n_1\log p} + C_2\log p}{n_1^2}}\cdot \left(\sqrt{p_1-r_1} + \sqrt{r_1} + C_3\sqrt{\log p}\right) \\
\leq & C_4\widetilde{\sigma} \cdot \sqrt{\frac{p_1}{n_1}}
\end{split}$$ with probability at least $1 - p^{-C}$.
Combining - with , we have the following inequality holds with probability at least $1 - p^{-C}$, $$\begin{split}
& \left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_1^{(0)}, {{\mathbf{U}}}_1\right)\right\|\\
\leq & \frac{\sigma_{r_1}({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)\|{{\mathbf{U}}}_{1\perp}^\top \widetilde{{\mathbf{A}}}_1 P_{({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1)^\top}\|}{\sigma_{r_1}^2({{\mathbf{U}}}_1^\top \widetilde{{\mathbf{A}}}_1) - \sigma_{r_1+1}^2(\widetilde{{\mathbf{A}}}_1)}\wedge 1\\
\leq & \frac{\left((1-c)\sigma_{r_1}({{\mathbf{A}}}_1) + \widetilde{\sigma}\sqrt{p_{-1}/n_1}\right)\cdot C_1\widetilde{\sigma} \sqrt{p_1/n_1}}{\left((1-c)\sigma_{r_1}({{\mathbf{A}}}_1) + \widetilde{\sigma}\sqrt{p_{-1}/n_1}\right)^2 - \frac{\widetilde{\sigma}^2}{n_1} \cdot \left(p_{-1} + C_2\sqrt{p_{-1}p_1} + C_3\sqrt{p_{-1}\log p} + C_4p_1 + C_5\log p\right)}\wedge 1\\
\end{split}$$ Since $n_1 \geq Cp^{3/2}\widetilde\sigma^2/\lambda_0^2$ for large constant $C>0$, we have $$\begin{split}
& \left((1-c)\sigma_{r_1}({{\mathbf{A}}}_1) + \widetilde{\sigma}\sqrt{p_{-1}/n_1}\right)^2 - \frac{\widetilde{\sigma}^2}{n_1} \cdot \left(p_{-1} + C_1\sqrt{p_{-1}p_1} + C_2\sqrt{p_{-1}\log p} + C_3p_1 + C_4\log p\right) \\
\geq & (1-c)^2\sigma_{r_1}^2({{\mathbf{A}}}_1) + 2(1-c)\sigma_{r_1}({{\mathbf{A}}}_1)\widetilde\sigma\sqrt{p_{-1}/n_1} - \frac{C_2\widetilde\sigma^2}{n_1}\left(\sqrt{p_1p_2p_3} + \sqrt{p_{-1}\log p} + C_3p_1 + C_4\log p\right)\\
\geq & c\sigma_{r_1}^2({{\mathbf{A}}}_1)
\end{split}$$ and additionally, $$\begin{split}
\left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_1^{(0)}, {{\mathbf{U}}}_1\right)\right\| \leq & \left(\frac{C_1\widetilde{\sigma}\sqrt{p_1/n_1} \cdot \sigma_{r_1}({{\mathbf{A}}}_1) + \widetilde{\sigma}^2 \sqrt{p_1p_2p_3}/n_1}{\sigma_{r_1}^2({{\mathbf{A}}}_1)}\right) \wedge 1.
\end{split}$$ with probability at least $1 - p^{-C}$. Similar inequalities also hold for $\left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_2^{(0)}, {{\mathbf{U}}}_2\right)\right\|$ and $\left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_3^{(0)}, {{\mathbf{U}}}_3\right)\right\|$. Based on these arguments, we conclude that holds. further implies that $$\label{ineq:regression-prob-0}
\begin{split}
e_0 := & \max_k\left\|\widetilde{{{\mathbf{U}}}}^{(0)\top}_{k\perp}\mathcal{M}_k({{\mathbfcal{A}}})\right\| = \max_k \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^{(0)\top} {{\mathbf{U}}}_k{{\mathbf{U}}}_k^\top \mathcal{M}_k({{\mathbfcal{A}}})\right\| \\
\leq & \max_k \|\widetilde{{{\mathbf{U}}}}^{(0)\top}_{k\perp} {{\mathbf{U}}}_k\|\cdot \|{{\mathbf{U}}}_k^\top\mathcal{M}_k({{\mathbfcal{A}}})\| \leq \max_k \|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k^{(0)}, {{\mathbf{U}}}_k)\| \cdot \|{{\mathbf{U}}}_k^\top \mathcal{M}_k({{\mathbfcal{A}}})\|\\
\leq & \max_k C\|{{\mathbf{A}}}_k\|\left(\frac{\widetilde{\sigma}\sqrt{p_k/n_1}}{\sigma_{r_k}({{\mathbf{A}}}_k)} + \frac{\widetilde{\sigma}^2\sqrt{p_1p_2p_3}/n_1}{\sigma_{r_k}^2({{\mathbf{A}}}_k)}\right)\\
\leq & C_1\kappa \left(\frac{\widetilde{\sigma}p^{1/2}}{n_1^{1/2}} + \frac{\widetilde{\sigma}^2p^{3/2}}{\lambda_0n_1}\right)
\end{split}$$ with probability at least $1 - p^{-C}$.
2. Then we develop the error bound for $\widetilde{{{\mathbf{U}}}}_k$ after enough number of iterations in this step. In particular, we aim to apply Theorem 1 in [@zhang2019HOOI] to give an error bound for the output $\widetilde{{{\mathbf{U}}}}_k$ from the high-order order orthogonal iteration (HOOI). To this end, we verify the conditions in Theorem 1 in [@zhang2019HOOI] in this step. Defining $$\label{eq:def-Z-T-tilde-T}
{{\mathbfcal{Z}}}= \widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}},\quad {{\mathbfcal{T}}}= {{\mathbfcal{A}}}+ {{\mathbfcal{Z}}}\times_1 P_{{{\mathbf{U}}}_1}\times_2 P_{{{\mathbf{U}}}_2}\times_3 P_{{{\mathbf{U}}}_3},\quad \widetilde{{{\mathbfcal{T}}}} = \widetilde{{{\mathbfcal{A}}}}.$$ Then, $$\label{eq:tilde-T-T}
\widetilde{{{\mathbfcal{T}}}} - {{\mathbfcal{T}}}= {{\mathbfcal{Z}}}- {{\mathbfcal{Z}}}\times_1 P_{{{\mathbf{U}}}_1} \times_2 P_{{{\mathbf{U}}}_2} \times_3 P_{{{\mathbf{U}}}_3}.$$ In order to apply Theorem 1 in [@zhang2019HOOI], we develop the following upper bounds under the assumptions of Theorem \[th:upper\_bound\_regression\].
- Since $\mathcal{M}_1\left((\widetilde{{\mathbfcal{A}}}- {{\mathbfcal{A}}})\times_1{{\mathbf{U}}}_1^\top\times_2{{\mathbf{U}}}_2^\top\times_3{{\mathbf{U}}}_3^\top\right)$ is a $r_1$-by-$(r_2r_3)$ matrix, Lemma \[lm:concentration-Gaussian-ensemble\] implies $$\begin{split}
& \left\|\mathcal{M}_1\left((\widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}})\times_1 {{\mathbf{U}}}_1^\top \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top\right)\right\|\\
= & \left\|{{\mathbf{U}}}_1^\top \mathcal{M}_1\left(\widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right)({{\mathbf{U}}}_3 \otimes {{\mathbf{U}}}_2)\right\|\\
= & \Bigg\|\frac{1}{n_1} \sum_{i=1}^{n_1} \left(\left\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{U}}}_3 \otimes {{\mathbf{U}}}_2), {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_{1}({{\mathbf{U}}}_3 \otimes {{\mathbf{U}}}_2) \right\rangle +\varepsilon_i^{(1)}\right) {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}\left({{\mathbf{U}}}_3 \otimes {{\mathbf{U}}}_2\right)\\
& - {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)\Bigg\|\\
\overset{\text{Lemma \ref{lm:concentration-Gaussian-ensemble}}}{\leq} & C_1\sqrt{\frac{\log p \cdot (r_1+r_2r_{3})\widetilde{\sigma}^2}{n_1}}\\
\end{split}$$ with probability at least $1 - p^{-C}$. Similar results also hold for $\mathcal{M}_2(\cdot)$ and $\mathcal{M}_3(\cdot)$. Then $$\begin{split}
\lambda_k({{\mathbfcal{T}}}) := & \sigma_{r_k}\left(\mathcal{M}_k({{\mathbfcal{T}}})\right)\\
\overset{\eqref{eq:tilde-T-T}}{\geq} & \sigma_{r_k} \left(\mathcal{M}_k({{\mathbfcal{A}}})\right) - \left\|\mathcal{M}_k\left( (\widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}) \times_1 P_{{{\mathbf{U}}}_1} \times_2 P_{{{\mathbf{U}}}_2} \times_3 P_{{{\mathbf{U}}}_3} \right)\right\|\\
\geq & \lambda_k - C_1\sqrt{\frac{\log p \cdot (r_k+r_{k+1}r_{k+2})\widetilde{\sigma}^2}{n_1}}\geq (1-c)\lambda_0
\end{split}$$ with probability at least $1 - p^{-C}$.
- Next, we consider $$\tau_{0k}: = \left\|\mathcal{M}_k(\widetilde{{{\mathbfcal{T}}}} - {{\mathbfcal{T}}}) \left({{\mathbf{U}}}_{k+2}\otimes {{\mathbf{U}}}_{k+1}\right)\right\|,\quad k=1,2,3.$$ In particular, $$\begin{split}
& \left\|\mathcal{M}_1(\widetilde{{{\mathbfcal{T}}}} - {{\mathbfcal{T}}}) \left({{\mathbf{U}}}_{3}\otimes {{\mathbf{U}}}_{2}\right)\right\| \\
\overset{\eqref{eq:tilde-T-T}}{=} & \left\|\mathcal{M}_1({{\mathbfcal{Z}}}- \llbracket{{\mathbfcal{Z}}}; P_{{{\mathbf{U}}}_1}, P_{{{\mathbf{U}}}_2}, P_{{{\mathbf{U}}}_3}\rrbracket)({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)\right\|\\
= & \left\|\mathcal{M}_1\left(\left(\widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}- \llbracket\widetilde{{{\mathbfcal{A}}}}-{{\mathbfcal{A}}}; P_{{{\mathbf{U}}}_1}, P_{{{\mathbf{U}}}_2}, P_{{{\mathbf{U}}}_3}\rrbracket\right)\times_{2} {{\mathbf{U}}}_{2}^\top\times_{3} {{\mathbf{U}}}_{3}^\top\right)\right\|\\
= & \Big\|\mathcal{M}_1\left((\widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}})\times_1(P_{{{\mathbf{U}}}_1}+P_{{{\mathbf{U}}}_{1\perp}}) \times_{2} {{\mathbf{U}}}_{2}^\top \times_{3} {{\mathbf{U}}}_{3}^\top\right)\\
& - \mathcal{M}_1\left((\widetilde{{{\mathbfcal{A}}}}-{{\mathbfcal{A}}}) \times_1 P_{{{\mathbf{U}}}_1}\times_2 {{\mathbf{U}}}_{2}^\top \times_3 {{\mathbf{U}}}_{3}^\top\right)\Big\|\\
= & \left\|\mathcal{M}_1\left((\widetilde{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}})\times_1P_{{{\mathbf{U}}}_{1\perp}} \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top\right)\right\|\\
= & \left\|{{\mathbf{U}}}_{1\perp}^\top (\widetilde{{{\mathbf{A}}}}_1 - {{\mathbf{A}}}_1) \cdot ({{\mathbf{U}}}_3\otimes{{\mathbf{U}}}_2)\right\|\\
\leq & \Big\|\frac{1}{n_1}\sum_{i=1}^{n_1}\left(\langle {{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)} ({{\mathbf{U}}}_{3}\otimes {{\mathbf{U}}}_{2}), {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1 ({{\mathbf{U}}}_{3}\otimes {{\mathbf{U}}}_{2})\rangle +\varepsilon_i^{(1)}\right){{\mathbf{U}}}_{1\perp}^\top{{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)\\
& - {{\mathbf{U}}}_{1\perp}^\top {{\mathbf{A}}}_1 ({{\mathbf{U}}}_{3}\otimes {{\mathbf{U}}}_{2}) \Big\|\\
\overset{\text{Lemma \ref{lm:concentration-independent}}}{\leq} & \widetilde{\sigma}\sqrt{\frac{n_1+C_1\sqrt{n_1\log p}}{n_1^2}}\left(\sqrt{p_1-r_1} + \sqrt{r_2 r_3} + C_2\sqrt{\log p}\right) \leq C_3\widetilde{\sigma}\sqrt{\frac{p_1}{n_1}},
\end{split}$$ with probability at least $1 - p^{-C}$. Thus, $$\label{ineq:regression-prob-1}
\begin{split}
{\mathbb{P}}\left(\tau_{0k} \leq C_1\widetilde{\sigma} \sqrt{p_k/n_1},~~ k=1,2,3\right) \geq 1 - p^{-C}.
\end{split}$$
- Next we consider the upper bound of $$\label{eq:def-tau_1}
\begin{split}
\tau_1 := & \max_{k}\Big\{\max_{\substack{{{\mathbf{V}}}\in \mathbb{R}^{(p_{k+1}-r_{k+1})\times r_{k+1}}\\\|{{\mathbf{V}}}\|\leq 1}}\left\|\mathcal{M}_k(\widetilde{{{\mathbfcal{T}}}}-{{\mathbfcal{T}}}) \cdot \left\{({{\mathbf{U}}}_{k+2, \perp} {{\mathbf{V}}}) \otimes {{\mathbf{U}}}_{k+1}\right\}\right\|,\\
& \quad\quad\quad\quad \max_{\substack{{{\mathbf{V}}}\in \mathbb{R}^{(p_{k+2}-r_{k+2})\times r_{k+2}}\\\|{{\mathbf{V}}}\|\leq 1}}\left\|\mathcal{M}_k(\widetilde{{\mathbfcal{T}}}- {{\mathbfcal{T}}}) \cdot \left\{{{\mathbf{U}}}_{k+2} \otimes ({{\mathbf{U}}}_{k+1, \perp}{{\mathbf{V}}})\right\}\right\| \Big\}.\\
\end{split}$$ Note that $$\begin{split}
& \mathcal{M}_1\left(\widetilde{{{\mathbfcal{T}}}}-{{\mathbfcal{T}}}\right) ({{\mathbf{U}}}_{3\perp}{{\mathbf{V}}})\otimes {{\mathbf{U}}}_2\\
= & \left(\mathcal{M}_1({{\mathbfcal{Z}}}) - \mathcal{M}_1\left({{\mathbfcal{Z}}}\times_1 P_{{{\mathbf{U}}}_1} \times_2 P_{{{\mathbf{U}}}_2} \times_3 P_{{{\mathbf{U}}}_3}\right)\right) ({{\mathbf{U}}}_{3\perp}{{\mathbf{V}}})\otimes {{\mathbf{U}}}_2 \\
= & \mathcal{M}_1({{\mathbfcal{Z}}}) ({{\mathbf{U}}}_{3\perp} {{\mathbf{V}}})\otimes {{\mathbf{U}}}_2 = \frac{1}{n_1}\sum_{i=1}^{n_1} y_i^{(1)} {{\mathbf{X}}}_{i1}^{(1)}(({{\mathbf{U}}}_{3\perp}{{\mathbf{V}}})\otimes {{\mathbf{U}}}_2),
\end{split}$$ $$y_i^{(1)} = \langle{{\mathbfcal{X}}}_i^{(1)}, {{\mathbfcal{A}}}\rangle+\varepsilon_i^{(1)} = \langle{{\mathbf{U}}}_1^\top {{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2), {{\mathbf{U}}}_1^\top{{\mathbf{A}}}_1({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2) \rangle+\varepsilon_i^{(1)}.$$ Since ${{\mathbf{U}}}_{3\perp}$ and ${{\mathbf{U}}}_3$ are orthogonal, $y_i^{(1)}$ and ${{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{U}}}_{3\perp}\otimes {{\mathbf{U}}}_2)$ are independently Gaussian distributed. Thus, conditioning on fixed values of $\{y_i^{(1)}\}_{i=1}^{n_1}$, $$\frac{1}{n_1}\sum_{i=1}^{n_1} y_i^{(1)} {{\mathbf{X}}}_{i1}^{(1)}({{\mathbf{U}}}_{3\perp}\otimes {{\mathbf{U}}}_2)\bigg| \|{{\mathbf{y}}}^{(1)}\|_2^2$$ is a $p_1$-by-$((p_2-r_2)r_3)$ random matrix with i.i.d. Gaussian entries with mean zero and variance $\|{{\mathbf{y}}}^{(1)}\|_2^2/n_1^2$. By Lemma 5 in [@zhang2017tensor], $$\label{ineq:regression-prob-4}
\begin{split}
& {\mathbb{P}}\Bigg(\max_{{{\mathbf{V}}}\in \mathbb{R}^{(p_2-r_2)\times r_2}}\left\|\mathcal{M}_1\left({{\mathbfcal{Z}}}({{\mathbf{U}}}_{3\perp}{{\mathbf{V}}}\otimes {{\mathbf{U}}}_2)\right)\right\| \\
& \quad \quad \geq \frac{C\|{{\mathbf{y}}}^{(1)}\|_2}{n_1}\left(\sqrt{p_1} + \sqrt{r_2r_3} + \sqrt{1+t}(\sqrt{p_2r_2}+\sqrt{p_3r_3})\right)\Bigg|\|{{\mathbf{y}}}^{(1)}\|_2^2\Bigg) \\
\leq & C\exp\left(-Ct(p_2r_2+p_3r_3)\right).
\end{split}$$ Note that $\|{{\mathbf{y}}}^{(1)}\|_2^2\sim \widetilde{\sigma}^2\chi^2_{n_1}$, we have $$\label{ineq:regression-prob-5}
\begin{split}
{\mathbb{P}}\left(\|{{\mathbf{y}}}^{(1)}\|_2^2 \geq \widetilde{\sigma}^2(n_1 + 2\sqrt{n_1t} + 2t)\right) \leq \exp(-t).
\end{split}$$ Combining (with $t = pr/(p_2r_2+p_3r_3)$), (with $t = Cpr$), and the fact that $n_1\geq Cpr$ for large constant $C>0$, we have $$\begin{split}
& {\mathbb{P}}\left(\max_{\substack{{{\mathbf{V}}}\in \mathbb{R}^{(p_3-r_2)\times r_1}\\\|{{\mathbf{V}}}\|\leq 1}}\left\|\mathcal{M}_1\left(\widetilde{{{\mathbfcal{T}}}}-{{\mathbfcal{T}}}\right) ({{\mathbf{U}}}_{3\perp}{{\mathbf{V}}})\otimes {{\mathbf{U}}}_2\right\| \geq C\widetilde{\sigma}\sqrt{\frac{pr}{n_1}}\right) \leq C\exp\left(-cpr\right).
\end{split}$$ By symmetry, we have similar results for other terms in the right hand side of and the following conclusion, $$\label{ineq:regression-prob-2}
\begin{split}
& {\mathbb{P}}\left(\tau_1 \geq C\widetilde{\sigma}\sqrt{\frac{pr}{n_1}}\right) \leq C\exp(-cpr).
\end{split}$$
- Based on essentially the same argument as the previous step, we can also show $$\label{ineq:regression-prob-3}
\begin{split}
\tau_2 := & \max_k \max_{\substack{{{\mathbf{V}}}\in \mathbb{R}^{(p_{k+1}-r_{k+1})\times r_{k+1}}: \|{{\mathbf{V}}}\|\leq 1;\\ {{\mathbf{V}}}' \in \mathbb{R}^{(p_{k+2}-r_{k+2})\times r_{k+2}}: \|{{\mathbf{V}}}'\|\leq 1}} \left\|\mathcal{M}_k({{\mathbfcal{Z}}})\left\{({{\mathbf{U}}}_{k+1 \perp}{{\mathbf{V}}}) \otimes ({{\mathbf{U}}}_{k+2 \perp}{{\mathbf{V}}}')\right\}\right\|\\
\leq & C\widetilde{\sigma}\sqrt{\frac{pr}{n_1}}
\end{split}$$ with probability at least $1 - C\exp(-cpr)$.
Now, when the statements in , , all hold, given $n_1\geq \frac{\widetilde{\sigma}^2}{\lambda_0^2}(\kappa pr\vee p^{3/2})$ for large enough constant $C>0$, we have $n_1 \geq \frac{C\widetilde{\sigma}^2}{\lambda_0^2}p^{4/3}r^{1/3}$ (by Hölder’s inequality) and the condition $$\begin{split}
& \frac{\tau_1}{\lambda({{\mathbfcal{T}}})} + \max_k\frac{4\tau_2(4\tau_{0k}+e_0)}{\lambda^2({{\mathbfcal{T}}})} \\
\leq & \frac{C_1\widetilde{\sigma}\sqrt{pr/n_1}}{\lambda_0} + \frac{C_2\widetilde{\sigma}\sqrt{pr/n_1}\left(\widetilde{\sigma}\sqrt{p/n_1}+\kappa\widetilde{\sigma}\sqrt{p/n_1} + \kappa\widetilde{\sigma}^2 p^{3/2}/(\lambda_0n_1)\right)}{\lambda_0^2}\\
\leq & \frac{C_1\widetilde{\sigma}p^{1/2}r^{1/2}}{\lambda_0n_1^{1/2}} + \frac{C_2\widetilde{\sigma}^2\kappa pr^{1/2}}{\lambda_0^2n_1} + \frac{C_3\kappa\widetilde{\sigma}^3p^2r^{1/2}}{\lambda_0^3n_1^{3/2}}\leq 1
\end{split}$$ holds. Namely, the condition in Theorem 1 in [@zhang2019HOOI] holds when the events of , , occur.
3. In this step, we try to establish the estimation errors for $\widetilde{{{\mathbf{U}}}}_k$ and $\widetilde{{{\mathbf{W}}}}_k$. First, Theorem 1 in [@zhang2019HOOI] and , , imply $$\begin{split}
& \left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k\right)\right\| \leq \frac{C\tau_{0k}}{\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{T}}}))}\leq \frac{C\widetilde{\sigma} \sqrt{p_k/n_1}}{\lambda_k},\quad k=1,2,3,
\end{split}$$ $$\begin{split}
\text{and} \quad & \left\|\llbracket\widetilde{{{\mathbfcal{T}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket - {{\mathbfcal{T}}}\right\|_{{{\rm HS}}} \leq C\widetilde{\sigma}\sqrt{\frac{p_1r_1+p_2r_2+p_3r_3+r_1r_2r_3}{n_1}}
\end{split}$$ with probability at least $1 - p^{-C}$. Moreover, $$\begin{split}
& \left\|{{\mathbfcal{T}}}- {{\mathbfcal{A}}}\right\|_{{{\rm HS}}} \overset{\eqref{eq:def-Z-T-tilde-T}}{=} \left\|\left(\widetilde{{{\mathbfcal{A}}}} -{{\mathbfcal{A}}}\right)\times_1 {{\mathbf{U}}}_1^\top \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top\right\|_{{{\rm HS}}}\\
= & \Big\|\frac{1}{n_1}\sum_{i=1}^{n_1}\left(\left\langle {{\rm vec}}({{\mathbfcal{X}}}_i\times_1 {{\mathbf{U}}}_1^\top \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top), {{\rm vec}}({{\mathbfcal{A}}}\times_1 {{\mathbf{U}}}_1^\top \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top) \right\rangle+\varepsilon_i\right)\\
& \quad \cdot {{\rm vec}}({{\mathbfcal{X}}}_i\times_1 {{\mathbf{U}}}_1^\top \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top) - {{\rm vec}}({{\mathbfcal{A}}}\times_1 {{\mathbf{U}}}_1^\top \times_2 {{\mathbf{U}}}_2^\top \times_3 {{\mathbf{U}}}_3^\top)\Big\|_2\\
\overset{\text{Lemma \ref{lm:concentration-Gaussian-ensemble}}}{\leq} & C\sqrt{\frac{\widetilde{\sigma}^2}{n_1}}\left(\sqrt{r_1r_2r_3} + \sqrt{\log p}\right)
\end{split}$$ with probability at least $1 - p^{-C}$. Combing the previous two inequalities, we have $$\label{ineq:regression-prob-6}
\begin{split}
& \left\|\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}} \\
\leq & \left\|\llbracket\widetilde{{{\mathbfcal{T}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket - {{\mathbfcal{T}}}\right\|_{{{\rm HS}}} + \left\|{{\mathbfcal{A}}}- {{\mathbfcal{T}}}\right\|_{{{\rm HS}}}\\
\leq & C\widetilde{\sigma}\sqrt{\frac{p_1r_1+p_2r_2+p_3r_3 + r_1r_2r_3}{n_1}}\asymp C\widetilde\sigma \sqrt{m/n_1}
\end{split}$$ with probability at least $1 - p^{-C}$. Then, for $k=1,2,3$, $$\begin{split}
& \|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\|_F \leq \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top \left(P_{\widetilde{{{\mathbf{U}}}}_k}\widetilde{{{\mathbf{A}}}}_k(P_{\widetilde{{{\mathbf{U}}}}_{k+2}}\otimes P_{\widetilde{{{\mathbf{U}}}}_{k+1}}) -{{\mathbf{A}}}_k\right)\right\|_F\\
\leq & \left\|P_{\widetilde{{{\mathbf{U}}}}_k}\widetilde{{{\mathbf{A}}}}_k(P_{\widetilde{{{\mathbf{U}}}}_{k+2}}\otimes P_{\widetilde{{{\mathbf{U}}}}_{k+1}}) -{{\mathbf{A}}}_k\right\| = \left\|\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}} \leq C\widetilde{\sigma}\sqrt{m/n_1}
\end{split}$$ with probability at least $1 - p^{-C}$.
Next, we are in the position of evaluating the estimation errors of $\widetilde{{{\mathbf{W}}}}_k$. Denote $\widetilde{{{\mathbfcal{S}}}} = \widetilde{{{\mathbfcal{A}}}}\times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top$, $\widetilde{{{\mathbf{V}}}}_k = {{\rm SVD}}_{r_k}\left(\mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}})^\top\right)$, we know $$\begin{split}
\widetilde{{{\mathbf{W}}}}_k = & (\widetilde{{{\mathbf{U}}}}_{k+2}\otimes \widetilde{{{\mathbf{U}}}}_{k+1})\widetilde{{{\mathbf{V}}}}_k = {{\rm SVD}}_{r_k}\left((\widetilde{{{\mathbf{U}}}}_{k+2}\otimes \widetilde{{{\mathbf{U}}}}_{k+1})\mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}})^\top\right)\\
= & {{\rm SVD}}_{r_k}\left(\mathcal{M}_k\left(\widetilde{{{\mathbfcal{S}}}}\times_{(k+1)}\widetilde{{{\mathbf{U}}}}_{k+1}\times_{(k+2)}\widetilde{{{\mathbf{U}}}}_{k+2}\right)^\top\right)\\
= & {{\rm SVD}}_{r_k}\left(\mathcal{M}_k\left(\widetilde{{{\mathbfcal{S}}}}\times_{(k+1)}\widetilde{{{\mathbf{U}}}}_{k+1}\times_{(k+2)}\widetilde{{{\mathbf{U}}}}_{k+2}\right)^\top \widetilde{{{\mathbf{U}}}}_k^\top\right)\\
= & {{\rm SVD}}_{r_k}\left(\mathcal{M}_k\left(\widetilde{{{\mathbfcal{S}}}}\times_{k} \widetilde{{{\mathbf{U}}}}_k\times_{(k+1)}\widetilde{{{\mathbf{U}}}}_{k+1}\times_{(k+2)}\widetilde{{{\mathbf{U}}}}_{k+2}\right)^\top\right) \\
= & {{\rm SVD}}_{r_k}\left(\mathcal{M}_k\left(\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket\right)^\top\right).
\end{split}$$ On the other hand, ${{\mathbf{W}}}_k = {{\rm SVD}}_{r_k}({{\mathbf{A}}}_k^\top) = {{\rm SVD}}_{r_k}\left(\mathcal{M}_k({{\mathbfcal{A}}})^\top\right)$. By Lemma \[lm:SVD-projection\], $$\begin{split}
\|{{\mathbf{A}}}_k \widetilde{{{\mathbf{W}}}}_{k\perp}\|_F \leq & 2\left\|\mathcal{M}_{k}(\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket) - \mathcal{M}_k({{\mathbfcal{A}}})\right\|_F \\
= & 2\left\|\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}} \overset{\eqref{ineq:regression-prob-6}}{\leq} C\widetilde{\sigma}\sqrt{\frac{m}{n_1}}
\end{split}$$ with probability at least $1 - p^{-C}$. Therefore, we also have $$\left\|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\right\|_F \leq \|\widetilde{{{\mathbf{W}}}}_{k\perp}^\top{{\mathbf{W}}}_k\|_F \leq \frac{\|\widetilde{{{\mathbf{W}}}}_{k\perp}^\top{{\mathbf{W}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}^\top {{\mathbf{A}}}_k^\top\|_F}{\sigma_{r_k}(\widetilde{{{\mathbf{W}}}}_{k\perp}^\top {{\mathbf{A}}}_k^\top)} \leq C\sqrt{\frac{\widetilde{\sigma}^2m}{\lambda_k^2n_1}}$$ with probability at least $1 - p^{-C}$.
To summarize the progress in this step, we have established the following probabilistic inequalities for $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3$ and $\widetilde{{{\mathbf{W}}}}_1, \widetilde{{{\mathbf{W}}}}_2, \widetilde{{{\mathbf{W}}}}_3$, $$\label{ineq:conclustion-HOOI}
\begin{split}
& \left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k\right)\right\| \leq \frac{C\widetilde{\sigma} \sqrt{p_k/n_1}}{\lambda_k}, \quad \left\|\sin\Theta\left(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k\right)\right\|_F \leq \frac{C\widetilde{\sigma} \sqrt{m/n_1}}{\lambda_k},\quad k=1,2,3,
\end{split}$$ $$\label{ineq:conclustion-HOOI-2}
\begin{split}
& \left\|\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\right\|_F \leq C\widetilde{\sigma}\sqrt{m/n_1},\quad \left\|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}\right\|_F \leq C\widetilde{\sigma}\sqrt{m/n_1},\quad k=1,2,3,
\end{split}$$ with probability at least $1 - p^{-C}$.
4. For the rest of the proof, we assume and hold. Next, we move on to evaluate the estimation error bound for $\widehat{{{\mathbfcal{A}}}}$. The focus now shifts from the first batch of samples $({{\mathbfcal{X}}}^{(1)}, {{\mathbf{y}}}^{(1)})$ to the second one $({{\mathbfcal{X}}}^{(2)}, y^{(2)})$. Denote $$\label{ineq:regression-theta}
\begin{split}
\theta_k := \left\|\sin\Theta\left(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k\right)\right\| \overset{\eqref{ineq:conclustion-HOOI}}{\leq} \frac{C\widetilde{\sigma}\sqrt{p_k/n_1}}{\lambda_k},\quad k=1,2,3;\\
\end{split}$$ $$\label{ineq:regression-xi}
\begin{split}
\xi_k := \|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}\|_F \overset{\eqref{ineq:conclustion-HOOI-2}}{\leq} C\widetilde{\sigma}\sqrt{m/n_1}, \quad k=1,2,3;
\end{split}$$ $$\label{ineq:regression-eta}
\begin{split}
\eta_k := & \left\|\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\right\|_F \overset{\eqref{ineq:conclustion-HOOI-2}}{\leq} C\widetilde{\sigma}\sqrt{m/n_1},\quad k=1,2,3;
\end{split}$$ $$\label{ineq:regression-hat-sigma}
\widehat{\sigma}^2 := \left\|P_{\widetilde{{{\mathbf{U}}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2^2 + \sigma^2.$$ By Lemma \[lm:projection\_remainder\], $$\begin{split}
\|P_{\widetilde{{{\mathbf{U}}}}_{\perp}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2\leq & \frac{C\widetilde{\sigma}^4mp}{n_1^2\lambda_0^2} + \frac{C_1\widetilde{\sigma}^6mp^2}{\lambda_0^4n_1^3}.
\end{split}$$ Provided that $m = r_1r_2r_3+\sum_k(p_k-r_k)r_k$ and $n_1\geq \frac{C\widetilde{\sigma}^2p}{\lambda_0^2}$, we know $$\label{ineq:hat-sigma}
\begin{split}
& \|P_{\widetilde{{{\mathbf{U}}}}_{\perp}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2\leq \frac{C\widetilde{\sigma}^4mp}{n_1^2\lambda_0^2}, \quad \widehat{\sigma}^2\leq \sigma^2 + \frac{C\widetilde{\sigma}^4 mp}{n_1^2\lambda_0^2}.
\end{split}$$
5. In this step, we evaluate two crucial quantities for applying the oracle inequality (Theorem \[th:upper\_bound\_general\]). Recall the importance sketching covariates are defined as $$\begin{split}
& \widetilde{{{\mathbf{X}}}} = \left[\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_1} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_2} ~~ \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_3}\right] \in \mathbb{R}^{n_2\times m}, \\
& \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n\times (r_1r_2r_3)},\quad \left(\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\right)_{[i,:]} =
{{\rm vec}}\left({{\mathbfcal{X}}}_i^{(2)} \times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3^\top\right),\\
& \widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k} \in \mathbb{R}^{n\times (p_k-r_k)r_k}, \quad \left(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k}\right)_{[i,:]} = {{\rm vec}}\left(\widetilde{{{\mathbf{U}}}}_{k\perp}^\top \mathcal{M}_k\left({{\mathbfcal{X}}}_{i}^{(2)} \times_{k+1} \widetilde{{{\mathbf{U}}}}_{k+1}^\top \times_{k+2} \widetilde{{{\mathbf{U}}}}_{k+2}^\top\right) \widetilde{{{\mathbf{V}}}}_k\right).
\end{split}$$ When ${{\mathbfcal{X}}}_i^{(2)}$ are i.i.d. Gaussian matrices and independent of $\widetilde{{{\mathbf{U}}}}_k$, $\widetilde{{{\mathbf{V}}}}_k$, $\widetilde{{{\mathbf{W}}}}_k$, $\widetilde{{\mathbf{X}}}$ can be seen as an orthogonal projection of ${{\mathbfcal{X}}}_i^{(2)}$ and has i.i.d. Gaussian entries. Thus, by Proposition 5.35 in [@vershynin2010introduction], $$\begin{split}
{\mathbb{P}}\left(\sigma_{\min}(\widetilde{{\mathbf{X}}}^\top\widetilde{{\mathbf{X}}}) = \sigma_{\min}^2(\widetilde{{{\mathbf{X}}}}) \geq \left(\sqrt{n_2}-\sqrt{m}-t\right)^2\right) \geq 1 - \exp(-t^2/2).
\end{split}$$ By definition, $\widetilde{\varepsilon}\in \mathbb{R}^{n_2}$ is independent of $\widetilde{{\mathbf{X}}}$, and $$\begin{split}
\widetilde\varepsilon_j = \langle{{\mathbfcal{X}}}_j^{(2)}, P_{\widetilde{{\mathbf{U}}}_{\perp}}{{\mathbfcal{A}}}\rangle + \varepsilon_j \sim N\left(0, \left\|P_{\widetilde{{{\mathbf{U}}}}_{\perp}}{{\rm vec}}({{\mathbfcal{A}}}) \right\|_{2}^2+\sigma^2\right) = N(0, \widehat\sigma^2).
\end{split}$$ Then, $\|\widetilde{{\boldsymbol{\varepsilon}}}\|_2^2\sim \widehat\sigma^2\chi^2_{n_2}$ and $\|\widetilde{{\mathbf{X}}}^\top\widetilde{\boldsymbol{\varepsilon}}\|_2^2 \Big| \|{\boldsymbol{\varepsilon}}\|_2^2 \sim \|{\boldsymbol{\varepsilon}}\|_2^2\chi^2_{m}$. Based on $\chi^2$ distribution tail bound [@laurent2000adaptive Lemma 1] and $n_2\geq C(p^{3/2} + r^3) \geq Cm$, $$\label{ineq:XXXepsilon-bound}
\begin{split}
& \left\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top\widetilde{{\boldsymbol{\varepsilon}}}\right\|_2^2 \\
\leq & \frac{\widehat{\sigma}^2\left(n_2+2\sqrt{n_2C_1\log(p)} + 2C_2\log(p)\right)\left(m+2\sqrt{mC_3\log(p)}+2C\log(p)\right)}{\left(\sqrt{n_2}-\sqrt{m}-C_4\log(p)\right)^4}\\
\leq & \frac{\widehat{\sigma}^2m}{n_2}\frac{\left(1 + 2\sqrt{\frac{C\log p}{n_2}}+2\frac{\log p}{n_2}\right)\left(1+2\sqrt{\frac{t}{m}}+2\frac{t}{m}\right)}{\left(1 - \sqrt{\frac{m}{n_2}} - \frac{C_1\log(p)}{\sqrt{n_2}}\right)^4}\\
= & \frac{\widehat{\sigma}^2m}{n_2}\left(1 + C_1\sqrt{\frac{m}{n_2}} + C_2\sqrt{\frac{\log p}{m}}\right).
\end{split}$$ with probability at least $1 - p^{-C}$.
We assume holds. It remains to check $\left\|\widehat{{\mathbf{D}}}_k(\widehat{{{\mathbf{B}}}}_k\widetilde{{\mathbf{V}}}_k)^{-1}\right\|$. Similarly as the proof of Theorem \[th:upper\_bound\_general\], we define $$\begin{split}
\widetilde{{{\mathbfcal{B}}}} = & \left\llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket = \left\llbracket {{\mathbfcal{S}}}\times_1 {{\mathbf{U}}}_1\times_2 {{\mathbf{U}}}_2\times_3 {{\mathbf{U}}}_3; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top , \widetilde{{{\mathbf{U}}}}_3^\top \right\rrbracket \in \mathbb{R}^{r_1\times r_2\times r_3};\\
\widetilde{{{\mathbf{B}}}}_k = & \mathcal{M}_k(\widetilde{{{\mathbfcal{B}}}}) \in \mathbb{R}^{r_k\times (r_{k+1}r_{k+2})}, \quad k=1,2,3,\\
\widetilde{{{\mathbf{D}}}}_1 = & \widetilde{{{\mathbf{U}}}}_{1\perp}^\top \mathcal{M}_1({{\mathbfcal{A}}}\times_2 \widetilde{{{\mathbf{U}}}}_2^\top \times_3 \widetilde{{{\mathbf{U}}}}_3) \widetilde{{{\mathbf{V}}}}_1 \overset{\text{Lemma \ref{lm:Kronecker-vectorization-matricization}}}{=} \widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1 \widetilde{{{\mathbf{W}}}}_1 \in \mathbb{R}^{(p_1-r_1)\times r_1},\\
\widetilde{{{\mathbf{D}}}}_2 = & \widetilde{{{\mathbf{U}}}}_{2\perp}^\top \mathcal{M}_2({{\mathbfcal{A}}}\times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_3 \widetilde{{{\mathbf{U}}}}_3) \widetilde{{{\mathbf{V}}}}_2 = \widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{A}}}_2 \widetilde{{{\mathbf{W}}}}_2 \in \mathbb{R}^{(p_2-r_2)\times r_2},\\
\widetilde{{{\mathbf{D}}}}_3 = & \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \mathcal{M}_3({{\mathbfcal{A}}}\times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_2) \widetilde{{{\mathbf{V}}}}_3 = \widetilde{{{\mathbf{U}}}}_{3\perp}^\top {{\mathbf{A}}}_3 \widetilde{{{\mathbf{W}}}}_3 \in \mathbb{R}^{(p_3-r_3)\times r_3}.
\end{split}$$ By the proof of Theorem \[th:upper\_bound\_general\], we have $$\label{ineq:B-B+D-D}
\begin{split}
& \left\|\widehat{{\mathbfcal{B}}}- \widetilde{{{\mathbfcal{B}}}}\right\|_{{{\rm HS}}}^2 + \sum_{k=1}^3\left\|\widehat{{\mathbf{D}}}_k-\widetilde{{\mathbf{D}}}_k\right\|_F^2 \overset{\eqref{th:hat_B-B}}{\leq} \left\|(\widetilde{{\mathbf{X}}}^\top\widetilde{{\mathbf{X}}})^{-1}\widetilde{{\mathbf{X}}}^\top\widetilde\varepsilon\right\|_2^2 \\
\overset{\eqref{ineq:XXXepsilon-bound}}{\leq} & \frac{\widehat{\sigma}^2m}{n_2}\left(1 + C_1s\sqrt{\frac{\log m}{n_2}} + C_2\sqrt{\frac{\log p}{m}}\right),
\end{split}$$ $$\label{ineq:tilde-DBV}
\|\widetilde{{\mathbf{D}}}_k(\widetilde{{\mathbf{B}}}_k\widetilde{{\mathbf{V}}}_k)^{-1}\| \overset{\eqref{ineq:tilde-D-tilde-B-tilde-V}
}{\leq} C\max_k\left\{\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\|, \|\sin\Theta({{\mathbf{W}}}_k, {{\mathbf{W}}}_k)\|\right\} \leq \frac{C\widetilde{\sigma}\sqrt{m/n_1}}{\lambda_k},$$ $$\sigma_{\min}(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k) = \sigma_{\min}(\widetilde{{{\mathbf{U}}}}_k^\top {{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_k) \overset{\eqref{ineq:sigma_min-tildeUAW}}{\geq} \lambda_k\left(1 - \frac{C\widetilde{\sigma}^2m}{\lambda_k^2n_1}\right) \geq \lambda_k(1-c)$$ for some constant $0<c<1$. This additionally means $$\label{ineq:hat-B-tilde-V}
\sigma_{\min}\left(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\right) \geq \sigma_{\min}(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k) - \|\widehat{{{\mathbf{B}}}}_k - {{\mathbf{B}}}_k\| \overset{\eqref{ineq:B-B+D-D}}{\geq} \lambda_k\left(1 - \frac{C\widetilde{\sigma}^2m}{\lambda_k^2n_1}\right) - \frac{C\widehat{\sigma}^2m}{n_2} \geq (1-c)\lambda_k.$$ It is easy to check that the following equality, $$(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1} = (\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1} + (\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1} \left(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k - \widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k\right) (\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}.$$ Thus, $$\label{ineq:delta-upper-bound}
\begin{split}
\rho := \left\|\widehat{{{\mathbf{D}}}}_k(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\| \leq & \left\|(\widehat{{{\mathbf{D}}}}_k-\widetilde{{{\mathbf{D}}}}_k)(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\| + \left\|\widetilde{{{\mathbf{D}}}}_k(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\|\\
\leq & \frac{C\left\|\widehat{{{\mathbf{D}}}}_k -\widetilde{{{\mathbf{D}}}}_k\right\|}{\lambda_k} + \left\|\widetilde{{{\mathbf{D}}}}_k(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\| \\
& + \left\|\widetilde{{{\mathbf{D}}}}_k(\widetilde{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\right\| \cdot \left\|(\widetilde{{{\mathbf{B}}}}_k-\widehat{{{\mathbf{B}}}}_k)\widetilde{{{\mathbf{V}}}}_k\right\| \cdot \|(\widehat{{{\mathbf{B}}}}_k\widetilde{{{\mathbf{V}}}}_k)^{-1}\|\\
\overset{\eqref{ineq:B-B+D-D}\eqref{ineq:tilde-DBV}\eqref{ineq:hat-B-tilde-V}}{\leq} & \frac{C\widetilde{\sigma}}{\lambda_k} \sqrt{\frac{m}{n_1}} + \frac{C\widehat{\sigma}}{\lambda_k}\sqrt{\frac{m}{n_2}}.
\end{split}$$
6. Finally, we apply the oracle inequality, i.e., Theorem \[th:upper\_bound\_general\], and obtain the final upper bound for $\widehat{{\mathbfcal{A}}}$. We have shown that the conditions of Theorem \[th:upper\_bound\_general\] holds if , , and hold. Then Theorem \[th:upper\_bound\_general\] implies $$\begin{split}
\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \leq & (1+C\theta+C\rho)\left\|(\widetilde{{{\mathbf{X}}}}^\top\widetilde{{{\mathbf{X}}}})^{-1}\widetilde{{{\mathbf{X}}}}^\top\widetilde{{\boldsymbol{\varepsilon}}}\right\|_2^2\\
\overset{\eqref{ineq:regression-theta}\eqref{ineq:XXXepsilon-bound}\eqref{ineq:delta-upper-bound}}{\leq} & \frac{\widehat{\sigma}m}{n_2}\left(1 + C_1\sqrt{\frac{m}{n_2}} + C_2\sqrt{\frac{\log p}{m}} + \frac{C_3\widetilde{\sigma}}{\lambda_0}\sqrt{\frac{m}{n_1}} + \frac{C_4\widehat{\sigma}}{\lambda_0}\sqrt{\frac{m}{n_2}}\right)\\
\overset{\eqref{ineq:hat-sigma}}{\leq} & \frac{m}{n_2}\left(\sigma^2+\frac{C_1\widetilde{\sigma}^4mp}{n_1^2\lambda_0^2}\right)\left(1 + C_2\sqrt{\frac{m}{n_2}} + C_3\sqrt{\frac{\log p}{m}} + \frac{C_4\widetilde{\sigma}}{\lambda_0}\sqrt{\frac{m}{n_1}} + \frac{C_5\widehat{\sigma}}{\lambda_0}\sqrt{\frac{m}{n_2}}\right)\\
\leq & \frac{m}{n_2}\left(\sigma^2 + \frac{C_1\widetilde{\sigma}^4mp}{n_1^2\lambda_0^2}\right)\left(1 + C_2\sqrt{\frac{\log p}{m}} + C_3\sqrt{\frac{m\widetilde{\sigma}^2}{(n_1\wedge n_2)\lambda_0^2}}\right)
\end{split}$$ with probability at least $1 - p^{-C}$. Here, the last inequality is due to $n_1\wedge n_2\geq C\widetilde{\sigma}^2(p^{3/2}+r^3)/\lambda_0^2$ and $\widehat{\sigma}=\|{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2+\sigma^2 \geq \lambda_0$. $\square$
Proof of Theorem \[th:lower-bound-regression\] {#sec:proof_lower-bound-regression}
----------------------------------------------
In this theorem, we provide an estimation error lower bound for low-rank tensor regression. The central idea is to carefully transform the original high-dimensional low-rank tensor regression model to the unconstrained dimension-reduced linear regression model , then apply the classic Bayes risk of linear regression (Lemma \[lm:linear-regression-lower-bound\]) to finalize the desired lower bound on estimation error.
Since $r_1, r_2$, and $r_3$ satisfy $r_k \leq r_{k+1}r_{k+2}$ for $k=1,2,3$, the $r_1$-by-$r_2$-by-$r_3$ tensor with i.i.d. normal entries has full Tucker rank with probability 1. Thus, we can set ${{\mathbfcal{S}}}_0\in \mathbb{R}^{r_1\times r_2\times r_3}$ as a fixed tensor with full Tucker rank, i.e., ${{\rm rank}}({{\mathbfcal{S}}}_0) = (r_1, r_2, r_3)$. Let $T>0$ be a large to-be-specified constant. Define $${{\mathbfcal{A}}}_0\in \mathbb{R}^{p_1\times p_2\times p_3}, \quad ({{\mathbfcal{A}}}_0)_{[1:r_1, 1:r_2, 1:r_3]} = T{{\mathbfcal{S}}}_0, \quad ({{\mathbfcal{A}}}_0)_{[1:r_1, 1:r_2, 1:r_3]^c} = 0.$$ Suppose ${{\mathbf{U}}}_k\in \mathbb{O}_{p_k, r_k}$ and ${{\mathbf{W}}}_k\in \mathbb{O}_{p_{-k}, r_k}$ are the left and right singular subspaces of $\mathcal{M}_k({{\mathbfcal{A}}}_0)$, respectively; ${{\mathbf{V}}}_k\in \mathbb{O}_{r_{k+1}r_{k+2}, r_k}$ is the right singular subspace of $\mathcal{M}_k({{\mathbfcal{S}}}_0)$. Then by definition of ${{\mathbfcal{A}}}_0$, $${{\mathbf{U}}}_k = \begin{bmatrix}
{{\mathbf{I}}}_{r_k}\\
\boldsymbol{0}_{(p_k-r_k) \times r_k}
\end{bmatrix}, \quad k=1,2,3.$$ Next, for to-be-specified values $\tau, T>0$, we introduce a prior distribution $\bar{P}_{\tau, T}$ on the class of $\mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}$: the $p_1$-by-$p_2$-by-$p_3$ random tensor $\bar{{{\mathbfcal{A}}}} \sim \bar{P}_{\tau,T}$ if and only if it can be generated based on the following process.
1. Generate an $r_1$-by-$r_2$-by-$r_3$ tensor ${{\mathbfcal{B}}}\overset{iid}{\sim} N(0, \tau^2)$ and assign $\bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, 1:r_3]} = T{{\mathbfcal{S}}}_0 + {{\mathbfcal{B}}}$.
2. Suppose $\mathcal{M}_k(\bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, 1:r_3]}) = \bar{{{\mathbf{A}}}}_{0k}\in \mathbb{R}^{r_k\times r_{-k}}$ and $\bar{{{\mathbf{V}}}}_k = {{\rm SVD}}_{r_k}(\bar{{{\mathbf{A}}}}_{0k}^\top)\in \mathbb{O}_{r_{-k}, r_k}$. Assign $$\mathcal{M}_1\left(\bar{{{\mathbfcal{A}}}}_{[(r_1+1):p_1, 1:r_2, 1:r_3]}\right) = {{\mathbf{B}}}_{1}\cdot \bar{{{\mathbf{V}}}}_1^\top,$$ $$\mathcal{M}_2\left(\bar{{{\mathbfcal{A}}}}_{[1:r_1, (r_2+1):p_2, 1:r_3]}\right) = {{\mathbf{B}}}_{2}\cdot \bar{{{\mathbf{V}}}}_2^\top,$$ $$\mathcal{M}_3\left(\bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, (r_3+1):p_3]}\right) = {{\mathbf{B}}}_{3}\cdot \bar{{{\mathbf{V}}}}_3^\top,$$ where all entries of ${{\mathbf{B}}}_{1}\in \mathbb{R}^{(p_1-r_1)\times r_1}, {{\mathbf{B}}}_{2}\in\mathbb{R}^{(p_2-r_2)\times r_2}, {{\mathbf{B}}}_{3}\in\mathbb{R}^{(p_3-r_3)\times r_3}$ are independently drawn from $N(0,\tau^2)$.
3. The other blocks of $\bar{{{\mathbfcal{A}}}}$ are calculated as follows, $$\begin{split}
& \bar{{{\mathbfcal{A}}}}_{[(r_1+1):p_1, (r_2+1):p_2, 1:r_3]} = \bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, 1:r_3]} \times_1 \left({{\mathbf{B}}}_{1}(\bar{{{\mathbf{A}}}}_{01}\bar{{{\mathbf{V}}}}_1)^{-1}\right) \times_2 \left({{\mathbf{B}}}_{2}(\bar{{{\mathbf{A}}}}_{02}\bar{{{\mathbf{V}}}}_2)^{-1}\right), \\
& \bar{{{\mathbfcal{A}}}}_{[(r_1+1):p_1, 1:r_2, (r_3+1):p_3]} = \bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, 1:r_3]} \times_1 \left({{\mathbf{B}}}_{1}(\bar{{{\mathbf{A}}}}_{01}\bar{{{\mathbf{V}}}}_1)^{-1}\right) \times_3 \left({{\mathbf{B}}}_{3}(\bar{{{\mathbf{A}}}}_{03}\bar{{{\mathbf{V}}}}_3)^{-1}\right), \\
& \bar{{{\mathbfcal{A}}}}_{[1:r_1, (r_2+1):p_2, (r_3+1):p_3]} = \bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, 1:r_3]} \times_2 \left({{\mathbf{B}}}_{2}(\bar{{{\mathbf{A}}}}_{02}\bar{{{\mathbf{V}}}}_2)^{-1}\right) \times_3 \left({{\mathbf{B}}}_{3}(\bar{{{\mathbf{A}}}}_{03}\bar{{{\mathbf{V}}}}_3)^{-1}\right), \\
& \bar{{{\mathbfcal{A}}}}_{[(r_1+1):p_1, (r_2+1):p_2, (r_3+1):p_3]} \\
& \quad = \bar{{{\mathbfcal{A}}}}_{[1:r_1, 1:r_2, 1:r_3]} \times_1 \left({{\mathbf{B}}}_{1}(\bar{{{\mathbf{A}}}}_{01}\bar{{{\mathbf{V}}}}_1)^{-1}\right) \times_2 \left({{\mathbf{B}}}_{2}(\bar{{{\mathbf{A}}}}_{02}\bar{{{\mathbf{V}}}}_2)^{-1}\right) \times_3 \left({{\mathbf{B}}}_{3}(\bar{{{\mathbf{A}}}}_{03}\bar{{{\mathbf{V}}}}_3)^{-1}\right).
\end{split}$$
One can check by comparing each block that $\bar{{{\mathbfcal{A}}}}$ satisfies $$\begin{split}
& \bar{{{\mathbfcal{A}}}} = \left\llbracket T{{\mathbfcal{S}}}_0 + {{\mathbfcal{B}}}; \bar{{{\mathbf{L}}}}_1, \bar{{{\mathbf{L}}}}_2, \bar{{{\mathbf{L}}}}_3\right\rrbracket, \quad \text{where}\quad \bar{{{\mathbf{L}}}}_k = \begin{bmatrix}
{{\mathbf{I}}}_{r_k}\\
{{\mathbf{B}}}_{k}(\bar{{{\mathbf{A}}}}_{0k}\bar{{{\mathbf{V}}}}_k)^{-1}
\end{bmatrix},\quad k=1,2,3.
\end{split}$$ Thus, ${{\rm rank}}(\bar{{{\mathbfcal{A}}}}) \leq (r_1, r_2, r_3)$ and $\bar{{{\mathbfcal{A}}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}$. Then we consider another distribution $P^\ast_{\tau, T}$ on the whole tensor space $\mathbb{R}^{p_1\times p_2\times p_3}$, $$\begin{split}
{{\mathbfcal{A}}}^\ast \sim P^\ast_{\tau,T}, \quad \text{such that}\quad & {{\mathbfcal{A}}}^\ast_{[1:r_1, 1:r_2, 1:r_3]} = T{{\mathbfcal{S}}}_0 + {{\mathbfcal{B}}}, \\
& \mathcal{M}_1\left({{\mathbfcal{A}}}^\ast_{[(r_1+1):p_1, 1:r_2, 1:r_3]}\right) = {{\mathbf{B}}}_{1}\cdot {{\mathbf{V}}}_{1}^\top; \\
& \mathcal{M}_2\left({{\mathbfcal{A}}}^\ast_{[1:r_1, (r_2+1):p_2, 1:r_3]}\right) = {{\mathbf{B}}}_{2}\cdot {{\mathbf{V}}}_{2}^\top; \\
& \mathcal{M}_3\left({{\mathbfcal{A}}}^\ast_{[(r_1+1):p_1, 1:r_2, 1:r_3]}\right) = {{\mathbf{B}}}_{3}\cdot {{\mathbf{V}}}_{3}^\top; \\
& \text{the other blocks of ${{\mathbfcal{A}}}^\ast$ are set to zero}.
\end{split}$$ Here, ${{\mathbfcal{B}}}, {{\mathbf{B}}}_{1}, {{\mathbf{B}}}_{2}, {{\mathbf{B}}}_{3} \overset{iid}{\sim}N(0, \tau^2)$. Suppose $\bar{{{\mathbfcal{A}}}}\sim \bar{P}_{\tau,T}$ and ${{\mathbfcal{A}}}^\ast\sim P^\ast_{\tau,T}$. Recall that ${{\mathbf{V}}}_k = {{\rm SVD}}_{r_k}(\mathcal{M}_k({{\mathbfcal{S}}}_0)^\top))$ and $\bar{{{\mathbf{V}}}}_k = {{\rm SVD}}_{r_k}(\mathcal{M}_k({{\mathbfcal{S}}}_0+{{\mathbfcal{B}}}/T)^\top)$. As $T\to \infty$, we must have $$\bar{{{\mathbf{V}}}}_k \overset{d}{\to} {{\mathbf{V}}}_k \quad \text{and}\quad (\bar{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}_0) \overset{d}{\to} ({{\mathbfcal{A}}}^\ast - {{\mathbfcal{A}}}_0).$$
Next, we move on to the regular tensor regression model $$y_i = \langle {{\mathbfcal{X}}}_i, {{\mathbfcal{A}}}\rangle + \varepsilon_i, \quad i=1,\ldots, n.$$ For convenience, we divide ${{\mathbfcal{X}}}_i$ and ${{\mathbfcal{A}}}$ into eight blocks and denote them separately as $$\begin{split}
& {{\mathbfcal{X}}}_{i, s_1s_2s_3} = ({{\mathbfcal{X}}}_i)_{[I_{1, s_1}, I_{2, s_2}, I_{3, s_3}]},\quad {{\mathbfcal{A}}}_{s_1s_2s_3} = {{\mathbfcal{A}}}_{[I_{1, s_1}, I_{2, s_2}, I_{3, s_3}]}, \text{ for } s_1, s_2, s_3 \in \{1, 2\}, \\
\text{where}\quad & I_{k, 1} = \{1,\ldots, r_k\}, \quad I_{k, 2} = \{r_k+1,\ldots, p_k\}, \quad k=1,2,3.
\end{split}$$ If ${{\mathbfcal{A}}}^\ast \sim P_{\tau, T}^\ast$, ${{\mathbfcal{A}}}^\ast_{122}, {{\mathbfcal{A}}}^\ast_{212}, {{\mathbfcal{A}}}^\ast_{221}, {{\mathbfcal{A}}}^\ast_{222}$ are all zeros. Then, $$\begin{split}
y_i = & \langle {{\mathbfcal{X}}}_i, {{\mathbfcal{A}}}^\ast\rangle + \varepsilon_i = \sum_{s_1, s_2, s_3=1}^2\langle{{\mathbfcal{X}}}_{i,s_1s_2s_3}, {{\mathbfcal{A}}}^\ast_{s_1s_2s_3}\rangle + \varepsilon_i\\
= & \langle ({{\mathbfcal{X}}}_{i, 111}, T{{\mathbfcal{S}}}_0 + {{\mathbfcal{B}}}\rangle + \langle\mathcal{M}_1({{\mathbfcal{X}}}_{i, 211}), {{\mathbf{B}}}_{1}{{\mathbf{V}}}_1^\top\rangle\\
& + \langle\mathcal{M}_2({{\mathbfcal{X}}}_{i, 121}), {{\mathbf{B}}}_{2}{{\mathbf{V}}}_2^\top\rangle + \langle\mathcal{M}_3({{\mathbfcal{X}}}_{i,112}), {{\mathbf{B}}}_{3} {{\mathbf{V}}}_3^\top\rangle + \varepsilon_i\\
= & \langle {{\mathbfcal{X}}}_i, {{\mathbfcal{A}}}_0 \rangle + \varepsilon_i + \langle {{\rm vec}}({{\mathbfcal{X}}}_{i, 111}), {{\rm vec}}({{\mathbfcal{B}}})\rangle + \langle \mathcal{M}_1({{\mathbfcal{X}}}_{i, 211}){{\mathbf{V}}}_1, {{\mathbf{B}}}_{1}\rangle \\
& + \langle \mathcal{M}_2({{\mathbfcal{X}}}_{i, 121}){{\mathbf{V}}}_2, {{\mathbf{B}}}_{2}\rangle + \langle \mathcal{M}_3({{\mathbfcal{X}}}_{i, 112}){{\mathbf{V}}}_3, {{\mathbf{B}}}_{3}\rangle\\
:= & \langle {{\mathbfcal{X}}}_i, {{\mathbfcal{A}}}_0\rangle + \langle \bar{{{\mathbf{X}}}}_i, {{\mathbf{b}}}\rangle + \varepsilon_i,
\end{split}$$ where $$\begin{split}
\bar{{{\mathbf{X}}}}_i = & \begin{bmatrix}
{{\rm vec}}\left({{\mathbfcal{X}}}_{i,111}\right)\\
{{\rm vec}}\left(\mathcal{M}_1({{\mathbf{X}}}_{i, 211}){{\mathbf{V}}}_1\right)\\ {{\rm vec}}\left(\mathcal{M}_2({{\mathbf{X}}}_{i, 121}){{\mathbf{V}}}_2\right)\\ {{\rm vec}}\left(\mathcal{M}_3({{\mathbf{X}}}_{i, 112}){{\mathbf{V}}}_3\right)
\end{bmatrix} \in \mathbb{R}^{m},\quad \bar{{{\mathbf{X}}}} = \begin{bmatrix}
\bar{{{\mathbf{X}}}}_1^\top\\
\vdots \\
\bar{{{\mathbf{X}}}}_n^\top
\end{bmatrix}\in \mathbb{R}^{n\times m},\quad
{{\mathbf{b}}}= \begin{bmatrix}
{{\rm vec}}({{\mathbfcal{B}}})\\ {{\rm vec}}({{\mathbf{B}}}_{1})\\ {{\rm vec}}({{\mathbf{B}}}_{2})\\ {{\rm vec}}({{\mathbf{B}}}_{3})
\end{bmatrix} \in \mathbb{R}^{m}.
\end{split}$$ Suppose the parameter ${{\mathbfcal{A}}}^\ast$ is drawn from the prior distribution $P_{\tau, T}^\ast$. Then, ${{\mathbf{b}}}\overset{iid}{\sim}N(0, \tau^2)$. Note that $\bar{{\mathbf{X}}}_i$ is an orthogonal projection of ${{\mathbfcal{X}}}_i$, so $\bar{{{\mathbf{X}}}}_i\overset{iid}{\sim}N(0, 1)$. Now, $y_i, \bar{{{\mathbf{X}}}}_i, \bar{{{\mathbf{b}}}}$ can be related by the following regression model, $$\label{eq:dimension-reduced-regression}
\begin{split}
& y_i - \langle {{\mathbfcal{X}}}_i, {{\mathbfcal{A}}}_0\rangle = \bar{{{\mathbf{X}}}}_i^\top {{\mathbf{b}}}+ \varepsilon_i,\quad i=1,\ldots, n;\\
& {{\mathbf{b}}}\overset{iid}{\sim}N(0, \tau^2),\quad \varepsilon \overset{iid}{\sim}N(0,\sigma^2).
\end{split}$$ By the construction of ${{\mathbfcal{A}}}^\ast$ and the setting that ${{\mathbfcal{S}}}_0$ is fixed, the estimation of ${{\mathbfcal{A}}}^\ast$ is equivalent to the estimation ${{\mathbf{b}}}$. By Lemma \[lm:linear-regression-lower-bound\], the Bayes risk of estimating ${{\mathbf{b}}}$ (and the Bayes risk of estimating ${{\mathbfcal{A}}}^\ast$ if ${{\mathbfcal{A}}}^\ast\sim P_{\tau, T}$) is $$\left\|\widehat{{\mathbfcal{A}}}^\ast - {{\mathbfcal{A}}}^\ast\right\|_{{{\rm HS}}}^2\Big|\{\bar{{\mathbf{X}}}_i\}_{i=1}^n = \left\|\widehat{{{\mathbf{b}}}} - {{\mathbf{b}}}\right\|_2^2\Big|\{\bar{{\mathbf{X}}}_i\}_{i=1}^n = {{\rm tr}}\left(\left(\frac{{{\mathbf{I}}}_m}{\tau^2}+\frac{\bar{{\mathbf{X}}}^\top\bar{{\mathbf{X}}}}{\sigma^2}\right)^{-1}\right).$$ Here, $\widehat{{\mathbfcal{A}}}^\ast$ and $\widehat{{{\mathbf{b}}}}$ are the posterior mean of ${{\mathbfcal{A}}}^\ast$ and ${{\mathbf{b}}}$, respectively.
Since $\bar{P}_{\tau, T}\to P_{\tau, T}$ and $\bar{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}_0 \to {{\mathbfcal{A}}}^\ast - {{\mathbfcal{A}}}_0$ as $T\to \infty$, we have $$\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - \bar{{{\mathbfcal{A}}}}\right\|_{{{\rm HS}}}^2 \Big| \{\bar{{\mathbf{X}}}_i\}_{i=1}^n \to \mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}}^\ast - {{\mathbfcal{A}}}^\ast\right\|_{{{\rm HS}}}^2 \Big| \{\bar{{\mathbf{X}}}_i\}_{i=1}^n = {{\rm tr}}\left(\left(\frac{{{\mathbf{I}}}_m}{\tau^2} + \frac{\bar{{{\mathbf{X}}}}^\top\bar{{\mathbf{X}}}}{\sigma^2}\right)^{-1}\right),$$ where $\widehat{{\mathbfcal{A}}}$ is the posterior mean of $\bar{{{\mathbfcal{A}}}}$ if $\bar{{\mathbfcal{A}}}\sim \bar{P}_{\tau, T}$. Since $\bar{{{\mathbfcal{A}}}} \sim \bar{P}_{\tau, T}$ and $\bar{P}_{\tau, T}$ is the distribution on $\mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}$, we have the following estimation lower bound, $$\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}} \left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2\Big| \{\bar{{\mathbf{X}}}_i\}_{i=1}^n \geq {{\rm tr}}\left(\left(\frac{{{\mathbf{I}}}_m}{\tau^2} + \frac{\bar{{\mathbf{X}}}^\top \bar{{\mathbf{X}}}}{\sigma^2}\right)^{-1}\right).$$
Finally, since $(\bar{{{\mathbf{X}}}}^\top \bar{{{\mathbf{X}}}})^{-1}$ is inverse Wishart distributed and[^6] $${{\rm tr}}(\mathbb{E}(\bar{{{\mathbf{X}}}}^\top \bar{{{\mathbf{X}}}})^{-1}) =\left\{\begin{array}{ll} \frac{1}{n-m-1}{{\rm tr}}({{\mathbf{I}}}_m) = \frac{m}{n-m-1} & n> m+1;\\
\infty & n\leq m+1.
\end{array}\right.$$ By letting $\tau \to \infty$, we finally obtain $$\begin{split}
& \inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}} \left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \limsup_{\tau\to \infty}\mathbb{E}{{\rm tr}}\left(\left(\frac{{{\mathbf{I}}}_m}{\tau^2} + \frac{\bar{{\mathbf{X}}}^\top \bar{{\mathbf{X}}}}{\sigma^2}\right)^{-1}\right) \\
= & {{\rm tr}}\left(\frac{\sigma^2 {{\mathbf{I}}}_m}{n-m-1}\right) = \left\{\begin{array}{ll}
\frac{m\sigma^2}{n-m-1}, & \text{if } n>m+1;\\
+\infty & \text{if }n\leq m+1.
\end{array}\right.
\end{split}$$ $\square$
Proof of Theorem \[th:upper\_bound\_sparse\_tensor\_regression\] {#sec:proof_upper_bound_sparse_tensor_regression}
----------------------------------------------------------------
In this theorem, we aim to establish an estimation error upper bound for sparse ISLET in sparse low-rank tensor regression problem. After introducing some necessary notations, we develop the estimation error bounds for sketching directions $\widetilde{{{\mathbf{U}}}}_k$ and $\widetilde{{{\mathbf{W}}}}_k$ in Steps 1 and 2. In Step 3, we give error bounds for a number of intermediate terms. In Step 4, we prove upper bounds for key quantities $\rho, \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbf{B}}}\right\|_2^2, \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2$, and $\max_{i = 1,\ldots, p_k} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k, [:, G_i^k]})^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n\right\|_2^2$. Finally, we plug in these values to Theorem \[th:upper\_bound\_sparse\_general\] to finalize the proof.
We first introduce a number of notations that will be used in the proof. Similarly as the proof of Theorem \[th:upper\_bound\_regression\], denote $${{\mathbf{A}}}_k = \mathcal{M}_k({{\mathbfcal{A}}}), \quad {{\mathbf{S}}}_k = \mathcal{M}_k({{\mathbfcal{S}}}),$$ $$\widetilde{{{\mathbf{A}}}}_k = \mathcal{M}_k(\widetilde{{{\mathbfcal{A}}}}), \quad \widetilde{{{\mathbfcal{S}}}} = \llbracket\widetilde{{{\mathbfcal{A}}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket, \quad \widetilde{{{\mathbf{S}}}}_k = \mathcal{M}_k(\widetilde{{{\mathbfcal{S}}}}), \quad {{\mathbf{X}}}_{jk} = \mathcal{M}_k({{\mathbfcal{X}}}_j), \quad k=1,2,3.$$ Recall $$\widetilde{\sigma}^2 = \|{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2 + \sigma^2,\quad \lambda_k = \sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{A}}})),$$ $$\label{eq:m_s}
m_s = r_1r_2r_3 + \sum_{k\in J_s} s_k(r_k + \log (p_k)) + \sum_{k \notin J_s} p_kr_k,$$ and $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_2$ are the output from Step 1. We also denote $$I_k = \left\{i: {{\mathbf{U}}}_{k, [i,:]} \neq 0\right\},\quad k=1, 2, 3,$$ $$\label{eq:zeta}
\zeta_j = ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)^\top{{\rm vec}}({{\mathbfcal{X}}}_j^{(1)}) = {{\rm vec}}(\llbracket{{\mathbfcal{X}}}_j^{(1)}; {{\mathbf{U}}}_1^\top, {{\mathbf{U}}}_2^\top, {{\mathbf{U}}}_3^\top\rrbracket) \in \mathbb{R}^{r_1r_2r_3},\quad j=1,\ldots, n_1,$$ $$\label{eq:sigma_zeta}
\widetilde{\sigma}^2_\zeta = \frac{1}{n_1}\sum_{j=1}^{n_1}\left(\varepsilon_j^{(1)} + \zeta_j^\top {{\rm vec}}({{\mathbfcal{S}}})\right)^2.$$
1. In this first step, we develop the perturbation bound for $\widetilde{{{\mathbf{U}}}}_k$ and $\widetilde{{{\mathbf{W}}}}_k$. First, $\widetilde{{{\mathbf{A}}}}$ can be decomposed as $$\label{eq:tilde-A-decomposition}
\begin{split}
\widetilde{{{\mathbfcal{A}}}} = & \frac{1}{n_1}\sum_{j=1}^{n_1} y_j^{(1)} {{\mathbfcal{X}}}_j^{(1)} = \frac{1}{n_1}\sum_{j=1}^{n_1} \left(\varepsilon_j^{(1)} + \langle {{\mathbfcal{X}}}_j^{(1)}, {{\mathbfcal{A}}}\rangle \right) {{\mathbfcal{X}}}_j^{(1)} \\
= & \frac{1}{n_1}\sum_{j=1}^{n_1} \left(\varepsilon_j^{(1)} + \langle\llbracket{{\mathbfcal{X}}}_j^{(1)}; {{\mathbf{U}}}_1^\top, {{\mathbf{U}}}_2^\top, {{\mathbf{U}}}_3^\top\rrbracket, {{\mathbfcal{S}}}\rangle \right) {{\mathbfcal{X}}}_j^{(1)}\\
= & \frac{1}{n_1}\sum_{j=1}^{n_1}\left(\varepsilon_j^{(1)}+\langle \llbracket{{\mathbfcal{X}}}_j^{(1)}; {{\mathbf{U}}}_1^\top, {{\mathbf{U}}}_2^\top, {{\mathbf{U}}}_3^\top \rrbracket, {{\mathbfcal{S}}}\rangle\right) \llbracket {{\mathbfcal{X}}}_j^{(1)}; P_{{{\mathbf{U}}}_1}, P_{{{\mathbf{U}}}_2}, P_{{{\mathbf{U}}}_3} \rrbracket\\
& ~~ + \frac{1}{n_1}\sum_{j=1}^{n_1}\left(\varepsilon_j^{(1)}+\langle \llbracket{{\mathbfcal{X}}}_j^{(1)}; {{\mathbf{U}}}_1^\top, {{\mathbf{U}}}_2^\top, {{\mathbf{U}}}_3^\top \rrbracket, {{\mathbfcal{S}}}\rangle\right) P_{({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)_\perp}[{{\mathbfcal{X}}}_j^{(1)}]\\
:= & {{\mathbfcal{H}}}+ {{\mathbfcal{R}}}.
\end{split}$$ In particular, ${{\mathbfcal{H}}}$ is fully determined by $\zeta_j$ and $\varepsilon_j^{(1)}$; ${{\mathbfcal{H}}}$ is of Tucker rank-$(p_1,p_2,p_3)$ and has loadings ${{\mathbf{U}}}_1,{{\mathbf{U}}}_2, {{\mathbf{U}}}_3$. By Lemma \[lm:concentration-Gaussian-ensemble\], $$\label{ineq:sparse-regression-0}
\begin{split}
& \left\|\mathcal{M}_1({{\mathbfcal{H}}}) - {{\mathbf{A}}}_1\right\| = \left\|{{\mathbf{U}}}_1^\top\mathcal{M}_1({{\mathbfcal{H}}}) ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2) - {{\mathbf{U}}}_1^\top{{\mathbf{A}}}_1({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)\right\|\\
= & \left\|\frac{1}{n_1}\sum_{j=1}^{n_1}\left(\varepsilon_j^{(1)}+\langle \llbracket{{\mathbfcal{X}}}_j^{(1)}; {{\mathbf{U}}}_1^\top, {{\mathbf{U}}}_2^\top, {{\mathbf{U}}}_3^\top \rrbracket, {{\mathbfcal{S}}}\rangle\right) {{\mathbf{U}}}_1^\top{{\mathbf{X}}}_{jk}^{(1)} ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2) - {{\mathbf{U}}}_1^\top{{\mathbf{A}}}_1({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)\right\|\\
= & \left\|\frac{1}{n_1}\sum_{j=1}^{n_1}\left(\varepsilon_j^{(1)}+\left\langle {{\mathbf{U}}}_1^\top{{\mathbf{X}}}_{j1}^{(1)}({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2), {{\mathbf{S}}}_1 \right\rangle\right) {{\mathbf{U}}}_1^\top{{\mathbf{X}}}_{j1}^{(1)} ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2) - {{\mathbf{S}}}_1\right\|\\
\leq & \sqrt{\frac{(r_1+r_2r_3)\widetilde{\sigma}^2\log p}{n_1}}
\end{split}$$ with probability at least $1 - p^{-C}$. Similar inequalities also hold for $\|\mathcal{M}_2({{\mathbfcal{H}}}) - {{\mathbf{A}}}_2\|$ and $\|\mathcal{M}_3({{\mathbfcal{H}}}) - {{\mathbf{A}}}_3\|$. Provided that $\lambda_0 = \min_{k=1,2,3} \sigma_{r_k}({{\mathbf{A}}}_k)$ satisfies $\lambda_0^2 \geq C\widetilde{\sigma}^2(r_1r_2+r_2r_3+r_3r_1)/n_1$, we have $$\label{ineq:sparse-regression-1}
\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{H}}})) \geq \sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{A}}})) - \left\|\mathcal{M}_k({{\mathbfcal{H}}}) - {{\mathbf{A}}}_k\right\| \geq (1-c)\lambda_k$$ with probability at least $1 - p^{-C}$.
Recall the definition of $\zeta_j$ and $\widetilde{\sigma}^2_\zeta$ in . For any $j = 1,\ldots, n_1$, $\varepsilon_j^{(1)} + \zeta_j^\top {{\rm vec}}({{\mathbfcal{S}}}) \sim N(0, \sigma^2 + \|{{\mathbfcal{S}}}\|_{{{\rm HS}}}^2) \sim N(0, \sigma^2 + \|{{\mathbfcal{A}}}\|_{{{\rm HS}}}^2) \sim N(0, \widetilde{\sigma}^2)$, which means $\widetilde{\sigma}_\zeta^2 \sim \frac{\widetilde{\sigma}^2}{n_1}\chi^2_{n_1}.$ By the tail bound of $\chi^2$ distribution [@laurent2000adaptive Lemma 1], $$\label{ineq:tilde-sigma-eta-sigma}
\left|\widetilde{\sigma}_{\zeta}^2 - \widetilde{\sigma}^2\right| \leq C\widetilde{\sigma}^2\left(\sqrt{\frac{\log p}{n_1}} + \frac{\log p}{n_1}\right)\leq C\widetilde{\sigma}^2\sqrt{\frac{\log p}{n_1}}$$ with probability at least $1 - p^{-C}$.
Since ${{\rm vec}}({{\mathbfcal{X}}}_j^{(1)})$ has i.i.d. Gaussian entries and $({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)$ is orthogonal to $({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)_\perp$, we have that $({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)^\top {{\rm vec}}({{\mathbfcal{X}}}_j^{(1)})$ is independent of $({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)^\top_\perp {{\rm vec}}({{\mathbfcal{X}}}_j^{(1)})$ and ${{\mathbfcal{R}}}$ (defined in ) is Gaussian distributed conditioning on fixed values of $\zeta_j$ and $\varepsilon_j^{(1)}$: $$\begin{split}
& {{\rm vec}}({{\mathbfcal{R}}}) \bigg| \{\varepsilon_j^{(1)}, \zeta_j \}_{j=1}^{n_1} \text{ has same distribution as } P_{({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2\otimes {{\mathbf{U}}}_1)_\perp} {{\rm vec}}({{\mathbfcal{R}}}_0), \\
& \text{ where } {{\mathbfcal{R}}}_0 \in \mathbb{R}^{p_1\times p_2\times p_3}, \quad {{\mathbfcal{R}}}_0 \overset{iid}{\sim}N\left(0, \frac{\widetilde{\sigma}^2_\zeta}{n_1}\right).
\end{split}$$ Particularly, ${{\mathbfcal{R}}}_{[I_1, I_2, I_3]^c}\Big|\{\varepsilon_j^{(1)}, \zeta_j\}_{j=1}^{n_1}\overset{iid}{\sim} N(0, \widetilde{\sigma}_\zeta)^2$, i.e., ${{\mathbfcal{R}}}$ is i.i.d. Gaussian outside of the support of ${{\mathbfcal{A}}}$.
2. The rest of this proof will be conditioning on the fixed value of $\{\varepsilon_j^{(1)}, \zeta_j\}_{j=1}^{n_1}$ that satisfies , , and . Provided , , and $$n_1 \geq \frac{C\widetilde{\sigma}^2}{\lambda_0^2}\left(s_1s_2s_3\log p + \sum_{k=1}^3(s_k^2r_k^2 + r_{k+1}^2r_{k+2}^2)\right),$$ we have the following signal-noise-ratio assumption for denoising problem: $\widetilde{{{\mathbf{A}}}} = {{\mathbfcal{H}}}+ {{\mathbfcal{R}}}$, $$\min_k \sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{H}}})) \geq \frac{C\widetilde{\sigma}_\zeta}{\sqrt{n_1}}\left((s_1s_2s_3\log p)^{1/2} + \sum_{k=1}^3(s_kr_k+r_{k+1}r_{k+1})\right).$$ By [@zhang2017optimal-statsvd Theorem 4] (with mild modifications to the proof to accommodate the fact that ${{\mathbfcal{R}}}_{[I_1, I_2, I_3]}$ here is projection of i.i.d. Gaussian but not exactly i.i.d. Gaussian), the STAT-SVD with the tuning parameter $\widehat{\sigma}={\rm Med}(|{{\rm vec}}(\widetilde{{{\mathbfcal{A}}}})|/0.6744)$ (where 0.6744 is the $75\%$ quantile of standard Gaussian) yields $$\label{ineq:sparse-regression-2}
\begin{split}
\left\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\right\|_F \leq & \frac{C\widetilde{\sigma}_\zeta\sqrt{(s_kr_k + s_k\log(p_k))/n_1}}{\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{H}}}))}\\
\overset{\eqref{ineq:sparse-regression-1}\eqref{ineq:tilde-sigma-eta-sigma}}{\leq} & \frac{C\widetilde{\sigma}\sqrt{(s_kr_k+s_k\log(p_k))/n_1}}{\lambda_k}, \quad k\in J_s,
\end{split}$$ $$\begin{split}
\left\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\right\|_F \leq & \frac{C\widetilde{\sigma}_\zeta\sqrt{p_kr_k/n_1}}{\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{H}}}))} \overset{\eqref{ineq:sparse-regression-1}\eqref{ineq:tilde-sigma-eta-sigma}}{\leq} \frac{C\widetilde{\sigma}\sqrt{p_kr_k/n_1}}{\lambda_k}, \quad k\notin J_s,
\end{split}$$ $$\label{ineq:sparse-regression-3}
\begin{split}
\text{and} \quad \left\|\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket - {{\mathbfcal{H}}}\right\|_{{{\rm HS}}}^2 \leq & \frac{C\widetilde{\sigma}_\eta^2}{n_1}\Big(r_1r_2r_3 + \sum_{k\in J_s}s_k(r_k+\log p) + \sum_{k\notin J_s} p_kr_k\Big) \\
\overset{\eqref{eq:m_s}}{\leq} & \frac{C\widetilde{\sigma}^2m_s}{n_1}
\end{split}$$ with probability at least $1 - p^{-C}$, where $\widetilde{{{\mathbf{U}}}}_1, \widetilde{{{\mathbf{U}}}}_2, \widetilde{{{\mathbf{U}}}}_3$ are the outcomes of STAT-SVD procedure. Since the leading right singular vectors of $\mathcal{M}_k\left(\llbracket\widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket\right)$ and $\mathcal{M}_k({{\mathbfcal{A}}})$ are $\widetilde{{{\mathbf{W}}}}_k$ and ${{\mathbf{W}}}_k$, respectively, we have $$\begin{split}
& \left\|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\right\|_F = \left\|\widetilde{{{\mathbf{W}}}}_{k\perp}^\top {{\mathbf{W}}}_k\right\|_F \leq \frac{\|\widetilde{{{\mathbf{W}}}}_{k\perp}^\top {{\mathbf{W}}}_k{{\mathbf{W}}}_k^\top \mathcal{M}_k({{\mathbfcal{H}}})^\top \|_F}{\sigma_{r_k}\left({{\mathbf{W}}}_k^\top\mathcal{M}_k({{\mathbfcal{H}}})^\top\right)}\\
= & \frac{\|\widetilde{{{\mathbf{W}}}}_{k\perp}^\top \mathcal{M}_k({{\mathbfcal{H}}})^\top \|_F}{\sigma_{r_k}\left(\mathcal{M}_k({{\mathbfcal{H}}})\right)} \overset{\text{Lemma \ref{lm:SVD-projection}}}{\leq} \frac{\left\|\mathcal{M}_k\left(\llbracket\widetilde{{{\mathbf{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket\right)-\mathcal{M}_k({{\mathbf{A}}})\right\|_F}{\sigma_{r_k}(\mathcal{M}_k({{\mathbfcal{H}}}))}\\
\overset{\eqref{ineq:sparse-regression-1}\eqref{ineq:sparse-regression-3}}{\leq} & C\widetilde{\sigma}\frac{\sqrt{m_s/n_1}}{\lambda_k},\quad k=1,2,3.
\end{split}$$ $$\begin{split}
\left\|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}\right\|_F = & \left\|{{\mathbf{A}}}_k{{\mathbf{W}}}_k{{\mathbf{W}}}_k^\top\widetilde{{{\mathbf{W}}}}_{k\perp}\right\|_F \leq \left\|{{\mathbf{W}}}_k^\top \widetilde{{{\mathbf{W}}}}_{k\perp}\right\|_F\cdot \|{{\mathbf{A}}}_k\|\\
= & \left\|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\right\|_F\cdot \|{{\mathbf{A}}}_k\| \leq C\kappa\widetilde{\sigma}\sqrt{m_s/n_1}.
\end{split}$$
Since $\widetilde{{{\mathbf{U}}}}_k$ and ${{\mathbf{U}}}_k$ are the leading left singular values of $\mathcal{M}_k\left(\llbracket \widetilde{{{\mathbfcal{A}}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{\mathbf{U}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket\right)$ and ${{\mathbf{A}}}_k$, respectively, $$\begin{split}
\left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\right\|_F = & \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{U}}}_k{{\mathbf{U}}}_k^\top{{\mathbf{A}}}_k\right\|_F \leq \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{U}}}_k\right\|_F\cdot \left\|{{\mathbf{U}}}_k^\top {{\mathbf{A}}}_k\right\| =\left\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\right\|_F\cdot \|{{\mathbf{A}}}_k\|\\
\leq & \left\{\begin{array}{ll} \frac{C\widetilde{\sigma}\sqrt{(s_kr_k+s_k\log(p_k))/n_1}}{\lambda_k} \cdot \|{{\mathbf{A}}}_k\| \leq C\kappa\widetilde{\sigma}\sqrt{(s_kr_k+s_k\log(p_k))/n_1}, & k\in J_s;\\
\frac{C\widetilde{\sigma}\sqrt{p_kr_k/n_1}}{\lambda_k} \cdot \|{{\mathbf{A}}}_k\| \leq C\kappa\widetilde{\sigma}\sqrt{p_kr_k/n_1}, & k\notin J_s.
\end{array}\right.
\end{split}$$ In summary, in the previous two steps, we have shown $$\label{ineq:STAT-SVD-conclusion}
\begin{split}
& \left\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\right\|_F \leq \left\{\begin{array}{ll}
\frac{C\widetilde{\sigma}\sqrt{(s_kr_k+s_k\log(p_k))/n_1}}{\lambda_k}, & k\in J_s;\\
\frac{C\widetilde{\sigma}\sqrt{p_kr_k/n_1}}{\lambda_k}, & k\notin J_s,
\end{array}\right. \\
& \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\right\|_F \leq C\kappa\widetilde{\sigma}\sqrt{m_s/n_1},\\
& \left\|\sin\Theta(\widetilde{{{\mathbf{W}}}}_k, {{\mathbf{W}}}_k)\right\|_F \leq \frac{C\widetilde{\sigma}\sqrt{m_s/n_1}}{\lambda_k},\\
& \left\|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}\right\|_F\leq C\kappa\widetilde{\sigma}\sqrt{m_s/n_1},\quad \text{for}\quad k=1,2,3
\end{split}$$ with probability at least $1 - p^{-C}$.
3. Next, we move on to analyze the second batch of samples $\{{{\mathbfcal{X}}}_j^{(2)}, \varepsilon_j^{(2)}\}_{j=1}^{n_2}$. We first introduce the following notations, $$\widehat{\sigma}_{{{\mathbfcal{B}}}}^2 = \sigma^2 + \left\|P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2^2,\quad \widehat{\sigma}^2_{{{\mathbf{E}}}_k} = \sigma^2 + \left\|P_{\left(\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes {{\mathbf{I}}}_{p_k})\right)_\perp} {{\rm vec}}({{\mathbfcal{A}}})\right\|_2^2.$$ In this step, we give an upper bound for $\widehat{\sigma}_{{{\mathbfcal{B}}}}^2$ and $\widehat{\sigma}_{{{\mathbf{E}}}_k}^2$ given holds. Note that $$\begin{split}
& \left\|P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 \\
= & \left\|{{\rm vec}}({{\mathbfcal{A}}}) - P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 = \left\|{{\mathbfcal{A}}}- \llbracket {{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket\right\|_{{{\rm HS}}} \\
= & \left\|\llbracket{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}+P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}, P_{\widetilde{{{\mathbf{U}}}}_2}+P_{\widetilde{{{\mathbf{U}}}}_{2\perp}}, P_{\widetilde{{{\mathbf{U}}}}_3}+P_{\widetilde{{{\mathbf{U}}}}_{3\perp}}\rrbracket- \llbracket {{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\rrbracket\right\|_{{{\rm HS}}}\\
\leq & \left\|{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}, P_{\widetilde{{{\mathbf{U}}}}_2}, P_{\widetilde{{{\mathbf{U}}}}_3}\right\|_{{{\rm HS}}} + \left\|{{\mathbfcal{A}}}; {{\mathbf{I}}}_{p_1}, P_{\widetilde{{{\mathbf{U}}}}_{2\perp}}, P_{\widetilde{{{\mathbf{U}}}}_3}\right\|_{{{\rm HS}}} + \left\|{{\mathbfcal{A}}}; {{\mathbf{I}}}_{p_1}, {{\mathbf{I}}}_{p_2}, P_{\widetilde{{{\mathbf{U}}}}_{3\perp}}\right\|_{{{\rm HS}}}\\
\leq & \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1\right\|_F + \left\|\widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{A}}}_2\right\|_F + \left\|\widetilde{{{\mathbf{U}}}}_{3\perp}^\top {{\mathbf{A}}}_3\right\|_F\\
\overset{\eqref{ineq:STAT-SVD-conclusion}}{\leq} & C\kappa\widetilde{\sigma}\sqrt{m_s/n_1},
\end{split}$$ $$\begin{split}
& \left\|P_{\left(\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes {{\mathbf{I}}}_{p_k})\right)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 = \left\|{{\rm vec}}({{\mathbfcal{A}}}) - P_{\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes {{\mathbf{I}}}_{p_k})}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 \\
= & \left\|{{\mathbf{A}}}_k P_{\widetilde{{{\mathbf{W}}}}_{k\perp}}\right\|_F = \left\|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}\right\|_F \leq C\kappa\widetilde{\sigma}\sqrt{m_s/n_1}.
\end{split}$$ Therefore, $$\label{ineq:sigma_B-sigma_E}
\widehat{\sigma}_{{{\mathbfcal{B}}}}^2 \leq \sigma^2 + \frac{Cm_s\kappa^2\widetilde{\sigma}^2}{n_1}, \quad \widehat{\sigma}_{{{\mathbf{E}}}_k}^2 \leq \sigma^2 + \frac{Cm_s\kappa^2\widetilde{\sigma}^2}{n_1},\quad k=1,2,3.$$
4. In this step, we analyze the estimation error for $\widehat{{{\mathbfcal{B}}}}$ and $\widehat{{{\mathbf{E}}}}_k$ under the assumption that hold (which further means holds). Recall the partial linear models on importance sketching covariates (see - ; also see the proof of Theorem \[th:upper\_bound\_sparse\_general\]), $$\begin{split}
y^{(2)} = \widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}{{\rm vec}}(\widetilde{{{\mathbfcal{B}}}}) + \widetilde{\varepsilon}_{{{\mathbfcal{B}}}},
\end{split}$$ $$y^{(2)} = \widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k} {{\rm vec}}(\widetilde{{{\mathbf{E}}}}_k) + \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}, \quad k=1,2,3,$$ where the covariates, parameters, and noises of these two regressions are $$\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}\in \mathbb{R}^{n_2\times (r_1r_2r_3)}, \quad (\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})_{i\cdot} = {{\rm vec}}\left({{\mathbfcal{X}}}_i^{(2)} \times_1 \widetilde{{{\mathbf{U}}}}_1 \times_2 \widetilde{{{\mathbf{U}}}}_2 \times_3 \widetilde{{{\mathbf{U}}}}_3\right);$$ $$\begin{split}
\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}\in \mathbb{R}^{n_2\times (p_kr_k)}, \quad (\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})_{i\cdot} = & {{\rm vec}}\left( {{\mathbf{X}}}_{ik}^{(2)}\left(\widetilde{{{\mathbf{U}}}}_{k+2}\otimes \widetilde{{{\mathbf{U}}}}_{k+1}\right)\widetilde{{{\mathbf{V}}}}_k\right) \\
= & {{\rm vec}}\left({{\mathbf{X}}}_{ik}^{(2)}\widetilde{{{\mathbf{W}}}}_k\right), \quad k=1,2,3;
\end{split}$$ $$\begin{split}
&\widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\in \mathbb{R}^n, \quad (\widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}})_j = \left\langle {{\rm vec}}\left({{\mathbfcal{X}}}^{(2)}_j\right); P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)_{\perp}}{{\rm vec}}({{\mathbfcal{A}}})\right\rangle + \varepsilon_j^{(2)}, \\
& \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k} \in \mathbb{R}^n, \quad
(\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k})_j = \left\langle {{\rm vec}}\left({{\mathbfcal{X}}}^{(2)}_{j}\right), P_{\left(\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes{{\mathbf{I}}}_{p_k})\right)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\rangle + \varepsilon_j^{(2)}, \quad k=1,2,3;
\end{split}$$ $$\begin{split}
& {{\rm vec}}(\widetilde{{{\mathbfcal{B}}}}) = {{\rm vec}}(\llbracket {{\mathbfcal{A}}}; \widetilde{{{\mathbf{U}}}}_1^\top, \widetilde{{{\mathbf{U}}}}_2^\top, \widetilde{{{\mathbf{U}}}}_3^\top\rrbracket) = (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1) {{\rm vec}}({{\mathbfcal{A}}}) \in \mathbb{R}^{r_1r_2r_3};\\
\text{and} \quad & \widetilde{{{\mathbf{E}}}}_k = \mathcal{M}_k\left({{\mathbfcal{A}}}\times_{k+1}\widetilde{{{\mathbf{U}}}}_{k+1}^\top\times_{k+2}\widetilde{{{\mathbf{U}}}}_{k+2}^\top\right)\widetilde{{{\mathbf{V}}}}_{k} = {{\mathbf{A}}}_k {{\mathbf{W}}}_k\in \mathbb{R}^{p_k\times r_k},\quad k=1,2,3.
\end{split}$$ These quantities satisfy the following properties.
- Based on the proof of Theorem \[th:upper\_bound\_sparse\_general\], $\widetilde{{{\mathbf{E}}}}_k, k\in J_s$ are group-wise sparse, $$\begin{split}
\left\|{{\rm vec}}(\widetilde{{{\mathbf{E}}}}_k)\right\|_{0, 2} = \sum_{i=1}^{p_k} 1_{\left\{({{\rm vec}}(\widetilde{{{\mathbf{E}}}}_k))_{G_i^k} \neq 0\right\}} \leq s_k,
\end{split}$$ where $G_i^k = \{i+p_k, \ldots, i+p_k(r_k-1)\}, i=1,\ldots, p_k, k\in J_s$.
- Conditioning on fixed values of $\widetilde{{{\mathbf{U}}}}_k \widetilde{{{\mathbf{V}}}}_k, \widetilde{{{\mathbf{W}}}}_k$, the noise distribution satisfies $$\begin{split}
& \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\Big| \widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k, \widetilde{{{\mathbf{W}}}}_k \overset{iid}{\sim} N\left(0, \sigma^2 + \left\|P_{(\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes {{\mathbf{U}}}_1)_\perp}[{{\mathbfcal{A}}}]\right\|_{{{\rm HS}}}\right) \sim N(0, \widehat{\sigma}^2_{{{\mathbfcal{B}}}}); \\
& \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\Big| \widetilde{{{\mathbf{U}}}}_k, \widetilde{{{\mathbf{V}}}}_k, \widetilde{{{\mathbf{W}}}}_k \overset{iid}{\sim} N\left(0, \sigma^2 + \left\|P_{\left(\mathcal{R}_k(\widetilde{{{\mathbf{W}}}}_k\otimes {{\mathbf{I}}}_{p_k})\right)_\perp}[{{\mathbfcal{A}}}]\right\|_{{{\rm HS}}}\right) \sim N(0, \widehat{\sigma}^2_{{{\mathbf{E}}}_k}).
\end{split}$$
- Note that $\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}$ is an $n_2$-by-$(r_1r_2r_3)$ matrix with i.i.d. Gaussian entries. Similarly to the argument in Step 5 in the proof of Theorem \[th:upper\_bound\_regression\], $$\begin{split}
& \left\|\left(\widetilde{{{\mathbf{X}}}}^\top_{{{\mathbfcal{B}}}}\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}\right)^{-1}\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\right\|_2^2 \\
\leq & \frac{\widehat{\sigma}^2_{{\mathbfcal{B}}}\left(n_2+2\sqrt{n_2C\log(p)}+2C\log(p)\right)\left(r_1r_2r_3 + 2\sqrt{Cr_1r_2r_3\log(p)}+2C\log(p)\right)}{\left(\sqrt{n_2}-\sqrt{r_1r_2r_3} - C\log(p)\right)^4}\\
\leq & \frac{\widehat{\sigma}_{{\mathbfcal{B}}}^2}{n_2}\frac{\left(1+2\sqrt{\frac{C\log p}{n_2}}+2\frac{\log p}{n_2}\right)Cm_s}{\left(1 - \sqrt{\frac{r_1r_2r_3}{n_2}} - C\sqrt{\frac{\log(p)}{n_2}}\right)^4} \leq \frac{C\widehat{\sigma}_{{\mathbfcal{B}}}^2m_s}{n_2}.
\end{split}$$ with probability at least $1 - p^{-C}$. Here, the second last inequality is due to $\sqrt{r_1r_2r_3\log(p)} \leq \frac{1}{2}\left(r_1r_2r_3 + \log(p)\right) \leq m_s$ and the last inequality is due to $n_2 \geq Cm_s$. By the proof of Theorem \[th:upper\_bound\_sparse\_general\], $$\label{ineq:X_BX_B^-1epsilon}
\begin{split}
& \left\|\widehat{{{\mathbfcal{B}}}} - \widetilde{{{\mathbfcal{B}}}}\right\|_{{{\rm HS}}}^2 \overset{\eqref{eq:hat_B-B-identity}}{=} \left\|\left(\widetilde{{{\mathbf{X}}}}^\top_{{{\mathbfcal{B}}}}\widetilde{{{\mathbf{X}}}}_{{{\mathbfcal{B}}}}\right)^{-1}\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\right\|_2^2 \leq \frac{Cm_s\widehat{\sigma}^2_{{{\mathbfcal{B}}}}}{n_2}.
\end{split}$$ with probability at least $1 - p^{-C}$. Similarly, we can show for $k\notin J_s$, the least square estimator $\widehat{{{\mathbf{E}}}}_k$ satisfies $$\label{ineq:X_EkX_Ek^-1epsilon}
\left\|\widehat{{{\mathbf{E}}}}_k - {{\mathbf{E}}}_k\right\|_F^2 \overset{\eqref{eq:hat_Ek-Ek-identity}}{=} \left\|\left(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}\right)^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2 \leq \frac{Cm_s\widehat{\sigma}_{{{\mathbf{E}}}_k}^2}{n_2}.$$
- By Lemma \[lm:GE->GRIP\] and $n_2\geq Cm_s$ for large constant $C>0$, $\widetilde{{{\mathbf{X}}}}_{{{\mathbf{D}}}_k}$ satisfies group restricted isometry property with $\delta = 1/4$ with probability at least $1 - \exp(-cn)$.
Next, since $\widetilde\varepsilon_{{{\mathbf{E}}}_k}\overset{iid}{\sim}N_{n_2}\left(0, \widehat{\sigma}_{{{\mathbf{E}}}_k}^2\right)$ and $(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i)^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k} \bigg| \|\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2^2 \sim N_{r_k}\left(0, \|\widetilde{\varepsilon}_{{{\mathbf{E}}}_j}\|_2^2\right)$, we know $$\|\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2^2 \sim \widehat{\sigma}_{{{\mathbf{E}}}_k}^2\chi^2_{n_2}\quad \text{and}\quad \|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i)^\top\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2^2 \bigg| \|\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2^2 \sim \|\widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\|_2^2\cdot\chi_{r_k}^2$$ By the tail bound of $\chi^2$ distribution, $$\begin{split}
\left\|(\widetilde{{{\mathbf{X}}}}^i_{{{\mathbf{E}}}_k})^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2 \leq & \widehat{\sigma}_{{{\mathbf{E}}}_k}^2\left(n_2+2\sqrt{n_2C\log(p)}+2C\log(p)\right)\left(r_k + 2\sqrt{r_kC\log(p)} + 2C\log(p)\right)\\
\leq & Cn_2\widehat{\sigma}_{{{\mathbf{E}}}_k}^2(r_k+\log(p))
\end{split}$$ with probability at least $1 - p^{-C}$. Since $\log(p_k)\asymp \log(p)$, we have $$\label{ineq:maxtildeX_Evarepsilon}
\max_{1\leq i\leq p_k} \left\|(\widetilde{{{\mathbf{X}}}}^i_{{{\mathbf{E}}}_k})^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2 \leq Cn_2\widehat{\sigma}^2_{{{\mathbf{E}}}_k}(r_k+\log(p_k))$$ with probability at least $1 - p^{-C}$.
- Similarly as the Step 5 in the proof of Theorem \[th:upper\_bound\_regression\], one can show $$\left\|\widehat{{{\mathbf{E}}}}_k (\widetilde{{{\mathbf{U}}}}_k^\top \widehat{{{\mathbf{E}}}}_k)^{-1}\right\| \leq 1 + \frac{C_1\kappa\widetilde{\sigma}}{\lambda_k}\sqrt{\frac{m_s}{n_1}} + \frac{C_2\kappa\widetilde{\sigma}}{\lambda_k}\sqrt{\frac{m_s}{n_2}}\leq 1+c,\quad k=1,2,3$$ for constant $0<c<1/2$.
By previous arguments, we have shown the conditions of Theorem \[th:upper\_bound\_sparse\_general\] hold with probability at least $1 - p^{-C}$ under the scenario of Theorem \[th:upper\_bound\_sparse\_tensor\_regression\]. Finally, Theorem \[th:upper\_bound\_sparse\_general\] implies $$\begin{split}
& \left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \\
\leq & \left(1+\frac{C_1\kappa\widetilde{\sigma}}{\lambda_0}\sqrt{\frac{m_s}{n_1\wedge n_2}}\right)\Bigg(\left\|(\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}})^{-1}\widetilde{{{\mathbf{X}}}}_{{\mathbfcal{B}}}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{\mathbfcal{B}}}\right\|_2^2 + C_2\sum_{k\in J_s} s_k \max_{1\leq i\leq p_k} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^i)^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}/n_2\right\|_2^2\\
& \quad\quad + \sum_{k\notin J_s} \left\|(\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k})^{-1}\widetilde{{{\mathbf{X}}}}_{{{\mathbf{E}}}_k}^\top \widetilde{{\boldsymbol{\varepsilon}}}_{{{\mathbf{E}}}_k}\right\|_2^2\Bigg)\\
\overset{(a)}{\leq} & C\left(\frac{m_s(\widehat{\sigma}_{{{\mathbfcal{B}}}}^2 + \widehat{\sigma}_{{{\mathbf{E}}}_k}^2)}{n_2} + C\sum_{k=1}^3 \frac{s_k(r_k+\log(p_k))\widehat{\sigma}_{{{\mathbf{E}}}_k}^2}{n_2}\right) \\
\overset{(b)}{\leq} & \frac{C_1m_s}{n_2}\left(\sigma^2 + \frac{C_2m_s\kappa^2\widetilde{\sigma}^2}{n_1}\right)
\end{split}$$ with probability at least $1 - p^{-C}$. Here, (a) is due to , , and ; (b) is due to . $\square$
Proof of Theorem \[th:lower\_bound\_sparse\_tensor\_regression\] {#sec:proof-lower_bound_sparse_tensor_regression}
----------------------------------------------------------------
This theorem gives a lower bound on the estimation error of sparse low-rank tensor regression. In order to prove the desired lower bound, we only need to prove the forthcoming and , respectively. To prove each inequality, we first construct a series of tensor parameters ${{\mathbfcal{A}}}^{(j)}$ that satisfy: (1) there are sufficient distances between ${{\mathbfcal{A}}}^{(j)}$ and ${{\mathbfcal{A}}}^{(l)}$ for any $j\neq l$; (2) the Kullback-Leiber divergence between the resulting observations, $\{y_i^{(j)}, {{\mathbfcal{X}}}_i^{(j)}\}_{i=1}^n$ and $\{y_i^{(l)}, {{\mathbfcal{X}}}_i^{(l)}\}_{i=1}^n$, are close. Finally, the lower bound is proved by an application of the generalized Fano’s Lemma.
In order to prove this theorem, we only need to show $$\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{s}}, {\boldsymbol{r}}}}\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \max\left\{\frac{cr_1r_2r_3\sigma^2}{n}, \max_{l=1,2,3}\frac{c\sigma^2\left(s_lr_l+s_l\log(ep_l/s_l)\right)}{n}\right\}.$$
1. If $$r_1r_2r_3 = \max\left\{r_1r_2r_3, \max_{k=1,2,3} \left(s_kr_k+s_k\log(ep_k/s_k)\right)\right\},$$ we only need to prove $$\label{ineq:lower-sparse-to-show-1}
\begin{split}
\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{s}}, {\boldsymbol{r}}}}\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \frac{cr_1r_2r_3\sigma^2}{n},
\end{split}$$ for $r_1r_2r_3\geq 9$ in order to finish the proof of this theorem. Construct ${{\mathbfcal{S}}}_0$ as an $r_1$-by-$r_2$-by-$r_3$ tensor with i.i.d. Gaussian entries. Since $r_k\geq r_{k+1}r_{k+2}$ for $k=1,2,3$, ${{\mathbfcal{S}}}_0$ has Tucker rank-$(r_1, r_2, r_3)$ with probability one. Let ${{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3$ be arbitrary fixed orthogonal matrices that satisfy $${{\mathbf{U}}}_k\in \mathbb{O}_{p_k, r_k}, \quad \|{{\mathbf{U}}}_k\|_{0, 2} = \sum_{i=1}^{p_k}1_{\{({{\mathbf{U}}}_k)_{[i,:]}\neq 0\}}\leq s_k, \quad k=1,2,3.$$ By Varshamov-Gilbert bound [@massart2007concentration Lemma 4.7], we can find ${{\mathbfcal{B}}}^{(1)},\ldots, {{\mathbfcal{B}}}^{(N)}\subseteq \{-1, 1\}^{r_1\times r_2\times r_3}$ such that $$\forall j\neq l, \quad \|{{\mathbfcal{B}}}^{(j)}-{{\mathbfcal{B}}}^{(l)}\|_{{{\rm HS}}}^2 = 2\sum_{i_1, i_2}|{{\mathbfcal{B}}}^{(j)}_{[i_1,i_2]} - {{\mathbfcal{B}}}^{(l)}_{[i_1,i_2]}| \geq 2r_1r_2r_3 \quad \text{and}\quad N\geq \exp(r_1r_2r_3/8).$$ On the other hand, $$\label{ineq:thm6-1}
\|{{\mathbfcal{B}}}^{(j)} - {{\mathbfcal{B}}}^{(l)}\|_{{{\rm HS}}}^2 \leq 2\|{{\mathbfcal{B}}}^{(j)}\|_{{{\rm HS}}}^2 + 2\|{{\mathbfcal{B}}}^{(l)}\|_{{{\rm HS}}}^2 \leq 4r_1r_2r_3.$$ Since $r_1r_2r_3\geq 9$, $N\geq 3$. Then we construct $${{\mathbfcal{A}}}^{(j)} = \llbracket {{\mathbfcal{S}}}_0 + \tau {{\mathbfcal{B}}}_j; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket, \quad j=1,\ldots, N,$$ where $\tau>0$ is a constant to be determined a little while later. By such the configuration, ${{\mathbfcal{A}}}^{(1)},\ldots, {{\mathbfcal{A}}}^{(N)}\subseteq \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{s}}, {\boldsymbol{r}}}$. Now, the Kullback–Leibler divergence between the samples generated from ${{\mathbfcal{A}}}^{(j)}$ and the samples generated from ${{\mathbfcal{A}}}^{(l)}$ satisfy $$\begin{split}
& D_{KL}\left(\{{{\mathbfcal{X}}}_i, y_i^{(j)}\}_{i=1}^n\Big|\Big| \{{{\mathbfcal{X}}}_i, y_i^{(l)}\}_{i=1}^n\right) \overset{\text{Lemma \ref{lm:regression-KL}}}{=} \frac{n}{2\sigma^{2}}\left\|{{\mathbfcal{A}}}^{(j)} - {{\mathbfcal{A}}}^{(l)}\right\|_{{{\rm HS}}}^{2} \\
\leq & \frac{n}{2\sigma^{2}}\left\|\tau{{\mathbfcal{B}}}^{(j)} - \tau{{\mathbfcal{B}}}^{(l)}\right\|_{{{\rm HS}}}^{2} \overset{\eqref{ineq:thm6-1}}{\leq} \frac{n}{2\sigma^{2}}(4\tau^2 r_1r_2r_3)
\end{split}$$ and $$\forall j\neq l,\quad \left\|{{\mathbfcal{A}}}^{(j)} - {{\mathbfcal{A}}}^{(l)}\right\|_{{{\rm HS}}}^2 = \left\|\tau{{\mathbfcal{B}}}^{(j)} - \tau{{\mathbfcal{B}}}^{(l)}\right\|_{{{\rm HS}}}^2 \geq 2\tau^2 r_1r_2r_3.$$ By generalized Fano’s lemma, $$\begin{split}
\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{s}}, {\boldsymbol{r}}}}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq & \inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \left\{{{\mathbfcal{A}}}^{(1)},\ldots, {{\mathbfcal{A}}}^{(N)}\right\}}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2\\
\geq & \tau^2r_1r_2r_3\left(1 - \frac{2\tau^2r_1r_2r_3n/\sigma^2+\log(2)}{\log(N)}\right).
\end{split}$$ By setting $\tau^2 = \sigma^2\log(N/2.5)/(2r_1r_2r_3n)$, we have $$\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{s}}, {\boldsymbol{r}}}}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq c\tau^2r_1r_2r_3 = \frac{c\sigma^2r_1r_2r_3}{n},$$ which has shown if $r_1r_2r_3\geq 9$.
2. If $$s_kr_k+s_k\log(ep_k/s_k) = \max\left\{r_1r_2r_3, \max_{l=1,2,3}\left(s_l r_l +s_k \log(ep_l/s_l)\right)\right\},$$ we only need to prove $$\label{ineq:lower-sparse-to-show-2}
\begin{split}
\inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}}\mathbb{E}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \frac{c\sigma^2\left(s_kr_k+s_k\log(ep_k/s_k)\right)}{n},
\end{split}$$ provided that $s_kr_k+s_k\log(ep_k/s_k)\geq C$ for large constant $C>0$. Without loss of generality we assume $k=1$.
To this end, we randomly generate an orthogonal matrix ${{\mathbf{S}}}\in \mathbb{O}_{r_2r_3, r_1}$ and construct ${{\mathbfcal{S}}}\in \mathbb{R}^{r_1\times r_2\times r_3}$ such that $\mathcal{M}_1({{\mathbfcal{S}}}) = {{\mathbf{S}}}^\top$. We also construct ${{\mathbf{U}}}_2$ and ${{\mathbf{U}}}_3$ as fixed orthogonal matrices that satisfies $\|{{\mathbf{U}}}_2\|_{0, 2} \leq s_2$ and $\|{{\mathbf{U}}}_3\|_{0, 2}\leq s_3$. By Lemma \[lm:sparse-Varshamov-Gilbert\], there exists $\{{{\mathbf{U}}}_1^{(k)}\}_{k=1}^N \subseteq \{1, 0, -1\}^{p_1\times r_1}$ such that $$\label{ineq:U_1-condition}
\begin{split}
& \|{{\mathbf{U}}}_1^{(j)}\|_{0, 2} = \sum_{i=1}^{p_1} 1_{\left\{({{\mathbf{U}}}_1^{(j)})_{[i,:]}\neq 0\right\}} \leq s_1, \quad j=1,\ldots, N,\\
& \left\|{{\mathbf{U}}}_1^{(j)} - {{\mathbf{U}}}_1^{(l)}\right\|_{1,1} = \sum_{i,j} \left|({{\mathbf{U}}}_1^{(j)})_{ij} - ({{\mathbf{U}}}_1^{(l)})_{ij}\right| > s_1r_1/2,\quad 1\leq j \neq l \leq N,
\end{split}$$ and $N\geq \exp\left(c(s_1r_1+s_1\log(ep_1/s_1))\right)$. We further let $${{\mathbfcal{A}}}^{(j)} = \llbracket \tau{{\mathbfcal{S}}}; {{\mathbf{U}}}_1^{(j)}, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3\rrbracket, \quad j=1,2,\ldots, N,$$ where $\tau$ is a fixed and to-be-determined value. By such the construction, for any $1\leq j \neq l \leq N$, $$\begin{split}
\left\|{{\mathbfcal{A}}}^{(j)} - {{\mathbfcal{A}}}^{(l)}\right\|_{{{\rm HS}}}^2 = & \tau^2\left\|{{\mathbf{U}}}_1^{(j)}\mathcal{M}_1({{\mathbfcal{S}}}){{\mathbf{U}}}_3^\top\otimes {{\mathbf{U}}}_2^\top - {{\mathbf{U}}}_1^{(l)}\mathcal{M}_1({{\mathbfcal{S}}}){{\mathbf{U}}}_3^\top\otimes {{\mathbf{U}}}_2^\top \right\|_F^2 \\
= & \tau^2\left\|{{\mathbf{U}}}_1^{(j)}{{\mathbf{S}}}^\top {{\mathbf{U}}}_3^\top\otimes {{\mathbf{U}}}_2^\top - {{\mathbf{U}}}_1^{(l)}{{\mathbf{S}}}^\top{{\mathbf{U}}}_3^\top\otimes {{\mathbf{U}}}_2^\top \right\|_F^2 = \tau^2\left\|{{\mathbf{U}}}_1^{(j)} - {{\mathbf{U}}}_1^{(l)}\right\|_F^2\\
& \quad \text{(since all entries of ${{\mathbf{U}}}_1^{(j)}, {{\mathbf{U}}}_1^{(l)}\in \{-1, 0, 1\}$)}\\
\geq & \tau^2\left\|{{\mathbf{U}}}_1^{(j)} - {{\mathbf{U}}}_1^{(l)}\right\|_{1,1} > \tau^2 s_1r_1/2,
\end{split}$$ $$\begin{split}
\text{and}\quad & D_{KL}\left(\{{{\mathbfcal{X}}}_i, y_i^{(j)}\}_{i=1}^n\Big|\Big| \{{{\mathbfcal{X}}}_i, y_i^{(l)}\}_{i=1}^n\right) = \frac{n}{2\sigma^{2}}\left\|{{\mathbfcal{A}}}^{(j)} - {{\mathbfcal{A}}}^{(l)}\right\|_{{{\rm HS}}}^{2}\\
= & \frac{n}{2\sigma^2}\tau^2\left\|{{\mathbf{U}}}_1^{(j)} - {{\mathbf{U}}}_1^{(l)}\right\|_{F}^2 \leq \frac{n\tau^2}{2\sigma^2} 2\left(\|{{\mathbf{U}}}_1^{(j)}\|_2^2 + \|{{\mathbf{U}}}_1^{(l)}\|_2^2\right) \leq \frac{n\tau^2}{2\sigma^2}\cdot 4s_1r_1.
\end{split}$$ By setting $\tau^2 = \sigma^2\log(N/2.5)/(2ns_1r_1)$, we have $$\begin{split}
& \inf_{\widehat{{{\mathbfcal{A}}}}}\sup_{{{\mathbfcal{A}}}\in \mathcal{A}_{{\boldsymbol{p}}, {\boldsymbol{r}}}}\left\|\widehat{{{\mathbfcal{A}}}} - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}}^2 \geq \frac{\tau^2s_1r_1}{4}\left(1 - \frac{\frac{2n\tau^2s_1r_1}{\sigma^2} - \log(2)}{\log(N)}\right) \\
\geq & \frac{2\sigma^2\log(N/2.5)}{4ns_1r_1}\cdot \frac{s_1r_1}{4}\cdot c \geq \frac{c\sigma^2\left(s_1r_1 + s_1\log(ep_1/s_1)\right)}{n},
\end{split}$$ which has shown .
In summary of the previous two parts, we have finished the proof of this theorem. $\square$
Technical Lemmas
================
\[lm:Kronecker-vectorization-matricization\] Suppose ${{\mathbf{A}}}\in \mathbb{R}^{p_1\times p_2}$, ${{\mathbfcal{A}}}\in\mathbb{R}^{p_1\times p_2\times \ldots \times p_d}$, ${{\mathbf{B}}}_k\in \mathbb{R}^{p_k\times r_k}$, ${{\mathbf{B}}}'_k \in \mathbb{R}^{r_k\times d_k}$, $k=1,\ldots, d$. Then, $$\label{eq:Kronecker-0}
({{\mathbf{B}}}_1\otimes \cdots \otimes{{\mathbf{B}}}_d)\cdot ({{\mathbf{B}}}_1'\otimes \cdots \otimes{{\mathbf{B}}}_d') = ({{\mathbf{B}}}_1{{\mathbf{B}}}_1')\otimes \cdots \otimes({{\mathbf{B}}}_d{{\mathbf{B}}}_d'),$$ $$\label{eq:kronecker-1}
{{\rm vec}}\left({{\mathbf{B}}}_1^\top {{\mathbf{A}}}{{\mathbf{B}}}_2 \right) = ({{\mathbf{B}}}_2^\top\otimes {{\mathbf{B}}}_1^\top){{\rm vec}}({{\mathbf{A}}}),$$ $$\label{eq:kronecker-2}
{{\rm vec}}\left(\llbracket{{\mathbfcal{A}}}; {{\mathbf{B}}}_1^\top, \ldots, {{\mathbf{B}}}_d^\top\rrbracket\right) = ({{\mathbf{B}}}_d^\top\otimes \cdots \otimes {{\mathbf{B}}}_1^\top){{\rm vec}}({{\mathbfcal{A}}}),$$ $$\label{eq:kronecker-3}
\mathcal{M}_k\left(\llbracket {{\mathbfcal{A}}}; {{\mathbf{B}}}_1^\top, \ldots, {{\mathbf{B}}}_d^\top \rrbracket\right) = {{\mathbf{B}}}_k^\top \mathcal{M}_k({{\mathbfcal{A}}}) \left({{\mathbf{B}}}_d\otimes \cdots \otimes {{\mathbf{B}}}_{k+1}\otimes {{\mathbf{B}}}_{k-1}\otimes \cdots \otimes {{\mathbf{B}}}_1\right).$$ Finally, for any ${{\mathbf{V}}}_k\in \mathbb{R}^{r_{-k}\times r_k}$, $$\label{eq:kronecker-4}
\begin{split}
& {{\rm vec}}\left({{\mathbf{B}}}_k^\top \mathcal{M}_k\left(\left\llbracket{{\mathbfcal{A}}}; {{\mathbf{B}}}_1^\top, \ldots, {{\mathbf{B}}}_{k-1}^\top, {{\mathbf{B}}}_{k+1}^\top,\ldots, {{\mathbf{B}}}_d^\top \right\rrbracket\right){{\mathbf{V}}}_k\right)\\
= & {{\mathbf{V}}}_k^\top\left({{\mathbf{B}}}_d^\top \otimes \cdots \otimes {{\mathbf{B}}}_{k+1}^\top \otimes {{\mathbf{B}}}_{k-1}^\top \otimes \cdots \otimes{{\mathbf{B}}}_1^\top \right)\otimes({{\mathbf{B}}}_k^\top) \cdot {{\rm vec}}(\mathcal{M}_k({{\mathbfcal{A}}}))\\
\end{split}$$
[**Proof of Lemma \[lm:Kronecker-vectorization-matricization\].**]{} See [@kolda2009tensor; @kolda2006multilinear] for the proof of , and . We shall also note that is the order-2 case of . Finally, $$\begin{split}
& {{\rm vec}}\left({{\mathbf{B}}}_k^\top \mathcal{M}_k\left(\left\llbracket{{\mathbfcal{A}}}; {{\mathbf{B}}}_1^\top, \ldots, {{\mathbf{B}}}_{k-1}^\top, {{\mathbf{B}}}_{k+1}^\top,\ldots, {{\mathbf{B}}}_d^\top \right\rrbracket\right){{\mathbf{V}}}_k\right)\\
\overset{\eqref{eq:kronecker-1}}{=} & ({{\mathbf{V}}}_k^\top\otimes {{\mathbf{B}}}_k^\top ){{\rm vec}}\left(\mathcal{M}_k\left(\left\llbracket{{\mathbfcal{A}}}; {{\mathbf{B}}}_1^\top, \ldots, {{\mathbf{B}}}_{k-1}^\top, {{\mathbf{I}}}_{p_k}, {{\mathbf{B}}}_{k+1}^\top,\ldots, {{\mathbf{B}}}_d^\top \right\rrbracket\right)\right)\\
\overset{\eqref{eq:kronecker-3}}{=} & ({{\mathbf{V}}}_k^\top\otimes {{\mathbf{B}}}_k^\top) {{\rm vec}}\left(\mathcal{M}_k({{\mathbfcal{A}}})({{\mathbf{B}}}_d\otimes \cdots \otimes {{\mathbf{B}}}_{k+1} \otimes {{\mathbf{B}}}_{k-1}\otimes \cdots \otimes {{\mathbf{B}}}_1)\right)\\
\overset{\eqref{eq:kronecker-1}}{=} & ({{\mathbf{V}}}_k^\top\otimes {{\mathbf{B}}}_k^\top) \left({{\mathbf{B}}}_d^\top\otimes \cdots \otimes {{\mathbf{B}}}_{k+1}^\top \otimes {{\mathbf{B}}}_{k-1}^\top\otimes \cdots \otimes {{\mathbf{B}}}_1^\top \otimes {{\mathbf{I}}}\right) {{\rm vec}}(\mathcal{M}_k({{\mathbfcal{A}}}))\\
= & {{\mathbf{V}}}_k^\top\left({{\mathbf{B}}}_d^\top \otimes \cdots \otimes {{\mathbf{B}}}_{k+1}^\top \otimes {{\mathbf{B}}}_{k-1}^\top \otimes \cdots \otimes{{\mathbf{B}}}_1^\top \right)\otimes({{\mathbf{B}}}_k^\top) \cdot {{\rm vec}}(\mathcal{M}_k({{\mathbfcal{A}}}))\\
\end{split}$$ $\square$
\[lm:concatenation-singular-value\] Suppose ${{\mathbf{A}}}\in \mathbb{R}^{p\times r}$ and ${{\mathbf{U}}}\in \mathbb{O}_{p, m}$. Then, $$\sigma_r^2({{\mathbf{A}}}) \geq \sigma_r^2({{\mathbf{U}}}^\top{{\mathbf{A}}}) + \sigma_r^2({{\mathbf{U}}}_{\perp}^\top{{\mathbf{A}}}),\quad
\left\|{{\mathbf{A}}}\right\|^2 \leq \left\|{{\mathbf{U}}}^\top {{\mathbf{A}}}\right\|^2 + \left\|{{\mathbf{U}}}_{\perp}^\top {{\mathbf{A}}}\right\|^2.$$
[**Proof of Lemma \[lm:concatenation-singular-value\].**]{} Let ${{\mathbf{v}}}$ be the right singular vector associated with the $r$-th singular value of ${{\mathbf{A}}}$. Then $\|{{\mathbf{A}}}{{\mathbf{v}}}\|_2 = \sigma_r({{\mathbf{A}}})\|{{\mathbf{v}}}\|_2 = \sigma_r({{\mathbf{A}}})$ and $$\begin{split}
\sigma_r^2({{\mathbf{A}}}) = & \|{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 = \|P_{{{\mathbf{U}}}}{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 + \|P_{{{\mathbf{U}}}_{\perp}}{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 =\|{{\mathbf{U}}}^\top {{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 + \|{{\mathbf{U}}}_{\perp}^\top {{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 \\
\geq & \sigma_r^2({{\mathbf{U}}}^\top {{\mathbf{A}}})\|{{\mathbf{v}}}\|_2^2 + \sigma_r^2({{\mathbf{U}}}_{\perp}^\top {{\mathbf{A}}})\|{{\mathbf{v}}}\|_2^2 = \sigma_r^2({{\mathbf{U}}}^\top {{\mathbf{A}}}) + \sigma_r^2({{\mathbf{U}}}_{\perp}^\top {{\mathbf{A}}}).
\end{split}$$ On the other hand, $$\begin{split}
\|{{\mathbf{A}}}\|^2 = & \max_{{{\mathbf{v}}}: \|{{\mathbf{v}}}\|_2\leq 1} \|{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 = \max_{{{\mathbf{v}}}: \|{{\mathbf{v}}}\|_2\leq 1} \left(\|P_{{\mathbf{U}}}{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 + \|P_{{{\mathbf{U}}}_{\perp}}{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2\right) \\
\leq & \max_{{{\mathbf{v}}}: \|{{\mathbf{v}}}\|_2\leq 1} \|P_{{\mathbf{U}}}{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 + \max_{{{\mathbf{v}}}: \|{{\mathbf{v}}}\|_2\leq 1}\|P_{{{\mathbf{U}}}_\perp}{{\mathbf{A}}}{{\mathbf{v}}}\|_2^2 = \|{{\mathbf{U}}}^\top {{\mathbf{A}}}\|^2 + \|{{\mathbf{U}}}_{\perp}^\top {{\mathbf{A}}}\|^2.
\end{split}$$ $\square$
The following lemma establish a deterministic upper bound for $\|\widehat{{\mathbf{F}}}\widehat{{\mathbf{G}}}^{-1}\widehat{{\mathbf{H}}}- {{\mathbf{F}}}{{\mathbf{G}}}^{-1}{{\mathbf{H}}}\|$ in terms of $\|\widehat{{\mathbf{F}}}-{{\mathbf{F}}}\|_F, \|\widehat{{\mathbf{G}}}-{{\mathbf{G}}}\|_F, \|\widehat{{\mathbf{H}}}-{{\mathbf{H}}}\|_F$ and its more general high-order form. This result serves as a key technical lemma for the theoretical analysis of the oracle inequalities.
\[lm:FGH\] Suppose ${{\mathbf{F}}}, \widehat{{{\mathbf{F}}}}\in \mathbb{R}^{p_1\times r}, {{\mathbf{G}}}, \widehat{{{\mathbf{G}}}}\in \mathbb{R}^{r\times r}, {{\mathbf{H}}}, \widehat{{{\mathbf{H}}}}\in \mathbb{R}^{r\times p_2}$. If ${{\mathbf{G}}}$ and $\widehat{{{\mathbf{G}}}}$ are invertible, $\|{{\mathbf{F}}}{{\mathbf{G}}}^{-1}\|\leq \lambda_1$, $\|{{\mathbf{G}}}^{-1}{{\mathbf{H}}}\|\leq \lambda_2$, and $\|\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}}\| \leq \lambda_2$, we have $$\label{ineq:FGH1}
\left\|\widehat{{{\mathbf{F}}}}\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}} - {{\mathbf{F}}}{{\mathbf{G}}}^{-1}{{\mathbf{H}}}\right\|_F \leq \lambda_2\|\widehat{{{\mathbf{F}}}}-{{\mathbf{F}}}\|_F + \lambda_1\|\widehat{{{\mathbf{H}}}}-{{\mathbf{H}}}\|_F + \lambda_1\lambda_2\|\widehat{{{\mathbf{G}}}}-{{\mathbf{G}}}\|_F.$$ More generally for any $d\geq 1$, suppose $\widehat{{{\mathbfcal{F}}}}, {{\mathbfcal{F}}}\in \mathbb{R}^{r_1\times \cdots \times r_d}$ are order-$d$ tensors, ${{\mathbf{G}}}_k, \widehat{{{\mathbf{G}}}}_k \in \mathbb{R}^{r_k\times r_k}$ ${{\mathbf{H}}}_k, \widehat{{{\mathbf{H}}}}_k \in \mathbb{R}^{p_k\times r_k}$. If $\|{{\mathbf{H}}}_k{{\mathbf{G}}}_k^{-1}\|\leq\lambda_k$, $\|\widehat{{{\mathbf{H}}}}_k\widehat{{{\mathbf{G}}}}_k^{-1}\|\leq\lambda_k$, and $\|{{\mathbf{G}}}_k^{-1}\mathcal{M}_k({{\mathbfcal{F}}})\|\leq \pi_k$, we have $$\label{ineq:FGH2}
\begin{split}
& \left\|\llbracket \widehat{{{\mathbfcal{F}}}}; (\widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1}), \ldots, (\widehat{{{\mathbf{H}}}}_d\widehat{{{\mathbf{G}}}}_d^{-1}) \rrbracket - \llbracket {{\mathbfcal{F}}}; ({{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}), \ldots, ({{\mathbf{H}}}_d{{\mathbf{G}}}_d^{-1}) \rrbracket\right\|_{{{\rm HS}}}\\
\leq & \lambda_1\cdots \lambda_d \|\widehat{{{\mathbfcal{F}}}}-{{\mathbfcal{F}}}\|_{{{\rm HS}}} + \sum_{k=1}^d \pi_k\lambda_1\cdots \lambda_d\|\widehat{{{\mathbf{G}}}}-{{\mathbf{G}}}\|_F + \sum_{k=1}^d \pi_k \lambda_1\cdots \lambda_d/\lambda_k \|\widehat{{{\mathbf{H}}}}_k - {{\mathbf{H}}}_k\|_F.
\end{split}$$
[**Proof of Lemma \[lm:FGH\].**]{} First, it is easy to check the following identity for any non-singular matrices ${{\mathbf{G}}}$ and $\widehat{{{\mathbf{G}}}}$, $$\widehat{{{\mathbf{G}}}}^{-1} = {{\mathbf{G}}}^{-1} - {{\mathbf{G}}}^{-1} (\widehat{{{\mathbf{G}}}} - {{\mathbf{G}}}) \widehat{{{\mathbf{G}}}}^{-1}.$$ Thus, $$\begin{split}
& \left\|\widehat{{{\mathbf{F}}}}\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}} - {{\mathbf{F}}}{{\mathbf{G}}}^{-1}{{\mathbf{H}}}\right\|_F \\
\leq & \left\|(\widehat{{{\mathbf{F}}}} - {{\mathbf{F}}})\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}}\right\|_F + \left\|{{\mathbf{F}}}\left({{\mathbf{G}}}^{-1} - {{\mathbf{G}}}^{-1}(\widehat{{{\mathbf{G}}}}-{{\mathbf{G}}})\widehat{{{\mathbf{G}}}}^{-1}\right)\widehat{{{\mathbf{H}}}} - {{\mathbf{F}}}{{\mathbf{G}}}^{-1}{{\mathbf{H}}}\right\|_F\\
\leq & \left\|\widehat{{{\mathbf{F}}}} - {{\mathbf{F}}}\right\|_F \cdot \left\|\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}}\right\| + \left\|{{\mathbf{F}}}{{\mathbf{G}}}^{-1}\widehat{{{\mathbf{H}}}} - {{\mathbf{F}}}{{\mathbf{G}}}^{-1}{{\mathbf{H}}}\right\|_F + \left\|{{\mathbf{F}}}{{\mathbf{G}}}^{-1}(\widehat{{{\mathbf{G}}}}-{{\mathbf{G}}})\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}}\right\|_F\\
\leq & \left\|\widehat{{{\mathbf{F}}}} - {{\mathbf{F}}}\right\|_F \left\|\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}}\right\| + \left\|{{\mathbf{F}}}{{\mathbf{G}}}^{-1}\right\|\left\|\widehat{{{\mathbf{H}}}} - {{\mathbf{H}}}\right\|_F + \left\|{{\mathbf{F}}}{{\mathbf{G}}}^{-1}\right\|\left\|\widehat{{{\mathbf{G}}}}-{{\mathbf{G}}}\right\|_F\left\|\widehat{{{\mathbf{G}}}}^{-1}\widehat{{{\mathbf{H}}}}\right\|\\
\leq & \lambda_2\|\widehat{{{\mathbf{F}}}}-{{\mathbf{F}}}\|_F + \lambda_1\|\widehat{{{\mathbf{H}}}}-{{\mathbf{H}}}\|_F + \lambda_1\lambda_2\|\widehat{{{\mathbf{G}}}}-{{\mathbf{G}}}\|_F.
\end{split}$$ Then we consider the proof of . Define $$\begin{split}
& \widehat{\widetilde{{{\mathbf{F}}}}}_d = \mathcal{M}_d(\widehat{{{\mathbfcal{F}}}})\left(\widehat{{{\mathbf{H}}}}_{d-1}\widehat{{{\mathbf{G}}}}_{d-1}^{-1}\otimes \cdots \otimes \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1}\right)^\top,\\
& \widetilde{{{\mathbf{F}}}}_d = \mathcal{M}_d({{\mathbfcal{F}}})\left({{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}\otimes \cdots \otimes {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}\right)^\top.
\end{split}$$ We shall note that $$\begin{split}
& \left\|{{\mathbf{G}}}_d^{-1}\widetilde{{{\mathbf{F}}}}_d\right\| = \left\|{{\mathbf{G}}}_d^{-1}\mathcal{M}_d({{\mathbfcal{F}}})\left({{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}\otimes \cdots \otimes {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}\right)\right\|\\
\leq & \left\|{{\mathbf{G}}}_d^{-1}\mathcal{M}_d({{\mathbfcal{F}}})\right\|\cdot \|{{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}\| \cdots \|{{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}\| \leq \pi_d\lambda_1\cdots \lambda_{d-1},
\end{split}$$ $$\begin{split}
\left\|{{\mathbf{H}}}_d{{\mathbf{G}}}_d^{-1}\right\| \leq \lambda_d, \quad \|\widehat{{{\mathbf{H}}}}_d\widehat{{{\mathbf{G}}}}_d^{-1}\| \leq \lambda_d.
\end{split}$$ By the first part of this lemma and tensor algebra, $$\label{ineq:FGH-1}
\begin{split}
& \left\|\llbracket\widehat{{{\mathbfcal{F}}}}; \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1},\ldots,\widehat{{{\mathbf{H}}}}_{d}\widehat{{{\mathbf{G}}}}_{d}^{-1}\rrbracket - \llbracket{{\mathbfcal{F}}}; {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1},\ldots, {{\mathbf{H}}}_{d} {{\mathbf{G}}}_{d}^{-1}\rrbracket\right\|_{{{\rm HS}}}\\
= & \left\|\mathcal{M}_d\left(\llbracket\widehat{{{\mathbfcal{F}}}}; \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1},\ldots,\widehat{{{\mathbf{H}}}}_{d}\widehat{{{\mathbf{G}}}}_{d}^{-1}\rrbracket\right) - \mathcal{M}_d\left(\llbracket{{\mathbfcal{F}}}; {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1},\ldots, {{\mathbf{H}}}_{d} {{\mathbf{G}}}_{d}^{-1}\rrbracket\right)\right\|_F\\
\overset{\text{Lemma \ref{lm:Kronecker-vectorization-matricization}}}{=} & \left\|\widehat{{{\mathbf{H}}}}_d\widehat{{{\mathbf{G}}}}_d^{-1}\widehat{\widetilde{{{\mathbf{F}}}}}_d - {{\mathbf{H}}}_d{{\mathbf{G}}}_d^{-1}\widetilde{{{\mathbf{F}}}}_d\right\|_F \\
\leq & \lambda_d \|\widehat{\widetilde{{{\mathbf{F}}}}}_d - \widetilde{{{\mathbf{F}}}}_d\|_F + \lambda_1\cdots\lambda_d \pi_d \|\widehat{{{\mathbf{G}}}}_d - {{\mathbf{G}}}_d\|_F + \lambda_1\cdots \lambda_{d-1}\pi_d \|\widehat{{{\mathbf{H}}}}_d - {{\mathbf{H}}}_d\|_F.
\end{split}$$ Next, we analyze $\|\widehat{\widetilde{{{\mathbf{F}}}}}_d - \widetilde{{{\mathbf{F}}}}_d\|_F$. Define $$\begin{split}
& \widehat{\widetilde{{{\mathbf{F}}}}}_{d-1} = \mathcal{M}_{d-1}(\widehat{{{\mathbfcal{F}}}})\left({{\mathbf{I}}}_{r_d}\otimes\widehat{{{\mathbf{H}}}}_{d-2}\widehat{{{\mathbf{G}}}}_{d-2}^{-1}\otimes \cdots \otimes \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1}\right)^\top,\\
& \widetilde{{{\mathbf{F}}}}_{d-1} = \mathcal{M}_{d-1}({{\mathbfcal{F}}})\left({{\mathbf{I}}}_{r_d} \otimes{{\mathbf{H}}}_{d-2}{{\mathbf{G}}}_{d-2}^{-1}\otimes \cdots \otimes {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}\right)^\top.
\end{split}$$ Then by tensor algebra (Lemma \[lm:Kronecker-vectorization-matricization\]), $$\begin{split}
& \|\widehat{\widetilde{{{\mathbf{F}}}}}_d - \widetilde{{{\mathbf{F}}}}_d\|_{F} = \left\|\llbracket\widehat{{{\mathbfcal{F}}}}; \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1}, \ldots, \widehat{{{\mathbf{H}}}}_{d-1}\widehat{{{\mathbf{G}}}}_{d-1}^{-1}, {{\mathbf{I}}}_{r_d}\rrbracket - \llbracket{{\mathbfcal{F}}}; {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}, \ldots, {{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}, {{\mathbf{I}}}_{r_d}\rrbracket\right\|_{{{\rm HS}}}\\
= & \left\|\mathcal{M}_{d-1}\left(\llbracket\widehat{{{\mathbfcal{F}}}}; \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1}, \ldots, \widehat{{{\mathbf{H}}}}_{d-1}\widehat{{{\mathbf{G}}}}_{d-1}^{-1}, {{\mathbf{I}}}_{r_d}\rrbracket\right) - \mathcal{M}_{d-1}\left(\llbracket{{\mathbfcal{F}}}; {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1}, \ldots, {{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}, {{\mathbf{I}}}_{r_d}\rrbracket\right)\right\|_F\\
= & \left\|\widehat{{{\mathbf{H}}}}_{d-1}\widehat{{{\mathbf{G}}}}_{d-1}^{-1}\widehat{\widetilde{{{\mathbf{F}}}}}_{d-1} - {{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}\widetilde{{{\mathbf{F}}}}_{d-1}\right\|_F.
\end{split}$$ Similarly as the previous argument, one can show by the first part of this lemma that $$\begin{split}
& \|\widehat{\widetilde{{{\mathbf{F}}}}}_d - \widetilde{{{\mathbf{F}}}}_d\|_{F} = \left\|\widehat{{{\mathbf{H}}}}_{d-1}\widehat{{{\mathbf{G}}}}_{d-1}^{-1}\widehat{\widetilde{{{\mathbf{F}}}}}_{d-1} - {{\mathbf{H}}}_{d-1}{{\mathbf{G}}}_{d-1}^{-1}\widetilde{{{\mathbf{F}}}}_{d-1}\right\|_F\\
\leq & \lambda_{d-1}\|\widehat{\widetilde{{{\mathbf{F}}}}}_{d-1}-\widetilde{{{\mathbf{F}}}}_{d-1}\|_F + \lambda_1\cdots \lambda_{d-1}\pi_{d-1}\|\widehat{{{\mathbf{G}}}}_{d-1}-{{\mathbf{G}}}_{d-1}\|_F + \lambda_1\cdots \lambda_{d-2}\pi_{d-1}\|\widehat{{{\mathbf{H}}}}_{d-1}-{{\mathbf{H}}}_{d-1}\|_F.\\
\end{split}$$ Therefore, by and the previous inequality, $$\begin{split}
& \left\|\llbracket\widehat{{{\mathbfcal{F}}}}; \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1},\ldots,\widehat{{{\mathbf{H}}}}_{d}\widehat{{{\mathbf{G}}}}_{d}^{-1}\rrbracket - \llbracket{{\mathbfcal{F}}}; {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1},\ldots, {{\mathbf{H}}}_{d} {{\mathbf{G}}}_{d}^{-1}\rrbracket\right\|_{{{\rm HS}}}\\
\leq & \lambda_{d-1}\lambda_d\left\|\widehat{\widetilde{{{\mathbf{F}}}}}_{d-1}-\widetilde{{{\mathbf{F}}}}_{d-1}\right\|_F + \sum_{k=d-1, d}\lambda_1\cdots \lambda_d\pi_{k}\|\widehat{{{\mathbf{G}}}}_k-{{\mathbf{G}}}_k\|_F + \sum_{k = d-1, d} \frac{\lambda_1\cdots \lambda_d\pi_{k}}{\lambda_k}\left\|\widehat{{{\mathbf{H}}}}_{k}-{{\mathbf{H}}}_k\right\|_F.
\end{split}$$ We further introduce $\widehat{\widetilde{{{\mathbf{F}}}}}_{d-2}, \widetilde{{{\mathbf{F}}}}_{d-2}, \ldots, \widehat{\widetilde{{{\mathbf{F}}}}}_1, \widetilde{{{\mathbf{F}}}}_1$, repeat the previous argument for $d$ time, and can finally obtain $$\begin{split}
& \left\|\llbracket\widehat{{{\mathbfcal{F}}}}; \widehat{{{\mathbf{H}}}}_1\widehat{{{\mathbf{G}}}}_1^{-1},\ldots,\widehat{{{\mathbf{H}}}}_{d}\widehat{{{\mathbf{G}}}}_{d}^{-1}\rrbracket - \llbracket{{\mathbfcal{F}}}; {{\mathbf{H}}}_1{{\mathbf{G}}}_1^{-1},\ldots, {{\mathbf{H}}}_{d} {{\mathbf{G}}}_{d}^{-1}\rrbracket\right\|_{{{\rm HS}}} \\
\leq & \lambda_1\cdots \lambda_d \|\widehat{{{\mathbfcal{F}}}} - {{\mathbfcal{F}}}\|_{{{\rm HS}}} + \sum_{k=1}^d \lambda\cdots \lambda_d\pi_k \|\widehat{{{\mathbf{G}}}}_k-{{\mathbf{G}}}_k\|_F + \sum_{k=1}^d \frac{\lambda_1\cdots \lambda_d \pi_k}{\lambda_k} \|\widehat{{{\mathbf{H}}}}_k - {{\mathbf{H}}}_k\|_F,
\end{split}$$ which has finished the proof of this lemma. $\square$
The following lemma characterizes the concentration of Gaussian ensemble measurements, which will be extensively used in the proof of Theorem \[th:upper\_bound\_regression\].
\[lm:concentration-Gaussian-ensemble\] Suppose ${{\mathbf{A}}}\in \mathbb{R}^{a\times b}$ is a fixed matrix, ${{\mathbf{X}}}_1,\ldots, {{\mathbf{X}}}_n\in \mathbb{R}^{a\times b}$ are random matrices with i.i.d. standard Gaussian entries, and $\varepsilon_1,\ldots, \varepsilon_n\overset{iid}{\sim}N(0,\sigma^2)$. Let ${{\mathbf{E}}}= \frac{1}{n}\sum_{i=1}^n \left(\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle +\varepsilon_i\right) {{\mathbf{X}}}_i$. Then there exists a uniform constant $C>0$ such that, $$\label{ineq:concentration-target}
{\mathbb{P}}\left(\left\|{{\mathbf{E}}}- {{\mathbf{A}}}\right\| \geq C\sqrt{(a+b)(\|{{\mathbf{A}}}\|_F^2+\sigma^2)}\left(\sqrt{\frac{\log(a+b) + t}{n}} + \frac{\log(a+b) + t}{n}\right) \right) \leq \exp(-t)$$
[**Proof of Lemma \[lm:concentration-Gaussian-ensemble\].**]{} Denote ${{\mathbf{Z}}}_i = \left(\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i \rangle + \varepsilon_i\right){{\mathbf{X}}}_i$. It is easy to check that $ \mathbb{E} {{\mathbf{Z}}}_i = {{\mathbf{A}}}$. Then, $$\begin{split}
\mathbb{E} ({{\mathbf{Z}}}_i-{{\mathbf{A}}})({{\mathbf{Z}}}_i-{{\mathbf{A}}})^\top = & \mathbb{E} {{\mathbf{Z}}}_i{{\mathbf{Z}}}_i^\top -{{\mathbf{A}}}(\mathbb{E} {{\mathbf{Z}}}_i)^\top - (\mathbb{E}{{\mathbf{Z}}}_i) {{\mathbf{A}}}^\top + {{\mathbf{A}}}{{\mathbf{A}}}^\top = \mathbb{E}{{\mathbf{Z}}}_i{{\mathbf{Z}}}_i^\top- {{\mathbf{A}}}{{\mathbf{A}}}^\top \\
= & \mathbb{E} \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2 {{\mathbf{X}}}_i{{\mathbf{X}}}_i^\top + \sigma^2 \mathbb{E}{{\mathbf{X}}}_i{{\mathbf{X}}}_i^\top -{{\mathbf{A}}}{{\mathbf{A}}}^\top\\
= & \mathbb{E} \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2 {{\mathbf{X}}}_i{{\mathbf{X}}}_i^\top + \sigma^2\cdot b{{\mathbf{I}}}_{a} - {{\mathbf{A}}}{{\mathbf{A}}}^\top
\end{split}$$ Note that for any entry $({{\mathbf{X}}}_i)_{[j, k]}$, $\mathbb{E}({{\mathbf{X}}}_i)_{[j, k]}=0, \mathbb{E}({{\mathbf{X}}}_i)_{[j, k]}^2=1, \mathbb{E}({{\mathbf{X}}}_i)_{[j, k]}^3=0, \mathbb{E}({{\mathbf{X}}}_i)_{[j, k]}^4=3$. When $j\neq k$, $$\begin{split}
& \left(\mathbb{E} \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2 {{\mathbf{X}}}_i {{\mathbf{X}}}_i^\top\right)_{jk} = \mathbb{E} \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2 \sum_{l=1}^{b} ({{\mathbf{X}}}_i)_{[j, l]}({{\mathbf{X}}}_i)_{[k, l]}\\
= & \mathbb{E} \sum_{l=1}^b \left(2{{\mathbf{A}}}_{[j,l]} {{\mathbf{A}}}_{[k,l]} ({{\mathbf{X}}}_i)_{[i, l]} ({{\mathbf{X}}}_i)_{[k,l]}\right) ({{\mathbf{X}}}_i)_{[i,l]} ({{\mathbf{X}}}_i)_{[k,l]}\\
= & 2 \sum_{l=1}^b {{\mathbf{A}}}_{[j,l]}{{\mathbf{A}}}_{[k,l]} = 2({{\mathbf{A}}}{{\mathbf{A}}}^\top)_{[j,k]};
\end{split}$$ when $j = k$, $$\begin{split}
& \left(\mathbb{E} \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2 {{\mathbf{X}}}_i {{\mathbf{X}}}_i^\top\right)_{[j,j]} = \mathbb{E} \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2 \sum_{l=1}^b ({{\mathbf{X}}}_i)_{[j,l]}^2 \\
= & \mathbb{E} \sum_{j'=1}^{a}\sum_{l'=1}^{b} \left({{\mathbf{A}}}_{[j',l']}^2 ({{\mathbf{X}}}_i)_{[j',l']}^2\right) \cdot \sum_{l=1}^b ({{\mathbf{X}}}_i)_{[j,l]}^2 = \sum_{j'=1}^{a}\sum_{l'=1}^b ({{\mathbf{A}}}_{[j',l']}^2)\cdot b + 2\sum_{l=1}^b {{\mathbf{A}}}_{[j,l]}^2\\
= & b\|{{\mathbf{A}}}\|_F^2 + 2({{\mathbf{A}}}{{\mathbf{A}}}^\top)_{[j,j]}.
\end{split}$$ Therefore, $\mathbb{E}\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle^2{{\mathbf{X}}}_i{{\mathbf{X}}}_i^\top = 2{{\mathbf{A}}}{{\mathbf{A}}}^\top +b\|{{\mathbf{A}}}\|_F^2 {{\mathbf{I}}}_a$, and $$\label{eq:lm-expectation1}
\left\|\mathbb{E}({{\mathbf{Z}}}_i - {{\mathbf{A}}})({{\mathbf{Z}}}_i - {{\mathbf{A}}})^\top\right\| = \left\|2{{\mathbf{A}}}{{\mathbf{A}}}^\top +b\|{{\mathbf{A}}}\|_F^2{{\mathbf{I}}}_a + b\sigma^2{{\mathbf{I}}}_a - {{\mathbf{A}}}{{\mathbf{A}}}^\top\right\| = \|{{\mathbf{A}}}\|^2+b\|{{\mathbf{A}}}\|_F^2 + b\sigma^2.$$ Similarly, we can also show $$\label{eq:lm-expectation2}
\left\|\mathbb{E}({{\mathbf{Z}}}_i - {{\mathbf{A}}})^\top({{\mathbf{Z}}}_i - {{\mathbf{A}}})\right\| = \left\|2{{\mathbf{A}}}^\top {{\mathbf{A}}}+a\|{{\mathbf{A}}}\|_F^2{{\mathbf{I}}}_b + \sigma^2{{\mathbf{I}}}_a - {{\mathbf{A}}}^\top {{\mathbf{A}}}\right\| = \|{{\mathbf{A}}}\|^2+a\|{{\mathbf{A}}}\|_F^2 + a\sigma^2.$$ Next, we consider the spectral norm of ${{\mathbf{Z}}}_i$ and aim to show that $$\big\|\left\|{{\mathbf{Z}}}_i - {{\mathbf{A}}}\right\|\big\|_{\psi_1} = \inf_{u\geq 0}\left\{u:\mathbb{E}\exp\left(\frac{\|{{\mathbf{Z}}}_i-{{\mathbf{A}}}\|}{u}\right) \leq 2\right\} \leq C \left(\sqrt{a}+\sqrt{b}\right)\sqrt{\|{{\mathbf{A}}}\|_F^2+\sigma^2}$$ for uniform constant $C>0$. Note that $\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i \rangle +\varepsilon_i \sim N\left(0, \|{{\mathbf{A}}}\|_F^2 + \sigma^2\right)$, ${{\mathbf{X}}}_i$ is a random matrix, by Gaussian tail bound inequality and random matrix theory (Corollary 5.35 in [@vershynin2010introduction]), $$\label{ineq:lm-tail-probability}
\begin{split}
& {\mathbb{P}}\left(\left|\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle+\varepsilon_i\right| \geq t\sqrt{\|{{\mathbf{A}}}\|_F^2 + \sigma^2}\right) \leq 2\exp(-t^2/3),\\
& {\mathbb{P}}\left(\|{{\mathbf{X}}}_i\| \geq \sqrt{a}+\sqrt{b} + t\right) \leq \exp(-t^2/2).
\end{split}$$ We set $u = C_0\left(\sqrt{a}+\sqrt{b}\right)\sqrt{\|{{\mathbf{A}}}\|_F^2+\sigma^2}$ for large uniform constant $C_0\geq 80$. Thus, for any $x \geq 1$, $$\begin{split}
& {\mathbb{P}}\left(\|{{\mathbf{Z}}}_i - {{\mathbf{A}}}\| \geq xu\right) \leq {\mathbb{P}}\left(\| (\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle + \varepsilon_i) {{\mathbf{X}}}_i\| \geq xu-\|{{\mathbf{A}}}\|\right) \\
\leq & {\mathbb{P}}\left(\| (\langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle + \varepsilon_i ) {{\mathbf{X}}}_i\| \geq \frac{xC_0(\sqrt{a}+\sqrt{b})}{2}\sqrt{\|{{\mathbf{A}}}\|_F^2 + \sigma^2}\right)\\
\leq & {\mathbb{P}}\left( | \langle {{\mathbf{A}}}, {{\mathbf{X}}}_i\rangle + \varepsilon_i | \geq \sqrt{\frac{xC_0}{2}\cdot \left(\|{{\mathbf{A}}}\|_F^2 + \sigma^2\right)}\right) + {\mathbb{P}}\left(\|{{\mathbf{X}}}_i\| \geq \sqrt{\frac{xC_0}{2}}\cdot (\sqrt{a}+\sqrt{b})\right)\\
\overset{\text{\eqref{ineq:lm-tail-probability}}}{\leq} & 3\exp(- C_0x/6).
\end{split}$$ For any real valued function smooth $g$ and non-negative random variable $Y$ with density $f_Y$, the following identity holds, $$\mathbb{E} g(Y) = \int_0^\infty g'(y)P(Y\geq y) dy.$$ Thus, $$\begin{split}
& \mathbb{E}\exp\left(\frac{\|{{\mathbf{Z}}}_i - {{\mathbf{A}}}\|}{u}\right) = \int_0^\infty \exp\left(x\right) {\mathbb{P}}\left(\frac{\|{{\mathbf{Z}}}_i-{{\mathbf{A}}}\|}{u} \geq x\right) dx\\
\leq & \int_0^1 \exp(u) du + \int_{1}^\infty \exp(x) \cdot 3\exp(-C_0x/6)dx \\
\leq & \exp(1)-1 + \frac{3}{C_0/6-1} \leq 2,
\end{split}$$ which implies $\big\|\|{{\mathbf{Z}}}_i - {{\mathbf{A}}}\|\big\|_{\psi_1} \leq C_0\left(\sqrt{a}+\sqrt{b}\right)\sqrt{\|{{\mathbf{A}}}\|_F^2 + \sigma^2}$ for some uniform constant $C_0>0$.
Finally we apply the Bernstein-type matrix concentration inequality (c.f., Proposition 2 in [@koltchinskii2011nuclear] and Theorem 4 in [@koltchinskii2013remark]), $$\label{ineq:Z_i-A}
\begin{split}
\left\|\frac{1}{n}\sum_{i=1}^n {{\mathbf{Z}}}_i - {{\mathbf{A}}}\right\| & \leq C\max\Bigg\{ \sigma_Z\sqrt{\frac{t+\log(a+b)}{n}},\\
& (\sqrt{a}+\sqrt{b})\sqrt{\|{{\mathbf{A}}}\|_F^2+\sigma^2} \log\left(\frac{C(\sqrt{a}+\sqrt{b})\sqrt{\|{{\mathbf{A}}}\|_F^2 + \sigma^2}}{\sigma_Z}\right)\cdot \frac{t+\log(a+b)}{n}\Bigg\}
\end{split}$$ with probability at least $1 - \exp(-t)$. Here, $$\begin{split}
\sigma_Z := & \max\left\{\left\|\frac{1}{n}\sum_{i=1}^n\mathbb{E}({{\mathbf{Z}}}_i-{{\mathbf{A}}})({{\mathbf{Z}}}_i-{{\mathbf{A}}})^\top\right\|^{1/2}, \left\|\frac{1}{n}\sum_{i=1}^n\mathbb{E}({{\mathbf{Z}}}_i-{{\mathbf{A}}})^\top({{\mathbf{Z}}}_i-{{\mathbf{A}}})\right\|^{1/2}\right\}\\
= & \sqrt{\|{{\mathbf{A}}}\|^2 + (a\vee b) \left(\|{{\mathbf{A}}}\|_F^2 + \sigma^2\right)}.
\end{split}$$ Noting that $\sqrt{(a\vee b)(\|{{\mathbf{A}}}\|_F^2+\sigma^2)} \leq \sigma_Z \leq \sqrt{(a\vee b+1)(\|{{\mathbf{A}}}\|_F^2+\sigma^2)}$, implies . $\square$
\[lm:concentration-Gaussina-vector\] Suppose $x_1,\ldots, x_n\overset{iid}{\sim}N(0, {{\mathbf{I}}}_m)$ are i.i.d. $m$-dimensional random vectors, $\varepsilon_1,\ldots, \varepsilon_n\overset{iid}{\sim}N(0, \sigma^2)$, and $a\in \mathbb{R}^m$ is a fixed vector. Then $${\mathbb{P}}\left(\left\|\frac{1}{n}\sum_{i=1}^n\left(\langle {{\mathbf{x}}}_i, {{\mathbf{a}}}\rangle + \varepsilon_i\right) {{\mathbf{x}}}_i - {{\mathbf{a}}}\right\|_2 \leq \frac{C\sqrt{\|{{\mathbf{a}}}\|_2^2+\sigma^2}\left(\sqrt{n}+\sqrt{t}\right)\left(\sqrt{m}+\sqrt{t}\right)}{n} \right) \geq 1 - 5\exp(-t).$$
[**Proof of Lemma \[lm:concentration-Gaussina-vector\].**]{} Denote $${{\mathbf{x}}}_i = (x_{i1},\ldots, x_{im})^\top, \quad i=1,\ldots, n.$$ Since the distribution of Gaussian random vectors are invariant after orthogonal transformation, without loss of generality we assume ${{\mathbf{a}}}= (\theta, 0, \ldots, 0)$. Then $$\frac{1}{n}\left(\sum_{i=1}^n\langle {{\mathbf{x}}}_i, {{\mathbf{a}}}\rangle + \varepsilon_i\right) {{\mathbf{x}}}_i - {{\mathbf{a}}}= \begin{pmatrix}
\frac{1}{n}\sum_{i=1}^n (x_{i1}^2-1)\theta\\
\frac{1}{n}\sum_{i=1}^n x_{i1}\theta x_{i2}\\
\vdots\\
\frac{1}{n}\sum_{i=1}^n x_{i1}\theta x_{im}
\end{pmatrix} + \frac{1}{n}\sum_{i=1}^n \varepsilon_i {{\mathbf{x}}}_i := h + \frac{1}{n}\sum_{i=1}^n \varepsilon_i {{\mathbf{x}}}_i;$$ Note that $\sum_{i=1}^n x_{i1}^2 \sim \chi^2_n$, by tail bounds of $\chi^2$ (c.f., [@laurent2000adaptive Lemma 1]), $${\mathbb{P}}\left(n - 2\sqrt{nt} \leq \sum_{i=1}^n x_{i1}^2\right) \geq 1 - \exp(-t),\quad {\mathbb{P}}\left(\sum_{i=1}^n x_{i1}^2 \leq n + 2\sqrt{nt} + 2t \right) \geq 1 - \exp(-t).$$ Conditioning on the fixed value of $\xi := \sum_{i=1}^n x_{i1}^2$, we have $$\frac{1}{n}\sum_{i=1}^n x_{i1}\theta x_{ik}\Big|\xi \sim N\left(0, \frac{\theta^2\xi}{n^2}\right), \quad k=2,\ldots, n,$$ $$\|h\|_2^2 \Big| \xi \sim \left(\frac{\xi}{n} - 1\right)^2\theta^2 + \frac{\theta^2\xi}{n^2} \chi^2_{m-1}.$$ Thus, $$\begin{split}
& {\mathbb{P}}\left(\left\|h\right\|_2^2 \geq 4\theta^2\left(\sqrt{\frac{t}{n}} + \frac{t}{n}\right)^2 + \frac{\theta^2\left(n+2\sqrt{nt}+2t\right)\left(m-1+2\sqrt{(m-1)t}+2t\right)}{n^2} \right) \\
\leq & {\mathbb{P}}\left(\xi \geq n+2\sqrt{nt}+2t \right) + {\mathbb{P}}\left(\xi \leq n-2\sqrt{nt}\right) + {\mathbb{P}}\left(\frac{\theta^2\xi}{n^2}\chi_{m-1}^2 \geq \frac{\theta^2\xi(m-1+2\sqrt{(m-1)t}+2t)}{n^2}\right)\\
\leq & 3\exp(-t).
\end{split}$$ Conditioning on fixed values of $\|{\boldsymbol{\varepsilon}}\|_2^2 = \sum_i \varepsilon_i^2$, $$\left\|\frac{1}{n}\sum_{i=1}^n \varepsilon_i {{\mathbf{x}}}_i \right\|_2^2\Bigg| \|{\boldsymbol{\varepsilon}}\|_2^2 \sim \frac{\sigma^2\|{\boldsymbol{\varepsilon}}\|_2^2}{n^2}\chi^2_m.$$ Additionally, ${\mathbb{P}}\left(\|{\boldsymbol{\varepsilon}}\|_2^2 \geq \sigma^2(n+2\sqrt{nt}+2t)\right) \leq \exp(-t)$, which means $$\begin{split}
& {\mathbb{P}}\left(\left\|\frac{1}{n}\sum_{i=1}^n \varepsilon_i {{\mathbf{x}}}_i\right\|_2^2 \geq \frac{\sigma^2\left(n+2\sqrt{nt}+2t\right)\left(m+2\sqrt{mt}+2t\right)}{n^2} \right) \\
\leq & {\mathbb{P}}\left(\|{\boldsymbol{\varepsilon}}\|_2^2 \geq \sigma^2(n+2\sqrt{nt}+2t)\right) + {\mathbb{P}}\left(\left\|\frac{1}{n}\sum_{i=1}^n \varepsilon_i {{\mathbf{x}}}_i\right\|_2^2\Bigg| \|{\boldsymbol{\varepsilon}}\|_2^2 \geq \frac{\sigma^2\|{\boldsymbol{\varepsilon}}\|_2^2}{n^2}\left(m+2\sqrt{mt}+2t\right)\right)\\
\leq & 2\exp(-t).
\end{split}$$ Combining the previous two inequalities, we finally obtain $$\begin{split}
& {\mathbb{P}}\left(\left\|\frac{1}{n}\sum_{i=1}^n\left(\langle {{\mathbf{x}}}_i, {{\mathbf{a}}}\rangle + \varepsilon_i\right) {{\mathbf{x}}}_i - {{\mathbf{a}}}\right\|_2 \leq \frac{C\sqrt{\theta^2+\sigma^2}\left(\sqrt{n}+\sqrt{t}\right)\left(\sqrt{m}+\sqrt{t}\right)}{n}\right) \\
\geq & 1 - 5\exp(-t).
\end{split}$$ for constant $C>0$. $\square$
\[lm:concentration-independent\] Suppose ${{\mathbf{X}}}_1,\ldots, {{\mathbf{X}}}_n\in \mathbb{R}^{a\times b}$ ($a\leq b$) are i.i.d. standard Gaussian matrices, $\xi_1,\ldots, \xi_n \overset{iid}{\sim} N(0, \tau^2)$, and ${{\mathbf{E}}}= \frac{1}{n}\sum_{i=1}^n \xi_i {{\mathbf{X}}}_i$. Then the largest and smallest singular values of ${{\mathbf{E}}}$ satisfies the following tail probability, $${\mathbb{P}}\left(\sigma_{\max}^2({{\mathbf{E}}}) \geq \tau^2\frac{n + 2\sqrt{nx}+2x}{n^2}\left(\sqrt{a}+\sqrt{b}+\sqrt{2x}\right)^2 \right) \leq 2\exp(-x),$$ $${\mathbb{P}}\left(\sigma_{\min}^2({{\mathbf{E}}}) \leq \tau^2\frac{n - 2\sqrt{nx}}{n^2}\left(\sqrt{b}-\sqrt{a}-\sqrt{2x}\right)^2 \right) \leq 2\exp(-x).$$
[**Proof of Lemma \[lm:concentration-independent\].**]{} In the given setting, $\|\xi\|_2^2 = \sum_{i=1}^n \xi_i^2 \sim \tau^2\chi^2_n$, and $${{\mathbf{E}}}= \frac{1}{n}\sum_{i=1}^n \xi_i {{\mathbf{X}}}_i \Big| \|\xi\|_2 \overset{iid}{\sim} N\left(0, \frac{\|\xi\|_2^2}{n^2}\right).$$ By Corollary 5.35 in [@vershynin2010introduction], $$\label{ineq:random-matrix-tail-bound}
\begin{split}
{\mathbb{P}}\left(\sigma_{\max}^2({{\mathbf{E}}}) \geq \frac{\|\xi\|_2^2}{n^2}\left(\sqrt{a}+\sqrt{b}+\sqrt{2x} \right)^2\Big| \|\xi\|_2\right) \leq \exp(-x), \\
{\mathbb{P}}\left(\sigma_{\min}^2({{\mathbf{E}}}) \leq \frac{\|\xi\|_2^2}{n^2}\left(\sqrt{b}-\sqrt{a}-\sqrt{2x} \right)^2\Big| \|\xi\|_2\right) \leq \exp(-x).
\end{split}$$ By the tail bound of $\chi^2$ distribution (Lemma 1 in [@laurent2000adaptive]), $$\label{ineq:chi-square-tail-bound}
{\mathbb{P}}\left(\|\xi\|_2^2 \geq \tau^2\left(n + 2\sqrt{nx}+2x\right)\right) \leq e^{-x},\quad {\mathbb{P}}\left(\|\xi_2\|_2^2 \leq \tau^2\left(n - 2\sqrt{nx}\right)\right) \leq e^{-x}.$$ By and , we have $$\begin{split}
& {\mathbb{P}}\left(\sigma_{\max}^2({{\mathbf{E}}}) \geq \tau^2\frac{n + 2\sqrt{nx}+2x}{n^2}\left(\sqrt{a}+\sqrt{b}+\sqrt{2x}\right)^2\right)\\
\leq & {\mathbb{P}}\left(\sigma_{\max}^2({{\mathbf{E}}}) \geq \frac{\|\xi\|_2^2}{n^2}\left(\sqrt{a}+\sqrt{b}+\sqrt{2x}\right)^2 \text{~or~} \|\xi\|_2^2 \geq \tau^2\left(n + 2\sqrt{nx}+2x\right)\right)\\
\leq & \exp(-x) + \exp(-x) = 2\exp(-x);
\end{split}$$ $$\begin{split}
& {\mathbb{P}}\left(\sigma_{\min}^2({{\mathbf{E}}}) \leq \tau^2\frac{n - 2\sqrt{nx}}{n^2}\left(\sqrt{b}-\sqrt{a}-\sqrt{2x}\right)^2\right)\\
\leq & {\mathbb{P}}\left(\sigma_{\min}^2({{\mathbf{E}}}) \leq \frac{\|\xi\|_2^2}{n^2}\left(\sqrt{b}-\sqrt{a}-\sqrt{2x}\right)^2 \text{~or~} \|\xi\|_2^2 \leq \tau^2\left(n - 2\sqrt{nx}\right)\right)\\
\leq & \exp(-x) + \exp(-x) = 2\exp(-x).
\end{split}$$ $\square$
The next lemma provides an upper bound for the projection error after perturbation, which is useful in the singular subspace perturbation analysis in the proofs of the main results.
\[lm:SVD-projection\] Suppose ${{\mathbf{A}}}, {{\mathbf{Z}}}$ are two matrices of the same dimension and $\widehat{{{\mathbf{U}}}} = {{\rm SVD}}_r({{\mathbf{A}}}+{{\mathbf{Z}}})$. Then, $$\left\|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{A}}}\right\|\leq \sigma_{r+1}({{\mathbf{A}}})+2\|{{\mathbf{Z}}}\|,\quad \left\|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{A}}}\right\|_F\leq \sqrt{\sum_{k\geq r+1}\sigma_k^2({{\mathbf{A}}})} + 2\|{{\mathbf{Z}}}\|_F.$$ In particular when ${{\rm rank}}({{\mathbf{A}}})\leq r$, $$\left\|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{A}}}\right\|\leq 2\|{{\mathbf{Z}}}\|, \quad \left\|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{A}}}\right\|_F\leq 2\min\left\{\|{{\mathbf{Z}}}\|_F, \sqrt{r}\|{{\mathbf{Z}}}\|\right\}.$$
[**Proof of Lemma \[lm:SVD-projection\].**]{} Suppose ${{\mathbf{A}}}= \sum_k \sigma_k({{\mathbf{A}}}){{\mathbf{u}}}_k {{\mathbf{v}}}_k^\top$ is the singular value decomposition. Then, $$\begin{split}
\|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{A}}}\| \leq & \left\|P_{\widehat{{{\mathbf{U}}}}_\perp}({{\mathbf{A}}}+{{\mathbf{Z}}})\right\| + \|{{\mathbf{Z}}}\| = \sigma_{r+1}({{\mathbf{A}}}+ {{\mathbf{Z}}})+ \|{{\mathbf{Z}}}\|\\
= & \min_{{{\rm rank}}({{\mathbf{M}}})\leq r}\|{{\mathbf{A}}}+{{\mathbf{Z}}}-{{\mathbf{M}}}\| + \|{{\mathbf{Z}}}\|\\
\leq & \left\|{{\mathbf{A}}}+ {{\mathbf{Z}}}- \sum_{k=1}^{r}\sigma_k({{\mathbf{A}}}){{\mathbf{u}}}_k {{\mathbf{v}}}_k^\top\right\| + \|{{\mathbf{Z}}}\| = \left\|{{\mathbf{Z}}}+ \sum_{k\geq r+1} \sigma_k({{\mathbf{A}}}){{\mathbf{u}}}_k {{\mathbf{v}}}_k^\top\right\| + \|{{\mathbf{Z}}}\| \\
\leq & \sigma_{r+1}({{\mathbf{A}}}) + 2\|{{\mathbf{Z}}}\|.
\end{split}$$ $$\begin{split}
\|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{A}}}\|_F \leq & \left\|P_{\widehat{{{\mathbf{U}}}}_\perp}({{\mathbf{A}}}+{{\mathbf{Z}}})\right\|_F + \|P_{\widehat{{{\mathbf{U}}}}_\perp}{{\mathbf{Z}}}\|_F = \sqrt{\sum_{k\geq r+1}\sigma_{k}^2({{\mathbf{A}}}+{{\mathbf{Z}}})}+ \|{{\mathbf{Z}}}\|_F\\
= & \min_{{{\rm rank}}({{\mathbf{M}}})\leq r}\|{{\mathbf{A}}}+{{\mathbf{Z}}}-{{\mathbf{M}}}\|_F + \|{{\mathbf{Z}}}\|_F\\
\leq & \left\|{{\mathbf{A}}}+ {{\mathbf{Z}}}- \sum_{k=1}^r\sigma_k({{\mathbf{A}}}){{\mathbf{u}}}_k {{\mathbf{v}}}_k^\top\right\|_F + \|{{\mathbf{Z}}}\|_F \leq \sqrt{\sum_{k\geq r+1}\sigma^2_k({{\mathbf{A}}})} + 2\|{{\mathbf{Z}}}\|_F.
\end{split}$$ Finally, when ${{\rm rank}}({{\mathbf{A}}})\leq r$, ${{\rm rank}}(P_{\widehat{{{\mathbf{U}}}}_{\perp}}{{\mathbf{A}}}) \leq {{\rm rank}}({{\mathbf{A}}}) \leq r$, then $$\|P_{\widehat{{{\mathbf{U}}}}_{\perp}}{{\mathbf{A}}}\|_F \leq \min\left\{\sqrt{\sum_{k\geq r+1}\sigma_k^2({{\mathbf{A}}})} + 2\|{{\mathbf{Z}}}\|_F, \sqrt{r}\left\|P_{\widehat{{{\mathbf{U}}}}_{\perp}}{{\mathbf{A}}}\right\|\right\} \leq \min\left\{2\|{{\mathbf{Z}}}\|_F, 2\sqrt{r}\|{{\mathbf{Z}}}\|\right\}.$$ $\square$
The Lemma \[lm:projection-difference\] below provides a inequality for tensors after tensor-matrix product projections.
\[lm:projection-difference\] Suppose ${{\mathbfcal{A}}}\in \mathbb{R}^{p_1\times \cdots \times p_d}$ is an order-$d$ tensor and $\widetilde{{{\mathbf{U}}}}_k \in \mathbb{O}_{p_k, r_k}$, $k=1,\ldots, d$, are orthogonal matrices. Let $\|\cdot\|_\bullet$ be a tensor norm that satisfies sub-multiplicative inequality, i.e., $\|{{\mathbfcal{A}}}\times_k {{\mathbf{B}}}\|_\bullet \leq \|{{\mathbfcal{A}}}\|_\bullet \cdot \|{{\mathbf{B}}}\|$ for any tensor ${{\mathbfcal{A}}}$ and matrix ${{\mathbf{B}}}$ (in particular, the tensor Hilbert-Schmitt norm satisfies this condition), we have $$\left\|\llbracket{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\rrbracket - {{\mathbfcal{A}}}\right\|_{\bullet} \leq \sum_{k=1}^d \left\|{{\mathbfcal{A}}}\times_k P_{\widetilde{{{\mathbf{U}}}}_{k\perp}}\right\|_{\bullet}.$$ Specifically, $$\left\|\llbracket{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\rrbracket - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}} = \left\|P_{(\widetilde{{{\mathbf{U}}}}_d\otimes \cdots \otimes \widetilde{{{\mathbf{U}}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 \leq \sum_{k=1}^d \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top\mathcal{M}_k({{\mathbfcal{A}}})\right\|_F.$$
[**Proof of Lemma \[lm:projection-difference\].**]{} Note that $$\begin{split}
{{\mathbfcal{A}}}= & \left\llbracket {{\mathbfcal{A}}}; \left(P_{\widetilde{{{\mathbf{U}}}}_1} + P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}\right), \ldots, \left(P_{\widetilde{{{\mathbf{U}}}}_d} + P_{\widetilde{{{\mathbf{U}}}}_{d\perp}}\right)\right\rrbracket \\
= & \left\llbracket {{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\right\rrbracket + \left\llbracket {{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\right\rrbracket + \left\llbracket {{\mathbfcal{A}}}; {{\mathbf{I}}}_{p_1}, P_{\widetilde{{{\mathbf{U}}}}_{2\perp}}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\right\rrbracket\\
& + \cdots + \left\llbracket {{\mathbfcal{A}}}; {{\mathbf{I}}}_{p_1}, {{\mathbf{I}}}_{p_2}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_{d\perp}}\right\rrbracket.
\end{split}$$ Additionally, $\|P_{\widetilde{{{\mathbf{U}}}}_k}\|\leq 1, \|P_{\widetilde{{{\mathbf{U}}}}_{k\perp}}\|\leq 1$. Thus, $$\begin{split}
\left\|\llbracket{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\rrbracket - {{\mathbfcal{A}}}\right\|_\bullet \leq & \left\|\left\llbracket {{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_{1\perp}}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\right\rrbracket\right\|_\bullet + \left\|\left\llbracket {{\mathbfcal{A}}}; {{\mathbf{I}}}_{p_1}, P_{\widetilde{{{\mathbf{U}}}}_{2\perp}}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\right\rrbracket\right\|_\bullet\\
& + \cdots + \left\|\left\llbracket {{\mathbfcal{A}}}; {{\mathbf{I}}}_{p_1}, {{\mathbf{I}}}_{p_2}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_{d\perp}}\right\rrbracket\right\|_\bullet\\
\leq & \sum_{k=1}^d \left\|{{\mathbfcal{A}}}\times_k P_{\widetilde{{{\mathbf{U}}}}_{k\perp}}\right\|_\bullet.
\end{split}$$ Specifically for the Hilbert-Schmitt norm, $$\begin{split}
& \left\|P_{(\widetilde{{{\mathbf{U}}}}_d\otimes \cdots \otimes \widetilde{{{\mathbf{U}}}}_1)_\perp}{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 = \left\|P_{(\widetilde{{{\mathbf{U}}}}_d\otimes \cdots \otimes \widetilde{{{\mathbf{U}}}}_1)}{{\rm vec}}({{\mathbfcal{A}}})-{{\rm vec}}({{\mathbfcal{A}}})\right\|_2 \\
\leq & \left\|\llbracket{{\mathbfcal{A}}}; P_{\widetilde{{{\mathbf{U}}}}_1}, \ldots, P_{\widetilde{{{\mathbf{U}}}}_d}\rrbracket - {{\mathbfcal{A}}}\right\|_{{{\rm HS}}} \leq \sum_{k=1}^d \left\|{{\mathbfcal{A}}}\times_k P_{\widetilde{{{\mathbf{U}}}}_{k\perp}}\right\|_{{{\rm HS}}} = \sum_{k=1}^d \left\|\mathcal{M}_k\left({{\mathbfcal{A}}}\times_k P_{\widetilde{{{\mathbf{U}}}}_{k\perp}}\right)\right\|_F\\
= & \left\|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top\mathcal{M}_k({{\mathbfcal{A}}})\right\|_{F}.
\end{split}$$ Therefore, we have finished the proof of lemma \[lm:projection-difference\]. $\square$
The next Lemma \[lm:projection\_remainder\] introduces a useful inequality for the tensor projected orthogonal to a Cross structure (i.e., $\widetilde{{{\mathbf{U}}}}$ in the statement below).
\[lm:projection\_remainder\] Suppose ${{\mathbfcal{A}}}= \llbracket{{\mathbfcal{S}}}; {{\mathbf{U}}}_1, {{\mathbf{U}}}_2, {{\mathbf{U}}}_3 \rrbracket$ is a rank-$(r_1, r_2 ,r_3)$ tensor. ${{\mathbf{U}}}_k\in \mathbb{O}_{p_k, r_k}$ and ${{\mathbf{W}}}_k\in \mathbb{O}_{p_{k+1}p_{k+2}, p_k}$ are the left and right singular subspaces of $\mathcal{M}_k({{\mathbfcal{A}}}) := {{\mathbf{A}}}_k$, respectively. Suppose $\widetilde{{{\mathbf{U}}}}_k\in \mathbb{O}_{p_k, r_k}$ and $$\widetilde{{{\mathbf{W}}}}_1 = (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2)\widetilde{{{\mathbf{V}}}}_1\in \mathbb{O}_{p_1, r_1}, \quad \widetilde{{{\mathbf{W}}}}_2 = (\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_1)\widetilde{{{\mathbf{V}}}}_2\in \mathbb{O}_{p_2, r_2}, \quad \widetilde{{{\mathbf{W}}}}_3 = (\widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1)\widetilde{{{\mathbf{V}}}}_3\in \mathbb{O}_{p_3, r_3}$$ are sample estimates of ${{\mathbf{U}}}$ and ${{\mathbf{W}}}_k$, respectively. Assume $\widetilde{{{\mathbf{U}}}}_k$ and $\widetilde{{{\mathbf{W}}}}_k$ satisfy $$\|\sin\Theta(\widetilde{{{\mathbf{U}}}}_k, {{\mathbf{U}}}_k)\| \leq \theta_k, \quad \|\widetilde{{{\mathbf{U}}}}_{k\perp}^\top {{\mathbf{A}}}_k\|_F \leq \eta_k, \quad \|{{\mathbf{A}}}_k\widetilde{{{\mathbf{W}}}}_{k\perp}\|_F \leq \xi_k, \quad k=1,2,3.$$ Let $$\widetilde{{{\mathbf{U}}}} = \begin{bmatrix}
\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1, ~~ \mathcal{R}_1(\widetilde{{{\mathbf{W}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_{1 \perp}), ~~ \mathcal{R}_2(\widetilde{{{\mathbf{W}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{2\perp}), ~~ \mathcal{R}_3(\widetilde{{{\mathbf{W}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_{3\perp})
\end{bmatrix},$$ where $\mathcal{R}_k(\cdot)$ is the row-permutation operator that matches the row indices of $\widetilde{{{\mathbf{W}}}}_k\otimes \widetilde{{{\mathbf{U}}}}_{k\perp}$ to ${{\rm vec}}({{\mathbfcal{A}}})$ and the actual definitions of $\mathcal{R}_k$ are provided in Section \[sec:row-permutation-operator\] in the supplementary materials. Recall $\widetilde{{{\mathbf{U}}}}_\perp$ is the orthogonal complement of ${{\mathbf{U}}}$. Then, $$\begin{split}
\|P_{\widetilde{{{\mathbf{U}}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2 \leq & \sum_{k=1,2,3} \left(\theta_k^2\xi_k^2 + \min\{\theta_{k+1}^2\eta_{k+2}^2, \theta_{k+2}^2\eta_{k+1}^2\}\right)\\
& + \min\{\eta_1^2\theta_2^2\theta_3^2, \theta_1^2\eta_2^2\theta_3^2, \theta_1^2\theta_2^2\eta_3^2\}.
\end{split}$$
[**Proof of Lemma \[lm:projection\_remainder\].**]{} Since $$\widetilde{{{\mathbf{U}}}} = \begin{bmatrix}
\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_1, ~~ \mathcal{R}_1(\widetilde{{{\mathbf{W}}}}_1\otimes \widetilde{{{\mathbf{U}}}}_{1\perp}), ~~ \mathcal{R}_2(\widetilde{{{\mathbf{W}}}}_2\otimes \widetilde{{{\mathbf{U}}}}_{2\perp}), ~~ \mathcal{R}_3( \widetilde{{{\mathbf{W}}}}_3 \otimes\widetilde{{{\mathbf{U}}}}_{3\perp})
\end{bmatrix}\in \mathbb{O}_{p_1p_2p_3, m},$$ where $m = r_1r_2r_3 + (p_1-r_1)r_1 + (p_2-r_2)r_2 + (p_3-r_3)r_3$. Denote $$\label{eq:tilde-U-1k}
\begin{split}
\widetilde{{{\mathbf{U}}}}_{11} =
\mathcal{R}_1\left(\left((\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})\widetilde{{{\mathbf{V}}}}_{1\perp}\right) \otimes \widetilde{{{\mathbf{U}}}}_{1\perp}\right) \in \mathbb{O}_{p_1p_2p_3, (p_1-r_1)(r_{2}r_{3}-r_1)},\\
\widetilde{{{\mathbf{U}}}}_{12} = \mathcal{R}_2\left(\left((\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{1})\widetilde{{{\mathbf{V}}}}_{2\perp}\right) \otimes \widetilde{{{\mathbf{U}}}}_{2\perp}\right) \in \mathbb{O}_{p_1p_2p_3, (p_2-r_2)(r_{1}r_{3}-r_2)},\\
\widetilde{{{\mathbf{U}}}}_{13} =\mathcal{R}_3\left(\left((\widetilde{{{\mathbf{U}}}}_{2}\otimes \widetilde{{{\mathbf{U}}}}_{1})\widetilde{{{\mathbf{V}}}}_{3\perp}\right) \otimes \widetilde{{{\mathbf{U}}}}_{3\perp}\right) \in \mathbb{O}_{p_1p_2p_3, (p_3-r_3)(r_{2}r_{1}-r_3)},\\
\end{split}$$ $$\label{eq:tilde-U-2k}
\begin{split}
\widetilde{{{\mathbf{U}}}}_{21} = \widetilde{{{\mathbf{U}}}}_{3\perp} \otimes \widetilde{{{\mathbf{U}}}}_{2\perp} \otimes \widetilde{{{\mathbf{U}}}}_{1} \in \mathbb{O}_{p_1p_2p_3, r_1(p_2-r_2)(p_3-r_3)};\\
\widetilde{{{\mathbf{U}}}}_{22} = \widetilde{{{\mathbf{U}}}}_{3\perp} \otimes \widetilde{{{\mathbf{U}}}}_{2} \otimes \widetilde{{{\mathbf{U}}}}_{1 \perp} \in \mathbb{O}_{p_1p_2p_3, r_2(p_1-r_1)(p_3-r_3)};\\
\widetilde{{{\mathbf{U}}}}_{23} = \widetilde{{{\mathbf{U}}}}_{3} \otimes \widetilde{{{\mathbf{U}}}}_{2\perp} \otimes \widetilde{{{\mathbf{U}}}}_{1 \perp} \in \mathbb{O}_{p_1p_2p_3, r_3(p_1-r_1)(p_2-r_2)};\\
\end{split}$$ $$\label{eq:tilde-U-3}
\widetilde{{{\mathbf{U}}}}_{3\ast} =
\widetilde{{{\mathbf{U}}}}_{3\perp} \otimes \widetilde{{{\mathbf{U}}}}_{2\perp} \otimes \widetilde{{{\mathbf{U}}}}_{1\perp} \in \mathbb{O}_{p_1p_2p_3, (p_1-r_1)(p_2-r_2)(p_3-r_3)}.$$ Then it is not hard to verify that $[\widetilde{{{\mathbf{U}}}}_{11}, \widetilde{{{\mathbf{U}}}}_{12}, \widetilde{{{\mathbf{U}}}}_{13}, \widetilde{{{\mathbf{U}}}}_{21}, \widetilde{{{\mathbf{U}}}}_{22}, \widetilde{{{\mathbf{U}}}}_{23}, \widetilde{{{\mathbf{U}}}}_{3\ast}]$ forms an orthogonal complement of $\widetilde{{{\mathbf{U}}}}$. Thus, we have the following decomposition, $$\begin{split}
\|P_{\widetilde{{{\mathbf{U}}}}_{\perp}} {{\rm vec}}({{\mathbfcal{A}}})\|_2^2 = \sum_{k=1,2,3} \|P_{\widetilde{{{\mathbf{U}}}}_{1k}} {{\rm vec}}({{\mathbfcal{A}}})\|_2^2 + \sum_{k=1,2,3} \|P_{\widetilde{{{\mathbf{U}}}}_{2k}} {{\rm vec}}({{\mathbfcal{A}}})\|_2^2 + \|P_{\widetilde{{{\mathbf{U}}}}_{3\ast}} {{\rm vec}}({{\mathbfcal{A}}})\|_2^2.
\end{split}$$ We analyze each term separately as follows.
- Note that $$\begin{bmatrix}
(\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})\widetilde{{{\mathbf{V}}}}_1, ~~ (\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})\widetilde{{{\mathbf{V}}}}_{1\perp}, ~~ (\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})_\perp
\end{bmatrix}$$ is a square orthogonal matrix, we know $$\begin{bmatrix}
(\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})\widetilde{{{\mathbf{V}}}}_{1\perp}, ~~ (\widetilde{{{\mathbf{U}}}}_{3}\otimes \widetilde{{{\mathbf{U}}}}_{2})_\perp
\end{bmatrix}$$ is an orthogonal complement to $\widetilde{{{\mathbf{W}}}}_1$. Given the left and right singular subspaces of ${{\mathbf{A}}}_1$ are ${{\mathbf{U}}}_1$ and ${{\mathbf{W}}}_1$, we have $$\begin{split}
& \|P_{\widetilde{{{\mathbf{U}}}}_{11}}{{\rm vec}}({{\mathbfcal{A}}})\|_F^2 \overset{\eqref{eq:tilde-U-1k}}{=} \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1\left((\widetilde{{{\mathbf{U}}}}_3\otimes \widetilde{{{\mathbf{U}}}}_2)\widetilde{{{\mathbf{V}}}}_{1\perp}\right)\right\|_F^2\\
\leq & \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_{1\perp}\right\|_F^2 = \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{U}}}_1 {{\mathbf{U}}}_1^\top {{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_{1\perp}\right\|_F^2\leq \|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{U}}}_1\|^2\cdot \|{{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_{1\perp}\|_F^2 \\
\leq & \|\sin\Theta(\widetilde{{{\mathbf{U}}}}_1, {{\mathbf{U}}}_1)\|^2 \cdot \|{{\mathbf{A}}}_1\widetilde{{{\mathbf{W}}}}_{1\perp}\|_F^2 \leq \theta_1^2\xi_1^2.
\end{split}$$ Similar inequalities also hold for $\|P_{\widetilde{{{\mathbf{U}}}}_{12}}{{\rm vec}}({{\mathbfcal{A}}})\|_F^2$ and $\|P_{\widetilde{{{\mathbf{U}}}}_{13}}{{\rm vec}}({{\mathbfcal{A}}})\|_F^2$.
- $$\begin{split}
& \|P_{\widetilde{{{\mathbf{U}}}}_{21}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2 = \|\widetilde{{{\mathbf{U}}}}_{21}^\top {{\rm vec}}({{\mathbfcal{A}}})\|_2^2 = \|{{\mathbfcal{A}}}\times_1 \widetilde{{{\mathbf{U}}}}_1^\top \times_2 \widetilde{{{\mathbf{U}}}}_{2\perp}^\top \times_3 \widetilde{{{\mathbf{U}}}}_{3\perp}^\top \|_{{{\rm HS}}}^2\\
= & \|\widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{A}}}_2 (\widetilde{{{\mathbf{U}}}}_{3\perp}\otimes \widetilde{{{\mathbf{U}}}}_1)\|_F^2 = \|\widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{U}}}_2{{\mathbf{U}}}_2^\top {{\mathbf{A}}}_2 (\widetilde{{{\mathbf{U}}}}_{3\perp}\otimes \widetilde{{{\mathbf{U}}}}_1)\|_F^2 \\
\leq & \|\widetilde{{{\mathbf{U}}}}_{2\perp}^\top {{\mathbf{U}}}_2\|^2 \cdot\|{{\mathbf{U}}}_2^\top {{\mathbf{A}}}_2 (\widetilde{{{\mathbf{U}}}}_{3\perp}\otimes \widetilde{{{\mathbf{U}}}}_1)\|_F^2\\
\leq & \|\sin\Theta(\widetilde{{{\mathbf{U}}}}_2, {{\mathbf{U}}}_2)\|^2\cdot \|{{\mathbf{A}}}_2(\widetilde{{{\mathbf{U}}}}_{3\perp}\otimes \widetilde{{{\mathbf{U}}}}_1)\|_F^2 = \theta_2^2 \cdot \|{{\mathbfcal{A}}}\times_1\widetilde{{{\mathbf{U}}}}_1^\top \times_3 \widetilde{{{\mathbf{U}}}}_{3\perp}^\top\|_{{{\rm HS}}}^2\\
= & \theta_2^2 \cdot \|\widetilde{{{\mathbf{U}}}}_{3\perp}^\top {{\mathbf{A}}}_3\|_F^2\leq \theta_2^2 \eta_3^2.
\end{split}$$
By symmetry, $\|P_{\widetilde{{{\mathbf{U}}}}_{21}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2\leq \theta_3^2\eta_2^2$. Similar inequalities also hold for $\|P_{\widetilde{{{\mathbf{U}}}}_{22}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2$ and $\|P_{\widetilde{{{\mathbf{U}}}}_{23}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2$. Therefore, $$\|P_{\widetilde{{{\mathbf{U}}}}_{2k}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2 \leq \min\{\theta_{k+1}^2\eta_{k+2}^2, \theta_{k+2}^2\eta_{k+1}^2\}, \quad \text{for} ~~ k=1,2,3.$$
- Similarly as the previous part, $$\begin{split}
\|P_{\widetilde{{{\mathbf{U}}}}_{3\ast}}{{\rm vec}}({{\mathbfcal{A}}})\|_2 \leq & \left\|{{\mathbfcal{A}}}\times_1 \widetilde{{{\mathbf{U}}}}_{1\perp}^\top \times_2 \widetilde{{{\mathbf{U}}}}_{2\perp}^\top \times_3 \widetilde{{{\mathbf{U}}}}_{3\perp}^\top\right\|_{{{\rm HS}}}\\
= & \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbfcal{A}}}_1 \left(\widetilde{{{\mathbf{U}}}}_{3\perp} \otimes \widetilde{{{\mathbf{U}}}}_{2\perp}\right)\right\|_F\\
= & \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1 ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2) ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)^\top (\widetilde{{{\mathbf{U}}}}_{3\perp}\otimes \widetilde{{{\mathbf{U}}}}_{2\perp})\right\|_F\\
\leq & \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1 ({{\mathbf{U}}}_3\otimes {{\mathbf{U}}}_2)\right\|_F \cdot \left\|({{\mathbf{U}}}_3^\top \widetilde{{{\mathbf{U}}}}_{3\perp})\otimes ({{\mathbf{U}}}_2^\top \widetilde{{{\mathbf{U}}}}_{2\perp})\right\| \\
\leq & \left\|\widetilde{{{\mathbf{U}}}}_{1\perp}^\top {{\mathbf{A}}}_1 \right\|_F\cdot \left\|({{\mathbf{U}}}_3^\top \widetilde{{{\mathbf{U}}}}_{3\perp})\right\|\cdot \left\| ({{\mathbf{U}}}_2^\top \widetilde{{{\mathbf{U}}}}_{2\perp})\right\|\leq \eta_1\theta_2\theta_3.
\end{split}$$ Similar upper bounds of $\theta_1\eta_2\theta_3$ and $\theta_1\theta_2\eta_3$ also hold. Thus, $$\|P_{\widetilde{{{\mathbf{U}}}}_{3\ast}}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2 \leq \min\{\eta_1^2\theta_2^2\theta_3^2, \theta_1^2\eta_2^2\theta_3^2, \theta_1^2\theta_2^2\eta_3^2\}.$$
In summary, $$\begin{split}
\|P_{{{\mathbf{U}}}_\perp}{{\rm vec}}({{\mathbfcal{A}}})\|_2^2 \leq & \sum_{k=1,2,3} \left(\theta_k^2\xi_k^2 + \min\{\theta_{k+1}^2\eta_{k+2}^2, \theta_{k+2}^2\eta_{k+1}^2\}\right)\\
& + \min\{\eta_1^2\theta_2^2\theta_3^2, \theta_1^2\eta_2^2\theta_3^2, \theta_1^2\theta_2^2\eta_3^2\}.
\end{split}$$ $\square$
The following lemma discusses the Bayes risk of regular linear regression. Though it is a standard result in statistical decision theory (c.f., Exercise 5.8, p. 403 in [@lehmann2006theory]), we present the proof here for completeness of statement.
\[lm:linear-regression-lower-bound\] Consider the linear regression model ${{\mathbf{y}}}= {{\mathbf{X}}}{\boldsymbol{\beta}}+\varepsilon$. Here, $\varepsilon\overset{iid}{\sim} N(0, \sigma^2)$; the parameter ${\boldsymbol{\beta}}$ is generated from a prior distribution: ${\boldsymbol{\beta}}\overset{iid}{\sim} N(0, \tau^2)$. We aim to estimate ${\boldsymbol{\beta}}$ based on $({{\mathbf{y}}}, {{\mathbf{X}}})$ with the minimal $\ell_2$ risk. Then, the Bayes estimator for ${\boldsymbol{\beta}}$ and the corresponding Bayes risk are $$\widehat{{\boldsymbol{\beta}}} = \left(\frac{\sigma^2 {{\mathbf{I}}}}{\tau^2} + {{\mathbf{X}}}^\top{{\mathbf{X}}}\right)^{-1}{{\mathbf{X}}}^\top \quad\text{and}\quad \mathbb{E}\left((\widehat{{\boldsymbol{\beta}}} - {\boldsymbol{\beta}})^2| {{\mathbf{X}}}\right) = {{\rm tr}}\left(\left(\frac{{{\mathbf{I}}}}{\tau^2} + \frac{{{\mathbf{X}}}^\top{{\mathbf{X}}}}{\sigma^2}\right)^{-1}\right).$$
[**Proof of Lemma \[lm:linear-regression-lower-bound\].**]{} When ${\boldsymbol{\beta}}\overset{iid}{\sim}N(0, \tau^2)$ and $\varepsilon\overset{iid}{\sim}N(0, \sigma^2)$, $$\begin{split}
p({\boldsymbol{\beta}}\Big| {{\mathbf{X}}}, {{\mathbf{y}}}) ~~ \varpropto ~~ & p({{\mathbf{y}}}|{{\mathbf{X}}}, {\boldsymbol{\beta}}) \cdot p({\boldsymbol{\beta}}) \\
\varpropto ~~ & \exp\left(-\|{{\mathbf{y}}}-{{\mathbf{X}}}{\boldsymbol{\beta}}\|_2^2/(2\sigma^2)\right)\cdot\exp(-{\boldsymbol{\beta}}^\top{\boldsymbol{\beta}}/(2\tau^2))\\
\varpropto ~~ & \exp\left(- \frac{{\boldsymbol{\beta}}^\top{\boldsymbol{\beta}}}{2\tau^2} - \frac{{\boldsymbol{\beta}}^\top {{\mathbf{X}}}^\top{{\mathbf{X}}}{\boldsymbol{\beta}}}{2\sigma^2} + \frac{{{\mathbf{y}}}^\top {{\mathbf{X}}}{\boldsymbol{\beta}}}{\sigma^2}\right)\\
\varpropto ~~ & \exp\left(- \frac{1}{2}\left\|\left(\frac{{{\mathbf{I}}}}{\tau^2} + \frac{{{\mathbf{X}}}^\top{{\mathbf{X}}}}{\sigma^2}\right)^{-1/2}\frac{{{\mathbf{X}}}^\top {{\mathbf{y}}}}{\sigma^2} - \left(\frac{{{\mathbf{I}}}}{\tau^2} + \frac{{{\mathbf{X}}}^\top{{\mathbf{X}}}}{\sigma^2}\right)^{1/2}{\boldsymbol{\beta}}\right\|_2^2 \right)
\end{split}$$ Thus, the posterior distribution of ${\boldsymbol{\beta}}$ is $${\boldsymbol{\beta}}\Big| {{\mathbf{X}}}, {{\mathbf{y}}}~~ \sim ~~ N\left(\left(\frac{\sigma^2 {{\mathbf{I}}}}{\tau^2} + {{\mathbf{X}}}^\top{{\mathbf{X}}}\right)^{-1}{{\mathbf{X}}}^\top {{\mathbf{y}}}, \left(\frac{{{\mathbf{I}}}}{\tau^2} + \frac{{{\mathbf{X}}}^\top{{\mathbf{X}}}}{\sigma^2}\right)^{-1}\right).$$ Then, the Bayes estimator, i.e., the posterior mean, and the corresponding Bayes risk are $$\widehat{\boldsymbol{\beta}}= \mathbb{E}({\boldsymbol{\beta}}|{{\mathbf{X}}}, {{\mathbf{y}}}) = \left(\frac{\sigma^2 {{\mathbf{I}}}}{\tau^2} + {{\mathbf{X}}}{{\mathbf{X}}}^\top\right)^{-1}{{\mathbf{X}}}^\top {{\mathbf{y}}}, \quad \mathbb{E}((\widehat{{\boldsymbol{\beta}}} - {\boldsymbol{\beta}})^2| {{\mathbf{X}}}, {{\mathbf{y}}}) = {{\rm tr}}\left(\left(\frac{{{\mathbf{I}}}}{\tau^2} + \frac{{{\mathbf{X}}}{{\mathbf{X}}}^\top}{\sigma^2}\right)^{-1}\right),$$ respectively. Thus, we have finished the proof of this lemma.$\square$
The following lemma provides a deterministic bound for the group Lasso estimator under group restricted isometry property.
\[lm:group-oracle-RIP\] Suppose ${{\mathbf{X}}}\in \mathbb{R}^{n\times pr}$, $\{G_1,\ldots, G_p\}$ is a partition of $\{1,\ldots, pr\}$ and $|G_1|=\cdots =|G_p|$. Assume ${{\mathbf{X}}}$ satisfies group restricted isometry condition, such that $$(1-\delta)n\|{\boldsymbol{\beta}}\|_2^2 \leq \|{{\mathbf{X}}}{\boldsymbol{\beta}}\|_2^2 \leq (1+\delta)n\|{\boldsymbol{\beta}}\|_2^2,\quad \forall {\boldsymbol{\beta}}~\text{such that}~ \sum_{i=1}^m 1_{\{{\boldsymbol{\beta}}_{G_i}\neq 0\}}\leq 2s.$$ Suppose ${{\mathbf{y}}}= {{\mathbf{X}}}{\boldsymbol{\beta}}+ \varepsilon$ and $\sum_{i=1}^p 1_{\{{\boldsymbol{\beta}}_{G_i}\neq 0\}}\leq s.$ Consider the following group Lasso estimator $$\label{eq:hat-beta-group-lasso}
\widehat{{\boldsymbol{\beta}}} = {\mathop{\rm arg\min}}_{{{\boldsymbol{\gamma}}}\in\mathbb{R}^{pr}}\left\{\frac{1}{2}\|{{\mathbf{y}}}- {{\mathbf{X}}}{{\boldsymbol{\gamma}}}\|_2^2 + \eta\sum_{i=1}^p \| {{\boldsymbol{\gamma}}}_{G_i}\|_2\right\}.$$ For $\eta \geq 3\max_{1\leq j\leq p}\|({{\mathbf{X}}}_{[:, G_j]})^\top \varepsilon\|_2$ and $\delta<2/7$, the optimal solution of yields $$\|\widehat{{\boldsymbol{\beta}}} - {\boldsymbol{\beta}}\|_2 \leq \frac{4\eta\sqrt{s/3}}{n(1-7\delta/2)}.$$
[**Proof of Lemma \[lm:group-oracle-RIP\].**]{} For convenience, define the $(2, \infty)$- and $(2, 1)$-norms of any vector $v\in \mathbb{R}^{pr}$ as $$\|{{\mathbf{v}}}\|_{2, \infty} = \max_{j=1,\ldots, p} \|{{\mathbf{v}}}_{G_j}\|_2\quad \text{and}\quad \|{{\mathbf{v}}}\|_{2, 1} = \sum_{j=1}^p \|{{\mathbf{v}}}_{G_j}\|_2.$$ Then, $\|\cdot\|_{2, \infty}$ and $\|\cdot \|_{2, 1}$ satisfies $\|{{\mathbf{v}}}\|_{2, \infty} \cdot \|w\|_{2, 1} \geq \langle {{\mathbf{v}}}, w\rangle$. We also define $J = \{j: {\boldsymbol{\beta}}_{G_j} \neq 0\}$ as the group support of ${\boldsymbol{\beta}}$, then $|J|\leq s$ based on the assumption. Suppose $h = \widehat{{\boldsymbol{\beta}}} - {\boldsymbol{\beta}}\in \mathbb{R}^{pr}$. By definition, $$\begin{split}
\frac{1}{2}\|{{\mathbf{y}}}- {{\mathbf{X}}}\widehat{{\boldsymbol{\beta}}}\|_2^2 + \eta \|\widehat{{\boldsymbol{\beta}}}\|_{2,1} \leq \frac{1}{2}\|{{\mathbf{y}}}-{{\mathbf{X}}}{\boldsymbol{\beta}}\|_2^2 + \eta\|{\boldsymbol{\beta}}\|_{2,1}.
\end{split}$$ Noting that $$\begin{split}
& \frac{1}{2}\left(\|{{\mathbf{y}}}- {{\mathbf{X}}}\widehat{{\boldsymbol{\beta}}}\|_2^2 - \|{{\mathbf{y}}}-{{\mathbf{X}}}{\boldsymbol{\beta}}\|_2^2\right) = \frac{1}{2}\left(\|\varepsilon-{{\mathbf{X}}}h\|_2^2 - \|{\boldsymbol{\varepsilon}}\|_2^2\right)\\
= & -\frac{1}{2} (2\varepsilon - {{\mathbf{X}}}h)^\top ({{\mathbf{X}}}h) \geq -\varepsilon^\top {{\mathbf{X}}}h \geq - \|{{\mathbf{X}}}^\top \varepsilon\|_{2, \infty} \cdot \|h\|_{2, 1} \\
= & - \|{{\mathbf{X}}}^\top \varepsilon\|_{2, \infty}(\|h_{J}\|_{2, 1} +\|h_{J^c}\|_{2, 1}),
\end{split}$$ $$\begin{split}
\eta\left(\|{\boldsymbol{\beta}}\|_{2,1} - \|\widehat{{\boldsymbol{\beta}}}\|_{2,1}\right) = \eta\left(\|{\boldsymbol{\beta}}_J\|_{2, 1} - \|\widehat{{\boldsymbol{\beta}}}_J\|_{2,1} - \|\widehat{{\boldsymbol{\beta}}}_{J^c}\|_{2,1}\right) \leq \eta\left(\|h_{J}\|_{2,1} - \|h_{J^c}\|_{2,1}\right),
\end{split}$$ we have $$\begin{split}
& -\|{{\mathbf{X}}}^\top \varepsilon\|_{2, \infty}(\|h_{J}\|_{2,1} + \|h_{J^c}\|_{2,1}) \leq \eta(\|h_{J}\|_{2,1} - \|h_{J^c}\|_{2,1}),\\
\Rightarrow \quad & \|h_{J^c}\|_{2,1} \leq \frac{\eta + \|{{\mathbf{X}}}^\top\varepsilon\|_{2,\infty}}{\eta - \|{{\mathbf{X}}}^\top \varepsilon\|_{2,\infty}}\|h_{J}\|_{2,1}.
\end{split}$$ Given $\eta\geq 3\|{{\mathbf{X}}}^\top \varepsilon \|_{2, \infty}$, we have $$\label{ineq:h_J^c<=h_J}
\|h_{J^c}\|_{2,1} \leq 2\|h_{J}\|_{2,1}.$$
Now we can sort all groups of $h$ by their $\ell_2$ norm and suppose $\|h_{G_{i_1}}\|_2\geq \cdots \geq \|h_{G_{i_p}}\|_2$, where $\{i_1,\ldots, i_p\}$ as a permutation of $\{1,\ldots, p\}$. Let $$h_{\max(s)} \in \mathbb{R}^{pr}, \quad (h_{\max(s)})_j = \left\{\begin{array}{ll}
h_j, & j\in G_{i_1}\cup \cdots \cup G_{i_s};\\
0, & \text{otherwise},
\end{array}\right.$$ Then $h_{\max(s)}$ is the vector $h$ with all but the $s$ largest groups in $\ell_2$ norm set to zero. We also denote $h_{-\max(s)} = h - h_{\max(s)}$. Then implies $$\label{ineq:h_max(s)<=h_-max(s)}
\|h_{-\max(s)}\|_{2,1} \leq \|h_{J^c}\|_{2,1} \leq 2\|h_{J}\|_{2,1} \leq 2\|h_{\max(s)}\|_{2,1}.$$
Let ${{\mathbf{v}}}\in \mathbb{R}^p$ with ${{\mathbf{v}}}_i = \|h_{G_i}\|_2, 1\leq i \leq p$ be the $\ell_2$ norms of each group of $h$. We can similarly define ${{\mathbf{v}}}_{\max(s)}$ as the vector ${{\mathbf{v}}}$ with all but the $s$ largest entries set to zero, and ${{\mathbf{v}}}_{-\max(s)} = {{\mathbf{v}}}- {{\mathbf{v}}}_{\max(s)}$. Then, $({{\mathbf{v}}}_{\max(s)})_i = \|(h_{\max(s)})_{G_i}\|_2$ and $({{\mathbf{v}}}_{-\max(s)})_i = \|(h_{-\max(s)})_{G_i}\|_2$. Let $$\alpha = \max\{\|h_{-\max(s)}\|_{2, \infty}, \|h_{-\max(s)}\|_{2, 1}/s\} = \max\{\|{{\mathbf{v}}}_{-\max(s)}\|_{\infty}, \|{{\mathbf{v}}}_{-\max(s)}\|_{1}/s\}.$$ By the polytope representation lemma (Lemma 1 in [@cai2014sparse]) with $\alpha$, one can find a finite series of vectors ${{\mathbf{v}}}^{(1)}, \cdots, {{\mathbf{v}}}^{(N)} \in \mathbb{R}^{p}$ and weights $\pi_1, \ldots, \pi_N$ such that $${{\rm supp}}({{\mathbf{v}}}^{(j)}) \subseteq {{\rm supp}}({{\mathbf{v}}}_{-\max(s)}), \quad \|{{\mathbf{v}}}^{(j)}\|_0 \leq s, \quad \|{{\mathbf{v}}}^{(j)}\|_\infty \leq \alpha, \quad \|{{\mathbf{v}}}^{(j)}\|_1 = \|{{\mathbf{v}}}_{-\max(s)}\|_1,$$ $${{\mathbf{v}}}_{-\max(s)} = \sum_{j=1}^N \pi_j {{\mathbf{v}}}^{(j)}, \quad 0\leq \pi_j \leq 1,\quad \text{and}\quad \sum_{j=1}^N \pi_j = 1.$$ Now we construct $$h^{(j)}\in \mathbb{R}^{pr}, \quad \text{where}\quad (h^{(j)})_{G_i} = \frac{(h_{-\max(s)})_{G_i}}{\|(h_{-\max(s)})_{G_i}\|_2}\cdot {{\mathbf{v}}}^{(j)}_i, \quad i=1,\ldots, p; j=1,\ldots, N.$$ Then $\{h^{(j)}\}_{j=1}^N$ satisfy $$\label{eq:h^(j)}
\begin{split}
& {{\rm supp}}(h^{(j)}) \subseteq {{\rm supp}}(h_{-\max(s)}), \quad \sum_{i=1}^p 1_{\{(h^{(j)})_{G_i}\neq 0\}} \leq s, \quad \|h^{(j)}\|_{2,\infty} \leq \alpha, \\
\quad & \|h^{(j)}\|_{2, 1} = \|h_{-\max(s)}\|_{2,1},\quad h_{-\max(s)} = \sum_{j=1}^N \pi_j h^{(j)}, \quad 0\leq \pi_j \leq 1,\quad \sum_{j=1}^N \pi_j = 1.
\end{split}$$ Therefore, $h_{\max(s)}$ and $h^{(j)}$ have distinct supports, $\sum_{i=1}^m 1_{(h_{\max(s)} + h^{(j)})_{G_i}\neq 0}\leq 2s$, $\|h_{\max(s)} + h^{(j)}\|_2^2 = \|h_{\max(s)}\|_2^2 + \|h^{(j)}\|_2^2$, and $$\begin{split}
\|h^{(j)}\|_2^2 \leq & \|h^{(j)}\|_{2,1}\cdot \|h^{(j)}\|_{2,\infty} \overset{\eqref{eq:h^(j)}}{\leq} \|h_{-\max(s)}\|_{2,1}\cdot \alpha \\
\overset{\eqref{ineq:h_max(s)<=h_-max(s)}}{\leq} & 2\|h_{\max(s)}\|_{2,1}\cdot \max\left\{\|h_{-\max(s)}\|_{2,\infty}, \|h_{-\max(s)}\|_{2,1}/s\right\}\\
\leq & 2\|h_{\max(s)}\|_{2,1}\cdot \max\left\{\min_{j: \|h_{G_j}\|_2\neq 0}\|h_{G_j}\|_2, 2\|h_{\max(s)}\|_{2,1}/s\right\}\\
\leq & 4\|h_{\max(s)}\|_{2,1}^2/s \leq 4\|h_{\max(s)}\|_2^2.
\end{split}$$ Thus, $$\begin{split}
& \left|\langle {{\mathbf{X}}}h_{\max(s)}, {{\mathbf{X}}}h_{-\max(s)}\rangle\right| \leq \sum_{j=1}^N \pi_j \left|\langle {{\mathbf{X}}}h_{\max(s)}, {{\mathbf{X}}}h^{(j)}\rangle\right|\\
= & \sum_{j=1}^N \frac{\pi_j}{4} \left|\|{{\mathbf{X}}}h_{\max(s)}+ {{\mathbf{X}}}h^{(j)}\|_2^2 - \|{{\mathbf{X}}}h_{\max(s)}- {{\mathbf{X}}}h^{(j)}\|_2^2\right|\\
\leq & \sum_{j=1}^N \frac{\pi_j}{4}\left(n(1+\delta)(\|h_{\max(s)}\|_2^2 + \|h^{(j)}\|_2^2) - n(1-\delta)(\|h_{\max(s)}\|_2^2 + \|h^{(j)}\|_2^2)\right)\\
\leq & \frac{\delta n}{2} \left(\|h_{\max(s)}\|_{2}^2 + 4\|h_{\max(s)}\|_{2}^2\right) = \frac{5\delta n}{2}\|h_{\max(s)}\|_2^2,
\end{split}$$ which means $$\label{ineq:inner-product}
\begin{split}
& \langle {{\mathbf{X}}}h_{\max(s)}, {{\mathbf{X}}}h\rangle =\|{{\mathbf{X}}}h_{\max(s)}\|_2^2 + \langle {{\mathbf{X}}}h_{\max(s)}, {{\mathbf{X}}}h_{-\max(s)}\rangle \\
\geq & n(1-\delta)\|h_{\max(s)}\|_2^2 - \frac{5\delta n}{2}\|h_{\max(s)}\|_2^2 = n(1-7\delta/2)\|h_{\max(s)}\|_2^2.
\end{split}$$
Next, by the KKT condition of $\widehat{{\boldsymbol{\beta}}}$ being the optimizer of , $$\|{{\mathbf{X}}}^\top(y - {{\mathbf{X}}}\widehat{{\boldsymbol{\beta}}})\|_{2, \infty} \leq \eta.$$ In addition, $\|{{\mathbf{X}}}^\top (y - {{\mathbf{X}}}{\boldsymbol{\beta}})\|_{2, \infty} = \|{{\mathbf{X}}}^\top \varepsilon\|_{2, \infty} \leq \eta/3$, which means $$\begin{split}
\langle {{\mathbf{X}}}h_{\max(s)}, {{\mathbf{X}}}h\rangle = & h_{\max(s)}^\top {{\mathbf{X}}}^\top {{\mathbf{X}}}h \leq \|h_{\max(s)}\|_{2,1}\cdot \|{{\mathbf{X}}}^\top {{\mathbf{X}}}h\|_{2,\infty} \\
\leq & \|h_{\max(s)}\|_{2,1}\cdot \left(\|{{\mathbf{X}}}^\top (y - {{\mathbf{X}}}\widehat{{\boldsymbol{\beta}}})\|_{2,\infty} + \|{{\mathbf{X}}}^\top (y - {{\mathbf{X}}}{\boldsymbol{\beta}})\|_{2,\infty}\right) \\
\leq & 4\eta/3\cdot \|h_{\max(s)}\|_{2,1} \leq 4\eta/3 \cdot \sqrt{s}\|h_{\max(s)}\|_2.
\end{split}$$ Combining the above inequality with , one has $$\frac{4\eta}{3}\sqrt{s}\|h_{\max(s)}\|_{2} \geq n(1-7\delta/2)\|h_{\max(s)}\|_2^2,$$ namely $$\|h_{\max(s)}\|_2 \leq \frac{\frac{4}{3}\eta\sqrt{s}}{n(1-7\delta/2)}.$$ Finally, $$\begin{split}
\|h_{-\max(s)}\|_2^2 \leq & \|h_{-\max(s)}\|_{2,1} \cdot \|h_{-\max(s)}\|_{2, \infty} \\
\leq & 2\|h_{\max(s)}\|_{2, 1}\cdot \min_{j: (h_{\max(s)})_{G_j}\neq 0} \|(h_{\max(s)})_{G_j}\|_2 \\
\leq & 2\|h_{\max(s)}\|_2^2.
\end{split}$$ Therefore, $$\|h\|_2 = \sqrt{\|h_{-\max(s)}\|_2^2 + \|h_{\max(s)}\|_2^2} \leq \sqrt{3}\|h_{\max(s)}\|_2 \leq \frac{4\eta\sqrt{s/3}}{n(1-7\delta/2)},$$ which has finished the proof of this lemma.$\square$
The next Lemma \[lm:GE->GRIP\] shows that the Gaussian Ensemble satisfies group restricted isometry property with high probability.
\[lm:GE->GRIP\] Suppose ${{\mathbf{X}}}\in \mathbb{R}^{n\times (pr)}$, $G_1, \ldots, G_p$ is a partition of $\{1,\ldots pr\}$ and $|G_1|=\cdots |G_p| = r$. If ${{\mathbf{X}}}\overset{iid}{\sim}N(0, 1)$ and $n \geq C(sr/\delta + s\log(ep/s))$ for large constant $C>0$, ${{\mathbf{X}}}$ satisfies the following group restricted isometry (GRIP) $$\label{ineq:lm-GRIP}
n(1 - \delta)\|{\boldsymbol{\beta}}\|_2^2\leq \left\|{{\mathbf{X}}}{\boldsymbol{\beta}}\right\|_2^2 \leq n(1 + \delta)\|{\boldsymbol{\beta}}\|_2^2, \quad \forall {\boldsymbol{\beta}}\text{ such that } \sum_{i=1}^p 1_{\{{\boldsymbol{\beta}}_{{{\mathbf{G}}}_i} \neq 0\}} \leq s$$ with probability at least $1 - \exp(-cn)$.
[**Proof of Lemma \[lm:GE->GRIP\].**]{} First, the statement is equivalently to $$\label{ineq:GRIP-equivalent}
\begin{split}
& \forall \text{ distinct } i_1,\ldots, i_s \subseteq \{1,\ldots, p\}, \\
& n(1-\delta) \leq \sigma_{\min}^2({{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}) \leq \sigma_{\max}^2({{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}) \leq n(1+\delta).
\end{split}$$ Since ${{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}$ is an $n$-by-$sr$ matrix with i.i.d. Gaussian entries, by random matrix theory (c.f., [@vershynin2010introduction Corollary 5.35]), $$\begin{split}
& {\mathbb{P}}\left(\sqrt{n}-\sqrt{sr}-x \leq \sigma_{\min}({{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}) \leq \sigma_{\max}({{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}) \leq \sqrt{n}+\sqrt{sr}+x\right) \\
\geq & 1 - 2\exp(-x^2/2),
\end{split}$$ which means $$\begin{split}
& {\mathbb{P}}\left(\eqref{ineq:GRIP-equivalent} \text{ does not hold}\right)\\
\leq & \sum_{\substack{\text{distinct } i_1,\ldots, i_s\\
\subseteq \{1,\ldots, p\}}}{\mathbb{P}}\left(\left\{n(1-\delta) \leq \sigma_{\min}^2({{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}) \leq \sigma_{\max}^2({{\mathbf{X}}}_{[:, G_{i_1}\cup \cdots \cup G_{i_s}]}) \leq n(1+\delta)\right\}^c\right) \\
\leq & 2\binom{p}{s}\exp\left(-\left(\sqrt{n}-\sqrt{n(1-\delta)}-\sqrt{sr}\right)_+^2 \wedge\left(\sqrt{n(1+\delta)}-\sqrt{n}-\sqrt{sr}\right)_+^2 \right),
\end{split}$$
Provided that $n\geq C(sr/\delta+s\log(ep/s))$ for large constant $C>0$, we have $$\left(\sqrt{n}-\sqrt{n(1-\delta)}-\sqrt{sr}\right)_+^2 \wedge \left(\sqrt{n(1+\delta)} - \sqrt{n}-\sqrt{sr}\right)_+^2 \geq (1-c)n,$$ $$\begin{split}
(1-c)n\geq (1-c)Cs\log(ep/s) \geq (1-c)C\log\left(\binom{p}{s}\right).
\end{split}$$ Therefore, we have $${\mathbb{P}}\left(\eqref{ineq:GRIP-equivalent} \text{ does not hold}\right) \leq \exp\left(\log\left(2\binom{p}{s}\right)-(1-c)n\right) \leq \exp(-cn)$$ and have finished the proof of this lemma.$\square$
The next lemma gives the Kullback–Leibler divergence between two regression models with random designs, which will be used in the lower bound argument in this paper.
\[lm:regression-KL\] Consider two linear regression models ${{\mathbf{y}}}^{(1)} = {{\mathbf{X}}}{\boldsymbol{\beta}}^{(1)} + \varepsilon$ and $y^{(2)} = {{\mathbf{X}}}{\boldsymbol{\beta}}^{(2)} + \varepsilon$. Here, ${{\mathbf{y}}}^{(1)}, y^{(2)}\in \mathbb{R}^{n}$ and ${{\mathbf{X}}}\in \mathbb{R}^{n\times p}$, ${\boldsymbol{\beta}}^{(1)}, {\boldsymbol{\beta}}^{(2)}\in \mathbb{R}^p$, and $\varepsilon\in \mathbb{R}^n$. Assume ${{\mathbf{X}}}\overset{iid}{\sim}N(0, 1)$, $\varepsilon\overset{iid}{\sim}N(0, \sigma^2)$, and ${\boldsymbol{\beta}}^{(1)}, {\boldsymbol{\beta}}^{(2)}$ are fixed. Then, $$D_{KL}\left(\{{{\mathbf{X}}}, {{\mathbf{y}}}^{(1)}\} \Big|\Big|\{{{\mathbf{X}}}, y^{(2)}\} \right) = \frac{n}{2\sigma^2}\left\|{\boldsymbol{\beta}}^{(1)} - {\boldsymbol{\beta}}^{(2)}\right\|_2^2.$$
[**Proof of Lemma \[lm:regression-KL\].**]{} Denote the $j$-th row vector of ${{\mathbf{X}}}$ as $x_j$, i.e., ${{\mathbf{X}}}= [x_1^\top \cdots x_n^\top]^\top$. Then, $(x_1^\top, y_1^{(1)\top}), \ldots, (x_n^\top, y_n^{(1)\top})$ are i.i.d. distributed vectors, $y_j^{(1)} = x_j^\top {\boldsymbol{\beta}}^{(1)}+\varepsilon_j$, and $$\left(x_j^\top, y_j^{(1)}\right) \sim N\left(0, \Sigma_1\right), \quad \Sigma_1 = \begin{bmatrix}
{{\mathbf{I}}}_p & {\boldsymbol{\beta}}^{(1)}\\
{\boldsymbol{\beta}}^{(1)\top} & \|{\boldsymbol{\beta}}^{(1)}\|_2^2 + \sigma^2
\end{bmatrix}.$$ Similarly, $$\left(x_j^\top, y_j^{(2)}\right) \sim N(0, \Sigma_2), \quad \Sigma_2 = \begin{bmatrix}
{{\mathbf{I}}}_p & {\boldsymbol{\beta}}^{(2)}\\
{\boldsymbol{\beta}}^{(2)\top} & \|{\boldsymbol{\beta}}^{(2)}\|_2^2 + \sigma^2
\end{bmatrix}.$$ Additionally, $$\det(\Sigma_i) = \det\left(\begin{bmatrix}
{{\mathbf{I}}}_p & 0\\
-{\boldsymbol{\beta}}^{(i)\top} & 1
\end{bmatrix}\cdot \begin{bmatrix}
{{\mathbf{I}}}_p & {\boldsymbol{\beta}}^{(i)}\\
{\boldsymbol{\beta}}^{(i)\top} & \|{\boldsymbol{\beta}}^{(i)}\|_2^2 + \sigma^2
\end{bmatrix}\right) = \det\left(\begin{bmatrix}
{{\mathbf{I}}}_p & {\boldsymbol{\beta}}^{(i)} \\
0 & \sigma^2
\end{bmatrix}\right) = \sigma^2,\quad i=1,2,$$ $$\Sigma_i^{-1} = \begin{bmatrix}
{{\mathbf{I}}}_p + {\boldsymbol{\beta}}^{(i)}{\boldsymbol{\beta}}^{(i)\top} \sigma^{-2} & -{\boldsymbol{\beta}}^{(i)}\sigma^{-2}\\
-{\boldsymbol{\beta}}^{(i)\top} \sigma^{-2} & \sigma^{-2}
\end{bmatrix}, \quad i=1,2.$$ By the formula for multivariate normal distribution KL-divergence, $$\begin{split}
& D_{KL}\left(\left\{x_j^\top, y_j^{(1)}\right\}\Big|\Big| \left\{x_j^\top, y_j^{(2)}\right\}\right) \\
= & \frac{1}{2}\left({{\rm tr}}\left(\Sigma_2^{-1}\Sigma_1\right) - (p+1) + \log\left(\frac{\det(\Sigma_1)}{\det(\Sigma_2)}\right) \right)\\
= & \frac{\sigma^{-2}}{2}\left({{\rm tr}}\left({\boldsymbol{\beta}}^{(2)}{\boldsymbol{\beta}}^{(2)\top} - {\boldsymbol{\beta}}^{(2)}{\boldsymbol{\beta}}^{(1)^\top} - {\boldsymbol{\beta}}^{(1)\top}{\boldsymbol{\beta}}^{(2)}\right)+\|{\boldsymbol{\beta}}^{(1)}\|_2^2\right)\\
= & \frac{1}{2\sigma^2}\left\|{\boldsymbol{\beta}}^{(1)} - {\boldsymbol{\beta}}^{(2)}\right\|_2^2.
\end{split}$$ Therefore, $$\begin{split}
D_{KL}\left(\left\{x_j^\top, y_j^{(1)}\right\}_{j=1}^n\Big|\Big| \left\{x_j^\top, y_j^{(2)}\right\}_{j=1}^n\right) = & nD_{KL}\left(\left\{x_j^\top, y_j^{(1)}\right\}\Big|\Big| \left\{x_j^\top, y_j^{(2)}\right\}\right) \\
= & \frac{n}{2\sigma^2}\left\|{\boldsymbol{\beta}}^{(1)} - {\boldsymbol{\beta}}^{(2)}\right\|_2^2.
\end{split}$$ $\square$
The next lemma can be seen as a sparse version of Varshamov-Gilbert bound [@massart2007concentration Lemma 4.7]. This result is crucial in the proof of the lower bound argument in sparse tensor regression (Theorem \[th:lower\_bound\_sparse\_tensor\_regression\]).
\[lm:sparse-Varshamov-Gilbert\] There exists a series of matrices ${{\mathbf{A}}}^{(1)},\ldots, {{\mathbf{A}}}^{(N)} \in \{1, 0, -1\}^{p\times r}$, such that $$\label{eq:targeting-condition}
\|{{\mathbf{A}}}^{(k)}\|_{0, 2} := \sum_{i=1}^p1_{\left\{{{\mathbf{A}}}^{(k)}_{[i,:]}\neq 0\right\}} \leq s,\quad \|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\|_{1, 1} = \sum_{i=1}^p\sum_{j=1}^r \left|{{\mathbf{A}}}^{(k)}_{[i,j]} - {{\mathbf{A}}}^{(l)}_{[i,j]}\right| > sr/2$$ for all $k, l$, and $N\geq \exp\left(c(sr+s\log(ep/s))\right)$ for some uniform constant $c>0$.
[**Proof of Lemma \[lm:sparse-Varshamov-Gilbert\]**]{} First, if $p/s \leq C$ for some constant $C>0$, the lemma directly follows from the Varshamov-Gilbert bound by restricting on the top $s\times r$ submatrices of ${{\mathbf{A}}}_1,\ldots, {{\mathbf{A}}}_N$. Thus, without loss of generality, we assume $p\geq 10s$ throughout the rest of the proof.
Next for $k=1,\ldots, N$, we randomly draw $s$ elements from $\{1,\ldots, p\}$ without replacement, form $\Omega^{(k)}$ as a random subset of $\{1,\ldots, p\}$, and generate $${{\mathbf{A}}}^{(k)}\in \mathbb{R}^{p\times r}, \quad \left({{\mathbf{A}}}^{(k)}\right)_{ij} \left\{\begin{array}{ll}
\sim \text{Rademacher}, & i \in \Omega^{(k)};\\
=0, & i \notin \Omega^{(k)},
\end{array}\right.$$ for $k=1,2,\ldots, N$. Here, $A\sim$ Rademacher if $A$ is equally distributed on -1 and 1. By such the construction, $$\|{{\mathbf{A}}}^{(k)}\|_{0, 2} = \sum_{i=1}^p1_{\left\{{{\mathbf{A}}}^{(k)}_{[i,:]}\neq 0\right\}} \leq s.$$ For any $k\neq l$, $$\label{eq:A^{(k)}-A^{(l)}}
\begin{split}
\left\|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\right\|_{1,1} \sim & r|\Omega^{(k)}\backslash \Omega^{(l)}| + r|\Omega^{(l)}\backslash \Omega^{(k)}| + 2\cdot \text{Bin}\left(r\left|\Omega^{(k)}\cap \Omega^{(l)}\right|, 1/2\right)\\
= & 2sr - 2r|\Omega^{(l)}\cap \Omega^{(k)}| - 2\cdot\text{Bin}\left(r|\Omega^{(l)}\cap \Omega^{(k)}|, 1/2\right)\\
\sim & 2sr - 2\cdot\text{Bin}\left(r|\Omega^{(l)}\cap \Omega^{(k)}|, 1/2\right).
\end{split}$$ Here, we used the fact that $|\Omega^{(k)}\backslash \Omega^{(l)}| = |\Omega^{(k)}| - |\Omega^{(k)}\cap \Omega^{(l)}| = s - |\Omega^{(k)}\cap \Omega^{(l)}|$. Moreover, $|\Omega^{(l)}\cap \Omega^{(k)}|$ satisfies the following hyper-geometric distribution: $${\mathbb{P}}\left(\left|\Omega^{(l)}\cap \Omega^{(k)}\right| = t\right) = \frac{\binom{s}{t}{\binom{p-s}{s-t}}}{\binom{p}{s}}, \quad t = 0,\ldots, s.$$ Let $Z_{kl} = \left|\Omega^{(l)}\cap \Omega^{(k)}\right|$. Then for any $s/2 \leq t\leq s$, $$\label{ineq:P(Z=t)}
\begin{split}
{\mathbb{P}}\left(Z = t\right) = & \frac{\frac{s\cdots (s-t+1)}{t!}\cdot \frac{(p-s)\cdots (p-2s+t+1)}{(s-t)!}}{\frac{p\cdots (p-s+1)}{s!}} \leq \binom{s}{t} \cdot \left(\frac{s}{p-s+1}\right)^t\\
\leq & 2^s\left(\frac{s}{p-s+1}\right)^t \leq \left(\frac{4s}{p-s+1}\right)^t.
\end{split}$$ Next, by Bernstein’s inequality, $$\begin{split}
& {\mathbb{P}}\left(\left\|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\right\|_{1,1} \leq sr/2\Big| Z\right) \overset{\eqref{eq:A^{(k)}-A^{(l)}}}{=} {\mathbb{P}}\left(\text{Bin}\left(rZ, 1/2\right) \geq 3sr/4\Big|Z\right)\\
= & {\mathbb{P}}\left(2\text{Bin}(rZ, 1/2) - rZ \geq \frac{3sr}{2} - rZ\right) \\
\leq & \left\{\begin{array}{ll}
2\exp\left(-\frac{(3sr/2-Zr)^2}{rZ + (3sr/2 - Zr)/3}\right), & s/2 \leq Z\leq s;\\
0, & Z< s/2.
\end{array}\right.
\end{split}$$ Thus, $$\begin{split}
{\mathbb{P}}\left(\left\|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\right\|_{1,1} \leq sr/2 \right) \leq & \sum_{s/2\leq t \leq s} {\mathbb{P}}\left(\left\|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\right\|_{1,1} \leq sr/2\Big| Z=t\right)\cdot {\mathbb{P}}\left(Z=t\right)\\
\leq & \sum_{s/2\leq t \leq s} 2\exp\left(-\frac{(3sr/2-tr)^2}{rt + (3sr/2 - tr)/3}\right) \left(\frac{4s}{p-s+1}\right)^t\\
\leq & \sum_{s/2\leq t \leq s} 2\exp\left(-\frac{(3sr/2-sr)^2}{sr + (3sr/2 - sr)/3}\right) \left(\frac{4s}{p-s+1}\right)^t\\
\leq & \sum_{t\geq s/2} 2\exp\left(-sr/14\right) \cdot \left(4s/(p-r+1)\right)^t\\
\leq & 2\exp(-sr/14) 2\cdot \left(4s/(p-s+1)\right)^{s/2}\\
\leq & 4\exp\left(-c(sr + s\log(ep/s))\right)
\end{split}$$ for some uniform constant $c>0$. Finally, $$\begin{split}
& {\mathbb{P}}\left(\forall 1\leq k\neq l\leq N, \left\|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\right\|_{1,1} > sr/2\right)\\
\geq & 1 - \binom{N}{2}{\mathbb{P}}\left(\left\|{{\mathbf{A}}}^{(k)} - {{\mathbf{A}}}^{(l)}\right\|_{1,1} \leq sr/2 \right) \geq 1 - \frac{N^2}{2} \cdot 4\exp\left(-c(sr + s\log(ep/s))\right)
\end{split}$$ We can see if $N \leq \exp(c(sr+s\log(ep/s)))$ for some uniform constant $c>0$, the previous event happens with a positive probability, which means there exists fixed ${{\mathbf{A}}}^{(1)},\ldots, {{\mathbf{A}}}^{(N)}$ satisfying the targeting condition for some $N \geq \exp(c(sr+s\log(p/s)))$.$\square$
[^1]: Software package downloaded at <https://hua-zhou.github.io/TensorReg/>
[^2]: The convex regularization aims to minimize the following objective function $$\sum_{i}^n \frac{1}{2n} (y_i - \langle \boldsymbol{\mathcal{X}}_i, \boldsymbol{\mathcal{A}} \rangle )^2 + \lambda \sum_{k = 1}^3 ||\mathcal{M}_{k}(\boldsymbol{\mathcal{A}})||_*.$$ Here, $\|\cdot\|_\ast$ is the matrix nuclear norm.
[^3]: Available online at: <https://cran.r-project.org/web/packages/gglasso/index.html>.
[^4]: The optimization of NNM is implemented by accelerated proximal gradient method [@toh2010accelerated] using the software package available online at <http://www.math.nus.edu.sg/~mattohkc/NNLS.html>.
[^5]: Link: <http://neurobureau.projects.nitrc.org/ADHD200/Data.html>
[^6]: See <https://en.wikipedia.org/wiki/Inverse-Wishart_distribution> for expectation of inverse Wishart distribution.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We investigate the detectability thresholds of various modular structures in the stochastic block model. Our analysis reveals how the detectability threshold is related to the details of the modular pattern, including the hierarchy of the clusters. We show that certain planted structures are impossible to infer regardless of their fuzziness.'
author:
- Tatsuro Kawamoto
- Yoshiyuki Kabashima
bibliography:
- 'bib-generalSBM.bib'
title: '**Detectability thresholds of general modular graphs**'
---
Introduction {#Introduction}
============
Motivated by needs in data-driven science, a number of frameworks and algorithms for modular structure detection have been proposed in several fields in the last few decades [@GirvanNewman2002; @ShiMalik2000; @Leskovec2009; @Fortunato201075; @Leger2013]. Correspondingly, theoretical and experimental analyses of statistical significance of results are thus the subject of significant research interest. For example, although an algorithm suggests the partition of a graph following the application of some optimization process, if the graph is a typical instance of a uniform random graph, it is doubtful whether the effected partition contains any useful information in practice. Moreover, even when the graph is generated from a model with some planted structure, it may be indistinguishable from a uniform random graph if the planted structure is too fuzzy.
It is a challenging problem in general, and the basic strategy to solve it involves investigating the conditions whereby we can retrieve the planted structure for a specified random graph ensemble. To this end, the so-called stochastic block model [@holland1983stochastic], which we explain in detail below, is often considered. This random graph model has controllable noise strength $\epsilon$, i.e., $\epsilon = 0$ represents a graph that clearly realizes the planted structure, and $\epsilon = 1$ represents a uniform random graph. Above a certain critical value $\epsilon^{\ast}$, an algorithm cannot retrieve the planted structure better than chance. This critical value is called the *detectability threshold*, and a large number of studies have been devoted to it [@Reichardt2008; @Nadakuditi2012; @Krzakala2013; @KawamotoKabashimaPRE2015; @KawamotoKabashimaEPL2015; @Decelle2011; @Decelle2011a; @Radicchi2013; @Radicchi2014; @Hu2012; @Ronhovde2012; @VerSteeg2014; @ZhangPRE2014; @Ghasemian2016] for sparse graphs, including rigorous treatments [@Mossel2014; @Massoulie2014; @banks2016information]. Besides the distinguishability from a uniform random graph, the exact recovery in dense graphs has also been studied [@Condon2001; @Onsjo2006; @BickelChenPNAS2009; @Rohe2011; @yun2014community; @AbbeFOCS2015; @AbbeNIPS2015].
Nevertheless, a large portion [@FootnoteAbbe] of the research focuses on the community structure (assortative structure) and the disassortative structure. In this paper, we investigate the detectability threshold of more general structures. We show that according to the linear stability analysis of *belief propagation* (BP), the detectability threshold varies depending on the details of the modular structure.
Stochastic block model
======================
The stochastic block model is a random graph model with a planted modular structure: the graph of $N$ vertices consists of $q$ clusters, each of which of size $\gamma_{\sigma}N$ ($\sigma \in \{1, \dots, q\}$), and every pair of vertices is connected independently and randomly according to its cluster assignments. For example, if vertices $i$ and $j$ belong to clusters $\sigma$ and $\sigma^{\prime}$, respectively, they are connected with probability $\omega_{\sigma\sigma^{\prime}}$ ($\sigma, \sigma^{\prime} \in \{1, \dots, q\}$); matrix $\ket{\omega}$ is called the affinity matrix. For given $N$, $q$, $\ket{\gamma}$, and $\ket{\omega}$, we can generate random graph instances of the stochastic block model. In the case of the inverse problem, which is of interest to us in this paper, our goal is to infer the parameters $\ket{\gamma}$ and $\ket{\omega}$ as well as cluster assignments $\ket{\sigma}$ given a graph. The number of clusters $q$ is sometimes given as input; otherwise, it is determined by some model selection criterion. Throughout this paper, we treat $q$ as input and focus on sparse graphs, i.e., each element of $\ket{\omega}$ is scaled as $O(1/N)$ so that the average degree does not diverge as $N \to \infty$.
While there exist many types of modular structures, the simplest and most studied case is the community structure as illustrated in Fig. \[FigAffinityMatrices\](a); that is, the affinity matrix has large values for its diagonal elements, $\omega_{\sigma\sigma} = \omega_{\mathrm{in}}$, and small values for the remaining elements, $\omega_{\sigma\sigma^{\prime}} = \omega_{\mathrm{out}}$ ($\sigma \ne \sigma^{\prime}$). Although the elements of the affinity matrix can be arbitrary nonnegative numbers, we hereafter consider the case where they are either $\omega_{\mathrm{in}}$ or $\omega_{\mathrm{out}}$: that is, $$\begin{aligned}
\ket{\omega} = (\omega_{\mathrm{in}} - \omega_{\mathrm{out}}) W + \omega_{\mathrm{out}} \ket{1}\bra{1}, \label{AffinityMatrix}\end{aligned}$$ where $W$ is an indicator matrix, where $W_{\sigma \sigma^{\prime}} = 1$ represents a densely connected cluster pair (which we refer to as a bicluster), $W_{\sigma \sigma^{\prime}} = 0$ represents a sparsely connected bicluster, and $\ket{1}$ is the column vector with all elements equal to unity. This random graph ensemble can be regarded as a restricted version of the stochastic block model, or a generalized version of the planted partition model [@Condon2001].
This affinity matrix contains the above community structure as a special case, and can express arbitrary modular patterns. Note that the indicator matrix $W$ can be regarded as a cluster-wise adjacency matrix, i.e., each planted cluster represents a coarse-grained vertex and a densely connected bicluster represents a bundled edge (a densely connected cluster constitutes a self-loop). We refer to the graph with adjacency matrix equal to $W$ as a *module graph*. Note that some matrices represent the equivalent modular pattern; for example, Figs. \[FigAffinityMatrices\](c) and \[FigAffinityMatrices\](d) differ only by permutation. The average degree $c$ of this stochastic block model is $c = N \bra{\gamma} \ket{\omega} \ket{\gamma}$. By defining the strength of the modular structure by $\epsilon \equiv \omega_{\mathrm{out}}/\omega_{\mathrm{in}}$, we can express elements $\omega_{\mathrm{in}}$ and $\omega_{\mathrm{out}}$ as $$\begin{aligned}
&\omega_{\mathrm{in}} = \frac{c}{N} \left[ (1-\epsilon) \bra{\gamma} W \ket{\gamma} + \epsilon \right]^{-1},
&\omega_{\mathrm{out}} = \epsilon \, \omega_{\mathrm{in}}. \end{aligned}$$
Bayesian inference of the stochastic block model {#SBMinference}
================================================
We now consider the Bayesian inference of the modular structure using the stochastic block model. The prior probability $p(\ket{\sigma} \lvert \ket{\gamma})$ of cluster assignments is represented by a multinomial distribution of each planted cluster of fraction $\gamma_{\sigma}$, and the probability of independent and random connections between vertex pairs is represented by the product of Bernoulli distributions. Thus, the likelihood of the stochastic block model is $$\begin{aligned}
p(A, \ket{\sigma} \lvert \ket{\omega}, \ket{\gamma},q)
&= p(A \lvert \ket{\sigma}, \ket{\omega}, \ket{\gamma}) p(\ket{\sigma} \lvert \ket{\gamma}) \notag\\
&= \prod_{i} \gamma_{\sigma_{i}} \prod_{i<j} \omega_{\sigma_{i}\sigma_{j}}^{A_{ij}}\left( 1 - \omega_{\sigma_{i}\sigma_{j}} \right)^{1-A_{ij}}. \end{aligned}$$ Using the affinity matrix of (\[AffinityMatrix\]), its log-likelihood reads as $$\begin{aligned}
& \log p(A, \ket{\sigma} \lvert \ket{\omega}, \ket{\gamma},q)
= \sum_{i} \log \gamma_{\sigma_{i}} \notag\\
&\hspace{5pt} + \sum_{i<j} W_{\sigma_{i} \sigma_{j}} \left( A_{ij}\log\omega_{\mathrm{in}} + (1-A_{ij})\log(1-\omega_{\mathrm{in}}) \right) \notag\\
&\hspace{5pt} + \sum_{i<j} \left( 1 - W_{\sigma_{i} \sigma_{j}} \right) \left( A_{ij}\log\omega_{\mathrm{out}} + (1-A_{ij})\log(1-\omega_{\mathrm{out}}) \right). \label{SBMloglikelihood}\end{aligned}$$
Our task is to evaluate the marginal probability distributions of the cluster assignments of vertices and to determine the values of parameters ($\ket{\gamma}$ and $\ket{\omega}$), in order to maximize the marginal log-likelihood $$\begin{aligned}
\log \sum_{\ket{\sigma}} p(A, \ket{\sigma} \lvert \ket{\gamma}, \ket{\omega},q). \label{MarginalLogLikelihood}\end{aligned}$$ To this end, we employ the expectation-maximization (EM) algorithm, which does not maximize (\[MarginalLogLikelihood\]) directly, but repeats the maximization of its lower bound until convergence: In the E-step, the posterior distribution of cluster assignments $\ket{\sigma}$ is estimated according to the given parameter estimates $(\ket{\gamma}, \ket{\omega})$. In the M-step, $(\ket{\gamma}, \ket{\omega})$ are updated to maximize the average of (\[SBMloglikelihood\]) with respect to the posterior distribution determined in the E-step. While there are many other Bayesian inference methods [@Nowicki2001; @daudin08; @latouche12; @PeixotoPRE2014MonteCarlo], as we see below, the present method is suited for theoretical analysis.
Cluster inference and parameter learning
----------------------------------------
Let $\psi^{i}_{\sigma}$ be the marginal probability of cluster $\sigma$ for vertex $i$ calculated in the E-step ($\sum_{\sigma} \psi^{i}_{\sigma} = 1$), and $\ket{\psi}^{i}$ be its row vector. Unfortunately, the exact computation of $\ket{\psi}^{i}$ is demanding. To avoid this computational burden, we use BP [@Decelle2011a; @MezardMontanari2009], which is justified for sparse graphs. Using tree approximation, the marginal probability $\ket{\psi}^{i}$ can be estimated as $$\begin{aligned}
\ket{\psi}^{i}
&= \frac{1}{Z^{i}} \ket{\gamma} \circ
\prod_{k \in \partial i} \left[ \ket{1} + \overline{\omega}_{\mathrm{in}} \ket{\psi}^{k \to i} W \right] \circ
\exp \left[ -\overline{\omega}_{\mathrm{in}} \omega_{\mathrm{out}} \sum_{\ell} \ket{\psi}^{\ell} W \right], \label{CompleteMarginal}\end{aligned}$$ where $\ket{1}$ and $\ket{\psi}^{k \to i}$ are the $q$-dimensional unit row-vector and the marginal probability for vertex $k$ without the contribution from edge $(k, i)$, respectively. The latter is often referred to as the cavity bias. $\circ$ and $\partial i$ represent the element-wise product (Hadamard product) and the set of neighboring vertices of vertex $i$, respectively, and $Z^{i}$ is the normalization factor. We also define $$\begin{aligned}
\overline{\omega}_{\mathrm{in}}
&\equiv \frac{\omega_{\mathrm{in}} - \omega_{\mathrm{out}}}{\omega_{\mathrm{out}}}
= \epsilon^{-1} - 1.\end{aligned}$$
To obtain $\ket{\psi}^{i \to j}$, we compute the following iterative equation, i.e., the BP update equation. $$\begin{aligned}
\ket{\psi}^{i \to j}
&= \frac{1}{Z^{i \to j}} \ket{\gamma} \circ
\prod_{k \in \partial i \backslash j} \left[ \ket{1} + \overline{\omega}_{\mathrm{in}} \ket{\psi}^{k \to i} W \right] \notag\\
&\hspace{70pt} \circ \exp \left[ -\overline{\omega}_{\mathrm{in}} \omega_{\mathrm{out}} \sum_{\ell} \ket{\psi}^{\ell} W \right]. \label{GeneralBPequation1}\end{aligned}$$ Analogously to (\[CompleteMarginal\]), $Z^{i \to j}$ is the normalization factor. The BP update equation (\[GeneralBPequation1\]) can be formally written as $$\begin{aligned}
\ket{\psi}^{i \to j} &= \mathcal{F}^{i \to j}\left[ \ket{\psi}^{k \to i} W, \ket{\psi}^{\ell} W \right], \label{GeneralBPequation1-2}\end{aligned}$$ where $\mathcal{F}^{i \to j}$ is the non-linear operator representing the right-hand side of (\[GeneralBPequation1\]). Note that $\ket{\psi}^{i \to j} = \mathcal{F}^{i \to j}\left[ \ket{\psi}^{k \to i}, \ket{\psi}^{\ell} \right]$ is essentially equivalent to the so-called mod-bp [@ZhangMoore2014] (without degree correction). If we consider cavity biases $\ket{\Psi}^{i \to j}$ of the transformed basis $$\begin{aligned}
\ket{\Psi}^{i \to j} &\equiv \ket{\psi}^{i \to j} W, \end{aligned}$$ its update equation is $$\begin{aligned}
\ket{\Psi}^{i \to j} &= \mathcal{F}^{i \to j}\left[ \ket{\Psi}^{k \to i}, \ket{\Psi}^{\ell} \right] W. \label{GeneralBPequation2}\end{aligned}$$ We can transform back to the original basis by operating $W^{-1}$ if it exists, or by operating $\mathcal{F}^{i \to j}$.
In the M-step, the parameter estimates ($\hat{\ket{\gamma}}$ and $\hat{\ket{\omega}}$) are updated as $$\begin{aligned}
\hat{\gamma}_{\sigma} &= \frac{1}{N}\sum_{i=1}^{N} {\left\langle \delta_{\sigma \sigma_{i}} \right\rangle}, \label{hatgammaSBM}\\
\hat{\omega}_{\mathrm{in}} &= \frac{\sum_{i<j}A_{ij} {\left\langle W_{\sigma_{i}\sigma_{j}} \right\rangle}}{\sum_{i<j}{\left\langle W_{\sigma_{i}\sigma_{j}} \right\rangle}}, \label{hatomegainSBM}\\
\hat{\omega}_{\mathrm{out}} &= \frac{\sum_{i<j}A_{ij} \left( 1 - {\left\langle W_{\sigma_{i}\sigma_{j}} \right\rangle}\right)}{\sum_{i<j}\left( 1 - {\left\langle W_{\sigma_{i}\sigma_{j}} \right\rangle}\right)} \label{hatomegaoutSBM}, \end{aligned}$$ which can be readily obtained by the extremum conditions, where $\delta_{\sigma\sigma^{\prime}}$ is the Kronecker delta and ${\left\langle \cdots \right\rangle} = \sum_{\ket{\sigma}} \cdots p(\ket{\sigma} \lvert \hat{\ket{\gamma}},\hat{\ket{\omega}},A)$ represents the average with respect to cluster assignments based on previous parameter estimates. Using the marginal probability estimates $\{\ket{\psi}^{i}\}$ and cavity biases $\{\ket{\psi}^{i \to j}\}$, we obtain ${\left\langle \delta_{\sigma \sigma_{i}} \right\rangle} = \psi^{i}_{\sigma}$ and $$\begin{aligned}
{\left\langle W_{\sigma_{i}\sigma_{j}} \right\rangle}
&= \frac{\omega_{\mathrm{in}} \ket{\psi}^{i \to j} W \ket{\psi}^{j \to i \top}}{(\omega_{\mathrm{in}} - \omega_{\mathrm{out}}) \ket{\psi}^{i \to j} W \ket{\psi}^{j \to i \top} + \omega_{\mathrm{out}} }. \label{Westimate}\end{aligned}$$ Assuming that cluster assignments are narrowly peaked [@ZhangMartinNewman2015], we can approximate the denominator of (\[hatomegainSBM\]) as $$\begin{aligned}
\sum_{i<j}{\left\langle W_{\sigma_{i}\sigma_{j}} \right\rangle}
\approx \frac{1}{2} \sum_{i,j} \ket{\psi}^{i} W \ket{\psi}^{j \top}. \end{aligned}$$ Note that we do not directly maximize (\[MarginalLogLikelihood\]). Instead, by iteratively updating (\[GeneralBPequation1\]) and (\[hatgammaSBM\])–(\[hatomegaoutSBM\]), the algorithm reaches a local extremum of the approximated marginal likelihood, or the negative Bethe free energy, which is a good estimate of (\[MarginalLogLikelihood\]) when the graph is sparse and is exact when the graph is a tree.
Detectability threshold
=======================
We now analyze the detectability threshold for a given affinity matrix $W$. In the undetectable phase, BP converges to a trivial (uninformative) fixed point. When the graph reaches the detectable phase, the trivial fixed point becomes unstable, and BP converges to an informative fixed point instead. To see this stability, we first consider the propagation of perturbations on a vertex at the trivial fixed point. In the linear-response regime, it is dominated by the transfer matrix of (\[GeneralBPequation2\]) $$\begin{aligned}
T_{\sigma^{\prime} \sigma}
= \frac{\delta \Psi^{i \to j}_{\sigma}}{\delta \Psi^{k \to i}_{\sigma^{\prime}}}
&= \frac{\overline{\omega}_{\mathrm{in}}}{1 + \overline{\omega}_{\mathrm{in}} \Psi^{k \to i}_{\sigma^{\prime}}} \psi^{i \to j}_{\sigma^{\prime}} \left( W_{\sigma^{\prime} \sigma} - \Psi^{i \to j}_{\sigma} \right). \label{TransferMatrix}\end{aligned}$$ We neglect the contribution due to $\overline{\omega}_{\mathrm{in}} \omega_{\mathrm{out}} \sum_{\ell} \Psi^{\ell}_{\tilde{\sigma}}$, because $\omega_{\mathrm{out}} = O(1/N)$.
Although the effect of the perturbation of a single vertex may be vanishingly small at a distant vertex, if the effect from all connected vertices adds to $O(1)$, the trivial fixed point is unstable. Under tree approximation, this is achieved when $c \nu^{2} > 1$, where $\nu$ is the leading eigenvalue of the transfer matrix $T$; the equality condition yields the detectability threshold. Note that investigating the detectability threshold for an arbitrary structure is difficult because the trivial fixed point is not always known. In the following, hence, we analyze some solvable cases.
A solvable case
---------------
Let us consider the case where a fraction of clusters is equal in size, i.e., $\gamma_{\sigma} = 1/q$ for any $\sigma$, and the average degree of each cluster is also equal. That is, $$\begin{aligned}
&\sum_{\sigma^{\prime}} W_{\sigma \sigma^{\prime}} = a & (a = \text{const.}) \label{RegularCondition}\end{aligned}$$ for any $\sigma$. In other words, the module graph constitutes a regular graph. This is also assumed in Ref. [@Decelle2011a]. In this case, the factorized state, i.e., $\psi^{i \to j}_{\sigma} = 1/q$ for any $i \to j$ and $\sigma$, is the trivial BP fixed point. Therefore, the transfer matrix $T$ at this fixed point is $$\begin{aligned}
T = \frac{\overline{\omega}_{\mathrm{in}}}{ q + a \overline{\omega}_{\mathrm{in}} } \left( W - \frac{a}{q}\ket{1}\bra{1} \right). \end{aligned}$$ Because $\ket{1}/\sqrt{q}$ is the leading eigenvector of $W$ with eigenvalue $a$, $\nu$ can be written as $$\begin{aligned}
\nu = \frac{\overline{\omega}_{\mathrm{in}}}{ q + a \overline{\omega}_{\mathrm{in}} } \lambda_{2}, \end{aligned}$$ where $\lambda_{2}$ is the second leading eigenvalue of $W$ in magnitude. Thus, in terms of $\epsilon$, the detectability threshold is given by $$\begin{aligned}
\epsilon^{\ast} = \frac{|\lambda_{2}|\sqrt{c} - a}{|\lambda_{2}|\sqrt{c} - a + q}. \label{RegularDetectability}\end{aligned}$$
The stochastic block model with a community structure has $a=1$ and $\lambda_{2} = 1$, which reproduces a previously known result [@Decelle2011a]. The threshold (\[RegularDetectability\]) indicates that as the number of densely connected clusters increases, the difficulty in inferring the structure also increases. In particular, when $c < (a/\lambda_{2})^{2}$, it is statistically impossible to infer the planted structure better than chance for any $\epsilon$. This behavior is shown in Fig. \[DetectabilityW6\]; when $c=4$, no signal is retrieved even when the noise $\epsilon$ is (almost) zero.
The $\lambda_{2}$-dependency of the module graph in (\[RegularDetectability\]) is another notable feature. For graph $G$, the second eigenvalue $\lambda_{2}$ of an adjacency matrix is bounded from below and above by the (normalized) edge expansion $h(G)$ as $$\begin{aligned}
1 - 2 h(G) \le \lambda_{2} \le 1 - \frac{h(G)^{2}}{2}, \label{CheegerInequality}\end{aligned}$$ which is known as Cheeger’s inequality [@chung1996spectral]. The edge expansion $h(G)$ is a measure of a sparse cut, defined by $$\begin{aligned}
h(G) = \min_{S} \frac{\lvert E(S, V \backslash S)\rvert}{a \min\{\lvert S \rvert, \lvert V \backslash S \rvert\}}, \end{aligned}$$ where $S$ is a subset of vertex set $V$ of the graph, and $\lvert E(S, V \backslash S)\rvert$ is the number of edges between sets $S$ and $V \backslash S$. The inequality (\[CheegerInequality\]) indicates that the module graph with no satisfactory sparse cut \[large $h(G)$\] tends to have a small value of $\lambda_{2}$: that is, the planted structure is difficult to infer. Put another way, if the graph has a strong hierarchical modular structure [@FootnoteHierarchicalStructure], its inference tends to be relatively easy. Note also that as long as the second eigenvalue is strictly positive, the detectability threshold is always positive for a sufficiently large average degree.
One might think that a different detectability threshold can be obtained if we instead use the flipped indicator matrix $\widetilde{W} = \ket{1}\bra{1} - W$ to parametrize noise strength as $\tilde{\epsilon} \equiv \epsilon^{-1}$, even though the structure to infer is the same. However, one can straightforwardly confirm that this treatment also yields threshold $\tilde{\epsilon}^{\ast}$ equal to (\[RegularDetectability\]).
Another solvable case
---------------------
In the case where the factorized state is not a trivial BP fixed point, the calculation of the detectability threshold is difficult. Although it is rather a toy model example, there is another case where we can obtain the analytical expression for it.
Let $W$ be a matrix whose linearly independent columns are orthogonal to one another, e.g., Fig. \[FigAffinityMatrices\](c). We set the prior distribution $\ket{\gamma}$ so that $\ket{\gamma}W \propto \ket{1}^{\top}$, and keep it fixed, i.e., we skip (\[hatgammaSBM\]); for the structure in Fig. \[FigAffinityMatrices\](c), we set $\ket{\gamma} = (1/4, 1/2, 1/4)$, although the fractions of the planted clusters do not have this ratio. In this case, the factorized fixed point is a BP fixed point. For this example, the transfer matrix (\[TransferMatrix\]) reads $$\begin{aligned}
T =
\frac{\overline{\omega}_{\mathrm{in}}}{4(2 + \overline{\omega}_{\mathrm{in}})}
\begin{pmatrix}
1 & -1 & 1 \\
-2 & 2 & -2 \\
1 & -1 & 1
\end{pmatrix} \label{Texample}\end{aligned}$$ and the leading eigenvalue is $\nu = \overline{\omega}_{\mathrm{in}}(2 + \overline{\omega}_{\mathrm{in}})^{-1}$. The corresponding detectability threshold is $$\begin{aligned}
\epsilon^{\ast} = \frac{\sqrt{c}-1}{\sqrt{c}+1}. \label{DetectabilityThresholdW8}\end{aligned}$$ This threshold was compared with the numerical experiment in Fig. \[DetectabilityW8\].
Summary and Discussion
======================
In this paper, we analyzed the detectability thresholds of general modular structures in the restricted graph ensembles. Although our results do not cover arbitrary structures, our solvable case analyses provide deeper insight into the nature of detectability. We showed that some structures are statistically impossible to infer (using BP in Sec. \[SBMinference\]), no matter how small the noise $\epsilon$ is. We also revealed that detectability transition is connected to the hierarchical structure of clusters. Our results are not rigorous and may differ from the information-theoretic limits. Also, when the number of clusters is large, there often exists another phase called the *hard phase* [@Decelle2011a]. These points are left as open questions for future research.
acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank Jean-Gabriel Young for useful comments. This work was supported by JSPS KAKENHI No. 26011023 (T.K.) and No. 25120013 (Y.K.).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In a new simple application of the non-Abelian discrete symmetry $A_4$ to charged-lepton and neutrino mass matrices, we show that for the current experimental central value of $\sin^2 2 \theta_{13} \simeq 0.1$, leptonic $CP$ violation is necessarily large, i.e. $|\tan \delta_{CP}| > 1.3$.'
---
UCRHEP-T520\
May 2012
[**New Simple $A_4$ Neutrino Model for\
Nonzero $\theta_{13}$ and Large $\delta_{CP}$\
**]{}
The non-Abelian discrete symmetry $A_4$ was introduced [@mr01; @m02; @bmv03] to achieve the seemingly impossible, i.e. the existence of a lepton family symmetry consistent with the three very different charged-lepton masses $m_e$, $m_\mu$, $m_\tau$. It was subsequently shown [@m04] to be a natural theoretical framework for neutrino tribimaximal mixing, i.e. $\sin^2 \theta_{23} = 1$, $\tan^2 \theta_{12} = 0.5$, and $\theta_{13}=0$. This pattern was consistent with experimental data until recently, when the Daya Bay Collaboration reported [@daya12] the first precise measurement of $\theta_{13}$, i.e. $$\sin^2 2 \theta_{13} = 0.092 \pm 0.016({\rm stat}) \pm 0.005({\rm syst}),$$ followed shortly [@reno12] by the RENO Collaboration, i.e. $$\sin^2 2 \theta_{13} = 0.113 \pm 0.013({\rm stat}) \pm 0.019({\rm syst}).$$ This means that tribimaximal mixing is not a good description, and more importantly, leptonic $CP$ violation is now possible because $\theta_{13}
\neq 0$, just as hadronic $CP$ violation in the quark sector is possible because $V_{ub} \neq 0$.
In this paper, we show that $A_4$ is still a good symmetry for understanding this pattern, using a new simple variation of the original idea. As shown below, it predicts a correlation between $\theta_{13}$, $\theta_{23}$, and $\delta_{CP}$ in such a way that given the experimentally allowed ranges of values for $\theta_{13}$ and $\theta_{23}$, a lower bound on $|\tan \delta_{CP}|$ is obtained. In particular, for the central values of $\sin^2 2 \theta_{13} = 0.1$ and $\sin^2 2 \theta_{12} = 0.87$, we find $|\tan \delta_{CP}| > 1.3$ from $\sin^2 2 \theta_{23} > 0.92$.
The most general $3 \times 3$ Majorana neutrino mass matrix has six complex entries, i.e. twelve parameters. Three are overall phases of the mass eigenstates which are unobservable. The nine others are three masses, three mixing angles, and three phases: one Dirac phase $\delta_{CP}$, i.e. the analog of the one complex phase of the $3 \times 3$ quark mixing matrix, and two relative Majorana phases $\alpha_{1,2}$ for two of the three mass eigenstates. The existence of nonzero $\delta_{CP}$ or $\alpha_{1,2}$ means that $CP$ conservation is violated. It is one of the most important issues of neutrino physics yet to be explored experimentally.
Before showing how the $A_4$ model is constructed, consider first the end results. In the $A_4$ basis, the $3 \times 3$ charged-lepton mass matrix is $${\cal M}_l = {1 \over \sqrt{3}} \pmatrix{1 & 1 & 1 \cr 1 & \omega^2 & \omega
\cr 1 & \omega & \omega^2} \pmatrix{m_e & 0 & 0 \cr 0 & m_\mu & 0 \cr
0 & 0 & m_\tau},$$ where $\omega = e^{2 \pi i/3} = -1/2 + i \sqrt{3}/2$, and the Majorana neutrino mass matrix is $${\cal M}_\nu = \pmatrix{a & f & e \cr f & a & d \cr e & d & a}.$$ Consider now the tribimaximal basis, i.e. $$\pmatrix{\nu_e \cr \nu_\mu \cr \nu_\tau} = \pmatrix{\sqrt{2/3} & 1/\sqrt{3}
& 0 \cr -1/\sqrt{6} & 1/\sqrt{3} & -1/\sqrt{2} \cr -1/\sqrt{6} & 1/\sqrt{3}
& 1/\sqrt{2}} \pmatrix{\nu_1 \cr \nu_2 \cr \nu_3},$$ then $${\cal M}_\nu^{(1,2,3)} = \pmatrix{a+d & b & 0 \cr b & a & c \cr 0 & c & a-d},$$ where $b=(e+f)/\sqrt{2}$, $c=(e-f)/\sqrt{2}$. The advantage of using this basis is that the experimental values of the mixing angles are not too far from the tribimaximal pattern, so that the unitary matrix which diagonalizes ${\cal M}_\nu^{(1,2,3)}$ may be approximated by $$U_\epsilon \simeq \pmatrix{1 & \epsilon_{12} & \epsilon_{13} \cr
-\epsilon_{12}^*
& 1 & \epsilon_{23} \cr -\epsilon_{13}^* & -\epsilon_{23}^* & 1}.$$ Suppose the parameters $a,b,c,d$ are all real in Eq. (6), then for small $b,c$, we find $$\epsilon_{12} \simeq {b \over d}, ~~~ \epsilon_{23} \simeq {c \over d},
~~~ \epsilon_{13} \simeq 0.$$ This implies $$\tan^2 \theta_{12} \simeq (1 - 3 \sqrt{2} \epsilon_{12})/2, ~~~
\sin^2 2 \theta_{23} \simeq 1 - (8 \epsilon_{23}^2/3), ~~~
\sin \theta_{13} \simeq -\epsilon_{23}/\sqrt{3}.$$ We then have the prediction $$\sin^2 2 \theta_{23} \simeq 1 - 2 \sin^2 2 \theta_{13}.$$ Using the existing bound [@pdg10] of $\sin^2 2 \theta_{23} > 0.92$, this would require $\sin^2 2 \theta_{13} < 0.04$, which is of course ruled out by the recent data, i.e Eqs. (1) and (2). This result is however not negative, but rather very positive, because it says that $\epsilon_{23}$ must be complex, in which case the approximation becomes $$\sin^2 2 \theta_{23} \simeq 1 - 8 [Re(U_{e3})]^2.$$ Now the new data can be accommodated provided that leptonic $CP$ violation is large.
In analyzing Eq. (6), we note from Eq. (4) that whereas the parameter $a$ may be chosen real, the others $b,c,d$ must be kept complex. In fact, even in the tribimaximal limit $(b=c=0)$, $d$ is in general complex, as shown already some time ago [@m05].
We now show how Eqs. (3) and (4) are obtained. The symmetry $A_4$ is that of the even permutation of four objects. It has twelve elements and is the smallest group which admits an irreducible three-dimensional representation. Its character table is given below.
class $n$ $h$ $\chi_1$ $\chi_{1'}$ $\chi_{1''}$ $\chi_3$
------- ----- ----- ---------- ------------- -------------- ----------
$C_1$ 1 1 1 1 1 3
$C_2$ 4 3 1 $\omega$ $\omega^2$ 0
$C_3$ 4 3 1 $\omega^2$ $\omega$ 0
$C_4$ 3 2 1 0 0 –1
: Character table of $A_4$.
The basic multiplication rule of $A_4$ is $$\underline{3} \times \underline{3} = \underline{1} + \underline{1}' +
\underline{1}'' + \underline{3} + \underline{3}.$$ As first shown in Ref. [@mr01], for $(\nu_i,l_i) \sim \underline{3}$, $l^c_i \sim \underline{1}, \underline{1}', \underline{1}''$, and $\Phi_i = (\phi^0_i, \phi^-_i) \sim \underline{3}$, the charged-lepton mass matrix is given by $${\cal M}_l = \pmatrix{v_1 & 0 & 0 \cr 0 & v_2 & 0 \cr 0 & 0 & v_3}
\pmatrix{1 & 1 & 1 \cr 1 & \omega^2 & \omega \cr 1 & \omega & \omega^2}
\pmatrix{f_1 & 0 & 0 \cr 0 & f_2 & 0 \cr 0 & 0 & f_3},$$ where $v_i = \langle \phi_i^0 \rangle$. For $v_1=v_2=v_3=v/\sqrt{3}$, we then obtain Eq.(3) with $m_e = f_1 v$, $m_\mu = f_2 v$, $m_\tau = f_3 v$. The original $A_4$ symmetry is now broken to the residual symmetry $Z_3$, i.e. lepton flavor triality [@m10], with $e \sim 1$, $\mu \sim \omega^2$, $\tau \sim \omega$. This is a good symmetry of the Lagrangian as long as neutrino masses are zero. Exotic scalar decays are predicted and may be observable at the Large Hadron Collider (LHC) in some regions of parameter space [@ckmo11; @cdmw11].
To obtain nonzero neutrino masses, we add four Higgs triplets: $\Delta_0 \sim \underline{1}$, $\Delta_i \sim \underline{3}$ under $A_4$. Let $\langle \Delta^0_0 \rangle = u_0$, $\langle \Delta^0_1 \rangle = u_i$, then Eq. (4) is the automatic result. In previous studies, $e=f=0$ has to be enforced to get tribimaximal mixing, which is technically an unnatural condition, requiring usually the addition of extra symmetries and auxiliary fields. Free of this burden, nonzero and arbitrary $d,e,f$ are easily implemented. For large Higgs triplet masses, small vacuum expectation values are naturally induced [@ms98] by the soft trilinear $\tilde{\Phi}^\dagger \Delta \Phi$ terms. We simply assume that $A_4$ is broken completely by these terms to obtain different $u_{1,2,3}$. On the other hand, the tribimaximal requirement of $u_2=u_3=0$ is very difficult to maintain, because it is not protected against infinite radiative corrections, which is the field theory’s way of telling us that they should be nonzero and arbitrary in the first place. In retrospect, it should have been obvious that Eq. (4) is the more natural choice for the neutrino mass matrix in the $A_4$ basis.
The most general neutrino mass matrix in the tribimaximal basis is $${\cal M}_\nu^{(1,2,3)} = \pmatrix{m_1 & m_6 & m_4 \cr m_6 & m_2 & m_5
\cr m_4 & m_5 & m_3}.$$ To first order, $\theta_{13}$ and $\theta_{23}$ are sensitive to $m_4$ and $m_5$, whereas $\theta_{12}$ is sensitive to $m_6$. The case of $m_6=0$ was considered in the original proposal [@m04] of tribimaximal mixing using $A_4$, and updated recently [@mw11]. The case of $m_5 \simeq 0$ is realized in a supersymmetric $B-L$ gauge model with $T_7$ symmetry discussed recently [@ckmo11R; @ikm12]. The case of unbroken residual symmetries in a class of discrete symmetries has been discussed recently [@hs12], as well as a general perturbation of the tribimaximal limit [@br12]. Here we consider the simplest and perhaps the most compelling case of $m_4=0$, which does not correspond to any unbroken residual symmetry. The fact that $m_4=0$ is simply the result of not having Higgs triplets which transform as $\underline{1}'$ or $\underline{1}''$ under $A_4$.
The neutrino mixing matrix $U$ has 4 parameters: $s_{12}, s_{23}, s_{13}$ and $\delta_{CP}$ [@pdg10]. We choose the convention $U_{\tau 1}, U_{\tau 2},
U_{e3}, U_{\mu 3} \to -U_{\tau 1}, -U_{\tau 2},
-U_{e3}, -U_{\mu 3}$ to conform with that of the tribimaximal mixing matrix of Eq. (5), then $${\cal M}_\nu^{(1,2,3)} = U^T_{TB} U \pmatrix{e^{i\alpha_1} m'_1 & 0 & 0 \cr
0 & e^{i\alpha_2} m'_2 & 0 \cr 0 & 0 & m'_3} U^T U_{TB},$$ where $m'_{1,2,3}$ are the physical neutrino masses, with $$\begin{aligned}
m'_2 &=& \sqrt{{m'_1}^2 + \Delta m^2_{21}}, \\
m'_3 &=& \sqrt{{m'_1}^2 + \Delta m^2_{21}/2 + \Delta m_{32}^2}~~{\rm (normal
~hierarchy)}, \\
m'_3 &=& \sqrt{{m'_1}^2 + \Delta m^2_{21}/2 - \Delta m_{32}^2}~~{\rm
(inverted~hierarchy)}.\end{aligned}$$ If $U$ and $\alpha_{1,2}$ are known, then all $m_{1,2,3,4,5,6}$ are functions only of $m'_1$.
We now diagonalize ${\cal M}_\nu^{(1,2,3)}$ using $$U_\epsilon {\cal M}_\nu^{(1,2,3)} U_\epsilon^T = \pmatrix{e^{i \alpha'_1}
m'_1 & 0 & 0 \cr 0 & e^{i \alpha'_2} m'_2 & 0 \cr 0 & 0 & e^{i \alpha'_3}
m'_3},$$ from which we obtain $U' = U_{TB} U^T_\epsilon$. To obtain $U$ with the usual convention, we rotate the phases of the $\mu$ and $\tau$ rows so that $U'_{\mu 3} e^{-i \alpha'_3/2}$ is real and negative, and $U'_{\tau 3} e^{-i \alpha_3/2}$ is real and positive. These phases are absorbed by the $\mu$ and $\tau$ leptons and are unobservable. We then rotate the $\nu_{1,2}$ columns so that $U'_{e1} e^{-i \alpha_3/2} = U_{e1} e^{i \alpha''_1/2}$ and $U'_{e2} e^{-i \alpha_3/2} = U_{e2} e^{i \alpha''_2/2}$, where $U_{e1}$ and $U_{e2}$ are real and positive. The physical relative Majorana phases of $\nu_{1,2}$ are then $\alpha_{1,2} = \alpha'_{1,2} + \alpha''_{1,2}$. The three angles and the Dirac phase are extracted according to $$\tan^2 \theta_{12} = |U'_{e1}/U'_{e2}|^2, ~~~
\tan^2 \theta_{23} = |U'_{\mu 3}/U'_{\tau 3}|^2, ~~~
\sin \theta_{13} e^{-i\delta_{CP}} = U'_{e3} e^{-i \alpha'_3/2}.$$ The effective Majorana neutrino mass in neutrinoless double beta decay is then given by $$m_{ee} = |U_{e1}^2 e^{i \alpha_1} m'_1 + U_{e2}^2 e^{i \alpha_2}
m'_2 + U_{e3}^2 m'_3|.$$
Although $b,c,d$ are in general complex, the structure of this model is restricted by data such that $Im(b)$ is very small, so we will assume in the following that $b$ is real. As for $Im(d)$, it is also small and affects only $m_{ee}$ slightly and not $\delta_{CP}$, so we will also take $d$ to be real. The main feature here is the complexity of $c$. To first approximation, we find $$d \simeq -a, ~~~ U_{e3} \simeq -{Re(c) \over 2a} + i {Im(c) \over 4a},$$ allowing only the normal ordering of neutrino masses.
The special case $b=0$ is especially interesting. It may be maintained by an interchange symmetry [@m04; @mw11] such that $f = -e$. As such, it was considered in Ref. [@hs12]. In that case, Eq. (6) can be diagonalized exactly. Assuming that $a,d$ are real and $c$ complex, we find $$\begin{aligned}
\tan^2 \theta_{12} &=& {1 - 3 \sin^2 \theta_{13} \over 2}, \\
\tan^2 \theta_{23} &=& { \left(1- {\sqrt{2} \sin \theta_{13} \cos \delta'_{CP}
\over \sqrt{1-3 \sin^2 \theta_{13}}} \right)^2 + {2 \sin^2 \theta_{13}
\sin^2 \delta'_{CP} \over 1-3 \sin^2 \theta_{13}} \over
\left(1+ {\sqrt{2} \sin \theta_{13} \cos \delta'_{CP}
\over \sqrt{1-3 \sin^2 \theta_{13}}} \right)^2 + {2 \sin^2 \theta_{13}
\sin^2 \delta'_{CP} \over 1-3 \sin^2 \theta_{13}}},\end{aligned}$$ where $\delta'_{CP} = \delta_{CP} - \alpha'_3/2$. The phase $\alpha'_3$ is defined in Eq. (19) and depends on the specific values of Eq. (6). For $\sin \theta_{13} = 0.16$, corresponding to $\sin^2 2 \theta_{13} = 0.1$, this predicts $\tan^2 \theta_{12} = 0.46$. If $\delta_{CP}=0$ (which also implies that $\alpha'_3 = 0$), then this would also predict $\sin^2 2 \theta_{23} = 0.80$ which is of course ruled out. Using $\sin^2 2 \theta_{23} > 0.92$, we find in this case $|\tan \delta'_{CP}| > 1.2$.
For our numerical analysis, we set $$\begin{aligned}
&& \Delta m^2_{21} = 7.59 \times 10^{-5}~{\rm eV}^2, ~~~
\Delta m^2_{32} = 2.45 \times 10^{-3}~{\rm eV}^2, \\
&& \sin^2 2 \theta_{12} = 0.87, ~~~
\sin^2 2 \theta_{13} = 0.05~{\rm to}~0.15. \end{aligned}$$ We then diagonalize Eq. (6) exactly and scan for solutions satisfying the above experimental inputs. We do not assume $b=0$ or $\alpha'_3$ to be necessarily small. We find that solutions exist only for the normal ordering of neutrino masses, i.e. $m'_1 < m'_2 < m'_3$, as in the tribimaximal case [@m05]. In Fig. 1 we show $|\tan \delta_{CP}|$ as a function of $\sin^2 2 \theta_{13}$ from 0.05 to 0.15, for the central value of $\sin^2 2 \theta_{12} = 0.87$ and the two fixed values of $\sin^2 2 \theta_{23} = 0.92$ and 0.96. In Fig. 2 we plot the parameter $b$ as a function of $\sin^2 2 \theta_{13}$. It shows that for $\sin^2 2 \theta_{12} = 0.87$, it is indeed very small. Note that for $b=0$, we find $\sin^2 2 \theta_{13} = 0.08$, i.e. $\sin^2 \theta_{13} =
0.02$. Using Eq. (23), we recover exactly $\tan^2 \theta_{12} = 0.47$, i.e. $\sin^2 2 \theta_{12} = 0.87$, as expected. In Figs. 3 and 4 we plot the physical neutrino masses $m'_{1,2,3}$ together with the effective neutrino mass $m_{ee}$ in neutrinoless double beta decay as functions of $\sin^2 2 \theta_{13}$ for $\sin^2 2 \theta_{23} = 0.92$ and 0.96 respectively. Note that $m_{ee}$ is always smaller than $m'_1$ because $e^{i \alpha_1} = -1$ and $e^{i \alpha_2} = 1$ in Eq. (21). In Figs. 5 and 6 we plot the parameters $a,-d,Re(c),Im(c)$ as functions of of $\sin^2 2 \theta_{13}$ for $\sin^2 2 \theta_{23} = 0.92$ and 0.96 respectively.
In conclusion, neutrino tribimaximal mixing may be dead, but $A_4$ is alive and even getting healthier. In a new simple application, given the present allowed ranges of values for $\sin^2 2 \theta_{13}$ and $\sin^2 2 \theta_{23}$, we predict large $CP$ violation and a normal ordering of neutrino masses.
: The work of H.I. is supported by Grant-in-Aid for Scientific Research, No. 23.696, from the Japan Society of Promotion of Science. The work of E.M. is supported in part by the U. S. Department of Energy under Grant No. DE-AC02-06CH11357.
[99]{} E. Ma and G. Rajasekaran, Phys. Rev. [**D64**]{}, 113012 (2001). E. Ma, Mod. Phys. Lett. [**A17**]{}, 2361 (2002). K. S. Babu, E. Ma, and J. W. F. Valle, Phys. Lett. [**B552**]{}, 207 (2003). E. Ma, Phys. Rev. [**D70**]{}, 031901 (2004). Daya Bay Collaboration: F. P. An [*et al.*]{}, Phys. Rev. Lett. [**108**]{}, 171803 (2012) \[arXiv:1203.1669 \[hep-ex\]\]. RENO Collaboration: J. K. Ahn [*et al.*]{}, arXiv:1204.0626 \[hep-ex\]. Particle Data Group: K. Nakamura [*et al.*]{}, J. Phys. G: Nucl. Part. Phys. [**37**]{}, 075021 (2010). E. Ma, Phys. Rev. [**D72**]{}, 037301 (2005). E. Ma, Phys. Rev. [**D82**]{}, 037301 (2010). Q.-H. Cao, S. Khalil, E. Ma, and H. Okada, Phys. Rev. Lett. [**106**]{}, 131801 (2011). Q.-H. Cao, A. Damanik, E. Ma, and D. Wegman, Phys. Rev. [**D83**]{}, 093012 (2011). E. Ma and U. Sarkar, Phys. Rev. Lett. [**80**]{}, 5716 (1998). E. Ma and D. Wegman, Phys. Rev. Lett. [**107**]{}, 061803 (2011). Q.-H. Cao, S. Khalil, E. Ma, and H. Okada, Phys. Rev. [**D84**]{}, 071302(R) (2011). H. Ishimori, S. Khalil, and E. Ma, arXiv:1204.2705 \[hep-ph\]. D. Hernandez and A. Yu. Smirnov, arXiv:1204.0445 \[hep-ph\]. B. Brahmachari and A. Raychaudhuri, arXiv:1204.5619 \[hep-ph\].
{width="14cm"}
{width="14cm"}
{width="14cm"}
{width="14cm"}
{width="14cm"}
{width="14cm"}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- Daniel Hetterich
- Gabriel Schmitt
- Lorenzo Privitera
- Björn Trauzettel
title: |
Supplementary Material for:\
Strong frequency dependence of transport in the driven Fano-Anderson model
---
Details on the BCH expansion
----------------------------
Given the two-step form of our drive, we can derive an effective Hamiltonian in the high-frequency regime with a Baker-Campbell-Hausdorff (BCH) formula. If we choose $t_0=0$ as our initial time, the effective (Floquet) Hamiltonian ${\hat{H}}_F$ is defined via $${\hat{F}} = \hat{U}\left(\tau , 0\right) \equiv e^{-i {\hat{H}}_F \tau} = e^{-i {\hat{H}}_1 \frac{\tau}{2}} e^{-i {\hat{H}}_2 \frac{\tau}{2}} \;$$ where $H_{1/2}$ are $$\begin{aligned}
H_{1/2} = \sum_i (\frac{m}{\sqrt{L}} (a_i^{\dagger} a_1^{\phantom{\dagger}} + a_1^{\dagger} a_i^{\phantom{\dagger}})+\epsilon_i^{\phantom{\dagger}} a_i^{\dagger} a_i^{\phantom{\dagger}}) \pm A a_1^{\dagger} a_1^{\phantom{\dagger}} \;.
\end{aligned}$$ Using the standard BCH formula we obtain, up to second order in $1/\omega$,$$\begin{split}
H_F= \frac{H_1 + H_2}{2} +& i2\pi\frac{\lbrack H_1, H_2 \rbrack }{8\omega} + \\
+ \frac{(2\pi)^2}{96\omega^2} &(- \lbrack H_1 \lbrack H_1, H_2\rbrack \rbrack + \lbrack H_2 \lbrack H_1, H_2 \rbrack \rbrack ) \;.
\end{split}$$ The commutators are easily computed as $$\begin{aligned}
&\lbrack H_1 , H_2 \rbrack
= - 2 A \frac{m}{\sqrt{L}} \sum_i ( a_i ^{\dagger} a_1 ^{\phantom{\dagger}} - a_1 ^{\dagger} a_i ) \;,\\
&\lbrack H_1 \lbrack H_1, H_2\rbrack \rbrack - \lbrack H_2 \lbrack H_1, H_2 \rbrack \rbrack =
4 \frac{ A^2 m}{\sqrt{L}} (a_1^{\dagger} a_i^{\phantom{\dagger}} + \mathrm{h.c.} ) \;,
\end{aligned}$$ yielding the Floquet Hamiltonian $$\begin{aligned}
H_F =& \sum_i \left(\frac{m}{\sqrt{L}} \left(a_i^{\dagger} a_1^{\phantom{\dagger}} +\mathrm{h.c.}\right)+\epsilon_i^{\phantom{\dagger}} a_i^{\dagger} a_i^{\phantom{\dagger}} \right)+\\ & -\frac{2\pi}{4\omega} i A \frac{m}{\sqrt{L}} \sum_i \left( a_i ^{\dagger} a_1 ^{\phantom{\dagger}} - \mathrm{h.c.} \right) + \\&- \frac{ A^2 (2\pi)^2 }{24 \omega^2}\frac{m}{\sqrt{L}} \left(a_1^{\dagger} a_i^{\phantom{\dagger}} + \mathrm{h.c.}\right) + \mathcal{O}\left(\frac{1}{\omega^3}\right).
\end{aligned}$$ As we mention in the main text, the first order correction gives an imaginary hopping term, breaking time-reversal symmetry (TRS). However, this is only apparent, because this term depends on the initial time, while the long-time dynamics should not. Changing the initial time is equivalent to a gauge transformation [@Eckardt_NJP15] on ${\hat{H}}_F$. Indeed, if one, for example, moves the initial time at $\tau/2$, this term changes its sign. Many quantities must be gauge independent, e.g. should not depend on the initial time of the driving. One example are the quasienergies, which we have used as a measure of localization through their gap statistics. If one is instead interested in an expansion which is gauge invariant, a tripartition of the Floquet operator must be employed [@Goldman_PRX14; @Eckardt_NJP15]: ${\hat{F}}_{t_0} = {\hat{S}}^{\dagger}\exp \left(-i{\hat{\tilde{H}}}_{F} \tau \right){\hat{S}}^{\phantomsection{\dagger}}(t_0) $. ${\hat{S}}$ are called kick operators, while the new Hamiltonian ${\hat{\tilde{H}}}_{F}$ does not depend on the initial time $t_0$ anymore. For our model, this procedure essentially consists of keeping in ${\hat{\tilde{H}}}_F$ only the second order real term. On the other hand, the kick operators will contain, among others, the first order term. In this way, one can separate between long time effects and intraperiod ones. For simplicity, we display in the main text the BCH result, as the localization behaviour does not change qualitatively.
[99]{} A. Eckardt and E. Anisimovas, New Journal of Physics 17, 093039 (2015) N. Goldman and J. Dalibard, Phys. Rev. X 4, 031027 (2014)
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
bibliography:
- 'SciPost\_Example\_BiBTeX\_File.bib'
---
[ **Probing Lepton Universality with (Semi)-Leptonic **[$B$]{} decays**** ]{}
G. Banelli^1^, R. Fleischer^1,2^, R. Jaarsma^1^, G. Tetlalmatzi-Xolocotzi^1\*^.
[**1**]{} Nikhef, Science Park 105, NL-1098 XG Amsterdam, Netherlands\
[**2**]{} Faculty of Science, Vrije Universiteit Amsterdam,\
NL-1081 HV Amsterdam, Netherlands\
\* [email protected]
Abstract {#abstract .unnumbered}
========
[**The most recent measurements of the observables $R_{D^{(*)}}$ are in tension with the Standard Model offering hints of New Physics in $b\rightarrow c \ell \bar{\nu}_{\ell}$ transitions. Motivated by these results, in this work we present an analysis on their $b\rightarrow u \ell \bar{\nu}_{\ell}$ counterparts (for $\ell=e, ~\mu, ~\tau$). Our study has three main objectives. Firstly, using ratios of branching fractions, we assess the effects of beyond the Standard Model scalar and pseudoscalar particles in leptonic and semileptonic $B$ decays ($B^-\rightarrow \ell^- \bar{\nu}_{\ell}$, $\bar{B}\rightarrow \pi \ell \bar{\nu}_{\ell}$ and $\bar{B}\rightarrow \rho \ell \bar{\nu}_{\ell}$). Here a key role is played by the leptonic $B$ processes, which are highly sensitive to new pseudoscalar interactions. In particular, we take advantage of the most recent measurement of the branching fraction of the channel $B^-\rightarrow \mu^-\bar{\nu}_{\mu}$ by the Belle collaboration. Secondly, we extract the CKM matrix element $|V_{ub}|$ while accounting simultaneously for New Physics contributions. Finally, we provide predictions for the branching fractions of yet unmeasured leptonic and semileptonic $B$ decays.** ]{}
------------------------------------------------------------------------
------------------------------------------------------------------------
Introduction {#sec:intro}
============
Recent measurements of the observables $R_{D^{^{(*)}}}\equiv \mathcal{B}(B\rightarrow D^{(*)}\tau \bar{\nu}_{\tau})/\mathcal{B}(B\rightarrow D^{(*)}\ell' \bar{\nu}_{\tau})$, with $\ell'= e,~\mu$, have caused a lot of excitement in the high-energy physics community. As a matter of fact, the combined measurements of the BaBar, LHCb and Belle collaborations show a $3.9~\sigma$ deviation with respect to the expected value from the Standard Model (SM) [@Amhis:2016xyh]. If this effect is confirmed by forthcoming experimental determinations, it will indicate the presence of New Physics (NP) in the exclusive decays $\bar{B}\rightarrow D^{(*)}\tau \bar{\nu}_{\tau}$, which are caused by the quark-level transition $b\rightarrow c \tau \bar{\nu}_{\tau}$ [@Svjetlana:2018]. Motivated by these results, we investigate the presence of NP effects in $b\rightarrow u \ell \bar{\nu}_{\ell}$ processes, where $\ell=e,~\mu,~\tau$. To derive the relevant constraints, our study involves the interplay of the leptonic decay channels $B^-\rightarrow \ell^- \bar{\nu}_{\ell}$ with the semileptonic transitions $\bar{B}\rightarrow \pi \ell \bar{\nu}_{\ell}$ (analogous to $\bar{B}\rightarrow D \ell \bar{\nu}_{\ell}$) and $\bar{B}\rightarrow \rho \ell \bar{\nu}_{\ell}$ (analogous to $\bar{B}\rightarrow D^{*} \ell \bar{\nu}_{\ell}$).\
The different decays to be included in this study are sensitive to NP scalar, pseudoscalar, vector and tensor interactions [@Hou:1992sy; @Sakaki:2013bfa]. The pseudoscalar components are special because, due to the structure of the equations for the branching fractions $\mathcal{B}(B^-\rightarrow \ell^{-}\bar{\nu}_{\ell})$, they lift the helicity suppression appearing in the corresponding SM expressions. This effect leads to interesting phenomenological predictions. Therefore, we focus on the pseudoscalar contributions and complement our NP analysis by including also scalar operators, which are their natural partners in terms of the Lorenz structure.\
We follow an effective theory approach, with the low-energy Hamiltonian
$$\label{Heff}
{\cal H}_{\rm eff}= \frac{4 G_{\rm F}}{\sqrt{2}}V_{ub}\left[C_{V_L} {\cal O}_{V_L}^\ell +
C_{S}^\ell{\cal O}_{S}^\ell
+
C_{P}^\ell{\cal O}_{P}^\ell \right] + \hbox{h.c.},$$
where
$$\begin{aligned}
\mathcal{O}^{\ell}_{V_L}=(\bar q \gamma^\mu P_L b)(\bar \ell \gamma_\mu P_L \nu_{\ell}),
\quad
{\cal O}^{\ell}_{S}= (\bar q b)(\bar \ell P_L \nu_{\ell}),\quad
{\cal O}^{\ell}_{P}= (\bar q \gamma_5 b)(\bar \ell P_L \nu_{\ell}),\end{aligned}$$
are the corresponding vector, scalar and pseudoscalar operators, respectively. The short-distance contributions are encoded in the Wilson coefficients $C_{V_L}$, $C^{\ell}_S$ and $C^{\ell}_P$. In the SM, only $C_{V_L}$ is non-vanishing and takes the value $C_{V_L}=1$. However, $C^{\ell}_S$ and $C^{\ell}_P$ may be non-zero in NP scenarios. An important example is the Two Higgs Doublet Model (2HDM) [@Hou:1992sy], where the scalar and pseudoscalar coefficients are related to each other as follows:
$$\begin{aligned}
C^{\ell}_{S}&=& C_P^\ell = -\tan^2\beta \Bigl(\frac{m_b m_l}{M^2_{H^{\pm}}}\Bigl).\end{aligned}$$
In general, the Wilson coefficients in Eq. (\[Heff\]) can be complex, with CP-violating NP phases. Unfortunately, direct CP asymmetries associated with leptonic and semileptonic processes cannot be used in their determination, because they vanish due to the absence of sizeable CP conserving phase differences. Hence, as discussed in Ref. [@Banelli:2018fnx], in order to search for new sources of CP violation, a different strategy based on correlations between magnitudes and phases of Wilson coefficients should be followed. For simplicity, our discussion in this report will be limited to real $C^{\ell}_S$ and $C^{\ell}_P$.\
It is important to bear in mind that the exclusive determination of the CKM matrix element $|V_{ub}|$ is done using semileptonic $B$ decays originating from the quark level transition $b\rightarrow u \ell \bar{\nu}_{\ell}$ assuming the SM. However, this value may be affected by NP effects. Therefore, in order to take into account these contributions we propose the following strategy:
1. Using combinations of ratios of branching fractions for leptonic and semileptonic $B$ decays, where $|V_{ub}|$ cancels, we determine the allowed regions for $C^{\ell}_S$ and $C^{\ell}_P$.
2. We then use these numerical ranges for $C^{\ell}_S$ and $C^{\ell}_P$ and evaluate the branching fraction for given leptonic or semileptonic $B$ processes, allowing us to finally extract $|V_{ub}|$ from data.
In the following sections, we will elaborate on this procedure and we will present our predictions for the branching fractions of leptonic and semileptonic $B$ decays which have not yet been measured. The full study can be found in Ref. [@Banelli:2018fnx].
Constraints from leptonic $ B$ decays
========================================
In the SM, the branching fraction of the process $B^{-}\rightarrow \ell^{-}\bar{\nu}_{\ell}$ is given by
$$\label{SM-Br}
{\mathcal B}(B^-\to\ell^-\bar\nu_\ell)|_{\rm SM}=\frac{G_{\rm F}^2}{8\pi}
|V_{ub}|^2M_{B^-}m_\ell^2\left(1-\frac{m_\ell^2}{M_{B^-}^2}\right)^2f_{B^-}^2\tau_{B^-}.$$
We observe that the branching ratio is proportional to the square of the mass of the lepton in the final state $m^2_{\ell}$. For electrons and muons, this factor very strongly suppresses the decay probabilities, which is referred as helicity suppression. Interestingly, it is also present in the rare decays $B_{d,s}\rightarrow \ell^+\ell^-$ [@Fleischer:2017ltw]. Since $m_{e}, m_{\mu} \ll m_{\tau}$, the helicity suppression is less effective for tau leptons.\
Leptonic $B$ decays are very clean channels, where all the non-perturbative hadronic information is encoded in the decay constant given by [@Aoki:2016frl; @Dowdall:2013tga] $$\begin{aligned}
f_{B^-}=0.186\pm 0.004.\end{aligned}$$
Assuming the SM, we use the CKMFitter value [@Charles:2004jd] $$\begin{aligned}
|V_{ub}|=(3.601\pm 0.098)\times 10^{-3},
\label{eq:fBs}\end{aligned}$$
together with Eq. (\[eq:fBs\]) to obtain $$\begin{aligned}
{\mathcal B}(B^-\to\tau^-\bar\nu_\tau) &=& (7.92\pm0.55)\times10^{-5}, \nonumber\\
{\mathcal B}(B^-\to\mu^-\bar\nu_\mu) &=& (3.56\pm0.25)\times10^{-7}, \label{eq:SMleptBrmu} \nonumber\\
{\mathcal B}(B^-\to e^-\bar\nu_e) &=& (8.33\pm0.58)\times10^{-12},\label{eq:SMleptBr}\end{aligned}$$
where due the tiny value of the mass of the electron, the helicity suppression leads to a extremely small branching fraction for the channel $B^-\to e^-\bar\nu_e$.\
The experimental results reported by BaBar, Belle and LHCb lead to $$\begin{aligned}
\label{eq:leptonicExp}
{\mathcal B}(B^-\to\tau^-\bar\nu_\tau)&=& (1.09\pm0.24)\times10^{-4},\nonumber\hbox{\cite{PhysRevD.98.030001}}\\
{\mathcal B}(B^-\to\mu^- \bar\nu_\mu)&=& (6.46 \pm 2.74)\times10^{-7},\hbox{\cite{Sibidanov:2017vph}}\nonumber\\
{\mathcal B}(B^- \to e^- \bar\nu_e) &<& 9.8 \times 10^{-7} \, \mbox{(90\% C.L.)}\hbox{\cite{Satoyama:2006xn}}.\end{aligned}$$
Here the measurement corresponding to the channel $B^-\to\mu^- \bar\nu_\mu$ was reported recently by the Belle collaboration with a $2.4~\sigma$ excess over background and will be key during our phenomenological study.\
For leptonic $B$ decays, the hadronic matrix element of the $O_S$ operator in Eq. (\[Heff\]) vanishes. Consequently, is does not receive scalar NP contributions. Once pseudoscalar NP effects are taken into account, Eq. (\[SM-Br\]) gets modified as
$$\begin{aligned}
{\mathcal B}(B^-\to\ell^-\bar\nu_\ell)={\mathcal B}(B^-\to\ell^-\bar\nu_\ell)|_{\rm SM}
\left|1+ \frac{M_{B^-}^2}{m_\ell (m_b+m_u)} C_P^\ell \right|^2.
\end{aligned}$$
In the case of electrons and muons, the helicity suppression is lifted by the mass ratio
$$\begin{aligned}
M^2_{B^-}/\Bigl[m_{\ell} (m_b +m_u)\Bigl]\sim M_{B^-}/m_{\ell},\end{aligned}$$
thereby amplifying the effects of $C_P^\ell$.\
To constrain the pseudoscalar Wilson coefficients $C^{\mu}_P$ and $C^{\tau}_P$, we consider the observable
$$\begin{aligned}
\label{eq:leptonicratio}
R^{\tau}_{\mu}\propto
\frac{{\mathcal B}(B^-\to\tau^-\bar\nu_{\tau})}{{\mathcal B}(B^-\to \mu^-\bar\nu_{\mu})},\end{aligned}$$
where the normalization factor is chosen in such a way that we get $R^{\tau}_{\mu}=1$ in the SM. The main features of $R^{\tau}_{\mu}$ are the cancellation of the hadronic decay constant $f_{B^-}$ and of the CKM matrix element $|V_{ub}|$. By comparing the corresponding theoretical determination for this ratio with the experimental result, we obtain the regions shown in Fig. \[fig:a\]. Here we can see that, even though $R^{\tau}_{\mu}$ is already imposing strong constraints on the values that $C_P^\mu$ and $C_P^\tau$ can take, the arms of the resulting cross-shaped area extend to infinity. To improve our bounds on these pseudoscalar Wilson coefficients, we have to include more observables sensitive to $C^{\mu}_P$ and $C^{\tau}_P$. This topic will be discussed in the next section.
![Allowed regions in the $C^{\mu}_P$–$C^{\tau}_P$ plane following from the leptonic ratio $R^{\tau}_{\mu}$.[]{data-label="fig:a"}](leptonicmutau.pdf){width="60.00000%"}
Semileptonic **[$B$]{} decays**
===============================
To derive stronger constraints on $C^{\mu}_P$ and $C^{\tau}_P$, we consider the branching ratio $\mathcal{B}(\bar{B}\rightarrow \rho \ell \bar{\nu}_{\ell})$, which in analogy with the leptonic decays, does not depend on scalar NP contributions [@Sibidanov:2013rkk]. Due to the presence of the $\rho$ meson in the final state, the hadronic contributions are more complicated than those encountered in the leptonic $B$ decays. In the helicity basis, these effects are encoded in the form factors denoted as $H^{\rho}_{V,+}$, $H^{\rho}_{V,-}$, $H^{\rho}_{V,0}$, $H^{\rho}_{V,t}$ and $H^{\rho}_{S}$. The non-perturbative technique employed for their calculation depends on the value of the square of the four-momentum transferred $q^2$ to $\ell$ and $\bar{\nu}_{\ell}$. In the literature, two main approaches are usually considered for their determination:
- QCD sum rules for the low energy regime $0\leq q^2\leq q^2_{\rm max}$, where typically $q^2_{\rm max}\in [12, 16]~\rm{GeV}^2$.
- Lattice QCD calculations [@Lattice:2015tia; @Bowler:2004zb] are applied when $q^2$ is close to the maximal leptonic momentum transfer: $q^2_{\rm max}\leq q^2\leq (M_{B}-M_{\rho})^2$.
As experimental input we consider the measurements of ${\mathcal B}(\bar B^0\rightarrow \rho^+ \ell^- \bar{\nu}_{\ell})$ and ${\mathcal B}(B^-\rightarrow \rho^0 \ell^- \bar{\nu}_{\ell})$ reported by Belle in 2013, which include an admixture of electrons and muons in the final state [@Sibidanov:2013rkk]. Using the isospin symmetry, we combine these two measurements to obtain $\Braket{{\mathcal B}(\bar B\rightarrow \rho \ell^- \bar{\nu}_{\ell})}_{[\ell=~ e, \mu],~q^2\leq 12~\rm{GeV}^2}=(1.98 \pm 0.12) \times 10^{-4}$. We introduce the ratio
$$\begin{aligned}
\label{eq:leptonicoversemileptonicrho-theo}
\mathcal{R}^{\mu}_{\Braket{e, \mu}; \rho ~ [q^2\leq 12]~\rm{GeV}^2}&\equiv&\mathcal{B}(B^-\rightarrow \mu^- \bar{\nu})/
\Braket{{\mathcal B}(\bar B \rightarrow \rho \ell^- \bar{\nu}_{\ell})}_{[\ell=~ e, \mu],~q^2\leq 12~\rm{GeV}^2},\end{aligned}$$
where the CKM element $|V_{ub}|$ cancels.\
Since the experimental determinations do not yet provide independent information for electrons and muons, we can only obtain the allowed values for $C^{e}_P$ and $C^{\mu}_P$ if we correlate these Wilson coefficients through different assumptions. We start by testing the hypothesis of having universal NP interactions in electrons and muons, i.e. $C^{e}_P=C^{\mu}_P$, and explore the behaviour of the semileptonic decay in the range $q^2\leq 12~\rm{GeV}^2$, where an analytical parameterization from QCD sum rules is available [@Straub:2015ica]. Then, we proceed to determine the allowed regions in the $C^{\mu}_P$ $-$ $C^{\tau}_P$ plane by using the observables $R^{\tau}_{\mu}$ and $\mathcal{R}^{\mu}_{\Braket{e, \mu}; \rho ~ [q^2\leq 12]~\rm{GeV}^2}$ as constraints . Moreover, we include the leptonic ratio $R^{e}_{\mu}$, which is analogous to $R^{\tau}_{\mu}$ in Eq. (\[eq:leptonicratio\]), calculated from the experimental bound available for $\mathcal{B}(B^-\rightarrow e^- \bar{\nu}_e)$ presented in Eq. (\[eq:leptonicExp\]). The resulting plot is shown in Fig. \[fig:lepsemmutau\], where only the elliptical areas labelled as “1” and “2” are allowed. Even though solution “1” is compatible with the SM, the solution inside region “2”, corresponding to NP, is not excluded.
![Allowed regions in the $C^{\mu}_P$–$C^{\tau}_P$ plane utilizing the ratios $R^{e}_{\mu}$, $R^{\tau}_{\mu}$ and $\mathcal{R}^{\mu}_{\Braket{e, \mu}; \rho~[q^2\leq 12]~\rm{GeV}^2}$ under the assumption $C^e_P=C^{\mu}_P$.[]{data-label="fig:lepsemmutau"}](lepsemmutau.pdf){width="60.00000%"}
So far, our treatment has been focused on constraining the pseudoscalar Wilson coefficient $C^{\ell}_P$. In order to have sensitivity on the scalar Wilson coefficient $C^{\ell}_S$, we include the branching ratios of the processes $\bar B\rightarrow \pi \ell \bar{\nu}_{\ell}$. Unfortunately, the current experimental information does not provide independent measurements for electrons and muons. To incorporate these processes in our analysis, we make an average using the isospin symmetry to combine the independent contributions of $B^0$ and $B^-$ provided in Ref. [@PhysRevD.98.030001], yielding $\Braket{{\mathcal B}(\bar{B}\rightarrow \pi \ell \bar{\nu}_{\ell})}_{[\ell=~ e, \mu]}=(1.53 \pm 0.04)\times 10^{-4}$.\
Two more observables, which are sensitive to both $C^{\ell}_S$ and $C^{\ell}_P$, are now at our disposal:
$$\begin{aligned}
\label{eq:leptonicoversemileptonicpi}
\mathcal{R}^{\mu}_{\Braket{e, \mu}; \pi}&\equiv&\mathcal{B}(B^-\rightarrow \mu^- \bar{\nu})/
\Braket{{\mathcal B}(\bar B \rightarrow \pi \ell^- \bar{\nu}_{\ell})},\nonumber\\
\mathcal{R}^{\Braket{e, \mu};\rho~[q^2_{\rm{min}} \leq q^2 \leq q^2_{\rm{max}}]}_{\Braket{e, \mu};\pi}
&\equiv&
\Braket{ \mathcal{B}(\bar{B}\rightarrow \rho \ell^- \bar{\nu}_{\ell})}_{[\ell=e, \mu]}\Bigl|^{q^2_{\rm{max}}}_{q^2_{\rm{min}}}/
\Braket{\mathcal{B}(\bar{B}\rightarrow \pi \ell^- \bar{\nu}_{\ell})}_{[\ell=e, \mu]}.
\label{eq:semirhosemipi}\end{aligned}$$
Including furthermore $\mathcal{R}^{\mu}_{\Braket{e, \mu}; \rho ~ [q^2\leq 12]~\rm{GeV}^2}$, introduced in Eq. (\[eq:leptonicoversemileptonicrho-theo\]), and making the assumptions $C^e_P=C^{\mu}_P$, $C^e_S=C^{\mu}_S$, we obtain the regions in the $C^{\mu}_S$$-$$C^{\mu}_P$ plane shown in Fig. \[fig:semilepoversemilep\]. We observe that $\mathcal{R}^{\Braket{e, \mu};\rho~[0 \leq q^2 \leq 12]~\rm{GeV}^2}_{\Braket{e, \mu};\pi}$ results in two horizontal bands, which are in tension with the SM at $(1$$-$$2)~\sigma$.
![The allowed regions in the $C^{\mu}_P$–$C^{\mu}_S$ plane following from the observables $\mathcal{R}^{\Braket{e, \mu};\rho~[0 \leq q^2 \leq 12]~{\rm GeV^2}}_{\Braket{e, \mu};\pi}$, $\mathcal{R}^{\mu}_{\Braket{e, \mu};\rho~[q^2 \leq 12]~{\rm GeV^2}}$ and $\mathcal{R}^{\mu}_{\Braket{e, \mu};\pi}$.[]{data-label="fig:semilepoversemilep"}](allCPCS.pdf){width="65.00000%"}
We have restricted our studies to the low $q^2$ regime. Let us now investigate whether the tension found with the SM persists for large $q^2$ values. To the best of our knowledge, the only available hadronic inputs for $B\rightarrow \rho$ transitions in the $12~\hbox{GeV}^2 \leq q^2$ range come from lattice determinations [@Bowler:2004zb] and were obtained in 2004. Due to the unstable nature of the $\rho$ meson, these calculations are rather challenging. To keep the non-perturbative uncertainties under control when evaluating the branching ratios, we use differential distributions in $q^2$ rather than fully integrated expressions. Therefore, for $12~\hbox{GeV}^2<q^2$ we introduce the following observable:
$$\begin{aligned}
\label{eq:dR}
d\mathcal{R}^{\Braket{e, \mu};\rho}_{\Braket{e, \mu};\pi}&=&
\frac{2\Braket{d\mathcal{B}(B^-\rightarrow \rho^0 \ell^{-}\bar{\nu}_\ell)/dq^2}_{[\ell=e, \mu]}}{\Braket{\mathcal{B}(\bar{B}\rightarrow \pi \ell^- \bar{\nu}_{\ell})}_{[\ell=e, \mu]}}.\end{aligned}$$
Interestingly, as can be seen in Fig. \[fig:dsemilepoversemilepbinbybinbeta6\], we find a mild tension with the SM for $q^2=17~\rm{GeV}^2$. To shed light on the origin of this feature, an updated determination of the form factors for the transition $B\rightarrow \rho$ in the high $q^2$ regime is needed.
![Regions in the $C^{\mu}_P$–$C^{\mu}_S$ plane following from the observable $d\mathcal{R}^{\Braket{e, \mu};\rho}_{\Braket{e, \mu};\pi}$ in the large $q^2$ regime for $q^2=17~\rm{GeV}^2$.[]{data-label="fig:dsemilepoversemilepbinbybinbeta6"}](qsq17beta60CPCSmu.pdf){width="45.00000%"}
Determination of **[$|V_{ub}|$]{} and predictions of the branching fractions for **[$B^- \to e^- \overline{\nu}_e$]{} and **[$\overline{B}\to \rho \tau^- \overline{\nu}_\tau$]{}****** {#Sec:Vubandpred}
=======================================================================================================================================================================================
Having the regions for $C^{\ell}_{P}$ and $C^{\ell}_{S}$ available, we can proceed with the extraction of $|V_{ub}|$. The strategy introduced at the end of Sec. \[sec:intro\] describes the basic ingredients required for the determination of this CKM matrix element. A more refined procedure, which accounts for possible correlations between our observables, is discussed in [@Banelli:2018fnx]. By applying this method, we obtain
$$\label{eq:vubSolUniversal}
|V_{ub}| = (3.31 \pm 0.32) \times 10^{-3}.$$
Although in agreement with the value reported by the CKMFitter collaboration [@Charles:2004jd], presented in Eq. (\[eq:fBs\]), the uncertainty in Eq. (\[eq:vubSolUniversal\]) is three times bigger. However, our target here is only to illustrate the application of our method which does not assume the SM during the extraction of $|V_{ub}|$. Future improvements in the precision of our observables will allow us to reduce the uncertainty in the CKM matrix element $|V_{ub}|$.\
Besides universal pseudoscalar NP interactions between electrons and muons, other possible scenarios can be considered. We have also explored the cases: $C^e_P=(1/10)C^{\mu}_P$ and $C^e_P=10 C^{\mu}_P$. In addition, we included the 2HDM, where the pseudocalar Wilson coefficients for electrons and tau leptons are correlated with $C^{\mu}_P$ as
$$\begin{aligned}
C^e_P=\frac{m_e}{m_{\mu}}\times C^{\mu}_P,&&
C^{\tau}_P=\frac{m_{\tau}}{m_{\mu}} \times C^{\mu}_P.\end{aligned}$$
In these scenarios, our method leads to the value for $|V_{ub}|$ in Eq. (\[eq:vubSolUniversal\]). Finally, we have considered the situation where NP affects only the 3rd generation of leptons, i.e. $C^{\tau}_P \neq 0$, while $C^e_P=C^{\mu}_P=0$. Here we obtain $|V_{ub}|=(4.85 \pm 1.03) \times 10^{-3}$, which is closer to the value of $|V_{ub}|$ following from inclusive determinations [@Charles:2004jd; @Aglietti:2004fz; @Aglietti:2006yb; @Aglietti:2007ik; @Lange:2005yw; @Bosch:2004th; @Bosch:2004cb; @Andersen:2005mj; @Gambino:2007rp].\
Finally, we use the ranges for the different Wilson coefficients to make predictions for the yet unmeasured branching ratios $\mathcal{B}(B^-\rightarrow e^- \bar{\nu}_{e})~~\hbox{and}~~
\mathcal{B}(\bar B\rightarrow \rho \tau \bar{\nu}_{\tau})$. In Fig. \[fig:prop-plot-e\], the predictions for $\mathcal{B}(B^-\rightarrow e^- \bar{\nu}_e)$ are illustrated. We would like to highlight that, for $C^e_P=10 C^{\mu}_P$, our analysis leads to a potential enhancement of $\mathcal{B}(B^-\rightarrow e^- \bar{\nu}_e)$ that may even saturate the current experimental bound. An interesting phenomenological prediction in the case $C^{e}_P=C^{\mu}_P$ is an enhancement by up to four orders of magnitude with respect to the SM value, thereby lying just a factor of 10 below the current experimental upper bound. An analogous effect in the case of $B_s \rightarrow e^+ e^-$ has been discussed in Ref. [@Fleischer:2017ltw].\
The values for $\mathcal{B}(\bar B\rightarrow \rho \tau^- \bar{\nu}_{\tau})$ following from our analysis are consistent with the SM picture at the $1~\sigma$ level. As discussed in [@Banelli:2018fnx], future measurements of this observable are a powerful tool to distinguish between solutions “1” and “2” in Fig. \[fig:lepsemmutau\]. Our full strategy is summarized in the flow chart in Fig. \[fig:flowchart\].
![Illustration of the possible enhancement of $\mathcal{B}(B^-\rightarrow e^- \bar{\nu}_{e})$ for the scenarios discussed in the text. The blue line gives the current experimental upper bound on $\mathcal{B}(B^-\rightarrow e^- \bar{\nu}_{e})$, whereas the red horizontal line on the bottom represents the SM value. The red regions indicate the values of the branching ratio that may be obtained.[]{data-label="fig:prop-plot-e"}](brBtoenuPropPlot.pdf){width="60.00000%"}
![Flowchart illustrating our strategy.[]{data-label="fig:flowchart"}](flowchart.pdf){width="100.00000%"}
Outlook
=======
Leptonic and semileptonic decays originating from $b\rightarrow u \ell \bar{\nu}_{\ell}$ transitions are very interesting and powerful channels to unveil potential NP contributions. They are the counterparts of the $b\rightarrow c \ell \bar{\nu}_{\ell}$ processes, where recent experimental results for the associated observables $R_{D^{(*)}}$ show effects which may originate from physics beyond the SM.\
We have presented a study which analyses the effects of scalar and pseudoscalar particles in leptonic and semileptonic decays arising from the transition $b\rightarrow u \ell \bar{\nu}_{\ell}$. Central to our analysis is the high sensitivity of the leptonic $B$ decays to the presence of NP pseudoscalar contributions. To take advantage of this feature, we have used the most recent Belle measurement of the branching fraction of the channel $B^-\rightarrow \mu^- \bar{\nu}_{\mu}$.\
We have developed a strategy with three main goals. Firstly, we have obtained the corresponding short-distance NP contributions, utilizing ratios of leptonic and semileptonic processes where the CKM element $|V_{ub}|$ cancels. Secondly, we have determined the value of $|V_{ub}|$, considering simultaneously the presence of NP effects. Finally, we have made predictions for branching fractions of non-yet measured $B$ decays. In particular, we have addressed the decay channels $B^- \to e^- \bar{\nu}_e$ and $\bar{B}\to \rho \tau \bar{\nu}_\tau$, and found NP effects that may be within the reach of the Belle II collaboration and future updates of the LHCb experiment.\
Our analysis includes the semileptonic processes $\bar{B}\rightarrow \pi \ell \bar{\nu}_{\ell}$ and $\bar{B}\rightarrow \rho \ell \bar{\nu}_{\ell}$ with $\ell=e,~\mu,~\tau$, where the current experimental data does not provide information on electrons and muons separately. In order to test universality in light leptons of different flavours, it would be desirable that experimental collaborations provide independent measurements for $\ell=e$ and $\ell=\mu$. In addition, a better understanding of the behaviour of the branching fraction of the process $\bar{B}\rightarrow \rho \ell \bar{\nu}_{\ell}$ requires an update of the corresponding non-perturbative contributions in the high $q^2$ regime. Following these lines we will be able to take further advantage of these semileptonic $B$ decays to search for NP. This will complement the ongoing searches in their counterparts $\bar{B}\rightarrow D^{(*)} \ell \bar{\nu}_{\ell}$, which enter the observables $R_{D^{(*)}}$.\
#### Funding information
This research project was supported by the Netherlands Foundation for Fundamental Research of Matter (FOM) programme 156, “Higgs as Probe and Portal”, and by the National Organisation for Scientific Research (NWO).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Lanthanide elements play important roles as an opacity source in the ejected material from neutron star mergers. Accurate and complete atomic data are necessary to evaluate the opacities and to analyze the observed data. In this paper, we perform extended, [*ab-initio*]{} atomic calculations from Pr II (Z=59) to Gd II (Z=64). By using multiconfiguration Dirac-Hartree-Fock and relativistic configuration-interaction methods, implemented in the general-purpose relativistic atomic structure package GRASP2K, we calculate the energy levels and transition data of electric dipole transitions. These computations are based on strategies (with small variations) of Nd II published by @Nd_jonai. Accuracy of data is evaluated by comparing computed energy levels with the NIST database or other works. For the energy levels, we obtain the average relative accuracy of 8%, 12%, 6%, 8%, and 7% for Pr II, Pm II, Sm II, Eu II, and Gd II ions, respectively as compared with the NIST data. Accuracy of energy transfer to the wavelength as 3%, 14% and 11% for Pr II, Eu II and Gd II. Our computed E1 type transition probabilities are in good agreement with experimental values presented by other authors especially for strong transitions.'
author:
- Laima Radžiūtė
- Gediminas Gaigalas
- Daiji Kato
- Pavel Rynkun
- Masaomi Tanaka
bibliography:
- 'reference.bib'
title: 'Extended calculations of energy levels and transition rates for singly ionized lanthanide elements I: Pr - Gd'
---
Introduction {#sec:intro}
============
Atomic opacities of heavy elements have a wide impact to astrophysics. In particular, recent observations of gravitational waves and electromagnetic waves from a neutron star merger (GW170817, @abbott17MMA) highlight the needs for heavy-element opacities. In optical and infrared wavelengths, the electromagnetic counterpart of GW170817 shows characteristics of kilonova, emission powered by radioactive decays of newly synthesized $r$-process (or rapid neutron capture process) nuclei. To study the $r$-process nucleosynthesis from the observed emission, we need to accurately understand the opacities of lanthanide elements since properties of kilonova are mainly governed by bound-bound opacities of $r$-process elements and lanthanide elements give the largest contributions (@kasen13; @barnes13; @tanaka13).
Several works have been done to study the properties and opacities of lanthanide elements (@kasen13 [@fontes17; @tanaka18; @tanaka19]). However, atomic calculations to evaluate the total opacities are not necessarily accurate enough to give a wavelength and a transition probability of each transition [@tanaka19]. Recently, @watson19 reported identification of Sr in the spectra of kilonova associated with GW170817. In principle, other elements can also be identified in the spectra. However, the line list used for astrophysics is not neccesarily complete even for strong transitions, in particular, in infrared wavelengths. By these reasons, it is still not straightforward to fully decode the spectra of kilonova. Accurate atomic calculations of lanthanide elements, therefore, play an important role as a benchmark to give accurate atomic data [@Nd_jonai].
There are many semi-empirical works which provide accurate atomic data of the lanthanide elements. In these works, the Racah-Slater parametric method is used [@Racha-Slater]. This method is known to give an excellent agreement between calculated energies using fitted radial parameters and available experimental energies. However, correct level identification of experimental spectra is needed, which is not always available. On the other hand, [*ab-initio*]{} methods can provide complete atomic data set without any empirical parameter. Nevertheless, there are few applications of such [*ab-initio*]{} methods for lanthanide with spectroscopic accuracy. This is because systematic improvement of subtle correlation effects in complicated atomic structures of open-$4f$ shell is not studied thoroughly.
In our previous paper [@Nd_jonai], we have performed accurate calculations for Nd ions. In this paper, we extend our calculations to Pr II, Pm II, Sm II, Eu II, and Gd II. Namely, we perform energy spectrum computations for states of the following configurations: \[Xe\]$4f^N \{6s,5d,6p\}$ and \[Xe\]$4f^{N-1} \{5d6s,5d6p,6s6p, 5d^2\}$ for $N=3,5,6,7,8$. We also perform energy spectrum computations for states of \[Xe\]$4f^{N+1}$ configuration for Sm II and Eu II, and \[Xe\]$4f^{N-1} 6s^2$ configuration for Gd II. Levels up to 10 eV are computed since such low-lying energy levels play dominant roles in the opacities in the neutron star merger ejecta at typical temperature of 5,000 K [@Nd_jonai]. Using these results, electric dipole (E1) transitions data were computed between these states. In this paper, we aim at providing complete atomic data with the overall accuracy of about 10%. This accuracy is not high enough to directly compare with spectroscopic experiments, but it is adequate to evaluate the opacities (“opacity accuracy” rather than “spectroscopic accuracy”, @Nd_jonai). In fact, typical accuracy of complete atomic calculations [@kasen13; @tanaka18] is much lower than the accuracy presented in this paper. The calculations are done using multiconfiguration Dirac-Hartree-Fock (MCDHF) and relativistic configuration-interaction (RCI) methods [@grant; @topical_rev], which are implemented in the general-purpose relativistic atomic structure package GRASP2K [@graspV3]. We employ a strategy similar to the @Nd_jonai including electron correlation, which is suitable for series of rare earth ions. For low lying levels, higher accuracy can be achieved using computational schemes including more electron correlations as in @ErIII. In addtion, there is an advantage in the computation since large computational tasks can be split in to smaller tasks by using this method. In Section \[sec:method\], we describe our method and strategy of calculations. Then, we show results of energy level structure and transition probabilities in Sections \[sec:evaluation\] and \[sec:transitions\], respectively. Finally we give summary in Section \[sec:summary\].
Methods {#sec:method}
=======
Computational procedure
-----------------------
The computational methods used in this paper follow the methods used in @Nd_jonai. Therefore, we briefly outline the methods in this section. We refer the reader to [@topical_rev] for further details. We use the MCDHF method, based on Dirac-Coulomb Hamiltonian, in this work. The atomic state functions (ASFs) are expressed by a linear combination of symmetry adapted configuration state functions (CSFs). The CSFs are built from products of one-electron Dirac orbitals. The radial parts of the Dirac orbitals and the expansion coefficients are optimized to self-consistency in the relativistic self-consistent field procedure.
The spin-angular approach [@Gaigalas_1996; @Gaigalas_1997] is used in these computations. The approach is based on the second quantization in a coupled tensorial form, on the angular momentum theory in the orbital, spin, and quasispin spaces and on the reduced coefficients of fractional parentage. It allows us to study configurations with open $f$-shells without any restrictions.
[c c c c c c c c c ]{}\[ht!!\] Pr II & 927 & 1 218 && 29 129 & 45 045\
Nd II\*& 3 270 & 2 813 && 188 357 & 113 900\
Pm II & 5 206 & 4 568 && 380 588 & 518 957\
Sm II & 3 153 & 5 240 && 1 272 634 & 2 133 183\
Eu II & 1 306 & 1 241 && 1 501 949 & 2 201 859\
Gd II & 2 035 & 2 335 && 3 033 793 & 1 721 371\
In the following RCI calculations, the Breit interaction is included in the Hamiltonian. In the RCI calculation, the leading quantum electrodynamics corrections (QED), self-interaction and vacuum polarization are also included.
The label of the ASF is the same as the label of the dominating CSF. The ASFs are obtained as expansions over $jj$-coupled CSFs. To provide the ASFs in the $LSJ$ labeling system, transformation from a $jj$-coupled CSF basis to an $LSJ$-coupled CSF basis has been done [@jj2lsj_atoms]. Review on all these methods and on the GRASP2K package can be found in @topical_rev.
Computation of transition probabilities
---------------------------------------
For electric dipole transitions there are two forms of the transition operator: the length (Babushkin) and velocity (Coulomb) forms. Although the exact solutions of the Dirac-equation should give the same value of the transition moment [@gauge], they do not necessarily agree in numerical calculations. The quantity $dT = | A_{\rm l} - A_{\rm v} | / {\rm max} (A_{\rm l}, A_{\rm v})$ [@ekman] defines the accuracy of the computed transition rates, where $A_{\rm l}$ and $A_{\rm v}$ are the transition rates in length and velocity forms, respectively.
The calculation of the transition moment breaks down in the task of summing up reduced matrix elements between different CSFs. Using standard techniques, by assuming that both left and right hand CSFs are formed from the same orthonormal set of spin-orbitals, the reduced matrix elements can be evaluated. This constraint is severe, since a high-quality and compact wave function requires orbitals optimized for a specific electronic state (see for example @SF). To avoid the problems of having a single orthonormal set of spin-orbitals, the wave-function representations of the two states are transformed in a way that the orbital sets became biorthonormal [@biotra]. To evaluate the matrix elements of the transformed CSFs, standard methods as in @topical_rev are used.
Computational Schemes {#sec:scheme}
---------------------
To compute singly ionized lanthanide elements, the **strategy C** by [@Nd_jonai] is used. Details of this strategy and extension of it are given below. Active space method is used for computation of energy levels and E1 transitions. The configuration space is increased step by step, by increasing the number of layers (L), that is, a set of virtual orbitals. The virtual orbitals of the increased layer are optimized in the relativistic self-consistent field procedure, while all orbitals of inner layers are fixed. The scheme used to increase the active spaces of the CSF’s is presented below:\
AS$_{0L}$ = $\{6s, 6p, 5d\}$,\
AS$_{1L}$ = AS$_{0L}$ + $\{7s, 7p, 6d, 5f\}$,\
AS$_{2L}$ = AS$_{1L}$ + $\{8s, 8p, 7d, 6f, 5g\}$.\
The number of computed levels and CSFs in the final even and odd state expansions are presented in Table \[summary\]. Computations are performed for each configuration separately (single reference method). This method allows to split the large computations into several tasks. In each task, the wave function expansion for a single reference configuration is constructed by substitution of one and two electrons from the reference configuration. For configurations $4f^N 6s$, $4f^N 6p$ and $4f^N 5d$, single and/or double (SD) substitutions are allowed from $4f^N nl$ shells ($l=s,p,d$) to $AS_{0L,1L}$ and single (S) substitutions are allowed to $AS_{2L}$. For configurations $4f^{N-1} 5d6s$, $4f^{N-1} 5d6p$, $4f^{N-1} 6s6p$, and $4f^{N-1} 5d^2$, only S substitutions are allowed. For Sm II and Eu II ions, a new configuration $4f^{N+1}$, which was not taken into account in **the strategy C** of [@Nd_jonai], is computed. For this configuration, single, double, and triple (SDT) substitutions are allowed from $4f^{N+1}$ shell to $AS_{0L,1L}$ and SD substitutions are allowed to $AS_{2L}$. For configuration $4f^{N-1} nl n'l'$, two electrons are excited from $4f$ orbital, and for $4f^N nl$, only one electron is excited from $4f$ orbital. Therefore, to include compensated correlations, we need to make less excitations from the first configuration and more excitations from the second one. For example, if we do SD substitutions for $4f^{N-1} nl n'l'$ configuration, we need to make SDT substitutions for $4f^N nl$ configuration.
To compute energy levels, it is important to have correct core radial wave functions, that is, initial Dirac-Fock (DF) computations. Correct selection of the core stabilizes solution of self-consistent field computation. We find that core radial wave functions \[Xe\]$4f$ from the ground configuration \[Xe\]$4f^N6s$ are the best solution. Radial wave functions up to $4f, 5s, 5p$ orbital are taken from the ground configuration for these configurations $4f^{N-1} 5d6s$, $4f^{N-1} 5d6p$, $4f^{N-1} 6s6p$, and $4f^{N-1} 5d^2$. Meanwhile, the radial wave functions were computed for each configurations $4f^N nl$ ($l=s,p,d$) separately.
For neutral atoms and ions of lanthanide elements with different ground configurations, we suggest that their ground configuration radial wave functions are used as common core. For example, for neutral lanthanides, radial wave functions of the ground configurations \[Xe\]$4f^N6s^2$ can be used as common core.
![Comparison of energy levels for odd and even configurations of Pr II with experimental values. References (1) @Rosen_Pr; (2) @Blaise_Pr and @Blaise_Pr_b; (3) @Ginibre_Pr_1989_a; (4) @Ivarsoon_Pr; (5) @Furmann_Pr_a, [@Furmann_Pr_b] and [@Furmann_Pr_c]; (6) @Akhtar_Pr; (7) Our computed levels. NL is the number of levels. \[Pr\_II\_su\_kitais\_autoriai\]](Pr_II_su_kitais_autoriai.pdf){width="47.00000%"}
For Eu II and Gd II, wave function is investigated differently due to the rapid increase of the number of configuration state functions in the active space (see Table \[summary\]). For these ions, self-consistent field computations are performed not for all $J$ values but only for one $J$ value. Then, using computed radial wave functions, RCI computations are performed. For example, for the configuration of Eu, $4f^5 6s$ atomic states only with $J=4$ are computed and it is later used in the RCI computation for $J =0-13$. For all configurations, the lowest $J$ values are selected for computation of the radial wave functions. This computational method demands less computational resources.
{width="50.00000%"} {width="50.00000%"} {width="50.00000%"} {width="50.00000%"} {width="50.00000%"} {width="50.00000%"}
{width="50.00000%"} {width="50.00000%"} {width="50.00000%"} {width="50.00000%"} {width="50.00000%"} {width="50.00000%"}
[c c c c c c c c c cc ccccccc]{}\[ht!!\] $4f^N 6s$ & 6 & 7 && 15 & 27&& 13& 17&& 8 & 12&& 23 & 1 && 4 &13\
$4f^N 6p$ & 4 &12 && 13 & 23&& -& -&& - & -&& 10 & 6 && 6 &30\
$4f^N 5d$ &10 &33 && 8 & 47&& 10& 5&& 5 & 52&& 9 &10 && 3 &56\
$4f^{N-1} 5d6s$ & - & - && 2 & 14&& -& -&& - & -&& 15 & 8 &&19 &19\
$4f^{N-1} 5d6p$ & - & - && 6 & 12&& -& -&& - & -&& - & - && 2 &49\
$4f^{N-1} 6s6p$ & - &- && 4 & 13&& -& -&& - & -&& - & - && 8 & 6\
$4f^{N-1} 5d^2$ &15 &1 && 15 & 22&& -& -&& - & -&& 13 & 3 &&13 &39\
$4f^{N-1} 6s^2$ & & && & && & && & && & &&66 & 1\
$4f^{N+1}$ & & && & && & && 14 & 1&& - & - && &\
$4f^N 7s$ & & && & && & && & && 3 & 2 && &\
$4f^N 8s$ & & && & && & && & && 9 & 1 && &\
$4f^N 6d$ & & && & && & && & && 2 &10 && &\
all & 8 &53 &&10 &158&& 12&22 && 6 & 65&& 8 &41 &&7 &213\
In addition, some states of Rydberg series (up to 10 eV) are computed for Eu II. This includes 38 levels from configurations $4f^6\{7s,8s,6d,7d,7p,8p\}$. Radial wave functions for configurations $4f^6\{7s,8s\}$ up to $4f$ are taken from the ground configuration ($4f^6 6s$). For the rest configurations, radial wave functions are computed in the same manner as in the configurations $4f^6\{5d,6p\}$. This means that each configuration from $4f^6\{6d,7d,7p,8p\}$ has different radial wave functions. Active space generated in a similar manner as for the configurations $4f^6\{6s,5d,6p\}$. For example, active space for the configuration are generated by SD substitutions from $4f^6~8s$ to AS$_{0L}$ = $\{6p, 5d\}$ and AS$_{1L}$ = AS$_{0L}$ + $\{6s, 7p, 6d, 5f\}$ and by S substitutions to AS$_{2L}$ = AS$_{1L}$ + $\{7s, 8p, 7d, 6f, 5g\}$.
For Gd II, radial wave function is generated also for only one $J$ value. Radial wave functions of $4f^7 5d^2$, $4f^7 6s^2$, $4f^7 5d6s$, and $4f^8 6p$ are computed together, using radial wave function of configuration $4f^8 6s$ up to $4f$. Rest of configurations are computed in the same manner as for Eu II. The MCDHF calculations are then followed by RCI calculations by including the Breit interaction and leading QED effects. The same active space (AS$_{2L}$) is used for the RCI computations as well as for MCDHF computations.
Energy levels {#sec:evaluation}
=============
All levels for $Z=59 - 64$ ions are given in Figure \[Pr\_II\_pilnas\], and the energy data computed for Pr II, Pm II, Sm II, Eu II, and Gd II are given in machine-readable format in Tables \[Pr\_II\_Energies\], \[Pm\_II\_Energies\], \[Sm\_II\_Energies\], \[Eu\_II\_Energies\], and \[Gd\_II\_Energies\], respectively. This includes label, $J$ and $P$ values, and energy value. Levels are given in $LS$-coupling, although it is suitable only for the lowest states of configurations and determination the configuration is complicated for higher states [@Cowan]. For the labels, we use notation $4f^N~^{(2S+1)}_{Nr} L ~n'l' ~^{(2S'+1)}L'$. Intermediate quantum numbers define parent levels $4f^N~^{(2S+1)}_{Nr} L$, where $N$ is electron number in $4f$ shell, $(2S+1)$ is multiplicity, $Nr$ is a sequential index number representing the group labels $nWU$ for the term, and $L$ is orbital quantum number (see @senioriy_2 for more about $Nr$). More complicated configurations are presented in the similar way.
To evaluate the accuracy of our calculations, comparison with critically evaluated data is necessary. In this section, we first summarize the available data for energy levels of Pr II, Pm II, Sm II, Eu II, Gd II in the NIST database. Then, we compare calculated energy levels with these available data.
Available data {#sec:nist}
--------------
### Pr II {#Pr_II_energies}
@Ginibre_Pr_1989_a have investigated 105 odd and 187 even experimental energies based on Fourier transform (FT) spectroscopy in range 2 783 - 27 920 cm$^{-1}$. Also, the large amount levels were investigated by @Rosen_Pr, @Blaise_Pr, and @Blaise_Pr_b. They performed semi-empirical fitting procedure to assign for some levels labels in $LS$-coupling [@Ginibre_Pr_1989_b]. Later, [@Ivarsoon_Pr] presented improved 39 energy levels using FT spectroscopy in 2 800 - 8 000 Å region. [@Furmann_Pr_a; @Furmann_Pr_b; @Furmann_Pr_c] investigated 31 odd and 14 even levels, using laser-induced fluorescence spectroscopy (LIF) in a hollow cathode discharge lamp. More recently, @Akhtar_Pr have redetermined energy values of 227 levels (74 having odd and 153 even parity) and hyperfine structures of 477 transitions in the range of 3 260 - 11 700 Å. They corrected the energy levels from the works of @Ginibre_Pr_1989_a and [@Ivarsoon_Pr].
All of these levels are measured/reanalyzed in high accuracy. However, each work presents energy levels in a narrow range as shown in Figure \[Pr\_II\_su\_kitais\_autoriai\]. Therefore, the transitions between measured energy levels give too small amount of lines needed for computation of opacities in neutron star mergers. Data of these authors are summarized by [@Martin]. Since the NIST database [@NIST] includes the work by [@Martin], we only give comparison with the NIST database here.
### Pm II {#Pm_II_energies}
Pm II is one of the ions whose spectrum is not well investigated. Energy levels of two configurations $4f^5 6s$ and $4f^5 5d$ were investigated by [@Martin]. Five new levels of $4f^5 5d$ configuration were measured by [@Otto_Pm] with the collinear laser ion beam spectroscopy (CLIBS) method and were identified using Hartree-Fock method.
### Sm II
[@Albertson_Sm] have assigned terms of 40 even levels of the $4f^6 6s$ and $4f^6 5d$ configurations based on the Zeeman patterns of over 300 lines. [@Spector_Sm_Gd] have done semi-empirical computation of energy values and $LS$-composition of 55 levels for $4f^6(^7F) 5d$ configuration. Also, a large amount of work for energy levels was done by [@Blaise_Sm]: 325 levels for Sm II were obtained from the Zeeman effect measurement in the visible and the ultraviolet spectrum. Then, these energy levels were re-evaluated by [@Martin]. Attempt of identification of odd configurations for some levels was done by [@Rao_Sm] using isotope shifts data, which was carried out on a recording Fabry-Perot spectrometer. The hyperfine structure and isotope shift were also measured by collinear fast ion beam laser spectroscopy. These data were used to assign configurations to the 13 odd upper levels by [@Villemoes_Sm]. Note that some of them do not have identification by [@Martin].
### Eu II
156 levels of configurations $4f^7\{6s,7s,8s,5d,6d,6p\}$ and $4f^6 5d6s$, $4f^6 5d^2$ were resolved with the spark spectrum of arc by [@Russel]. This work is the extension of the analysis by [@Albertson] on 9 levels of $4f^7\{6s,5d,6p\}$ configurations. Then these energy levels were re-evaluated by [@Martin]. More recently, 13 new energy levels of $4f^7 6s$ configuration were suggested from hyperfine constant and isotope shift measurements [@Furmann_Eu].
![The difference between the lowest level of $4f^{N-1} 5d6s$ and the lowest level of $4f^{N} 6s$ for singly ionized lanthanide elements. Red circles are our computed theoretical values, while blue squares are values recommended by the NIST database. Black symbols indicate data from [@Martin_SD] and [@Cowan]: closed circles are predicted values, open circles are experimental values, and triangles are estimated data based on incomplete experimental data. \[Kowan\_type\]](Kowan_type.pdf){width="47.00000%"}
{width="47.00000%"} {width="47.00000%"} {width="47.00000%"} {width="47.00000%"} {width="47.00000%"} {width="47.00000%"} {width="47.00000%"} {width="47.00000%"}
### Gd II
[@Albertson_Gd] have investigated 9 odd and 11 even energy levels, have deduced quantum numbers from Zeeman effect pattern, and have established the ground configurations to be $4f^75d6s$. [@Venugopalan_Gd] and [@Venugopalan_II_Gd] have measured isotope shift of 33 spectroscopic lines, using photoelectric recording Fabry-Perot spectrometer. They suggested new configuration identification of 4 high energy levels (lying above 35 000 cm$^{-1}$): 35 362.630 cm$^{-1}$ ($J = 13/2$) as $4f^7 5d6s$; 35 822.697 cm$^{-1}$ ($J = 9/2$) as mix of two configurations $4f^75d6s$+$4f^86p$; 37 831.032 cm$^{-1}$ ($J = 11/2$) and 38 010.603 cm$^{-1}$ ($J = 11/2$) as $4f^86p$. [@Blaise_Gd] have done the analysis of the spark spectrum of Gd II of 178 new levels. Total 30 levels were ascribed to $4f^8 (^7F) 6p$ configuration by their strong transitions with the levels on the $4f^8 (^7F) 6s$ and $4f^8(^7F) 5d$ sub-configurations. [@Spector_Sm_Gd] have done semi-empirical computation of energy values and $LS$-composition of 57 levels for $4f^8(^7F) 5d$ configuration. 164 odd and 150 even parity energy levels of Gd II are listed by @Martin. [@Spector_Gd] have done extended analysis on levels of the configurations $4f^8(^7F)\{6s,6p,5d\}$ and measured new levels of $4f^8(^7F)\{6s,5d\}$ configurations and new odd levels.
[c c c c c c c c c ]{}\[ht!!\] Pr II & 1.0833& 1.0986 & 2.5036 & 2.5484 & 4.3130\
I07 & 1.0589& 1.0667 & & & 4.2924\
Nd II\*& 1.0291& 1.0440 & 2.4607 & 2.5085 & 4.2522\
I07 & 1.0054& 1.0190 & & & 4.2252\
Pm II & 0.9832& 0.9981 & 2.4223 & 2.4717 & 4.1948\
I07 & 0.9624& 0.9796 & & & 4.1608\
Sm II & 0.9442& 0.9590 & 2.3892 & 2.4400 & 4.1402\
I07 & 0.9249& 0.9392 & & & 4.1012\
Eu II & 0.9098& 0.9256 & 2.3642 & 2.4272 & 4.0870\
I07 & 0.8920& 0.8999 & & & 4.0438\
Gd II & 0.8797& 0.8929 & 2.3991 & 2.5041 & 3.6878\
I07 & 0.8218& 0.8221 & 2.4547 & 2.4846 & 3.7930\
Comparison of the energy levels
-------------------------------
The energy levels for each configuration are compared with those in the NIST database in Figure \[Pr\_II\_pilnas\]. Only the common configurations for Pr II - Gd II are presented in the figure. Although the energy levels in the NIST database sometimes include questionable identification of the configuration, this figure includes all levels.
To analyse the accuracy of our calculations as compared with the NIST data, we use an expression $\Delta E_{i}/E_{\rm NIST} = (E_{\rm NIST} - E_{i}) / E_{\rm NIST} \times 100\% $ . For the indicator of the accuracy for many levels, we use a value $\overline{\Delta E/E} = \sum \frac{|\Delta E_{i}/E_{\rm NIST}|}{N}$, where $N$ is the number of the compared levels. Summary of the accuracy for each configuration is given in Table \[summary\_accuracy\]. Levels with unquestionable identification are included in to the comparison. Empty space in Table \[summary\_accuracy\] means that configuration is not computed while a mark with “-” means that data are missing in the NIST database (or there is only one level). The last line (all) of the table presents averaged accuracy with unquestionable identification between our results and the NIST database.
Overall, we find that our calculations give good accuracy: 8%, 12%, 6%, 8%, and 7% for Pr II, Pm II, Sm II, Eu II, and Gd II ions, respectively. There is no clear trend with the atomic number $Z$. The accuracy depends on the configurations. For example, the degree of agreement for $4f^N 6s$, $4f^N 5d$ and $4f^N 6p$ configurations slightly differ. These variations of the differences is mainly caused by the different number of levels used for comparison. Note that the biggest deviation is found for level $^8S_{7/2}$ of configuration $4f^7~6s^2$ (66% difference for this 1 level).
[rlrrr]{}\[ht!!\] 1 & $4f^{3}(^4_1I)~6s~^{5}I $ & 4 & $-$ & 0.00\
2 & $4f^{3}(^4_1I)~6s~^{5}I $ & 5 & $-$ & 511.12\
3 & $4f^{3}(^4_1I)~6s~^{5}I $ & 6 & $-$ & 1558.26\
4 & $4f^{3}(^4_1I)~6s~^{3}I $ & 5 & $-$ & 1772.15\
5 & $4f^{3}(^4_1I)~6s~^{5}I $ & 7 & $-$ & 2773.88\
6 & $4f^{3}(^4_1I)~6s~^{3}I $ & 6 & $-$ & 3337.17\
7 & $4f^{3}(^4_1I)~5d~^{5}L $ & 6 & $-$ & 3506.00\
8 & $4f^{3}(^4_1I)~5d~^{5}K $ & 5 & $-$ & 3948.78\
9 & $4f^{3}(^4_1I)~6s~^{5}I $ & 8 & $-$ & 4104.37\
10 & $4f^{3}(^4_1I)~5d~^{5}L $ & 7 & $-$ & 4553.26\
11 & $4f^{3}(^4_1I)~5d~^{5}K $ & 6 & $-$ & 4898.65\
12 & $4f^{3}(^4_1I)~6s~^{3}I $ & 7 & $-$ & 4937.25\
13 & $4f^{2}(^3_1H)~5d^{2}(^3_2F)~^5L $ & 6 & $+$ & 4948.47\
14 & $4f^{3}(^4_1I)~5d~^{5}L $ & 8 & $-$ & 5703.77\
15 & $4f^{3}(^4_1I)~5d~^{5}K $ & 7 & $-$ & 5941.26\
16 & $4f^{2}(^3_1H)~5d^{2}(^3_2F)~^5L $ & 7 & $+$ & 6265.05\
17 & $4f^{2}(^3_1H)~5d^{2}(^3_2F)~^5I $ & 4 & $+$ & 6861.72\
18 & $4f^{3}(^4_1I)~5d~^{5}L $ & 9 & $-$ & 6944.26\
19 & $4f^{3}(^4_1I)~5d~^{5}K $ & 8 & $-$ & 7067.68\
20 & $4f^{2}(^3_1H)~5d^{2}(^3_2F)~^5K $ & 5 & $+$ & 7591.81\
As mentioned in Section \[sec:scheme\], computations of Eu II are performed in a slightly different manner: radial wave functions are computed only for one $J$ symmetry of the lowest ASF. To test the influence of such splitting, we compute configurations $4f^7{6s}$, $4f^7{5d}$, $4f^6 5d6s$, $4f^6 5d6p$, and $4f^6 5d^2$ in both ways. We find that the differences between two methods are small: the maximum averaged difference of energy levels per configuration is 0.5% for $4f^7{5d}$ configuration (614 levels) and the minimum difference is 0.02% for $4f^7{6s}$ (261 levels). Levels of Rydberg state of the configurations $4f^7\{7s,8s,6d\}$ for Eu II are also compared in Table \[summary\_accuracy\]. There is a good agreement for levels of configurations $4f^7\{7s,6d\}$ obtained in this research with values from NIST database.
Figure \[Statiska\] show the histogram of the relative difference compared to the NIST for all computed ions. This figure includes only the levels with the exact identification. Note that the number of the available energy levels has a large variation as summarized in Section \[sec:nist\]. The biggest numbers of levels are available for the Nd II and the Gd II in the NIST database, and thus, the distribution is close to the normal distribution for these ions.
The accuracy of our calculations can also be evaluated using Figure \[Kowan\_type\], which shows the energy difference between the lowest levels of $4f^{N-1} 5d6s$ and the lowest levels of $4f^{N} 6s$ configurations for singly ionized lanthanides. As shown in the figure, the overall agreement is very good. Our results and those in the NIST database give smaller energy differences than those in [@Martin_SD] and [@Cowan] for Nd II and Pm II ions. The increase of the energy difference is observed for Eu II by all the works, but our result shows a bigger increase than in @Martin_SD, [@Cowan] and the NIST data. Here it should be noted that, for the cases of Pr II and Sm II, the identification of $4f^{N-1} 5d6s$ configurations are questionable in the NIST database. More detailed investigation was done by [@Brewer_Sm] (see their Figure 1). They have estimated energies for lowest levels of configurations involving $4f$, $5d$, $6p$, and $6s$ shells for singly-triply ionized lanthanides and actinides. Their computations are based on the thermodynamic data of the metals. In a similar manner, the energy differences were also analysed by [@Vander_Sm]. In fact, our results are very close to the data of these authors.
[rlrrr]{} 1 &$ 4f^{5} (^6_0H)~6s~^7H $ & 2 & $-$ & 0.00\
2 &$ 4f^{5} (^6_0H)~6s~^7H $ & 3 & $-$ & 441.78\
3 &$ 4f^{5} (^6_0H)~6s~^7H $ & 4 & $-$ & 1073.15\
4 &$ 4f^{5} (^6_0H)~6s~^5H $ & 3 & $-$ & 1850.51\
5 &$ 4f^{5} (^6_0H)~6s~^7H $ & 5 & $-$ & 1858.38\
6 &$ 4f^{5} (^6_0H)~6s~^7H $ & 6 & $-$ & 2765.09\
7 &$ 4f^{5} (^6_0H)~6s~^5H $ & 4 & $-$ & 2830.31\
8 &$ 4f^{5} (^6_0H)~6s~^7H $ & 7 & $-$ & 3765.53\
9 &$ 4f^{5} (^6_0H)~6s~^5H $ & 5 & $-$ & 3914.91\
10 &$ 4f^{5} (^6_0H)~5d~^7K $ & 4 & $-$ & 4799.46\
11 &$ 4f^{5} (^6_0H)~6s~^7H $ & 8 & $-$ & 4836.02\
12 &$ 4f^{5} (^6_0H)~6s~^5H $ & 6 & $-$ & 5078.38\
13 &$ 4f^{5} (^6_0H)~5d~^7K $ & 5 & $-$ & 5507.08\
14 &$ 4f^{5} (^6_0H)~6s~^5H $ & 7 & $-$ & 6298.26\
15 &$ 4f^{5} (^6_0H)~5d~^7K $ & 6 & $-$ & 6322.77\
16 &$ 4f^{5} (^6_0F)~6s~^7F $ & 0 & $-$ & 6567.93\
17 &$ 4f^{5} (^6_0F)~6s~^7F $ & 1 & $-$ & 6678.52\
18 &$ 4f^{5} (^6_0F)~6s~^7F $ & 2 & $-$ & 6917.29\
19 &$ 4f^{5} (^6_0H)~5d~^7K $ & 7 & $-$ & 7232.66\
20 &$ 4f^{5} (^6_0F)~6s~^7F $ & 3 & $-$ & 7320.95\
Energy level distribution for each configuration {#E_l_d_e_c}
------------------------------------------------
Identification of energy levels is a complicated task for lanthanides due to a mix of configurations. Even assigning particular configuration labeling to some levels is complicated. The discussion below should give enlightenment on the inner structure of the energy spectrum. Energy levels have formed groups around parent level of $4f^N$ or $4f^{N-1}$ configurations with the same term of $f$ shell. Levels belonging to the different groups are separated by energy gaps. Below are given more details about these groups for each configuration.
Energy level structures for states of $4f^N \{6s,6p,5d\}$, $4f^{N-1} \{5d6s,5d6p,6s6p\}$, and $4f^{N-1} 5d^2$ configurations are presented in Figure \[all\_configurations\_energies\]. Cut off line 80 700 cm$^{-1}$ (10 eV) is given by the horizontal lines. For the Pr II and Nd II, computations are done up to ionizations limits: it is 85 745 cm$^{-1}$ for Pr II and 86 970 cm$^{-1}$ for Nd II according to the NIST database. The number of computed levels are displayed below the line and the number of levels above the line are left uncomputed. The sum of these numbers comprise a possible number of levels in $jj$-coupling. We find that the increase of the nuclear charge has a small effect on the positions of first level relative to the ground state for the configurations $4f^N 6p$ and $4f^N 5d$. The energy level structures of these configurations are influenced by the structure of core \[Xe\]$4f^N$. Similar to the system difference analysed by [@Cowan], the increase of the energy of first level relative to the ground state is found for the configurations of $4f^{N-1} 5d6s$, $4f^{N-1} 5d6p$, $4f^{N-1} 6s6p$, and $4f^{N-1} 5d^2$ (see Figure \[Kowan\_type\] for $4f^{N-1} 5d6s$). The highest density of the energy levels are found for $4f^{N-1} 5d^2$ and $4f^{N-1} 5d6p$ configurations.
[rlrrr]{} 1 &$ 4f^{6} (^7_0F)~6s~^8F $ & 1/2 & $ + $ & 0.00\
2 &$ 4f^{6} (^7_0F)~6s~^8F $ & 3/2 & $ + $ & 296.55\
3 &$ 4f^{6} (^7_0F)~6s~^8F $ & 5/2 & $ + $ & 765.67\
4 &$ 4f^{6} (^7_0F)~6s~^8F $ & 7/2 & $ + $ & 1372.98\
5 &$ 4f^{6} (^7_0F)~6s~^6F $ & 1/2 & $ + $ & 1853.48\
6 &$ 4f^{6} (^7_0F)~6s~^8F $ & 9/2 & $ + $ & 2084.20\
7 &$ 4f^{6} (^7_0F)~6s~^6F $ & 3/2 & $ + $ & 2285.57\
8 &$ 4f^{6} (^7_0F)~6s~^8F $ & 11/2 & $ + $ & 2870.66\
9 &$ 4f^{6} (^7_0F)~6s~^6F $ & 5/2 & $ + $ & 2918.77\
10 &$ 4f^{6} (^7_0F)~6s~^6F $ & 7/2 & $ + $ & 3688.53\
11 &$ 4f^{6} (^7_0F)~6s~^8F $ & 13/2 & $ + $ & 3709.70\
12 &$ 4f^{6} (^7_0F)~6s~^6F $ & 9/2 & $ + $ & 4548.72\
13 &$ 4f^{6} (^7_0F)~6s~^6F $ & 11/2 & $ + $ & 5467.12\
14 &$ 4f^{6} (^7_0F)~5d~^8H $ & 3/2 & $ + $ & 6571.55\
15 &$ 4f^{6} (^7_0F)~5d~^8H $ & 5/2 & $ + $ & 6913.83\
16 &$ 4f^{6} (^7_0F)~5d~^8H $ & 7/2 & $ + $ & 7375.65\
17 &$ 4f^{6} (^7_0F)~5d~^8H $ & 9/2 & $ + $ & 7942.79\
18 &$ 4f^{6} (^7_0F)~5d~^8D $ & 3/2 & $ + $ & 8488.11\
19 &$ 4f^{6} (^7_0F)~5d~^8H $ & 11/2 & $ + $ & 8600.85\
20 &$ 4f^{6} (^7_0F)~5d~^8D $ & 5/2 & $ + $ & 9058.38\
The lowest levels of $4f^N 6s$ and $4f^N 6p$ configuration form blocks of energy levels around the parent levels of $4f^N$ ($^4_1I^*$), ($^5_1I^*$), ($^6_0H$ and $^6_0F^*$), ($^7_0F^*$), ($^8_0S^*$, $^6_0P$, $^6_0I^*$, and $^6_0D^*$), and ($^7_0F^*$) for $Z=59 - 64$, respectively. After the levels with core configuration marked by “$^*$” above, there is an energy gap, except for $4f^3 6p$ of Pr II ion. Levels with the specific parent levels do not mix with others, except for the parent level states of $4f^5 6p$ of Pm II ion ($4f^5~^6_0H$ mix between $4f^5~^6_0F^*$). For $4f^N 5d$ configuration, the situation is different because of the strong interaction between $4f$ and $5d$ (Figure \[all\_configurations\_energies\]).
[rlrrr]{} 1 &$ 4f^7 (^8_0S)~6s~^9S $& 4 &$ - $& 0.00\
2 &$ 4f^7 (^8_0S)~6s~^7S $& 3 &$ - $& 2057.90\
3 &$ 4f^7 (^8_0S)~5d~^9D $& 2 &$ - $& 10657.96\
4 &$ 4f^7 (^8_0S)~5d~^9D $& 3 &$ - $& 10784.18\
5 &$ 4f^7 (^8_0S)~5d~^9D $& 4 &$ - $& 10964.35\
6 &$ 4f^7 (^8_0S)~5d~^9D $& 5 &$ - $& 11212.47\
7 &$ 4f^7 (^8_0S)~5d~^9D $& 6 &$ - $& 11551.69\
8 &$ 4f^7 (^8_0S)~5d~^7D $& 5 &$ - $& 18922.41\
9 &$ 4f^7 (^8_0S)~5d~^7D $& 4 &$ - $& 18964.72\
10 &$ 4f^7 (^8_0S)~5d~^7D $& 3 &$ - $& 19032.93\
11 &$ 4f^7 (^8_0S)~5d~^7D $& 2 &$ - $& 19095.86\
12 &$ 4f^7 (^8_0S)~5d~^7D $& 1 &$ - $& 19143.88\
13 &$ 4f^7 (^8_0S)~6p~^9P $& 3 &$ + $& 21378.10\
14 &$ 4f^7 (^8_0S)~6p~^9P $& 4 &$ + $& 21708.29\
15 &$ 4f^7 (^8_0S)~6p~^9P $& 5 &$ + $& 23385.62\
16 &$ 4f^7 (^8_0S)~6p~^7P $& 4 &$ + $& 23999.79\
17 &$ 4f^7 (^8_0S)~6p~^7P $& 3 &$ + $& 24276.27\
18 &$ 4f^7 (^8_0S)~6p~^7P $& 2 &$ + $& 24446.93\
19 &$ 4f^7 (^6_0P)~6s~^7P $& 4 &$ - $& 32530.49\
20 &$ 4f^7 (^6_0P)~6s~^7P $& 3 &$ - $& 32852.51\
For $4f^{N-1} 5d6s$ configuration, groups of energy levels are formed around the lowest parent levels for only two elements i.e., Eu II and Gd II. These parent levels are $4f^6~^7_0F$ and $4f^7~^8_0S^*$ for Eu II and Gd II, respectively. For $4f^{N-1} 5d6p$ configuration, only for Gd II has formed a group of energy levels around $4f^7~^8_0S^*$ parent level (Figure \[all\_configurations\_energies\]).
Levels of $4f^{N-1} 6s6p$ configuration do not form group of energy levels around the parent levels. For Eu II, all levels of $4f^6 6s6p$ and of $4f^6 5d6p$ configurations belong to the parent levels $4f^6~^7_0F$, because of the 10 eV cut off (Figure \[all\_configurations\_energies\]). For $4f^{N-1} 5d^2$ configuration, groups of energy levels forms around the lowest parent levels for Gd II $4f^8~^8_0S^*$ (Figure \[all\_configurations\_energies\]).
[rlrrr]{} 1 & $4f^7 (^8_0S)~5d~^9D 6s~^1D $ & 5/2 & $ - $ & 0.00\
2 & $4f^7 (^8_0S)~5d~^9D 6s~^1D $ & 7/2 & $ - $ & 225.02\
3 & $4f^7 (^8_0S)~5d~^9D 6s~^1D $ & 9/2 & $ - $ & 536.90\
4 & $4f^7 (^8_0S)~5d~^9D 6s~^1D $ & 11/2 & $ - $ & 959.98\
5 & $4f^7 (^8_0S)~5d~^9D 6s~^1D $ & 13/2 & $ - $ & 1536.81\
6 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 3/2 & $ - $ & 3026.08\
7 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 5/2 & $ - $ & 3173.49\
8 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 7/2 & $ - $ & 3382.26\
9 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 9/2 & $ - $ & 3654.80\
10 & $4f^7 (^8_0S)~5d~^9D 6s~^8D $ & 3/2 & $ - $ & 3767.49\
11 & $4f^7 (^8_0S)~5d~^9D 6s~^8D $ & 5/2 & $ - $ & 3965.84\
12 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 11/2 & $ - $ & 3994.44\
13 & $4f^7 (^8_0S)~5d~^9D 6s~^8D $ & 7/2 & $ - $ & 4269.62\
14 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 13/2 & $ - $ & 4405.08\
15 & $4f^7 (^8_0S)~5d~^9D 6s~^8D $ & 9/2 & $ - $ & 4713.57\
16 & $4f^7 (^8_0S)~5d^2 (^3_2F)~^1F $ & 15/2 & $ - $ & 4888.14\
17 & $4f^7 (^8_0S)~5d~^9D 6s~^8D $ & 11/2 & $ - $ & 5359.77\
18 & $4f^7 (^8_0S)~6s^2~^8S $ & 7/2 & $ - $ & 5713.88\
19 & $4f^8 (^7_0F)~6s~^8F $ & 13/2 & $ + $ & 7507.66\
20 & $4f^8 (^7_0F)~6s~^8F $ & 11/2 & $ + $ & 8691.73\
Radii of the orbitals of the configuration $4f^N 6s$ and $4f^{N-1} 5d6s$ are presented in Table \[radius\]. For higher $Z$, all orbitals contract (see Table \[radius\]). The exception is Eu II and Gd II: there is no big differences for $<5d\_>$ and $<5d>$ orbitals between Eu II and Gd II. Indeed, for Gd II, the radii for orbitals $<5d\_>$ and $<5d>$ show small increase with respect to Eu II. This may be caused by different computation of the radial wave functions (see section \[sec:scheme\]). Some of the radii are compared with computations by @Indelicato. Radii by @Indelicato differ from 1 to 8% from those computed in this paper. It is likely that these differences are caused by inclusion of Breit interaction into the self-consistent field procedure in the MCDHF computations.
![Comparison of transition probability between states of configurations $4f^3~6s$ - $4f^3~6p$ and between states of configurations $4f^3~6p$ and $4f^3~5d$ for Pr II. The top and bottom panels show a comparison between our results and results from the NIST database. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. The black and reds points show the values calculated with the length (Babushkin) and velocity (Coulomb) forms, respectively. \[Pr\_II\_6p-5d\_transitions\]](Pr_II_6p-6s_transitions.pdf "fig:"){width="47.00000%"} ![Comparison of transition probability between states of configurations $4f^3~6s$ - $4f^3~6p$ and between states of configurations $4f^3~6p$ and $4f^3~5d$ for Pr II. The top and bottom panels show a comparison between our results and results from the NIST database. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. The black and reds points show the values calculated with the length (Babushkin) and velocity (Coulomb) forms, respectively. \[Pr\_II\_6p-5d\_transitions\]](Pr_II_6p-5d_transitions.pdf "fig:"){width="47.00000%"}
E1 transitions {#sec:transitions}
==============
In this section, we show the results of our calculations of transition probabilities. The transition data computed for Pr II, Pm II, Sm II, Eu II, and Gd II are given in machine-readable format in Tables \[Pr\_II\_transitions\], \[Pm\_II\_transitions\], \[Sm\_II\_transitions\], \[Eu\_II\_transitions\], and \[Gd\_II\_transitions\]. The tables include identification of upper and lower levels in $LSJ$ coupling, transition energy, wavelength, line strength, weighted oscillator strength, and transition probabilities in length form. The numbers of transitions are 411 314, 7 104 005, 4 720 626, 467 724 (plus 13 154 transitions with Rydberg states, 480 878 in total), and 1 383 694 for Pr II, Pm II, Sm II, Eu II, and Gd II, respectively.
In the following sections, we compare the calculated transition probabilities with available data, except for Pm and Sm for which enough data are not available [^1].
Pr II {#prii}
-----
For Pr II, rather rich data are available in the NIST database. Therefore, it can be used as evaluation of our calculations. Comparison between the calculated E1 transitions probabilities and those in the NIST database is presented in Figure \[Pr\_II\_6p-5d\_transitions\]. Figure \[Pr\_II\_6p-5d\_transitions\] includes transitions between $4f^3~6s$ and $4f^3~6p$ and transitions between $4f^3~6p$ and $4f^3~5d$ with clear level identification. The same transitions in length and velocity form are connected with dashed lines. Transitions in the NIST database are based on FT spectroscopy by @Ivarsoon_Pr and measurements of branching fractions with use of a laser/fast-ion-beam method by @Li_Pr_transitions and lifetimes determined in a previous study with beam-laser method [@Li_Pr_transitions_lifetime].
We find that transition probabilities calculated in two forms agree better for the transitions between $4f^3~6s$ and $4f^3~6p$ than those between $4f^3~6p$ and $4f^3~5d$. Compared with the data by other authors, our transitions in velocity form gives a better agreement in the strong transition area. Therefore, hereafter we show transition probabilities computed in velocity form.
As for the transition wavelength, our calculations give a good agreement with the NIST data. Averaged agreement in the transition wavelength is 2% for the transitions between states of configurations $4f^3~6s$ and $4f^3~6p$, and 4% for the transitions between states of configurations $4f^3~6p$ and $4f^3~5d$ (see Figure \[Pr\_II\_lambda\]).
![Comparison of transition wavelengths for Pr II between our results ($\lambda_{cal}$) and NIST database recommended values ($\lambda_{NIST}$). The thick line corresponds to perfect agreement, while thin solid and dashed lines correspond to 10% and 20% deviations. \[Pr\_II\_lambda\]](Pr_II_lambda.pdf){width="47.00000%"}
![Comparison of transition probability of Eu II between our results ($A_{v}$) and NIST database. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. Transition probabilities are presented in the velocity (Coulomb) form. \[Eu\_transitions\_A\]](Eu_transitions_A.pdf){width="47.00000%"}
[r c c c c ]{}\[ht!!\] @Biemont & –& TR-LIF & HCL & 27\
@Zhang_Eu & 9& TR-LIF & HCL & 31\
@Lawer_Eu_A & 6& TR-LIF & FTS & 24\
@Wang_Eu & 30& TR-LIF & HCL & 18\
@Than_Eu & 11& TR-LIF & HCL & 24\
Eu II
-----
The NIST database presents 13 lines with transition probabilities which are compared with our calculations in Figure \[Eu\_transitions\_A\]. There is a very good agreement of transitions probabilities although the agreement in the transition wavelength is rather poor, about 14%.
![Comparison of transition probability of Eu II between our results ($A_{v}$) and data of other authors ($A_{exp}$): 1) @Zhang_Eu; 2) @Komarovskii; 3) @Than_Eu; 4) @Wang_Eu; 5) @Lawer_Eu_A; 6) @Karner and @Biemont. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. Transition probabilities are presented in the velocity (Coulomb) form. \[Eu\_transitions\_A\_authors\]](Eu_II_A_A_C_authors.pdf){width="47.00000%"}
It is worth comparing our results with more available measurements although the data are not always critically evaluated. Summary of experiments for Eu II is given in Table \[summary\_tran\]. Absolute transitions probabilities are measured experimentally through the measurements of lifetimes ($\tau$) and branching fractions (BF) by other authors. Measurements for the lifetime are done using time-resolved laser-induced fluorescence (TR-LIF) while branching factors are estimated from emission spectra of a hollow-cathode discharge lamp with Eu powder in the cathode (HCL) or Fourier transform spectrometer (FTS) data. Table \[summary\_tran\] includes the methods as well as the number of lifetimes measurements N$_{\tau}$ and the number of lines N$_{L}$.
Comparison with these measurement is given in Figure \[Eu\_transitions\_A\_authors\]. In this figure, only the levels with clear identifications are included. The most transitions are in the ranges of dashed lines showing the deviation by a factor of 2.0. However, we observe a relatively large deviation in the weak transitions: our calculations give a much smaller transition probabilities than those estimated from the experiments. This may suggest that our strategy of computations is not good enough for weak transitions. Another possible reason is that transitions other than E1, which we do not include in our calculations, may contribute to these weak lines.
![Comparison of transition wavelengths for Gd II between our results and experimental data by @Hartog. The thick line corresponds to perfect agreement, while thin solid and dashed lines correspond to 10% and 20% deviations. \[Gd\_II\_ambda\]](Gd_II_lambda.pdf){width="47.00000%"}
![Comparison of transition probability of Gd II between our results ($A_{v}$) and the results by @Hartog ($A_{exp}$). Three panels are divided according to the configurations involved in the transitions. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. Transition probabilities are presented in the velocity (Coulomb) form. \[Gd\_II\_transitions\_III\]](Gd_5d6p_transitions_A_C.pdf "fig:"){width="47.00000%"} ![Comparison of transition probability of Gd II between our results ($A_{v}$) and the results by @Hartog ($A_{exp}$). Three panels are divided according to the configurations involved in the transitions. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. Transition probabilities are presented in the velocity (Coulomb) form. \[Gd\_II\_transitions\_III\]](Gd_6s6p_transitions_A_C.pdf "fig:"){width="47.00000%"} ![Comparison of transition probability of Gd II between our results ($A_{v}$) and the results by @Hartog ($A_{exp}$). Three panels are divided according to the configurations involved in the transitions. The thick line corresponds to perfect agreement, while the thin solid and dashed lines correspond to deviations by factors of 1.5 and 2.0, respectively. Transition probabilities are presented in the velocity (Coulomb) form. \[Gd\_II\_transitions\_III\]](Gd_5d_transitions_A_C.pdf "fig:"){width="47.00000%"}
Gd II
-----
For Gd II, transitions probabilities are not presented in the NIST database [@NIST]. However, there are several experimental works to address the transition probabilities. For example, experimental transition probabilities are estimated by [@Corliss]. Also, @Wang_Gd_II_transitions have experimentally measured branching fractions of 12 levels for Gd II using the emission spectrum of a hollow cathode lamp. As a results, transition probabilities for 74 lines of Gd II were derived from a combination of the radiative lifetimes reported in the earlier literature and newly determined branching fractions. More recently, @Hartog have investigated absolute transition probabilities for 611 lines for Gd II, by using combination of LIF radiative lifetime measurements and branching fraction measurements. Identification of upper and lower energy levels is based on the work by @Martin. In Figure \[Gd\_II\_ambda\], wavelengths of 460 transitions from their experiments are compared with our calculations. For comparison, we include only the levels with clear identification. 66% of lines wavelengths are within 10% agreement range (solid lines) and 12% of wavelengths have more than 20% disagreement (dashed lines).
[llrrrrrr]{}\[ht!!\] $4f^2 (^3_1H)\,5d^2 (^3_2F)~^5D_{1}$ & $4f^3 (^4_1F)\,5d~^5D_{0}$ & 2769 & 36101 & 1.197D-01 & 1.007D-03 & 5.157D+03 & 0.872\
$4f^2 (^3_1H)\,5d^2 (^3_2F)~^5F_{1}$ & $4f^3 (^4_1F)\,5d~^5D_{0}$ & 553 & 180550 & 5.998D-02 & 1.009D-04 & 2.065D+01 & 0.997\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^5F_{1}$ & 936 & 106828 & 2.116D-01 & 6.018D-04 & 1.172D+02 & 0.266\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1H)\,5d^2 (^3_2F)~^3D_{1}$ & 1618 & 61775 & 6.203D-03 & 3.050D-05 & 1.777D+01 & 0.972\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^3P_{1}$ & 2826 & 35384 & 4.636D-03 & 3.979D-05 & 7.067D+01 & 0.992\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^5P_{1}$ & 3885 & 25734 & 6.263D-02 & 7.393D-04 & 2.482D+03 & 0.857\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^5D_{1}$ & 4313 & 23184 & 1.093D-02 & 1.432D-04 & 5.927D+02 & 0.915\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^5P_{1}$ & 4739 & 21100 & 5.712D-02 & 8.223D-04 & 4.106D+03 & 0.854\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^3S_{1}$ & 6623 & 15096 & 2.818D-03 & 5.670D-05 & 5.532D+02 & 0.395\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^1_1G)\,5d^2 (^3_2F)~^3D_{1}$ & 7603 & 13151 & 5.877D-03 & 1.357D-04 & 1.745D+03 & 0.883\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^1P_{1}$ & 7933 & 12604 & 3.494D-03 & 8.422D-05 & 1.178D+03 & 0.349\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2P)~^5F_{1}$ & 8216 & 12171 & 3.326D-03 & 8.300D-05 & 1.245D+03 & 0.964\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^1_2D)~^3P_{1}$ & 8397 & 11908 & 8.419D-03 & 2.147D-04 & 3.367D+03 & 0.148\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2F)~^3D_{1}$ & 8690 & 11507 & 4.930D-04 & 1.301D-05 & 2.185D+02 & 0.965\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^1_1G)\,5d^2 (^3_2F)~^3P_{1}$ & 11098 & 9010 & 1.194D-03 & 4.027D-05 & 1.102D+03 & 0.577\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2P)~^5D_{1}$ & 12485 & 8009 & 7.617D-03 & 2.888D-04 & 1.001D+04 & 0.750\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^1_2G)~^3P_{1}$ & 12982 & 7702 & 3.509D-03 & 1.383D-04 & 5.186D+03 & 0.992\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1H)\,5d^2 (^1_2G)~^3D_{1}$ & 13084 & 7642 & 5.366D-03 & 2.133D-04 & 8.119D+03 & 0.234\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1F)\,5d^2 (^3_2P)~^3D_{1}$ & 14340 & 6973 & 9.093D-05 & 3.961D-06 & 1.811D+02 & 0.991\
$4f^3 (^4_1F)\,5d~^5D_{0}$ & $4f^2 (^3_1H)\,5d^2 (^1_2G)~^3P_{1}$ & 15904 & 6287 & 4.553D-03 & 2.199D-04 & 1.237D+04 & 0.993\
As for the transition probabilities, we obtain reasonable agreement between our computed values and the LIF measurements (Figure \[Gd\_II\_transitions\_III\], colors of the points represent different configurations). In this figure, we include transitions with transition probabilities higher than 10$^3$ from @Hartog. At closer look, however, there is disagreement in particular for the two-electron-one-photon transitions between states of configurations $4f^7 5d6p$ and $4f^7 6s^2$ and $4f^7 6s6p$ and $4f^7 5d^2$. Our calculations underestimate the experimental values of these transitions. These transitions are due to mixing of configurations in the ASFs which allows one-electron-one-photon transitions (with one electron jump and $\Delta l\pm 1$). The calculated values can be changed significantly by a subtle change in degrees of mixing of the allowed configurations in the ASFs. On the other hand, agreement with @Hartog is much better for strong transitions.
[llrrrrrr]{} $4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1I)~5d^2 (^3_2F)~^7G_{1}$ & 7085 & 14112 & 1.596D-02 & 3.435D-04 & 3.835D+03 & 0.986\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1I)~5d^2 (^3_2F)~^7F_{1}$ & 10754 & 9298 & 4.139D-01 & 1.352D-02 & 3.477D+05 & 0.612\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1I)~5d^2 (^3_2F)~^5F_{1}$ & 12200 & 8196 & 1.145D-01 & 4.245D-03 & 1.405D+05 & 0.670\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1I)~5d^2 (^1_2G)~^5D_{1}$ & 18581 & 5381 & 4.990D-10 & 2.816D-11 & 2.162D-03 & 1.000\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_0S)~5d^2 (^3_2F)~^7F_{1}$ & 19182 & 5213 & 6.356D-03 & 3.703D-04 & 3.030D+04 & 0.962\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1I)~5d^2 (^1_2G)~^5F_{1}$ & 20262 & 4935 & 5.274D-05 & 3.246D-06 & 2.963D+02 & 0.738\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2F)~^7G_{1}$ & 20827 & 4801 & 2.122D-03 & 1.342D-04 & 1.295D+04 & 0.943\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2F)~^7D_{1}$ & 22329 & 4478 & 2.629D-02 & 1.783D-03 & 1.977D+05 & 0.619\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2F)~^7F_{1}$ & 22721 & 4401 & 8.370D-02 & 5.777D-03 & 6.632D+05 & 0.426\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2F)~^5F_{1}$ & 24509 & 4080 & 1.207D-03 & 8.991D-05 & 1.200D+04 & 0.521\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2P)~^7G_{1}$ & 25471 & 3926 & 3.671D-02 & 2.840D-03 & 4.097D+05 & 0.240\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_0S)~5d^2 (^3_2F)~^5F_{1}$ & 26213 & 3814 & 6.045D-04 & 4.814D-05 & 7.354D+03 & 0.736\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^1_2D)~^5P_{1}$ & 26337 & 3796 & 7.825D-06 & 6.260D-07 & 9.656D+01 & 0.927\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^1_2D)~^5P_{1}$ & 26949 & 3710 & 7.221D-06 & 5.911D-07 & 9.545D+01 & 0.994\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^1_2D)~^5D_{1}$ & 27588 & 3624 & 5.019D-04 & 4.206D-05 & 7.118D+03 & 0.864\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2P)~^7D_{1}$ & 28134 & 3554 & 1.602D-02 & 1.369D-03 & 2.409D+05 & 0.259\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2F)~^5D_{1}$ & 28353 & 3526 & 8.565D-03 & 7.376D-04 & 1.318D+05 & 0.111\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2P)~^7F_{1}$ & 28610 & 3495 & 2.125D-03 & 1.846D-04 & 3.361D+04 & 0.786\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2P)~^7F_{1}$ & 29032 & 3444 & 1.239D-02 & 1.093D-03 & 2.048D+05 & 0.553\
$4f^5 (^6_0H)~5d~^7F_{0}$ & $4f^4 (^5_1F)~5d^2 (^3_2F)~^5P_{1}$ & 29210 & 3423 & 2.067D-02 & 1.834D-03 & 3.480D+05 & 0.216\
[llrrrrrr]{} $4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^6_0I)~^6I_{11/2}$ & 34628 & 2887 & 2.376D-02 & 2.499D-03 & 1.666D+05 & 0.226\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^6_0F)~^6F_{11/2}$ & 44272 & 2258 & 6.084D-03 & 8.182D-04 & 8.914D+04 & 0.278\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^6_0F)~^6F_{11/2}$ & 47585 & 2101 & 6.743D-06 & 9.747D-07 & 1.226D+02 & 0.881\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_2H)~^4H_{11/2}$ & 51906 & 1926 & 5.479D-06 & 8.638D-07 & 1.293D+02 & 0.373\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_2H)~^4H_{11/2}$ & 52021 & 1922 & 1.221D-04 & 1.929D-05 & 2.902D+03 & 0.475\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_2H)~^4H_{11/2}$ & 52375 & 1909 & 5.185D-05 & 8.249D-06 & 1.257D+03 & 0.050\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_1K)~^4K_{11/2}$ & 54211 & 1844 & 1.917D-05 & 3.157D-06 & 5.158D+02 & 0.994\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_4H)~^4H_{11/2}$ & 55780 & 1792 & 8.194D-04 & 1.388D-04 & 2.401D+04 & 0.121\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_4H)~^4H_{11/2}$ & 56367 & 1774 & 1.286D-04 & 2.203D-05 & 3.891D+03 & 0.228\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_7G)~^4G_{11/2}$ & 56877 & 1758 & 1.260D-04 & 2.177D-05 & 3.915D+03 & 0.243\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_6G)~^4G_{11/2}$ & 58572 & 1707 & 8.167D-07 & 1.453D-07 & 2.771D+01 & 0.697\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_6G)~^4G_{11/2}$ & 58814 & 1700 & 7.015D-05 & 1.253D-05 & 2.409D+03 & 0.470\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_4I)~^4I_{11/2}$ & 59094 & 1692 & 2.094D-05 & 3.759D-06 & 7.297D+02 & 0.950\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_2I)~^4I_{11/2}$ & 61660 & 1621 & 9.523D-05 & 1.783D-05 & 3.769D+03 & 0.167\
$4f^6 (^7_0F)~5d~^8H_{11/2}$ & $4f^7 (^4_3K)~^4K_{11/2}$ & 62829 & 1591 & 5.517D-04 & 1.052D-04 & 2.310D+04 & 0.156\
$4f^6 (^7_0F)~5d~^8D_{11/2}$ & $4f^7 (^6_0I)~^6I_{11/2}$ & 31992 & 3125 & 2.834D-03 & 2.754D-04 & 1.567D+04 & 0.545\
$4f^6 (^7_0F)~5d~^8D_{11/2}$ & $4f^7 (^6_0F)~^6F_{11/2}$ & 41636 & 2401 & 3.545D-02 & 4.483D-03 & 4.320D+05 & 0.939\
$4f^6 (^7_0F)~5d~^8D_{11/2}$ & $4f^7 (^6_0F)~^6F_{11/2}$ & 44949 & 2224 & 6.156D-02 & 8.405D-03 & 9.440D+05 & 0.921\
$4f^6 (^7_0F)~5d~^8D_{11/2}$ & $4f^7 (^4_2H)~^4H_{11/2}$ & 49270 & 2029 & 1.859D-04 & 2.782D-05 & 3.755D+03 & 0.717\
$4f^6 (^7_0F)~5d~^8D_{11/2}$ & $4f^7 (^4_2H)~^4H_{11/2}$ & 49385 & 2024 & 2.738D-03 & 4.107D-04 & 5.568D+04 & 0.862\
[llrrrrrr]{} $4f^7 (^6_0P)\,5d~^7F_{0}$ & $4f^8 (^7_0F)~^7F_{1}$ & 9040 & 11061 & 1.303D-03 & 3.579D-05 & 6.504D+02 & 0.840\
$4f^7 (^6_0P)\,5d~^7F_{0}$ & $4f^8 (^5_3D)~^5D_{1}$ & 30106 & 3321 & 2.056D-02 & 1.880D-03 & 3.789D+05 & 0.794\
$4f^7 (^6_0P)\,5d~^7F_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 36281 & 2756 & 1.696D-03 & 1.869D-04 & 5.472D+04 & 0.992\
$4f^7 (^6_0P)\,5d~^7F_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 36869 & 2712 & 1.379D-03 & 1.544D-04 & 4.668D+04 & 0.765\
$4f^7 (^6_0P)\,5d~^7F_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 37397 & 2673 & 3.813D-03 & 4.332D-04 & 1.347D+05 & 0.714\
$4f^7 (^6_0P)\,5d~^5D_{0}$ & $4f^8 (^7_0F)~^7F_{1}$ & 2361 & 42339 & 1.176D-03 & 8.438D-06 & 1.046D+01 & 0.962\
$4f^7 (^6_0P)\,5d~^5D_{0}$ & $4f^8 (^5_3D)~^5D_{1}$ & 23427 & 4268 & 1.070D-03 & 7.619D-05 & 9.298D+03 & 0.231\
$4f^7 (^6_0P)\,5d~^5D_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 29602 & 3378 & 2.478D-02 & 2.228D-03 & 4.342D+05 & 0.363\
$4f^7 (^6_0P)\,5d~^5D_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 30191 & 3312 & 5.379D-02 & 4.933D-03 & 9.998D+05 & 0.579\
$4f^7 (^6_0P)\,5d~^5D_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 30718 & 3255 & 5.116D-02 & 4.773D-03 & 1.001D+06 & 0.980\
$4f^8 (^7_0F)~^7F_{1}$ & $4f^7 (^6_0D)\,5d~^7F_{0}$ & 1158 & 86316 & 2.027D-02 & 7.135D-05 & 6.388D+01 & 0.982\
$4f^7 (^6_0D)\,5d~^7F_{0}$ & $4f^8 (^5_3D)~^5D_{1}$ & 19907 & 5023 & 6.089D-03 & 3.681D-04 & 3.244D+04 & 0.891\
$4f^7 (^6_0D)\,5d~^7F_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 26082 & 3833 & 7.236D-01 & 5.733D-02 & 8.671D+06 & 0.880\
$4f^7 (^6_0D)\,5d~^7F_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 26670 & 3749 & 1.442D-02 & 1.168D-03 & 1.848D+05 & 0.970\
$4f^7 (^6_0D)\,5d~^7F_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 27198 & 3676 & 1.119D-01 & 9.245D-03 & 1.520D+06 & 0.847\
$4f^8 (^7_0F)~^7F_{1}$ & $4f^7 (^6_0D)\,5d~^5D_{0}$ & 5664 & 17653 & 9.848D-04 & 1.694D-05 & 3.627D+02 & 0.914\
$4f^7 (^6_0D)\,5d~^5D_{0}$ & $4f^8 (^5_3D)~^5D_{1}$ & 15400 & 6493 & 2.752D-02 & 1.287D-03 & 6.791D+04 & 0.063\
$4f^7 (^6_0D)\,5d~^5D_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 21576 & 4634 & 6.207D-02 & 4.068D-03 & 4.211D+05 & 0.731\
$4f^7 (^6_0D)\,5d~^5D_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 22164 & 4511 & 3.348D-02 & 2.254D-03 & 2.462D+05 & 0.022\
$4f^7 (^6_0D)\,5d~^5D_{0}$ & $4f^8 (^5_2F)~^5F_{1}$ & 22692 & 4406 & 4.103D-01 & 2.828D-02 & 3.238D+06 & 0.942\
[llrrrrrr]{} $4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^7_0F)\,6s~^8F_{11/2}$ & 7731 & 12933 & 1.622D-06 & 3.809D-08 & 1.265D-01 & 0.994\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^7_0F)\,6s~^6F_{11/2}$ & 8671 & 11531 & 3.131D-07 & 8.247D-09 & 3.447D-02 & 0.998\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_3G)\,6s~^6G_{11/2}$ & 33824 & 2956 & 1.511D-06 & 1.552D-07 & 9.875D+00 & 0.031\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_3G)\,6s~^4G_{11/2}$ & 34708 & 2881 & 1.063D-06 & 1.121D-07 & 7.510D+00 & 0.161\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_0L)\,6s~^6L_{11/2}$ & 35251 & 2836 & 9.906D-09 & 1.060D-09 & 7.327D-02 & 0.508\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_1H)\,6s~^6H_{11/2}$ & 39556 & 2528 & 1.645D-08 & 1.976D-09 & 1.719D-01 & 0.710\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_1H)\,6s~^4H_{11/2}$ & 40715 & 2456 & 6.683D-09 & 8.265D-10 & 7.616D-02 & 0.669\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_2F)\,6s~^6F_{11/2}$ & 41839 & 2390 & 1.807D-07 & 2.297D-08 & 2.235D+00 & 0.803\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_2I)\,6s~^6I_{11/2}$ & 44260 & 2259 & 3.334D-11 & 4.482D-12 & 4.881D-04 & 0.915\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_2I)\,6s~^4I_{11/2}$ & 45531 & 2196 & 1.077D-09 & 1.489D-10 & 1.716D-02 & 0.275\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_0K)\,6s~^6K_{11/2}$ & 47419 & 2108 & 8.083D-08 & 1.164D-08 & 1.455D+00 & 0.426\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_0K)\,6s~^4K_{11/2}$ & 49135 & 2035 & 4.864D-08 & 7.259D-09 & 9.743D-01 & 0.161\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_2G)\,6s~^6G_{11/2}$ & 49860 & 2005 & 1.753D-07 & 2.655D-08 & 3.669D+00 & 0.641\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_0K)\,6s~^4K_{11/2}$ & 50373 & 1985 & 7.274D-08 & 1.113D-08 & 1.569D+00 & 0.924\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_2G)\,6s~^4G_{11/2}$ & 50830 & 1967 & 5.492D-08 & 8.481D-09 & 1.218D+00 & 0.933\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_0K)\,6s~^4K_{11/2}$ & 52462 & 1906 & 1.719D-06 & 2.739D-07 & 4.191D+01 & 0.711\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^3_5K)\,6s~^4K_{11/2}$ & 54181 & 1845 & 1.796D-10 & 2.956D-11 & 4.824D-03 & 0.992\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^3_3K)\,6s~^4K_{11/2}$ & 55996 & 1785 & 5.269D-12 & 8.963D-13 & 1.562D-04 & 0.982\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^3_6G)\,6s~^4G_{11/2}$ & 56703 & 1763 & 6.780D-08 & 1.167D-08 & 2.087D+00 & 0.920\
$4f^7 (^8_0S)\,5d~^9D\,6s~^1D_{11/2}$ & $4f^8 (^5_2H)\,6s~^6H_{11/2}$ & 57987 & 1724 & 9.638D-09 & 1.697D-09 & 3.173D-01 & 0.574\
Summary {#sec:summary}
=======
We presented [*ab-initio*]{} atomic calculations of energy levels and E1 transitions from Pr II to Gd II ions based on the strategy developed for the calculations of Nd II [@Nd_jonai]. In total 2 145, 9 774, 8 393, 2 473, and 4 397 levels are presented for Pr II, Pm II, Sm II, Eu II, and Gd II, respectively. Some of the Rydberg states are also included to the computations for Eu II. By comparing with the NIST database and the results by other authors, we confirmed that our calculations achieve good accuracy. For the energy levels, the averaged accuracy compared with the NIST data are 8%, 12%, 6%, 8%, and 7% for Pr II, Pm II, Sm II, Eu II, and Gd II, respectively. These are the highest accuracies achieved for this kind of complete atomic calculations needed for opacity calculations. There is no clear dependence of accuracy on atomic number $Z$. This means that data of lanthanide set can be computed in similar way to the izoelectronic sequence. By using the results of atomic structure calculations, E1 transitions between levels are computed. We provide data for 411 314, 7 104 005, 4 720 626, 467 724, and 1 383 694 transitions for Pr II, Pm II, Sm II, Eu II, and Gd II, respectively. Transition probabilities are compared with NIST database as well as the results of other works. Our computed E1 type transition probabilities are in good agreement with presented in NIST database experimental values, especially in the area of strong transitions.
This research was funded by a grant (No. S-LJB-18-1) from the Research Council of Lithuania. This research was also supported by the JSPS Bilateral Joint Research Project. D.K. is grateful for the support by the NINS program of Promoting Research by Networking among Institutions (grant No. 01411702). The computations presented in this paper were performed at the High Performance Computing Center “HPC Sauletekis” of the Faculty of Physics at Vilnius University.
[^1]: For Sm II, there are transitions probabilities for 7 lines in the NIST database. Unfortunately upper levels do not have clear identification of the configuration. [@7630TP] have performed radiative lifetime measurements with time-resolved laser induced fluorescence (LIF) techniques for 47 levels and have performed relativistic Hartree-Fock (HFR) computations over the energy range [21 000]{} - [36 000]{} cm$^{-1}$, but again the identification of these levels is unclear. Large amount of data (958 lines) have been measured with the same method [@Lawler], but all upper levels do not have clear identification. Lifetimes of 82 levels in range 21 655.420 - 29 591.120 cm $^{-1}$ were investigated beam-laser method and transition probabilities were calculated using branching ratios for 35 transitions by @Scholl_sm.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Damped Lyman-$\alpha$ systems (DLAs) seen in the spectra of high-$z$ QSOs allow us to probe the physical conditions in protogalaxies. Our understanding of physical conditions in DLAs at high-$z$ is primarily based on the absorption lines of H$_2$ molecules and fine-structure transitions. Another important way of probing the thermal state of interstellar medium in these systems is by studying the 21-cm absorption in the spectra of background quasars. Here we report the main results of our GMRT survey to search for 21-cm absorption in a representative and unbiased sample of 35 DLA candidates at 1.10$\le$$z$$\le$1.45. Our sample of DLA candidates is drawn from the strong Mg [ii]{} systems in SDSS DR5 and has resulted in discovery of 9 new 21-cm absorbers. Prior to our survey only one 21-cm absorber was known in the redshift range: 0.7$\le$$z$$\le$2. This survey has allowed us to investigate the dependence of detectability of 21-cm absorption on the properties of UV absorption lines detected in SDSS spectra and estimate the number per unit redshift of 21-cm absorbers. Our GMRT survey provides a representative sample of systems that can be used in combination with various follow-up observations: (1) for investigating the physical conditions in the absorbing gas using spin temperature (T$_{\rm S}$) measurements, (2) for investigating the effect of metallicity and dust content on the detectability of 21-cm absorption, (3) for studying the morphology of the absorbing gas and (4) for probing the time evolution of various fundamental constants. Results from the first phase of our survey are presented in Gupta et al. (2007). Detailed description of the entire sample and results from the survey are presented in Gupta et al. (2009).'
author:
- 'N. Gupta$^1$, R. Srianand$^2$, P. Petitjean$^3$, P. Noterdaeme$^2$, D. J. Saikia$^4$'
title: '21-cm absorbers at intermediate redshifts.'
---
Introduction
============
Observations of high-$z$ galaxies suggest that the global comoving star-formation rate density peaks at $1\le z\le 2$ and then sharply decreases towards $z\sim0$ (e.g. Madau et al. 1996, Hopkins 2004). The determination of the mass density of the gas and its content (molecules, dust and cold H [i]{} gas) over the same redshift range provides an independent and complementary understanding of the redshift evolution of star-formation at similar epochs. While the H [i]{} content of galaxies can be best probed by the surveys of 21-cm emission, limited sensitivity of current radio telescopes does not allow them to reach beyond the local Universe (e.g., Zwaan et al. 2005). On the contrary, detection of H [i]{} in the spectra of distant QSOs in the form of damped Lyman-$\alpha$ absorption provides a luminosity unbiased way of probing the evolution of the H [i]{} content in the universe (Wolfe, Gawiser & Prochaska 2005). Our understanding of physical conditions in DLAs at $z\ge2$ is largely based on the analysis of H$_2$ and/or atomic fine-structure transitions (Ledoux, Petitjean & Srianand 2003). Unfortunately for the time being the above mentioned tracers can not be used to probe the physical state of the absorbing gas at $z\le1.8$ because the useful transitions are located below the atmospheric cut-off. It has been shown by Rao, Turnskek & Nestor (2006) that DLAs essentially have Mg [ii]{} rest equivalent width, W$_{\rm r}$(Mg [ii]{}$\lambda2796$)$\ge$0.6Å. Therefore, the search of 21-cm absorption in a sample of strong Mg [ii]{} absorbers is an unique way to probe the redshift evolution of physical conditions in DLAs like absorption systems at intermediate and low-$z$.
Our GMRT sample
===============
Our sample is drawn from the identification of strong Mg [ii]{} systems, W$_{\rm r}$(Mg [ii]{} $\lambda2796$)$\ge$1.0Å, by Prochter, Prochaska & Burles (2006, hereafter P06) in SDSS DR3 and by us using our automatic procedure for additional systems in DR5. We select the absorbers that are in the redshift range: 1.10$\le z_{abs}\le$1.45 such that the redshifted 21-cm frequency lies in the GMRT 610-MHz band. GMRT is the only radio telescope available at present in the [*relatively*]{} RFI-clean environment (say compared to Green Bank Telescope or Westerbork Synthesis Radio Telescope) for covering this redshift range. These absorbers are then cross-correlated with NVSS and FIRST surveys to select the Mg [ii]{} absorbers in front of compact radio sources brighter than 50mJy and hence suitable for the 21-cm absorption search. We plot in Fig. 1 the redshift distribution of the 35 Mg [ii]{} absorption systems observed as part of our GMRT survey along with the sample of Lane (2000). Latter is the only large survey at low-$z$ for which both detections and non-detections are systematically reported. It includes 62 systems observed with the WSRT and 10 other systems from the literature satisfying their selection criterion (see Lane 2000 for details). The detections shown as a hatched histogram include detections reported in Lane (2000) together with detections from better quality data by Kanekar & Chengalur (2003) and Curran et al. (2007) for systems that were originally reported as non-detections. In the same figure, the filled histogram shows the distribution of Mg [ii]{} systems in our GMRT sample. For equivalent width cutoff of $\sim$ 1[Å]{}, our GMRT sample has more than twice the number of systems investigated by Lane (2000).
[ ]{}
We observed 35 Mg [ii]{} systems with GMRT 610-MHz band using in total $\sim$400hrs of telescope time mostly spread over the years 2006-2008. For our survey, we have usually used the 1MHz baseband bandwidth split into 128 frequency channels yielding a spectral resolution of $\sim$4kms$^{-1}$. GMRT data were reduced using the NRAO Astronomical Image Processing System (AIPS) following the standard procedures.
Results and Discussion
======================
We present the results of a systematic Giant Metrewave Radio Telescope (GMRT) survey of 21-cm absorption in a representative and unbiased sample of 35 strong Mg [ii]{} systems in the redshift range: $z_{abs}\sim$1.10$-$1.45, 33 of which have W$_{\rm r}\ge$1Å. The survey using $\sim$400hrs of telescope time has resulted in 9 new 21-cm detections and good upper limits for the remaining 26 systems (Fig. \[mg2det\]). This is by far the largest number of systems detected in a single systematic survey in a narrow redshift range. Two of these systems also show 2175 [Å]{} dust feature at the redshift of the absorbers (Srianand et al. 2008). Results from the first phase of our survey are presented in Gupta et al. (2007). Detailed description of the entire sample and results from the survey are presented in Gupta et al. (2009). In the following we summarise the main results.
-14.0cm
(400,400)(0,0) (160,260) (340,260)
\[uv21\]
We study the dependence of detectability of 21-cm absorption on different properties of the UV absorption lines detected in the SDSS spectra (Fig. \[uv21\]). We find that if absorption systems are selected with a Mg [ii]{} doublet ratio, DR $<$ 1.3, and a ratio W$_{\rm r}$(Mg [i]{})/W$_{\rm r}$(Mg [ii]{}) $>$ 0.3, the success rate for 21-cm detection is very high (up to $90$%; see [*right*]{} panel of Fig. \[uv21\]). We notice that the detections found in a low-$z$ sample by Lane (2000) also obey these joint constraints (see left panel of Fig. \[uv21\]). In our sample, we find an apparent paucity of 21-cm absorption among systems with W$_{\rm r}$(Mg [ii]{}$\lambda$2796)$>$1.8 [Å]{}, the median W$_{\rm r}$ of our sample. This is contrary to what has been seen at low-$z$ (Lane 2000). Interestingly most of these high W$_{\rm r}$ systems have high DR and low values of W$_{\rm r}$(Mg [i]{})/W$_{\rm r}$(Mg [ii]{}). This strongly suggests that the equivalent width in these systems is dominated by velocity spread and not by line saturation.
We estimate the number of 21-cm absorption systems per unit redshift interval for a given limiting value of the integrated 21-cm optical depth and W$_{\rm r}$(Mg [ii]{}$\lambda$2796). We show that the fraction of Mg [ii]{} systems with 21-cm absorption and the number per unit redshift decrease from $z\sim0.5$ to $z\sim1.3$. The decrease is larger when we use higher equivalent width cutoff. Using a sub-sample of compact sources, with high frequency VLBA observations available, we show that this can not be accounted for by simple covering factor effects. As mentioned above and based on the available data, it appears that most likely the main reason behind this cosmological evolution is the decrease of the CNM covering factor (and volume filling factor) in the strong Mg [ii]{} absorbers. Indeed, it is known that the number of Mg [ii]{} systems per unit redshift increases with increasing redshift. The evolution is steeper for stronger systems (Steidel & Sargent, 1992 and P06 for recent reference). Using the data of Steidel & Sargent (1992), Srianand (1996) found that the strongest redshift evolution was seen among the Mg [ii]{} absorbers with W$_{\rm r}$(Fe [ii]{})$\lambda$2383/W$_{\rm r}$(Mg [ii]{})$\lambda$2796$<$0.5. This clearly means the physical conditions in strong Mg [ii]{} absorbers are different at high and low-$z$.
We have estimated the velocity spread of the 21-cm absorption systems using the apparent optical depth method (Ledoux et al. 2006). We do not find any statistically significant correlation between W$_{\rm r}$(Mg [ii]{}$\lambda$2796) and the 21-cm velocity width in our sample. A marginal correlation is found for the low-$z$ sample. The absence of correlation in the high-$z$ sample is related to the lack of 21-cm absorbers among Mg [ii]{} systems with W$_{\rm r}$(Mg [ii]{})$>$1.8 [Å]{} in the GMRT sample. This is probably due to a true evolution with redshift of the physical state of the Mg [ii]{} systems and consistent with the idea that the Mg [ii]{} equivalent width is mostly correlated with the overall kinematics of the gas in the absorbing system and not with the column density in the component associated with the cold gas. When high spectral resolution data are available, we note that the 21-cm absorption is not always associated with the strongest Mg [ii]{} component.
We selected systems with W$_{\rm r} \ge$1Å but detected by chance a 21-cm absorption in a system with W$_{\rm r}$ = 0.43Å(at $z_{\rm abs}$ = 1.3710 toward J0108$-$0037). Efforts are underway at GMRT to extend our survey to weaker (W$_{\rm r} \le$1Å ) Mg [ii]{} systems. This will be crucial for understanding the physical state of Mg [ii]{} systems and to determine the detectability of 21-cm absorption versus W$_{\rm r}$. Ideally one would like to estimate the number density of 21-cm absorbers and measure the cosmological evolution without preselection from the UV absorption lines. This can be achieved only by a blind survey of 21-cm absorption in front of radio loud QSOs. It will be possible to embark upon such a survey with the upcoming Square Kilometer Array (SKA) pathfinders such as the Australian Square Kilometer Array Pathfinder (ASKAP) and Karoo Array Telescope (meerKAT). In particular, the ASKAP with its instantaneous wide bandwidth of 300MHz and large field of view (30degree$^2$) is an ideal instrument for this (Johnston et al. 2008). An ASKAP survey with 150 pointings of 16 hrs each (i.e 2400 hrs in total) in the 700-1000 MHz frequency band would yield detection of $\sim$100 to 250 intervening 21-cm absorbers in the redshift range 0.4$\le z \le$1. As the energy of the 21-cm transition is proportional to $x=\alpha^2 G_p/ \mu$, high resolution optical and 21-cm spectra can be used together to probe the combined cosmological variation of these constants (Tubbs & Wolfe 1980). Our GMRT survey provides systems in a narrow redshift range in which this measurement can be done. Thus high resolution optical spectroscopy of the corresponding QSOs are suitable to perform this test at $z\sim1.3$.
We thank Rajaram Nityananda for useful discussions and DDTs, and the GMRT staff for their co-operation during the observations. GMRT is run by the National Centre for Radio Astrophysics of the Tata Institute of Fundamental Research. We acknowledge the use of SDSS spectra from the archive (http://www.sdss.org/) and the radio images from NVSS and FIRST surveys.
Curran S.J., Tzanavaris P., Murphy M.T., Webb J.K., Pihlstrom Y.M., 2007, MNRAS, 381, L6 Gupta N., Srianand R., Petitjean P., Khare P., Saikia D.J., York D.G., 2007, ApJ, 654, L111 Gupta N., Srianand R., Petitjean P., Noterdaeme P., Saikia D.J., 2009, MNRAS, Submitted Hopkins A.M., 2004, ApJ, 615, 209 Johnston S., et al., 2008, ExA, 22, 151 Kanekar N., Chengalur J.N., 2003, A&A, 399, 857 Lane W., 2000, PhD Thesis, University of Groningen Ledoux C., Petitjean P., Srianand R., 2003, MNRAS, 346, 209 Ledoux C., Petitjean P., Fynbo J.P.U., Moller P., Srianand R., 2006, A&A, 457, 71 Madau P., Ferguson H.C., Dickinson M.E., Giavalisco M., Steidel C.C., Fruchter A., 1996, MNRAS, 283, 1388 Prochter G.E., Prochaska J.X., Burles S.M., 2006, ApJ, 639, 766 Rao S.M., Turnshek D.A., Nestor D.B., 2006, ApJ, 636, 610 Srianand R., 1996, ApJ, 462, 643 Srianand R., Gupta N., Petitjean P., Noterdaeme P., Saikia D.J., 2008, MNRAS, 391, L69 Steidel C.C., Sargent W.L.W., 1992, ApJS, 80, 1 Tubbs A.D., Wolfe A.M. 1980, ApJ, 236, L105 Wolfe A.M., Gawiser E., Prochaska J.X., 2005, ARA&A, 43, 861 Zwaan M.A., Meyer M.J., Staveley-Smith L., Webster R.L., 2005, MNRAS, 359, L30
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We investigate local-global principles for multinorm equations over a global field. To this extent, we generalize work of Drakokhrust and Platonov to provide explicit and computable formulae for the obstructions to the Hasse principle and weak approximation for multinorm equations. We illustrate the scope of this technique by extending results of Bayer-Fluckiger–Lee–Parimala [@eva], Demarche–Wei [@demarche] and Pollio [@pollio].'
address: |
André Macedo\
Department of Mathematics and Statistics\
University of Reading\
Whiteknights, PO Box 220\
Reading RG6 6AX\
UK
author:
- André Macedo
title: On the obstruction to the Hasse principle for multinorm equations
---
Introduction
============
Let $K$ be a global field and let $L=(L_1,\dots,L_n)$ be an $n$-tuple ($n \geq 1$) of finite separable extensions of $K$. In this paper, we study the so-called *multinorm principle* for $L$, which is said to hold if, for any $c \in K^*$, the affine $K$-variety $$\label{eq:Xc}
X_c : \prod\limits_{i=1}^{n} N_{L_i/K}(\Xi_i)=c$$ satisfies the Hasse principle. In other words, $L$ satisfies the multinorm principle if, for all $c \in K^*$, the existence of points on $X_c$ over every completion of $K$ implies the existence of a $K$-point.
From a geometric viewpoint, $X_c$ defines a principal homogenous space under the *multinorm one torus* $T$, defined by the exact sequence of $K$-algebraic groups $$1 \to T \to \prod\limits_{i=1}^{n} R_{L_i/K} {\mathbb G_m}\xrightarrow{\prod_i N_{L_i/K}} {\mathbb G_m}\to 1,$$ where $R_{L_i/K} {\mathbb G_m}$ denotes the Weil restriction of ${\mathbb G_m}$ from $L_i$ to $K$. In this way, the Tate-Shafarevich group $\Sha(T)$ of $T$ is naturally identified with the *obstruction to the multinorm principle* for $L$, defined as $$\mathfrak{K}(L,K)=K^* \cap \prod\limits_{i=1}^{n} N_{L_i/K}({\mathbb{A}}^*_{L_i}) / \prod\limits_{i=1}^{n} N_{L_i/K}(L_i^*),$$ where ${\mathbb{A}}^{*}_{L_i}$ denotes the idèle group of $L_i$ and the multinorm principle holds if and only if $\mathfrak{K}(L,K)=1$.
In the toric case, the Hasse principle for principal homogeneous spaces is strikingly connected with *weak approximation*. This property is said to hold for a torus $T$ over $K$ if the *defect of weak approximation* $$A(T) =\prod\limits_v T(K_v)/\overline{T(K)}$$ is trivial (here $\overline{T(K)}$ denotes the closure of $T(K)$ in $\prod_v T(K_v)$ with respect to the product topology). In [@Vosk §11.6], Voskresenskiĭ showed the existence of an exact sequence $$\label{eq:Vosk}
0 \to A(T) \to \operatorname{H}^1(K,\operatorname{Pic}\overline{X})^{\vee} \to \Sha (T) \to 0,$$ where $X$ denotes a smooth compactification of $T$, $\overline{X}$ the base change of $X$ to an algebraic closure of $K$ and $\phantom{ }^\vee$ stands for the Pontryagin dual of an abelian group. Returning to the multinorm principle, when $n=1$ one recovers the classical *Hasse norm principle* (HNP), a topic that has been extensively studied in the literature (see e.g. [@Platonov §6.3] or [@MN19 §1] for a survey of known results). If $L/K$ is Galois, then there is an explicit description of the obstruction to the HNP (due to Tate in [@C-F p. 198]) in terms of the group cohomology of its local and global Galois groups. Drakokhrust later obtained (in [@Drak]) a more general description of this obstruction for an arbitrary extension $L/K$ in terms of generalized representation groups.
For $n > 1$, such a description has not yet been obtained. Nonetheless, multiple cases have been analyzed in the literature. For example, if $n=2$ it is known that the multinorm principle holds if
1. \[hur\] $L_1$ or $L_2$ is a cyclic extension of $K$ ([@Hur Proposition 3.3]);
2. \[pras\] $L_1/K$ is abelian, satisfies the HNP and $L_2$ is linearly disjoint from $L_1$ ([@Prar Proposition 4.2]);
3. \[PR\_res\] the Galois closures of $L_1/K$ and $L_2/K$ are linearly disjoint over $K$ ([@PR]).
Subsequent work of Demarche and Wei provided a generalization of the result in to $n$ extensions ([@demarche Theorems 1 and 6]), while also addressing weak approximation for the associated multinorm one torus. In [@pollio], Pollio computed the obstruction to the multinorm principle for a pair of abelian extensions and, in [@eva], Bayer-Fluckiger, Lee and Parimala provided sufficient and necessary conditions for the multinorm principle to hold assuming that one of the extensions $L_i/K$ is cyclic.
In this paper, we provide an explicit description of the obstructions to the multinorm principle and weak approximation for the multinorm one torus of $n$ arbitrary extensions. To achieve this, we generalize the concept (due to Drakokhrust and Platonov in [@DP]) of the *first obstruction to the Hasse principle* (see Section \[sec:1st\_obs\]). By then adapting work of Drakokhrust ([@Drak]), we obtain our main result (Theorem \[thm:main\_result\]), describing the obstructions to the multinorm principle and weak approximation for the multinorm one torus in terms of generalized representation groups of the relevant local and global Galois groups. The formulas given in Theorem \[thm:main\_result\] are effectively computable and we provide algorithms in GAP [@gap] for this effect (see Remark \[rem:finite\_time2\]).
We also apply our techniques to describe the validity of the local-global principles in three concrete examples (see Section \[sec:applications\]). We start by proving a result inspired by [@demarche Theorem 6] that compares the birational invariants $\operatorname{H}^1(K,\operatorname{Pic}\overline{X})$ and $\operatorname{H}^1(K,\operatorname{Pic}\overline{Y})$, where $Y$ is a smooth compactification of the norm one torus $S=R^1_{F/K} {{\mathbb G}}_m$ of the extension $F=\bigcap\limits_{i=1}^{n} L_i$. In particular, we show (Theorem \[thm:demarch\_wei\_thm\]) that under certain conditions there is an isomorphism $$\operatorname{H}^1(K,\operatorname{Pic}\overline{X}) \xrightarrow{\simeq}\operatorname{H}^1(K,\operatorname{Pic}\overline{Y}).$$
This results further allows us to compare the defect of weak approximation for $T$ with the defect of weak approximation for $S$ (Corollary \[cor:dem1\]).
Under the same assumptions, we also show (Theorem \[thm:pollio\]) the existence of isomorphisms $$\mathfrak{K}(L,K) \cong \mathfrak{K}(F/K) \textrm{ and } A(T) \cong A(S)$$ when all the extensions $L_i/K$ are abelian. This theorem generalizes Pollio’s result (in [@pollio]) on the obstruction to the multinorm principle for a pair of abelian extensions.
In Section \[sec:eva\] we complement [@eva Theorem 8.3] by providing a characterization (Theorem \[thm:eva\]) of weak approximation for the multinorm one torus of $n$ non-isomorphic cyclic extensions of prime degree $p$. More precisely, we show that both the multinorm principle and weak approximation for $T$ hold if $[L_1 \dots L_n : K]>p^2$. Otherwise, weak approximation holds if and only if the multinorm principle fails (a property that can be detected by precise local conditions, see Remark \[rem:eva\]). While preparing this paper, we became aware of the recent (and independent) work of Lee [@lee], who extends results of [@eva §8] to provide a description of the multinorm principle and weak approximation for the multinorm one torus of $n$ non-isomorphic cyclic extensions (and, in this way, obtains a result more general than Theorem \[thm:eva\]).
Notation {#notation .unnumbered}
--------
Given a global field $K$, we denote its set of places by $\Omega_K$. For $v \in \Omega_K$, we use the notation $K_v$ for the completion of $K$ at $v$ and, if $L$ is a Galois extension of $K$, we denote by $G_v$ a choice of decomposition group of $L/K$ at $v $.
Given a finite group $G$, a subgroup $H$ of $G$, a $G$-module $A$, an integer $q$ and a prime number $p$, we use the notation:
--------------------------------- -------------------------------------------------------------------------------------------------
$|G|$ the order of $G$
$Z(G)$ the center of $G$
$[H,G]$ the subgroup of $G$ generated by all commutators $[h,g]$ with $h \in H,g \in G$
$\Phi^{G}(H)$ the subgroup of $H$ generated by all commutators $[h,g]$ with $h \in H \cap g H g^{-1},g \in G$
$G^{ab}$ the abelianization $G/[G,G]$ of $G$
$G_p$ a Sylow $p$-subgroup of $G$
$\hat{\operatorname{H}}^q(G,A)$ the $q$-th Tate cohomology group
--------------------------------- -------------------------------------------------------------------------------------------------
We also often use the notation $G'$ for the derived subgroup $[G,G]$ of $G$. If $H$ is a normal subgroup of $G$, we write $H \trianglelefteq G$. For $x,y \in G$ we adopt the convention $[x,y]=x^{-1}y^{-1}xy$ and $x^y=y^{-1}xy$. If $G$ is abelian, we denote its $p$-primary part by $G_{(p)}$.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I would like to thank Prof. Eva Bayer-Fluckiger for a conversation that motivated this work and my supervisor Rachel Newton for useful discussions on the manuscript and for pointing out the recent preprint [@lee]. This work was supported by the FCT doctoral scholarship SFRH/BD/117955/2016.
The first obstruction to the multinorm principle {#sec:1st_obs}
================================================
In this section we define the concept of the first obstruction to the multinorm principle and present several of its properties. We fix a global field $K$, an $n$-tuple $L=(L_1,\dots,L_n)$ of finite separable extensions of $K$ and a Galois extension $N/K$ containing all the fields $L_1,\dots,L_n$. We denote $G=\operatorname{Gal}(N/K)$, $H_i=\operatorname{Gal}(N/L_i)$ for $i=1,\dots,n$ and $H=\langle H_1,\dots,H_n \rangle$, the subgroup of $G$ generated by all the $H_i$. Note that $H=\operatorname{Gal}(N/F)$, where $F=\bigcap\limits_{i=1}^{n} L_i$.
We define the *first obstruction to the multinorm principle for $L$ corresponding to $(N,L,K)$* as $$\mathfrak{F}(N,L,K)=K^* \cap \prod\limits_{i=1}^{n} N_{L_i/K}({\mathbb{A}}^{*}_{L_i}) / \prod\limits_{i=1}^{n} N_{L_i/K}(L_i^*)(K^* \cap N_{N/K}({\mathbb{A}}^{*}_{N})).$$
This notion generalizes the concept (introduced by Drakokhrust and Platonov in [@DP]) of the *first obstruction to the Hasse principle for $L/K$ corresponding to a tower of fields $N/L/K$*, defined as $\mathfrak{F}(N/L/K)=K^* \cap N_{L/K}({\mathbb{A}}^{*}_{L}) / N_{L/K}(L^*)(K^* \cap N_{N/K}({\mathbb{A}}^{*}_{N})).$
The first obstruction to the multinorm principle has various useful properties – for example, it is clear from the definition that the total obstruction to the multinorm principle $\mathfrak{K}(L,K)$ surjects onto $\mathfrak{F}(N,L,K)$ with equality if the Hasse norm principle holds for $N/K$. Moreover, this equality also happens if the first obstruction to the Hasse principle for some extension $L_i/K$ coincides with the total obstruction to the Hasse norm principle $\mathfrak{K}(L_{i}/K)=K^* \cap N_{L_i/K}({\mathbb{A}}^{*}_{L_i}) / N_{L_i/K}(L_i^*)$ (called the *knot group* of $L_i/K$):
\[lem:1stobs\_equal\_multiknot\] If $\mathfrak{K}(L_{i}/K)=\mathfrak{F}(N/L_{i}/K) $ for some $i =1,\dots,n$, then $\mathfrak{K}(L,K)=\mathfrak{F}(N,L,K)$.
The assumption translates into $K^* \cap N_{N/K}({\mathbb{A}}^*_N) \subset N_{L_{i}/K}(L_{i}^*) $. This implies that ${\prod\limits_{i=1}^{n}N_{L_i/K}(L_i^*)(K^* \cap N_{N/K}({\mathbb{A}}^*_N)) = \prod\limits_{i=1}^{n} N_{L_i/K}(L_i^*)}$ and hence $\mathfrak{K}(L,K)=\mathfrak{F}(N,L,K)$.
\[cor:square\_free\] If $[L_{i}:K]$ is square-free for some $i = 1,\dots,n$, then $\mathfrak{K}(L,K)=\mathfrak{F}(N,L,K)$.
By [@DP Corollary 1], if $[L_{i}:K]$ is square-free, then $ \mathfrak{K}(L_{i}/K)=\mathfrak{F}(N/L_{i}/K)$. Now apply Lemma \[lem:1stobs\_equal\_multiknot\].
More generally, one has the following criterion (extending [@DP Theorem 3]) for the equality $\mathfrak{K}(L,K)=\mathfrak{F}(N,L,K)$.
\[prop:1stobs\_cor\] Let $k_1,\dots,k_n$ be positive integers. For each $i=1,\dots,n$, choose a collection of $k_i$ subgroups $G_{i,1},\dots,G_{i,k_i}$ of $G$ and $k_i$ subgroups $H_{i,1},\dots,H_{i,k_i}$ such that $H_{i,j} \subset H_i \cap G_{i,j}$ for any $j=1,\dots, k_i$. Set $L_{i,j}=N^{H_{i,j}}$ and $K_{i,j}=N^{G_{i,j}}$ for all $i,j$. Suppose that the Hasse norm principle holds for all the extensions $L_{i,j}/K_{i,j}$ and that the map $$\bigoplus\limits_{i=1}^{n}\bigoplus\limits_{j=1}^{k_i} \operatorname{Cor}^{G}_{G_{i,j}}:\bigoplus\limits_{i=1}^{n}\bigoplus\limits_{j=1}^{k_i} \hat{\operatorname{H}}^{-3}(G_{i,j},{\mathbb{Z}}) \to \hat{\operatorname{H}}^{-3}(G,{\mathbb{Z}})$$
is surjective. Then $ \mathfrak{K}(L,K)=\mathfrak{F}(N,L,K)$.
The statement follows from an argument analogous to the one given by Drakokhrust and Platonov for the Hasse norm principle case, see [@DP Theorem 3].
A further trait of the first obstruction to the multinorm principle $\mathfrak{F}(N,L,K)$ is that it can be expressed in terms of the local and global Galois groups of the towers $N/L_i/K$ (in similar fashion to the first obstruction to the Hasse norm principle). In order to prove this, we mimic the work Drakokhrust and Platonov in [@DP §2]. We will use the following lemma:
[@DP Lemma 1]\[lem1DP\] Let $N/L/K$ be a tower of global fields with $N/K$ Galois. Set $G=\operatorname{Gal}(N/K)$ and $H=\operatorname{Gal}(N/L)$. Then, given a place $v$ of $K$, the set of places $w$ of $L$ above $v$ is in bijection with the set of double cosets in the decomposition $G = \bigcup\limits_{i=1}^{r_v} H x_i G_v$. If $w$ corresponds to $H x_{i} G_v$, then the decomposition group $H_w$ of the extension $N/L$ at ${w}$ equals $H \cap x_{i} G_v x_{i}^{-1}$.
In our situation, for any $v \in \Omega_K$ and $i=1,\dots,n$, let $G=\bigcup\limits_{k=1}^{r_{v,i}} H_i x_{i,k} G_v$ be a double coset decomposition. By the above lemma, $H_{i,w}:= H_i \cap x_{i,k} G_v x_{i,k}^{-1}$ is the decomposition group of $N/L_i$ at a place $w$ of $L_i$ above $v$ corresponding to the double coset $H_i x_{i,k} G_v$. Now consider the commutative diagram: $$\label{diag:1stobs_defn}
\xymatrix{
\bigoplus\limits_{i=1}^{n} {H}_i^{\textrm{ab}} \ar[r]^{{\psi}_1} & {G}^{\textrm{ab}}\\
\bigoplus\limits_{i=1}^{n}(\bigoplus\limits_{v \in \Omega_K} ( \bigoplus\limits_{w|v} {{H}_{i,w}^{\textrm{ab}} }) ) \ar[r]^{\ \ \ \ \ \ \ {\psi}_2} \ar[u]^{{\varphi}_1 }&\bigoplus\limits_{v \in \Omega_K}{{G}_v^{\textrm{ab}} }\ar[u]_{\varphi_2}
}$$
Here the superscript$\phantom{a}^{\textrm{ab}}$ above a group denotes its abelianization and the inside sum over $w|v$ runs over all the places $w$ of $L_i$ above $v$. Additionally, the maps $\varphi_1,\psi_1$ and $\varphi_2$ are induced by the inclusions $H_{i,w} \hookrightarrow H_i, H_i \hookrightarrow G$ and $G_v \hookrightarrow G$, respectively, while $\psi_2$ is obtained from the product of all conjugation maps $H_{i,w}^{ab} \to G_v^{ab}$ sending $h_{i,k} [H_{i,w},H_{i,w}]$ to $x_{i,k}^{-1} h_{i,k} x_{i,k} [G_v,G_v]$. We denote by ${\psi}_2^{v}$ (respectively, ${\psi}_2^{nr}$) the restriction of the map ${\psi}_2$ to the subgroup $\bigoplus\limits_{i=1}^{n} ( \bigoplus\limits_{w|v} {{H}_{i,w}^{\textrm{ab}} }) $ (respectively, $\bigoplus\limits_{i=1}^{n}(\bigoplus\limits_{\substack{v \in \Omega_K \\ v \text{ unramified}}} ( \bigoplus\limits_{w|v} {{H}_{i,w}^{\textrm{ab}} }) )$). With this notation set, we can now establish the main result of this section (generalizing [@DP Theorem 1]):
\[thm:thm1DP\_gen\] In the notation of diagram , we have $$\mathfrak{F}(N,L,K) \cong \ker\psi_1/\varphi_1(\ker\psi_2).$$
Diagram can be written as $$\label{diag:1stobs_defn2}
\xymatrix{
\bigoplus\limits_{i=1}^{n} \hat{\operatorname{H}}^{-2}({H}_i,{\mathbb{Z}}) \ar[r]^{{\psi}_1} & \hat{\operatorname{H}}^{-2}(G,{\mathbb{Z}})\\
\bigoplus\limits_{i=1}^{n}(\bigoplus\limits_{v \in \Omega_K} ( \bigoplus\limits_{w|v} { \hat{\operatorname{H}}^{-2}(H_{i,w},{\mathbb{Z}}) }) ) \ar[r]^{\ \ \ \ \ \ \ {\psi}_2} \ar[u]^{{\varphi}_1}&\bigoplus\limits_{v \in \Omega_K}{ \hat{\operatorname{H}}^{-2}(G_v,{\mathbb{Z}}) }\ar[u]_{\varphi_2}
}$$
By the local (respectively, global) Artin isomorphism, we have $\hat{\operatorname{H}}^{-2}(H_{i,w},{\mathbb{Z}}) \cong \hat{\operatorname{H}}^{0}(H_{i,w},N_w^*)$ and $\hat{\operatorname{H}}^{-2}(G_v,{\mathbb{Z}}) \cong \hat{\operatorname{H}}^{0}(G_v,N_v^*)$ (respectively, $\hat{\operatorname{H}}^{-2}({H}_i,{\mathbb{Z}}) \cong \hat{\operatorname{H}}^{0}(H_i,C_N)$ and $\hat{\operatorname{H}}^{-2}(G,{\mathbb{Z}}) \cong \hat{\operatorname{H}}^{0}(G,C_N)$, where $C_N$ is the idèle class group of $N/K$). Additionally, by [@C-F Proposition 7.3(b)] there are identifications $\bigoplus\limits_{v \in \Omega_K}( \bigoplus\limits_{w|v} \hat{\operatorname{H}}^{0}(H_{i,w},N_w^*)) \cong \hat{\operatorname{H}}^{0}(H_{i},{\mathbb{A}}^{*}_N)$ and $\bigoplus\limits_{v \in \Omega_K} \hat{\operatorname{H}}^{0}(G_v,N_v^*) \cong \hat{\operatorname{H}}^{0}(G,{\mathbb{A}}_{N}^{*})$. In this way, an argument analogous to the one given in [@DP §2] for the $n=1$ case shows that diagram induces the commutative diagram $$\label{diag:1stobs_defn3}
\xymatrix{
\bigoplus\limits_{i=1}^{n} \hat{\operatorname{H}}^{0}({H}_i,C_N) \ar[r]^{\ {\psi}_1} & \hat{\operatorname{H}}^{0}(G,C_N)\\
\bigoplus\limits_{i=1}^{n} \hat{\operatorname{H}}^{0}(H_i,{{\mathbb A}}^*_{N}) \ar[r]^{\ {\psi}_2} \ar[u]^{{\varphi}_1}& \hat{\operatorname{H}}^{0}(G,{{\mathbb A}}^*_{N}) \ar[u]_{\varphi_2}
}$$
where $\varphi_{1},\varphi_2$ are the natural projections and $\psi_{1},\psi_2$ are induced by the product of the norm maps $N_{L_i/K}$. Using the definition of the cohomology group $\hat{\operatorname{H}}^0$, this diagram is equal to $$\label{diag:1stobs_defn4}
\xymatrix{
\bigoplus\limits_{i=1}^{n} \frac{{\mathbb{A}}_{L_i}^{*}}{L_i^{*} N_{N/L_i}({\mathbb{A}}^{*}_{N})} \ar[r]^{{\psi}_1} & \frac{{\mathbb{A}}_{K}^{*}}{K^{*} N_{N/K}({\mathbb{A}}^{*}_{N})}\\
\bigoplus\limits_{i=1}^{n} \frac{{\mathbb{A}}_{L_i}^{*}}{ N_{N/L_i}({\mathbb{A}}^{*}_{N})} \ar[r]^{\ \ \ {\psi}_2} \ar[u]^{{\varphi}_1}&{ \frac{{\mathbb{A}}_{K}^{*}}{ N_{N/K}({\mathbb{A}}^{*}_{N})} }\ar[u]_{\varphi_2}
}$$
From diagram , it is clear that $$\ker \psi_1 = \{ (x_i L_i^* N_{N/L_i}({\mathbb{A}}^{*}_{N}))_{i=1}^{n} | \prod\limits_{i=1}^{n} N_{L_i/K}(x_i) \in K^* N_{N/K}({\mathbb{A}}^{*}_N)\}$$ and $$\varphi_1(\ker \psi_2) = \{ (x_i L_i^* N_{N/L_i}({\mathbb{A}}^{*}_{N}))_{i=1}^{n} | \prod\limits_{i=1}^{n} N_{L_i/K}(x_i) \in N_{N/K}({\mathbb{A}}^{*}_N)\}.$$
Now define $$\begin{aligned}
f \colon \ker \psi_1 / \varphi_1(\ker \psi_2) &\longrightarrow \mathfrak{F}(N,L,K) \\
(x_i L_i^* N_{N/L_i}({\mathbb{A}}^{*}_{N}))_{i=1}^{n} &\longmapsto x \prod\limits_{i=1}^{n} N_{L_i/K}({L_i}^{*}) (K^* \cap N_{N/K}({\mathbb{A}}^{*}_{N}) )\end{aligned}$$
where $x$ is any element of $ K^* \cap \prod\limits_{i=1}^{n} N_{L_i/K}({\mathbb{A}}^{*}_{L_i})$ such that $\prod\limits_{i=1}^{n} N_{L_i/K}(x_i) \in x N_{N/K}({\mathbb{A}}^{*}_N)$. It is straightforward to check that $f$ is well defined and an isomorphism.
\[rem:finite\_time\] Given the knowledge of the local and global Galois groups of the towers $N/L_i/K$, the first obstruction to the multinorm principle can be computed in finite time by employing Theorem \[thm:thm1DP\_gen\]. First, it is clear that the computation of the groups $\ker \psi_1$ and $\varphi_1(\ker \psi_2^{v})$ for the ramified places $v$ of $N/K$ is finite. Moreover, from the definition of the maps in diagram , it is clear that if $v_1,v_2 \in \Omega_K$ are such that $G_{v_1}=G_{v_2}$, then $\varphi_1(\ker \psi_2^{v_1}) = \varphi_1(\ker \psi_2^{v_2})$. This shows that the computation of $\varphi_1(\ker \psi_2^{nr})$ is also finite. On this account, we designed a function in GAP [@gap] (whose code is available in [@macedo_code]) that takes as input the Galois groups $G, H_i$ and the decomposition groups $G_v$ at the ramified places of $N/K$ and outputs the group $\mathfrak{F}(N,L,K)$.
We conclude this section by providing two results that further reduce the amount of calculations necessary to compute $\mathfrak{F}(N,L,K)$ via Theorem \[thm:thm1DP\_gen\]. These are inspired by the same properties of the first obstruction to the Hasse norm principle (in [@DP §3]) and proved in the same way.
[[@DP Lemma 2]]{} Let $v_1,v_2 \in \Omega_K$ be such that $G_{v_2} \subset G_{v_1}$. Then, in the notation of diagram , we have $$\varphi_1(\ker \psi_2^{v_2}) \subset \varphi_1(\ker \psi_2^{v_1}).$$
[[@DP Lemma 3]]{} Let $v_1,v_2 \in \Omega_K$ be such that $G_{v_1} M = G_{v_2} M$ for some subgroup $M \subset Z(G) \cap \bigcap\limits_{i=1}^{n} H_i$. Then, in the notation of diagram , we have $$\varphi_1(\ker \psi_2^{v_1})= \varphi_1(\ker \psi_2^{v_2}).$$
Generalized representation groups {#sec:gen_gps}
=================================
In this section we prove that the obstruction to the multinorm principle for $L$ can always be expressed in terms of the arithmetic of the extensions $L_i/K$ by using generalized representation groups (see Definition \[gen\_rep\_gp\_defn\] below) of $G=\operatorname{Gal}(N/K)$. Once again, many of the results in this section are inspired by and generalize Drakokhrust’s work [@Drak] on the Hasse norm principle.
\[gen\_rep\_gp\_defn\] Let $G$ be a finite group. A finite group $\overline{G}$ is called a *generalized representation group* of $G$ if there exists a central extension $$1 \to M \to \overline{G} \xrightarrow[]{\lambda} G \to 1$$ such that $M \cap [\overline{G},\overline{G}] \cong \hat{\operatorname{H}}^{-3}(G,{\mathbb{Z}})$. We call $M$ the base normal subgroup of $\overline{G}$. If in addition $M \subset [\overline{G},\overline{G}]$, we say that $\overline{G}$ is a *Schur covering group of $G$*.
\[prop:1st=knot\] There exists a Galois extension $P/K$ containing $N$ and such that $$\mathfrak{F}(P,L,K)=\mathfrak{K}(L,K).$$ Furthermore, this extension has the property that $\overline{G}=\operatorname{Gal}(P/K)$ is a generalized representation group of $G$ with base normal subgroup $\overline{M}=\operatorname{Gal}(P/N)$ and if $\overline{\lambda}:\overline{G} \to G$ is the associated projection map, we have $\operatorname{Gal}(P/L_i)=\overline{\lambda}^{-1}(H_i)$.
It follows from the proof of [@Drak Lemma 1] (see also [@opolka Satz 3]) that there exists a Galois extension $P/K$ such that the first obstruction to the Hasse norm principle $ \mathfrak{F}(P/L_i/K)$ coincides with the knot group $\mathfrak{K}(L_i/K)$ for all $L_i \in L$. Now apply Lemma \[lem:1stobs\_equal\_multiknot\]. The stated properties of $P/K$ are shown in the references given above.
As remarked in [@Drak], the extension $P/K$ is not uniquely determined and the computation of its arithmetic is not always easy. Nonetheless, one can still compute $\mathfrak{F}(P,L,K)$ by commencing with an arbitrary generalized represention group of $G$.
Let $\widetilde{G}$ be any generalized representation group of $G$ with projection map $\widetilde{\lambda}$ and base normal subgroup $\widetilde{M}$. For any subgroup $B$ of $G$, define $\widetilde{B}=\widetilde{\lambda}^{-1}(B)$. We will use the following auxiliary lemma:
\[lem:tau\_1\] There exists an isomorphism $$\tau:[\widetilde{G},\widetilde{G}] \xrightarrow[]{\simeq} [\overline{G},\overline{G}]$$
with the following properties:
1. \[prop\_tau1\] $\overline{\lambda}(\tau(a))=\widetilde{\lambda}(a)$ for every $a \in [\widetilde{G},\widetilde{G}]$;
2. \[prop\_tau2\] $\tau([\widetilde{g}_1,\widetilde{g}_2])=[\overline{g}_1,\overline{g}_2]$ for all $\widetilde{g}_1,\widetilde{g}_2 \in \widetilde{G}$ and $\overline{g}_1,\overline{g}_2 \in \overline{G}$ such that $\widetilde{\lambda}(\widetilde{g}_i)=\overline{\lambda}(\overline{g}_i)$.
For any subgroup $B$ of $G$, $\tau$ further identifies
- $[\widetilde{B},\widetilde{B}] \cong [\overline{B},\overline{B}]$ and
- $\widetilde{M} \cap [\widetilde{B},\widetilde{B}] \cong \overline{M} \cap [\overline{B},\overline{B}].$
The isomorphism $\tau$ is constructed in [@kar Theorems 2.4.6(iv) and 2.5.1(i)] and the stated properties are clear from this construction. The additional identifications follow from \[prop\_tau1\] and \[prop\_tau2\].
Let $R$ be the set of ramified places of $N/K$. For any $v \in \Omega_K$, set $$\widetilde{S}_v=\begin{cases}
\widetilde{G_v}\textrm{, if $v \in R$,}\\
\textrm{a cyclic subgroup of } \widetilde{G_v} \textrm{ such that } \widetilde{\lambda}(\widetilde{S}_v) = G_v \textrm{, otherwise.}
\end{cases}$$
Furthermore, by the Chebotarev density theorem we can (and do) choose the subgroups $\widetilde{S}_v$ for $v \not\in R$ in such a way that all the cyclic subgroups of $\widetilde{G_v}$ such that $\widetilde{\lambda}(\widetilde{S}_v) = G_v$ occur.
As pointed out in [@Drak p. 31], a double coset decomposition $\overline{G}=\bigcup\limits_{k=1}^{r_{v,i}} \overline{H_i} \overline{x}_{i,k} \overline{G_v}$ corresponds to a double coset decomposition $\widetilde{G}=\bigcup\limits_{k=1}^{r_{v,i}} \widetilde{H_i} \widetilde{x}_{i,k} \widetilde{S}_v$, where $\widetilde{x}_{i,k}$ are any elements of $\widetilde{G}$ such that $\widetilde{\lambda}(\widetilde{x}_{i,k} )=\overline{\lambda}(\overline{x}_{i,k} )$.
Consider the following diagram analogous to : $$\label{diag:1stobs_defn_generalized}
\xymatrix{
\bigoplus\limits_{i=1}^{n} \widetilde{H_i}^{\textrm{ab}} \ar[r]^{\widetilde{\psi}_1} & \widetilde{G}^{\textrm{ab}}\\
\bigoplus\limits_{i=1}^{n}(\bigoplus\limits_{v \in \Omega_K} ( \bigoplus\limits_{w|v} {\widetilde{H}_{i,w}^{\textrm{ab}} }) ) \ar[r]^{\ \ \ \ \ \ \ \widetilde{\psi}_2} \ar[u]^{\widetilde{\varphi}_1}&\bigoplus\limits_{v \in \Omega_K}{\widetilde{S}_v^{\textrm{ab}} }\ar[u]_{\widetilde{\varphi}_2}
}$$
where $\widetilde{H}_{i,w}=\widetilde{H_i} \cap \widetilde{x}_{i,k} \widetilde{S}_v \widetilde{x}_{i,k}^{-1}$ and all the maps are defined as in diagram .
We now prove the main result of this section, namely that the object $\ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2)$ does not depend on the choice of generalized representation group (and thus, by Theorem \[thm:thm1DP\_gen\] and Proposition \[prop:1st=knot\], it always coincides with $\mathfrak{K}(L,K)$). Before we show this, we need a lemma. To ease the notation, we often omit the cosets $\widetilde{H_i}'$ and $\overline{H_i}'$ when working with elements of $\ker \widetilde{\psi}_1$ or $\ker \overline{\psi}_1$.
\[lem:simpl\_inters\] For any indices $1 \leq i_1 < i_2 \leq n$ and any $m \in \widetilde{H_{i_1}} \cap\widetilde{H_{i_2}} $, we have $$h=(1,\dots,\underbrace{m}_{i_1\textrm{-th entry}}, 1,\dots, 1, \underbrace{m^{-1}}_{i_2\textrm{-th entry}}, 1,\dots, 1) \in \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}).$$
We construct a vector $\alpha \in \bigoplus\limits_{i=1}^{n}(\bigoplus\limits_{\substack{v \in \Omega_K \\ v \text{ unramified}}}( \bigoplus\limits_{w|v} {\widetilde{H}_{i,w}^{\textrm{ab}} }) )$ such that $ \widetilde{\psi}_2(\alpha)=1$ and $\widetilde{\varphi}_1(\alpha)=h$. Let $v$ be an unramified place of $K$ such that $\widetilde{S}_v = \langle m \rangle$. By definition, if $\widetilde{G}=\bigcup\limits_{k=1}^{r_{v,i}} \widetilde{H_i} \widetilde{x}_{i,k} \widetilde{S}_v$ is a double coset decomposition of $\widetilde{G}$, then $\widetilde{H}_{i,w}= \widetilde{H_i} \cap \widetilde{x}_{i,k} \widetilde{S}_v \widetilde{x}_{i,k}^{-1}$. Let us suppose, without loss of generality, that $\widetilde{x}_{i_1,{k_1}}=1=\widetilde{x}_{i_2,{k_2}}$ for some index $1 \leq k_1 \leq r_{v,i_1}$ (respectively, $1 \leq k_2 \leq r_{v,i_2}$) corresponding to a place $w_1 \in \Omega_{L_{i_1}}$ (respectively, $w_2 \in \Omega_{L_{i_2}}$) via Lemma \[lem1DP\]. In this way, we have $m \in \widetilde{H}_{i_1,w_1}$ and $m^{-1} \in \widetilde{H}_{i_2,w_2}$. Setting the $(i_1,v,w_1)$-th (respectively, $(i_2,v,w_2)$-th) entry of $\alpha$ to be equal to $m$ (respectively, $m^{-1}$) and all other entries equal to $1$, we obtain $
\widetilde{\psi}_2(\alpha)=1$ and $\widetilde{\varphi}_1(\alpha)=h$.
\[thm:main\_knot\]
In the notation of diagram , we have $$\mathfrak{K}(L,K) \cong \ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2).$$
By Theorem \[thm:thm1DP\_gen\] and Proposition \[prop:1st=knot\], we have $\mathfrak{K}(L,K) \cong \ker \overline{\psi}_1 / \overline{\varphi}_1(\ker \overline{\psi}_2)$, where the $\overline{\phantom{a}}$ notation is as in diagram with respect to the groups of Proposition \[prop:1st=knot\]. Therefore, it suffices to prove that $$\ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2) \cong \ker \overline{\psi}_1 / \overline{\varphi}_1(\ker \overline{\psi}_2).$$
Define $$\begin{aligned}
f \colon \ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2) &\longrightarrow \ker \overline{\psi}_1 / \overline{\varphi}_1(\ker \overline{\psi}_2) \\
(\widetilde{h}_1 ,\dots,\widetilde{h}_n ) &\longmapsto (\overline{h}_1 ,\dots, \overline{h}_n )\end{aligned}$$
where, for each $i=1,\dots,n$, the element $\overline{h}_i \in \overline{H_i}$ is selected as follows: take $\overline{h}_i \in \overline{H_i}$ such that $\overline{\lambda}(\overline{h}_i)=\widetilde{\lambda}(\widetilde{h}_i)$ (note that $\overline{h}_i$ is only defined modulo ${\overline{M}}=\ker \overline{\lambda}$). In this way, we have $\overline{\lambda}(\overline{h}_1 \dots \overline{h}_n)=\widetilde{\lambda}(\widetilde{h}_1 \dots \widetilde{h}_n)$. Additionally, by Lemma \[lem:tau\_1\]\[prop\_tau1\], $\overline{\lambda}(\tau(\widetilde{h}_1 \dots \widetilde{h}_n))=\widetilde{\lambda}(\widetilde{h}_1 \dots \widetilde{h}_n)$ and thus $$\label{tau_eq}
\tau(\widetilde{h}_1 \dots \widetilde{h}_n)=\overline{h}_1 \dots \overline{h}_n m$$ for some $m \in \overline{M}$. Changing $\overline{h}_n$ if necessary, we assume that $m=1$ so that $\overline{h}_1 \dots \overline{h}_n \in [\overline{G},\overline{G}]$ and therefore $ (\overline{h}_1 ,\dots, \overline{h}_n ) \in \ker \overline{\psi}_1$.
**Claim 1:** $f$ is well defined, i.e. it does not depend on the choice of the elements $\overline{h}_i$ and moreover ${f(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2)) \subset \overline{\varphi}_1(\ker \overline{\psi}_2)}$.
**Proof:** We first prove that $f$ does not depend on the choice of $\overline{h}_i$. Suppose that, for each $i=1,\dots,n$, we choose elements $\underline{h}_i \in \overline{H_i}$ satisfying $\widetilde{\lambda}(\widetilde{h}_i)=\overline{\lambda}(\underline{h}_i)$ and $\tau(\widetilde{h}_1\dots\widetilde{h}_n)=\underline{h}_1 \dots \underline{h}_n$. We show that $(\underline{h}_1,\dots,\underline{h}_n)=(\overline{h}_1,\dots,\overline{h}_n)$ in $\ker \overline{\psi}_1 / \overline{\varphi}_1(\ker \overline{\psi}_2)$. Writing $\underline{h}_i=\overline{h}_i m_i$ for some $m_i \in \overline{M}$, it suffices to prove that $(m_1,\dots,m_n) \in \overline{\varphi}_1(\ker \overline{\psi}_2)$. Since $\overline{h}_1\dots \overline{h}_n =\tau(\widetilde{h}_1 \dots\widetilde{h}_n)=\underline{h}_1 \dots \underline{h}_n$ and the elements $m_i$ are in $\overline{M} \subset Z(\overline{G})$, we obtain $m_1 \dots m_n=1$. As $\overline{M} \subset \bigcap\limits_{i=1}^{n} \overline{H_i}$, multiplying $(m_1,\dots,m_n)$ by $(m_2,m_2^{-1},1,\dots,1)$ (which lies in $\overline{\varphi}_1(\ker \overline{\psi}_2)$ by Lemma \[lem:simpl\_inters\]), we have $(m_1,\dots,m_n)\equiv(m_1 m_2, 1,m_3,\dots , m_n) \pmod{\overline{\varphi}_1(\ker \overline{\psi}_2)}$. Repeating this procedure, we obtain $(m_1,\dots,m_n) \equiv (m_1\dots m_n,\dots,1)=(1,\dots,1) \pmod{\overline{\varphi}_1(\ker \overline{\psi}_2)}$ and therefore $(m_1,\dots,m_n)$ is in $\overline{\varphi}_1(\ker \overline{\psi}_2)$, as desired.
We now show that $f(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2)) \subset \overline{\varphi}_1(\ker \overline{\psi}_2)$. It suffices to check that $f(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v})) \subset \overline{\varphi}_1(\ker \overline{\psi}_2^{v})$ for any $v \in \Omega_K$. For $i=1,\dots,n$, let $\widetilde{G}=\bigcup\limits_{k=1}^{r_{v,i}} \widetilde{H_i} \widetilde{x}_{i,k} \widetilde{S}_v$ be a double coset decomposition of $\widetilde{G}$ and recall that, by definition, the group $\widetilde{H}_{i,w}$ equals $\widetilde{H_i} \cap \widetilde{x}_{i,k} \widetilde{S}_v \widetilde{x}_{i,k}^{-1}$ if $w \in \Omega_{L_i}$ corresponds to the double coset $\widetilde{H_i} \widetilde{x}_{i,k} \widetilde{S}_v$. Let $\alpha = \bigoplus\limits_{i=1}^{n} \bigoplus\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k} \in \ker \widetilde{\psi}_2^{v}$, where $\widetilde{h}_{i,k} \in \widetilde{H}_{i,w}$ for all possible $i,k$. We thus have $$\label{assumpt_v0}
\widetilde{\psi}_2(\alpha)=\prod\limits_{i=1}^{n}\prod\limits_{k=1}^{r_{v,i}} \widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k} \widetilde{x}_{i,k} \in [\widetilde{S}_v,\widetilde{S}_v].$$
For any $i=1,\dots,n$ define $\widetilde{h}_i= \prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k}$. We need to show that $f(\widetilde{h}_1,\dots,\widetilde{h}_n)$ is in $\overline{\varphi}_1(\ker \overline{\psi}_2^v)$.
Set $x_{i,k}:=\widetilde{\lambda}(\widetilde{x}_{i,k}) \in G$ and $h_{i,k}:=\widetilde{\lambda}(\widetilde{h}_{i,k}) \in {H}_{i} \cap x_{i,k} G_v x_{i,k}^{-1}$ for all possible $i,k$. We have $\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} x_{i,k}^{-1} h_{i,k} x_{i,k} \in [G_v,G_v]$. Let $\overline{x}_{i,k} \in \overline{G}$ be such that $\overline{\lambda}(\overline{x}_{i,k})=x_{i,k}$ and $\overline{h}_{i,k} \in \overline{H}_{i} \cap \overline{x}_{i,k} \overline{G_v} \overline{x}_{i,k}^{-1}$ satisfying $\overline{\lambda}(\overline{h}_{i,k})=h_{i,k}$. Multiplying one of the $\overline{h}_{1,k}$ by an element of $\overline{M}$ if necessary, we can assure that $$\label{assumpt_v}
\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \overline{x}_{i,k}^{-1} \overline{h}_{i,k} \overline{x}_{i,k} \in [\overline{G_v},\overline{G_v}].$$
In particular, $\alpha':=\bigoplus\limits_{i=1}^{n}\bigoplus\limits_{k=1}^{r_{v,i}} \overline{h}_{i,k}$ is in $ \ker \overline{\psi}_2^{v}$. Defining $\overline{h}_i := \prod\limits_{k=1}^{r_{v,i}} \overline{h}_{i,k}$ for $i=1,\dots,n$, we get $\overline{\varphi_1}(\alpha')=(\overline{h}_1 , \dots, \overline{h}_n )$. We have $\widetilde{\lambda}(\widetilde{h}_i)=\overline{\lambda}(\overline{h}_i)$ by construction and therefore $$\tau(\widetilde{h}_1 \dots \widetilde{h}_n)=\overline{h}_1 \dots \overline{h}_n m$$ for some $m \in \overline{M}$. We prove that $m$ is also in $[\overline{G_v},\overline{G_v}]$ so that, by multiplying one of the elements $\overline{h}_{1,k}$ by $m^{-1} \in \overline{M} \cap [\overline{G_v},\overline{G_v}]$ if necessary (note that doing so does not change condition ), we obtain $f(\widetilde{h}_1,\dots,\widetilde{h}_n)=(\overline{h}_1,\dots,\overline{h}_n) $. As $(\overline{h}_1,\dots,\overline{h}_n)$ is in $\overline{\varphi}_1(\ker \overline{\psi}_2^{v})$, this proves the claim.
Note that $$\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k}=(\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k}) (\prod\limits_{i=n}^{1} \prod\limits_{k=r_{v,i}}^{1} \widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k}^{-1} \widetilde{x}_{i,k}) \widetilde{\psi}_2(\alpha).$$ Denote $(\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k}) (\prod\limits_{i=n}^{1} \prod\limits_{k=r_{v,i}}^{1} \widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k}^{-1} \widetilde{x}_{i,k})$ by $\beta$. Then $\beta \in [\widetilde{G},\widetilde{G}]$ and using an explicit description of $\beta$ as a product of commutators and Lemma \[lem:tau\_1\]\[prop\_tau2\], we deduce that $\tau(\beta)=\beta'$, where $\beta'=(\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \overline{h}_{i,k}) (\prod\limits_{i=n}^{1} \prod\limits_{k=r_{v,i}}^{1} \overline{x}_{i,k}^{-1} \overline{h}_{i,k}^{-1} \overline{x}_{i,k})$. Therefore, we have $$\prod\limits_{i=1}^{n} \overline{h}_{i}=\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \overline{h}_{i,k} \equiv \beta' = \tau(\beta) \equiv \tau(\prod\limits_{i=1}^{n} \widetilde{h}_{i}) \pmod{[\overline{G_v},\overline{G_v}]},$$ and thus $m \in [\overline{G_v},\overline{G_v}]$, as desired.
**Claim 2:** $f$ is a homomorphism.
**Proof:** Let $h=(\widetilde{h}_1 ,\dots,\widetilde{h}_n ),h'=(\widetilde{h}'_1 ,\dots,\widetilde{h}'_n ) \in \ker \widetilde{\psi}_1$ and write $f(h)=(\overline{h}_1 ,\dots, \overline{h}_n )$ and $f(h')=(\overline{h}'_1 ,\dots, \overline{h}'_n )$ for some elements $\overline{h}_i,\overline{h}'_i \in \overline{H_i}$. We have $f(h)f(h')=(\overline{h}_1\overline{h}'_1 ,\dots,\overline{h}_n\overline{h}'_n )$. On the other hand, ${hh'=(\widetilde{h}_1\widetilde{h}'_1 ,\dots,\widetilde{h}_n\widetilde{h}'_n )}$ and $$\tau(\widetilde{h}_1 \widetilde{h}_1' \dots \widetilde{h}_n \widetilde{h}_n') \equiv \tau((\widetilde{h}_1 \dots \widetilde{h}_n)( \widetilde{h}_1' \dots \widetilde{h}_n')) = (\overline{h}_1\dots \overline{h}_n) (\overline{h}_1' \dots \overline{h}_n') \equiv \overline{h}_1\overline{h}_1'\dots \overline{h}_n\overline{h}_n' \pmod{[\overline{G},\overline{G}]}.$$
Since $\widetilde{\lambda}(\widetilde{h}_i \widetilde{h}_i')=\overline{\lambda}(\overline{h}_i \overline{h}_i')$ for all $i=1,\dots,n$ and $(\overline{h}_1\dots \overline{h}_n) (\overline{h}_1' \dots \overline{h}_n') \in [\overline{G},\overline{G}]$, by the definition of $f$ it follows that $f(h h')= (\overline{h}_1\overline{h}_1',\dots, \overline{h}_n\overline{h}_n')=f(h)f( h')$.
**Claim 3:** $f$ is surjective.
**Proof:** For $i=1,\dots,n$, let $\overline{h}_i \in \overline{H_i}$ be such that $\overline{h}_1 \dots \overline{h}_n \in [\overline{G},\overline{G}]$. Take any elements $\widetilde{h}_i \in \widetilde{H_i}$ satisfying $\widetilde{\lambda}(\widetilde{h}_i)=\overline{\lambda}(\overline{h}_i)$. As above, by Lemma \[lem:tau\_1\]\[prop\_tau1\] this implies that there exists $m \in \overline{M}$ such that $$\tau(\widetilde{h}_1\dots \widetilde{h}_n)=\overline{h}_1 \dots \overline{h}_n m \in [\overline{G},\overline{G}].$$ Since $\overline{h}_1\dots \overline{h}_n \in [\overline{G},\overline{G}]$, we have $m \in \overline{M} \cap [\overline{G},\overline{G}]$. But $\overline{M} \cap [\overline{G},\overline{G}] = \tau(\widetilde{M} \cap [\widetilde{G},\widetilde{G}])$ by Lemma \[lem:tau\_1\]. Therefore $m = \tau(m')$ for some $m' \in \widetilde{M} \cap [\widetilde{G},\widetilde{G}]$ and thus $(\overline{h}_1,\dots,\overline{h}_n)=f(\widetilde{h}_1,\dots,\widetilde{h}_n m'^{-1})$.
**Claim 4:** $f$ is an isomorphism.
**Proof:** We have seen that $f$ is surjective. Now we can analogously define a surjective map from $\ker \overline{\psi}_1 / \overline{\varphi}_1(\ker \overline{\psi}_2)$ to $\ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2) $. It follows that the finite groups $\ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2) $ and $\ker \overline{\psi}_1 / \overline{\varphi}_1(\ker \overline{\psi}_2)$ have the same size and so $f$ is an isomorphism.
Using this theorem, one can also obtain descriptions of the birational invariant $\operatorname{H}^1(K,\operatorname{Pic}\overline{X})$ and the defect of weak approximation $A(T)$ for the multinorm one torus $T$:
\[thm:main\_result\] Let $T$ be the multinorm one torus associated to $L$ and let $X$ be a smooth compactification of $T$. In the notation of diagram , we have $$\Sha(T) \cong \ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2),$$ $$\operatorname{H}^1(K,\operatorname{Pic}\overline{X})\cong \ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{{nr}}),$$ $$A(T) \cong \widetilde{\varphi}_1(\ker \widetilde{\psi}_2) / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{{nr}}).$$
The first isomorphism is the statement of Theorem \[thm:main\_knot\] (recall that $\Sha(T)$ is canonically isomorphic to $\mathfrak{K}(L,K)$). The two remaining isomorphisms follow in the same way as in the Hasse norm principle case, see [@Drak p. 32–33].
\[rem:finite\_time2\]
As explained in Remark \[rem:finite\_time\], all the groups $ \ker \widetilde{\psi}_1, \widetilde{\varphi}_1(\ker \widetilde{\psi}_2)$ and $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ in Theorem \[thm:main\_result\] can be computed in finite time. To this extent, we assembled a function in GAP [@gap] (whose code is available in [@macedo_code]) that, given the relevant local and global Galois groups, outputs the obstructions to the multinorm principle and weak approximation for the multinorm one torus of a finite number of extensions by means of Theorem \[thm:main\_result\].
We end this section by generalizing Corollary \[cor:square\_free\] and proving that, in many situations, one can actually circumvent the use of generalized representation groups when computing the obstructions to the local-global principles.
Before we present this result, we need to introduce the notion of focal subgroups. For a moment, let $G$ be any finite group and let $H$ be a subgroup of $G$. The *focal subgroup of $H$ in $G$* is defined as $\Phi^{{G}}({H})=\langle [h,x] | h \in {H} \cap x {H} x^{-1}, x \in {G} \rangle$. In [@DP Theorem 2], it was proved that $${\varphi}_1(\ker {\psi}_2^{{nr}}) = \Phi^{G}(H) / [H,H]$$ in the setting of the first obstruction to the Hasse norm principle (case $n=1$). Returning to the multinorm context, this fact promptly implies that, in the notation of diagram , we have $$\label{eq:nr_hnp_inclusion}
(1,\dots,\underbrace{\Phi^{\widetilde{G}}(\widetilde{H_i})}_{i\textrm{-th entry}}, 1,\dots, 1) \subset \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}).$$ for every $i=1,\dots,n$.
\[prop:sq\_free\_mid\_gp\]
Suppose that there exists $j \in \{1,\dots,n\}$ such that, for every prime $p$ dividing $|\hat{\operatorname{H}}^{-3}(G,{\mathbb{Z}})|$, $p^2$ does not divide $[L_j:K]$. Then, in the notation of diagram , we have $$\Sha(T) \cong \ker {\psi}_1 / {\varphi}_1(\ker {\psi}_2),$$ $$\operatorname{H}^1(K,\operatorname{Pic}\overline{X})\cong \ker {\psi}_1 / {\varphi}_1(\ker {\psi}_2^{{nr}}),$$ $$A(T) \cong {\varphi}_1(\ker {\psi}_2) / {\varphi}_1(\ker {\psi}_2^{{nr}}).$$
We prove only that $\operatorname{H}^1(K,\operatorname{Pic}\overline{X})\cong \ker {\psi}_1 / {\varphi}_1(\ker {\psi}_2^{{nr}})$ (the other two isomorphisms can be obtained by a similar argument). Assume, without loss of generality, that $j=1$ and $\widetilde{G}$ is a Schur covering group of $G$ so that $ \widetilde{M}$ is contained in $ [\widetilde{G},\widetilde{G}]$ and $\widetilde{M} \cong \hat{\operatorname{H}}^{-3}(G,{\mathbb{Z}})$. We show that the map $$\begin{aligned}
\rho \colon\ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) &\longrightarrow \ker {\psi}_1 / {\varphi}_1(\ker {\psi}_2^{nr})\\
h=(\widetilde{h}_1 ,\dots,\widetilde{h}_n ) &\longmapsto (\widetilde{\lambda}(\widetilde{h}_1) ,\dots,\widetilde{\lambda}(\widetilde{h}_n))\end{aligned}$$
is an isomorphism, which proves the desired statement by Theorem \[thm:main\_result\].
We first verify that $\rho$ is well defined. It is enough to check that $\rho(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v})) \subset {\varphi}_1(\ker {\psi}_2^{v})$ for an unramified place $v$ of $N/K$. Note that if $\widetilde{G}=\bigcup\limits_{k=1}^{r_{v,i}} \widetilde{H_i} \widetilde{x}_{i,k} \widetilde{S}_v$ is a double coset decomposition of $\widetilde{G}$, then ${G}=\bigcup\limits_{k=1}^{r_{v,i}} {H}_i {x}_{i,k} G_v$ is a double coset decomposition of ${G}$, where ${x}_{i,k}=\widetilde{\lambda}(\widetilde{x}_{i,k})$. From this observation, it is straightforward to verify that $\rho(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v})) \subset {\varphi}_1(\ker {\psi}_2^{v})$.
We now prove that $\rho$ is surjective. Suppose that we are given, for $i=1,\dots,n$, elements $h_i \in H_i$ such that $h_1 \dots h_n \in [G,G]$. Since $\widetilde{M} \subset [\widetilde{G},\widetilde{G}]$, any choice of elements $\widetilde{h}_i \in \widetilde{H_i}$ such that $\widetilde{\lambda}(\widetilde{h}_i)=h_i$ will satisfy $\widetilde{h}_1 \dots \widetilde{h}_n \in [\widetilde{G},\widetilde{G}]$ and thus $({h}_1, \dots, {h}_n)=\rho(\widetilde{h}_1 ,\dots ,\widetilde{h}_n)$.
We finally show that $\rho$ is injective. Suppose that $(h_1,\dots,h_n)=\rho(h) \in {\varphi}_1(\ker {\psi}_2^{v})$ for some unramified place $v$ of $N/K$. Write $h_i=\varphi_1(\bigoplus\limits_{k=1}^{r_{v,i}} h_{i,k})$ for some elements $h_{i,k} \in H_i \cap x_{i,k} G_v x_{i,k}^{-1}$. As $\rho(h) \in {\varphi}_1(\ker {\psi}_2^{v})$, we obtain $\prod\limits_{i=1}^{n}\prod\limits_{k=1}^{r_{v,i}} x_{i,k}^{-1} h_{i,k} x_{i,k}=1$. Picking elements $\widetilde{h}_{i,k}\in\widetilde{\lambda}^{-1}(h_{i,k})$ and $\widetilde{x}_{i,k}\in\widetilde{\lambda}^{-1}(x_{i,k})$ for all possible $i,k$, we obtain $\prod\limits_{i=1}^{n}\prod\limits_{k=1}^{r_{v,i}} \widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k} \widetilde{x}_{i,k}=m$ for some $m \in \widetilde{M} = \ker \widetilde{\lambda}$. As $m \in Z(\widetilde{G}) \cap \bigcap\limits_{i=1}^{n} \widetilde{H_i}$, we have $(\widetilde{h}_1 m^{-1},\widetilde{h}_2, ,\dots,\widetilde{h}_n) \in \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$. Therefore, in order to prove that $h \in \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ it suffices to show that $(m^{-1},1,\dots,1) \in \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$. We prove that $m \in \Phi^{\widetilde{G}}(\widetilde{H_1})$, which completes the proof by .
**Claim:** If $p^2$ does not divide $[L_1:K]$ for every prime $p$ dividing $|\widetilde{M}|$, then $\widetilde{M} \subset \Phi^{\widetilde{G}}(\widetilde{H_1})$.
**Proof:** We show that $\widetilde{M}_{(p)} \subset \Phi^{\widetilde{G}}(\widetilde{H_1})$. We have $[L_1:K]=[G:H_1]$ and therefore $[G_p:(H_1)_{p}] = [\widetilde{G}_p :(\widetilde{H_1})_{p}]=1$ or $p$. In any case, $(\widetilde{H_1})_p \trianglelefteq \widetilde{G}_p$ and we can write $\widetilde{G}_p = \langle x_p \rangle . (\widetilde{H_1})_p $ for some $x_p \in \widetilde{G}_p$. Since $\widetilde{M}_{(p)} \subset \widetilde{G}_p \cap [\widetilde{G},\widetilde{G}] \cap Z(\widetilde{G})$ and $\widetilde{G}_p \cap [\widetilde{G},\widetilde{G}] \cap Z(\widetilde{G}) \subset [\widetilde{G}_p,\widetilde{G}_p]$ (this last inclusion follows from properties of the transfer map, e.g. [@Isaacs Lemma 5.5]), we have $\widetilde{M}_{(p)} \subset [\widetilde{G}_p ,\widetilde{G}_p ]$ and so it suffices to prove that $[\widetilde{G}_p ,\widetilde{G}_p ] \subset \Phi^{\widetilde{G}}(\widetilde{H_1})$. Let $z=[x_p^a h_1, x_p^b h_1']$ for some $a,b \in {\mathbb{Z}}$ and $h_1,h_1' \in (\widetilde{H_1})_p$. Using the commutator properties, we have $z=[x_p^a,h_1']^{h_1}[h_1,h_1'][h_1,x_p^b]^{h_1'}$. As $(\widetilde{H_1})_p \trianglelefteq \widetilde{G}_p $ and $\Phi^{\widetilde{G}}(\widetilde{H_1}) \trianglelefteq \widetilde{H_1}$, it follows that each one of the commutators above is in $\Phi^{\widetilde{G}}(\widetilde{H_1})$.
As a consequence we obtain the following result, which can be thought of as an analog of [[@DP Corollary 1]]{} for the birational invariant $\operatorname{H}^1(K,\operatorname{Pic}\overline{X})$.
Let $L/K$ be an extension of global fields and suppose that $[L:K]$ is square-free. Let $X$ be a smooth compactification of the norm one torus $R^1_{L/K} {{\mathbb G}}_m$. Then $$\operatorname{H}^1(K,\operatorname{Pic}\overline{X}) \cong \frac{H \cap [G,G]}{\Phi^{G}(H)}.$$
The conditions of Proposition \[prop:sq\_free\_mid\_gp\] are satisfied and hence $\operatorname{H}^1(K,\operatorname{Pic}\overline{X}) \cong \ker {\psi}_1 / {\varphi}_1(\ker {\psi}_2^{nr})$. The result then follows from [@DP Theorem 2].
Applications {#sec:applications}
============
In this section we employ the techniques developed so far in order to analyze the multinorm principle or weak approximation for the multinorm one torus in three different situations. Namely, we extend results of Bayer-Fluckiger–Lee–Parimala [@eva], Demarche–Wei [@demarche] and Pollio [@pollio]. The notation used throughout this section is as in Sections \[sec:1st\_obs\] and \[sec:gen\_gps\]. Additionally, we will make use of the norm one torus $S=R^1_{F/K} {{\mathbb G}}_m$ of the extension $F=\bigcap\limits_{i=1}^{n} L_i$ and we let $Y$ denote a smooth compactification of $S$. We start by establishing a few auxiliary lemmas to be used in later applications.
Preliminary results
-------------------
\[lem:incl\_unr\] In the notation of diagram , we have $$\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) \subseteq \{ (h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}') \in \ker \widetilde{\psi}_1 | h_1\dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H}) \}.$$
A proof of this lemma can be obtained by following the same strategy as in the proof of the analogous result for the Hasse norm principle (case $n=1$) in [@DP Theorem 2]. Nonetheless, as the details are slightly intricate, we include a proof here for the benefit of the reader.
Since $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) = \prod\limits_{\substack{v \in \Omega_K \\ v \text{ unramified}}} \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{ v })$, it suffices to prove that $$\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v}) \subseteq \{ (h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}') | h_1\dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H}) \}$$ for any unramified place $v$ of $N/K$. Let $\alpha \in \ker \widetilde{\psi}_2^{ v }$ and fix a double coset decomposition $\widetilde{G}=\bigcup\limits_{k=1}^{r_{v,i}} \widetilde{H_i} \widetilde{x}_{i,k} \widetilde{S}_v$. Write $\widetilde{S}_v=\langle g \rangle$ and $\alpha = \bigoplus\limits_{i=1}^{n} \bigoplus\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k} $ for some $g \in \widetilde{G}$, $\widetilde{h}_{i,k} = \widetilde{x}_{i,k} g^{e_{i,k}} \widetilde{x}_{i,k}^{-1} \in \widetilde{H_i} \cap \widetilde{x}_{i,k} \langle g \rangle \widetilde{x}_{i,k}^{-1}$ and some $e_{i,k} \in {\mathbb{Z}}$. By hypothesis, we have $1=\widetilde{\psi}_2(\alpha)=g^{\sum_{i,k} e_{i,k}}$ and therefore $$\sum\limits_{i,k} e_{i,k} \equiv 0 \pmod{m},$$ where $m$ is the order of $g$. Since $g^m=1$, by changing some of the $e_{i,k}$ if necessary, we can (and do) assume that $$\label{eq:sum=0}
\sum\limits_{i,k} e_{i,k} = 0.$$
Letting ${h_i} =\prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k} $ for any $1 \leq i \leq n$, we have $\widetilde{\varphi}_1(\alpha)=(h_1 \widetilde{H_1},\dots,h_n \widetilde{H_n}) \in \ker \widetilde{\psi}_1$. We prove that $$\prod\limits_{i=1}^{n} {h_i} =\prod\limits_{i=1}^{n}(\prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k})=\prod\limits_{i=1}^{n} (\prod\limits_{k=1}^{r_{v,i}} \widetilde{x}_{i,k} g^{e_{i,k}} \widetilde{x}_{i,k}^{-1}) \in \Phi^{\widetilde{G}}(\widetilde{H})$$ by induction on $s:=\sum\limits_{i=1}^{n} r_{v,i}$. The case $s=1$ is trivial and the case $s=2$ is solved in [@DP p. 308]. Now let $s > 2$ and set $d=\gcd(e_{i,k} | 1 \leq i \leq n, 1 \leq k \leq r_{v,i})$ and $f_{i,k}=\frac{e_{i,k}}{d}$. It follows that $\gcd(f_{i,k} | 1 \leq i \leq n, 1 \leq k \leq r_{v,i})=1$ and, since $\sum\limits_{i,k} f_{i,k} = 0$ by , we have ${\gcd(f_{i,k} | 1 \leq i \leq n, 1 \leq k \leq r_{v,i} \textrm{ and } (i,k) \neq (n,r_{v,n}))}=1$. Hence there exist $a_{i,k} \in {\mathbb{Z}}$ such that $\sum\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} f_{i,k} a_{i,k} = 1$. Consider the element $$\beta=\Big(\bigoplus\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} \widetilde{x}_{i,k} g^{e_{i,k}f_{n,r_{v,n}}a_{i,k}} \widetilde{x}_{i,k}^{-1}\Big) \oplus \widetilde{x}_{n,r_{v,n}} g^{-e_{n,r_{v,n}}} \widetilde{x}_{n,r_{v,n}}^{-1} \in \bigoplus\limits_{i=1}^{n}\Big( \bigoplus\limits_{k=1}^{r_{v,i}} {\widetilde{H}_{i,w} }\Big) .$$ Since $e_{i,k} f_{n,r_{v,n}}=e_{n,r_{v,n}} f_{i,k},$ we have $$\widetilde{\psi}_2(\beta)=g^{\Big(\sum\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} e_{i,k}f_{n,r_{v,n}}a_{i,k}\Big)-e_{n,r_{v,n}}}=g^{\Big(\sum\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} e_{n,r_{v,n}}f_{i,k}a_{i,k}\Big)-e_{n,r_{v,n}}}=1$$ and so $\beta \in \ker \widetilde{\psi}_2^{v}$.
Additionally, if $\widetilde{\varphi}_1(\beta) = (\widetilde{h}_1,\dots,\widetilde{h}_n)$, we have
$$\begin{split}
\prod\limits_{i=1}^{n} \widetilde{h}_i & = \left(\prod\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} \widetilde{x}_{i,k} g^{e_{i,k}f_{n,r_{v,n}}a_{i,k}} \widetilde{x}_{i,k}^{-1} \right) \widetilde{x}_{n,r_{v,n}} g^{-e_{n,r_{v,n}}} \widetilde{x}_{n,r_{v,n}}^{-1} = \\
& = \left(\prod\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} \widetilde{x}_{i,k} g^{e_{i,k}f_{n,r_{v,n}}a_{i,k}} \widetilde{x}_{i,k}^{-1} \right) \widetilde{x}_{n,r_{v,n}} g^{-e_{n,r_{v,n}} \sum\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} f_{i,k} a_{i,k}} \widetilde{x}_{n,r_{v,n}}^{-1} \equiv \\
& \equiv \left(\prod\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} \widetilde{x}_{i,k} g^{e_{i,k}f_{n,r_{v,n}}a_{i,k}} \widetilde{x}_{i,k}^{-1} \widetilde{x}_{n,r_{v,n}} g^{-e_{i,k}f_{n,r_{v,n}}a_{i,k}} \widetilde{x}_{n,r_{v,n}}^{-1} \right) \pmod{[\widetilde{H},\widetilde{H}]}
\end{split}$$
since the elements $\widetilde{x}_{i,k} g^{e_{i,k}} \widetilde{x}_{i,k}^{-1}$ (for all possible $i,k$) are in $\widetilde{H}$. Arguing similarly to the case $s=2$ (see [@DP p. 308]), we deduce that $\prod\limits_{i=1}^{n} \widetilde{h}_i \in \Phi^{\widetilde{G}}(\widetilde{H})$. Finally, consider the element $$\alpha'=\alpha \beta=\bigoplus\limits_{\substack{i,k\\(i,k) \neq (n,r_{v,n})}} \widetilde{x}_{i,k} g^{e_{i,k}(1+f_{n,r_{v,n}}a_{i,k})} \widetilde{x}_{i,k}^{-1} \in \bigoplus\limits_{i=1}^{n}\Big( \bigoplus\limits_{k=1}^{r_{v,i}} {\widetilde{H}_{i,w} }\Big).$$ It is clear that $\alpha' \in \ker \widetilde{\psi}_2^{v}$. By the induction hypothesis, if $\widetilde{\varphi}_1(\alpha')=(\widehat{h}_1,\dots,\widehat{h}_n)$ we have $\widehat{h}_1 \dots \widehat{h}_n \in \Phi^{\widetilde{G}}(\widetilde{H})$. Since $\widehat{h}_i \equiv h_i \widetilde{h}_i \pmod{[\widetilde{H},\widetilde{H}]}$ for all $i=1,\dots,n$, we conclude that ${h_1} \dots {h_n} \in \Phi^{\widetilde{G}}(\widetilde{H})$ as well.
\[lem:surject\_int\_Galois\]
1. \[lem:surject\_int\_Galois1\] There exists a surjection $f:\operatorname{H}^1(K,\operatorname{Pic}\overline{X}) \xrightarrow{} \operatorname{H}^1(K,\operatorname{Pic}\overline{Y})$. If in addition $$\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) \supseteq \{ (h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}') | h_1\dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H}) \}$$ (in the notation of diagram ), then $f$ is an isomorphism.
2. \[lem:surject\_int\_Galois2\] If $F/K$ is Galois, $f$ induces a surjection $\Sha(T) \twoheadrightarrow \Sha(S)$.
Consider the analog of diagram for the extension $F/K$ (note that this is the fixed field of the group ${H}$ inside $N/K$): $$\label{diag:F/K_v0}
\xymatrix{
\widetilde{H}^{\textrm{ab}} \ar[r]^{\widehat{\psi}_1} & \widetilde{G}^{\textrm{ab}}\\
\bigoplus\limits_{v \in \Omega_K} ( \bigoplus\limits_{w|v} {\widetilde{H}_{w}^{\textrm{ab}} }) \ar[r]^{\ \ \ \ \ \ \ \widehat{\psi}_2} \ar[u]^{\widehat{\varphi}_1}&\bigoplus\limits_{v \in \Omega_K}{\widetilde{S}_v^{\textrm{ab}} }\ar[u]_{\widehat{\varphi}_2}
}$$
Here all the maps with the $\widehat{\phantom{a}}$ notation are defined as in diagram with respect to the extension $F/K$. Now define $$\begin{aligned}
f \colon \ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) &\longrightarrow \ker \widehat{\psi}_1 / \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr}) \\
(\widetilde{h}_1 \widetilde{H_1}',\dots,\widetilde{h}_n \widetilde{H_n}')\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) &\longmapsto (\widetilde{h}_1 \dots\widetilde{h}_n[\widetilde{H},\widetilde{H}]) \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})\end{aligned}$$ Since $\widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})=\Phi^{\widetilde{G}}(\widetilde{H})/[\widetilde{H},\widetilde{H}]$ (see [@DP Theorem 2]), the map $f$ is well defined by Lemma \[lem:incl\_unr\]. Additionally, as the target group is abelian, it is easy to check that $f$ is a homomorphism and surjective. By Theorem \[thm:main\_result\] we have $\operatorname{H}^1(K,\operatorname{Pic}\overline{X}) \cong \ker \widetilde{\psi}_1 / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ and $\operatorname{H}^1(K,\operatorname{Pic}\overline{Y}) \cong \ker \widehat{\psi}_1 / \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$. The statement in the first sentence follows. Finally, if we assume $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) \supseteq \{ (h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}') | h_1\dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H}) \}$, then it is clear that $f$ is injective.
We now prove \[lem:surject\_int\_Galois2\]. By Theorem \[thm:main\_result\], it is enough to show that $f( \widetilde{\varphi}_1(\ker \widetilde{\psi}_2)) \subset \widehat{\varphi}_1(\ker \widehat{\psi}_2)$. Since $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2)=\prod\limits_{v \in \Omega_K} \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v})$, it suffices to verify $f(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v}))\subset \widehat{\varphi}_1(\ker \widehat{\psi}_2)$ for all $v \in \Omega_K$. Let $\alpha \in \ker \widetilde{\psi}_2^{ v }$ and write $\alpha = \bigoplus\limits_{i=1}^{n} \bigoplus\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k} $ for some $\widetilde{h}_{i,k} \in \widetilde{H_i} \cap \widetilde{x}_{i,k} \widetilde{S}_v \widetilde{x}_{i,k}^{-1}$. Hence, we obtain $\widetilde{\varphi}_1(\alpha)=(\widetilde{h}_1,\dots,\widetilde{h}_n)$, where $\widetilde{h}_i=\prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k}$, and we wish to show that $\prod\limits_{i=1}^{n} \widetilde{h}_i \in \widehat{\varphi}_1(\ker \widehat{\psi}_2)$. Since $F/K$ is Galois, $\widetilde{H}$ is a normal subgroup of $ \widetilde{G}$ and thus $\Phi^{\widetilde{G}}(\widetilde{H})=[\widetilde{H},\widetilde{G}]$. In this way, we have $$\prod\limits_{i=1}^{n} \widetilde{h}_i = \prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \widetilde{h}_{i,k} \equiv \prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k} \widetilde{x}_{i,k} = \widetilde{\psi}_2(\alpha) \pmod{\Phi^{\widetilde{G}}(\widetilde{H})}.$$
As $\Phi^{\widetilde{G}}(\widetilde{H})/[\widetilde{H},\widetilde{H}]=\widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$, it suffices to prove that $\widetilde{\psi}_2(\alpha) \in \widehat{\varphi}_1(\ker \widehat{\psi}_2^{v})$. For this, let $\widetilde{G}=\bigcup\limits_{j=1}^{r} \widetilde{H} \widetilde{y}_j \widetilde{S}_v$ be a double coset decomposition and suppose, without loss of generality, that $\widetilde{y}_{j_0}=1$ for some index $1 \leq j_0 \leq r$ corresponding to a place $w_0$ of $F$ via Lemma \[lem1DP\]. Therefore, we obtain $\widetilde{\psi}_2(\alpha)=\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} \widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k} \widetilde{x}_{i,k} \in \widetilde{H} \cap \widetilde{S}_v=\widetilde{H}_{w_0}$ since $\widetilde{x}_{i,k}^{-1} \widetilde{h}_{i,k} \widetilde{x}_{i,k} \in \widetilde{H}$ for all possible $i,k$. In this way, if $\beta \in \bigoplus\limits_{v \in \Omega_K} ( \bigoplus\limits_{w|v} {\widetilde{H}_{w}^{\textrm{ab}} }) $ is the vector with the $(v,w_0)$-th entry equal to $\widetilde{\psi}_2(\alpha)$ and all other entries equal to $1$, we have $\widehat{\psi}_2(\beta)=\widetilde{\psi}_2(\alpha) \in [\widetilde{S}_v,\widetilde{S}_v]$ (as $\alpha \in \ker \widetilde{\psi}_2^{ v }$) and so $\widetilde{\psi}_2(\alpha) =\widehat{\varphi}_1(\beta) \in \widehat{\varphi}_1(\ker \widehat{\psi}_2^{v}) $.
Multinorm principle for linearly disjoint extensions
----------------------------------------------------
In this subsection we prove a theorem similar to the main result of [@demarche], but with a slightly different hypothesis (and in some cases more general, see Remark \[rem:demarche\_different\] below).
\[thm:demarch\_wei\_thm\] For any non-empty subset $I \subset \{1,\dots,n\}$, let $L_I \subseteq N$ be the compositum of the fields $L_i$ $(i \in I)$ and let $E_{I}$ be the Galois closure of $L_I/K$. Suppose that there exist indices $i_0,j_0 \in \{1,\dots,n\}$ such that, for every $1 \leq i \leq n$, there is a partition $I_i \sqcup J_i = \{1,\dots,n\}$ with $i_0 \in I_i,j_0 \in J_i$ and $E_{I_i} \cap E_{J_i} \subseteq L_i$. Then $$\operatorname{H}^1(K,\operatorname{Pic}\overline{X}) \cong \operatorname{H}^1(K,\operatorname{Pic}\overline{Y}).$$
If $n=1$ there is nothing to show, so assume $n \geq 2$. By Lemma \[lem:surject\_int\_Galois\]\[lem:surject\_int\_Galois1\] it suffices to prove that $$\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) \supseteq \{ (h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}') | h_1\dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H}) \}.$$
Let $\alpha=(h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}')$ be such that $h_1 \dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H})$. Renaming the fields $L_i$ if necessary, we assume that $i_0 =1$ and $j_0=2$. Denoting $B_{I_i}=\operatorname{Gal}(N/E_{I_i}), B_{J_i}=\operatorname{Gal}(N/E_{J_i})$ for all $1 \leq i \leq n$, the hypothesis $E_{I_i} \cap E_{J_i} \subseteq L_i$ is equivalent to $ B_{I_i}B_{J_i} \supseteq H_i$ and thus $$\label{eq:inc_hyp}
\widetilde{H_i} \subseteq \widetilde{B_{I_i}} \widetilde{B_{J_i}}$$ with $1 \in I_i$, $2 \in J_i$ and $i \in I_i$ or $J_i$. If $n \geq 3$, this implies that for any $3 \leq i \leq n$ we can decompose $h_i = h_{1,i} h_{2,i}$ for some $h_{1,i} \in \widetilde{H_1} \cap \widetilde{H_i}$ and $h_{2,i} \in \widetilde{H_2} \cap \widetilde{H_i}$. Using Lemma \[lem:simpl\_inters\] as done in Claim 1 of the proof of Theorem \[thm:main\_knot\], we obtain $$\alpha \equiv ((\prod\limits_{3 \leq i \leq n} h_{1,i}) h_1,(\prod\limits_{3 \leq i \leq n} h_{2,i}) h_2,1,\dots,1)$$ modulo $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$. We can thus assume $\alpha$ to be of the form $(h_1',h_2',1,\dots,1)$ for some $h_1' \in \widetilde{H_1},h_2' \in \widetilde{H_2}$ such that $h_1' h_2' \in \Phi^{\widetilde{G}}(\widetilde{H})$. Note that implies that $\widetilde{H} =\langle \widetilde{H_i} \rangle \subset \widetilde{B_1} \widetilde{B_2}$, where $B_1=\operatorname{Gal}(N/E_{\{1\}})$ and $B_2=\operatorname{Gal}(N/E_{\{2\}})$. It thus follows that $\Phi^{\widetilde{G}}(\widetilde{H}) \subset \Phi^{\widetilde{G}}(\widetilde{B_1}\widetilde{B_2})=\Phi^{\widetilde{G}}(\widetilde{B_1})\Phi^{\widetilde{G}}(\widetilde{B_2}) $ and so $h_1' h_2' \in \Phi^{\widetilde{G}}(\widetilde{B_1}) \Phi^{\widetilde{G}}(\widetilde{B_2})$. Since $\Phi^{\widetilde{G}}(\widetilde{B_i}) \subset \Phi^{\widetilde{G}}(\widetilde{H_i})$ and recalling that $$(1,\dots,\underbrace{\Phi^{\widetilde{G}}(\widetilde{H_i})}_{i\textrm{-th entry}}, 1,\dots, 1) \subset \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$$ (see in Section \[sec:gen\_gps\]), we can multiply $h_1'$ and $h_2'$ by elements of $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ to attain $\alpha \equiv (h_1'',h_2'',1,\dots,1) \pmod{\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})}$ for some $h_1'' \in \widetilde{H_1},h_2'' \in \widetilde{H_2}$ such that $h_1'' h_2''=1$. Thus $h_2''=h_1''^{-1}$ and $\alpha = (h_1'',h_1''^{-1},1,\dots,1)$, which by Lemma \[lem:simpl\_inters\] is in $ \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$, as desired.
\[rem:demarche\_different\] It is easy to see that if there exists a partition $I \sqcup J=\{1,\dots,n\}$ such that $E_I \cap E_J=F$ (the assumption in [@demarche Theorem 6] when $F_i=E_I$ and $F_j=E_J$ for every $i \in I, j \in J$), the conditions of Theorem \[thm:demarch\_wei\_thm\] are satisfied. Therefore, our theorem applies to all the cases described in [@demarche Example 9(i)–(iii)]. Moreover, our hypothesis applies for $n$-tuples of fields for which the assumptions in [@demarche Theorem 6] might fail. For example, let $L=({\mathbb{Q}}(\sqrt{2},\sqrt{3}),{\mathbb{Q}}(\sqrt{2},\sqrt{5}),{\mathbb{Q}}(\sqrt{3},\sqrt{5}))$. It is easy to see that the assumptions of Theorem \[thm:demarch\_wei\_thm\] are satisfied, but [@demarche Theorem 6] does not apply to this tuple of fields. Indeed, Demarche and Wei’s hypothesis imply that there is a partition $I \sqcup J=\{1,\dots,n\}$ such that $L_I \cap L_J=F$, which does not exist in the example above.
As consequence of Theorem \[thm:demarch\_wei\_thm\] we also obtain versions of [@demarche Corollaries 7 and 8]:
\[cor:dem1\] Let $c\in K^*$. Assume the hypothesis of Theorem \[thm:demarch\_wei\_thm\] and suppose that the $K$-variety $ N_{F/K}(\Xi)=c$ satisfies weak approximation. Then the multinorm equation $\prod\limits_{i=1}^{n} N_{L_i/K}(\Xi_i)=c$ satisfies weak approximation if and only if it has a $K$-point.
\[cor:dem2\] Assume the hypothesis of Theorem \[thm:demarch\_wei\_thm\] and suppose that the Hasse principle and weak approximation hold for all norm equations $ N_{F/K}(\Xi)=c$, $c \in K^*$. Then the Hasse principle and weak approximation hold for all multinorm equations $\prod\limits_{i=1}^{n} N_{L_i/K}(\Xi_i)=c$.
Multinorm principle and weak approximation for abelian extensions
-----------------------------------------------------------------
In this subsection we generalize the main theorem of [@pollio] to $n$ abelian extensions under the conditions of Theorem \[thm:demarch\_wei\_thm\].
\[thm:pollio\] Let $L=(L_1,\dots,L_n)$ be an $n$-tuple of abelian extensions of $K$ and suppose that the conditions of Theorem \[thm:demarch\_wei\_thm\] are satisfied for $L$. Then $$\Sha(T) \cong \Sha(S),$$ $$A(T) \cong A(S).$$
Note that if $A(T) \cong A(S)$, then by Theorem \[thm:demarch\_wei\_thm\] and Voskresenskiĭ’s exact sequence we deduce that $|\Sha(T)|=|\Sha(S)|$. Since $\Sha(T)$ surjects onto $\Sha(S)$ by Lemma \[lem:surject\_int\_Galois\]\[lem:surject\_int\_Galois2\], we conclude that $\Sha(T) \cong \Sha(S)$. Therefore, it is enough to prove that $A(T) \cong A(S)$.
Let us again consider the analog of diagram for the extension $F/K$: $$\label{diag:F/K_v2}
\xymatrix{
\widetilde{H}^{\textrm{ab}} \ar[r]^{\widehat{\psi}_1} & \widetilde{G}^{\textrm{ab}}\\
\bigoplus\limits_{v \in \Omega_K} ( \bigoplus\limits_{w|v} {\widetilde{H}_{w}^{\textrm{ab}} }) \ar[r]^{\ \ \ \ \ \ \ \widehat{\psi}_2} \ar[u]^{\widehat{\varphi}_1}&\bigoplus\limits_{v \in \Omega_K}{\widetilde{S}_v^{\textrm{ab}} }\ar[u]_{\widehat{\varphi}_2}
}$$
As before, in this diagram all the maps with the $\widehat{\phantom{a}}$ superscript are defined as in diagram with respect to $F/K$. By Theorem \[thm:main\_result\], we have $A(T) \cong \widetilde{\varphi}_1(\ker \widetilde{\psi}_2)/ \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ (in the notation of diagram ) and $A(S) \cong \widehat{\varphi}_1(\ker \widehat{\psi}_2) / \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$ (in the notation of diagram ). Therefore it suffices to show that $ \widetilde{\varphi}_1(\ker \widetilde{\psi}_2) / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ is isomorphic to $ \widehat{\varphi}_1(\ker \widehat{\psi}_2) / \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$. For this, we again consider the natural map $$\begin{aligned}
f \colon\widetilde{\varphi}_1(\ker \widetilde{\psi}_2) / \widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) &\longrightarrow \widehat{\varphi}_1(\ker \widehat{\psi}_2) / \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr}) \\
(\widetilde{h}_1 \widetilde{H_1}',\dots,\widetilde{h}_n \widetilde{H_n}')\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) &\longmapsto (\widetilde{h}_1 \dots \widetilde{h}_n [\widetilde{H},\widetilde{H}])\widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr}) \end{aligned}$$
In the proof of Lemma \[lem:surject\_int\_Galois\]\[lem:surject\_int\_Galois2\] it was shown that $f(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2))\subset \widehat{\varphi}_1(\ker \widehat{\psi}_2)$. Additionally, recalling that $\widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr}) =\Phi^{\widetilde{G}}(\widetilde{H})/[\widetilde{H},\widetilde{H}]$ by [@DP Theorem 2], we have $f(\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}))= \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$ by Lemma \[lem:incl\_unr\] and the proof of Theorem \[thm:demarch\_wei\_thm\]. This shows that $f$ is well defined and injective.
Finally, let us check that $f$ is surjective. Fix a place $v$ of $K$ and a double coset decomposition $\widetilde{G}=\bigcup\limits_{j=1}^{r}\widetilde{H} \widetilde{y}_j \widetilde{G}_v$ and let $\alpha \in \widehat{\varphi}_1(\ker \widehat{\psi}_2^{v})$. We can write $\alpha = \widehat{\varphi}_1(\bigoplus\limits_{j=1}^{r} \widetilde{h}_j)=\prod\limits_{j=1}^{r} \widetilde{h}_j$ for some $\widetilde{h}_j \in \widetilde{H} \cap \widetilde{y}_j \widetilde{S}_v \widetilde{y}_j^{-1}$ such that $\beta:=\widehat{\psi}_2(\bigoplus\limits_{j=1}^{r} \widetilde{h}_j)=\prod\limits_{j=1}^{r} \widetilde{y}_j^{-1} \widetilde{h}_j \widetilde{y}_j$ is in $[\widetilde{S}_v,\widetilde{S}_v]$. Note that as $G$ is abelian, we have $[\widetilde{G},\widetilde{G}] \subset \widetilde{M}$ and therefore $[\widetilde{S}_v,\widetilde{S}_v] \subset \widetilde{M} \subset \widetilde{H}_i$ for every $1 \leq i \leq n$. In particular, we have $\beta \in \widetilde{H}_1 \cap \widetilde{S}_v$ and from this one readily checks that the $n$-tuple $(\beta,1,\dots,1) $ is in $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{v})$. Since $\widetilde{H} \trianglelefteq \widetilde{G}$, we have $\Phi^{\widetilde{G}}(\widetilde{H})=[\widetilde{H},\widetilde{G}]$ and thus $f(\beta,1,\dots,1)=\beta=\prod\limits_{j} \widetilde{y}_j^{-1} \widetilde{h}_j \widetilde{y}_j \equiv \prod\limits_{j} \widetilde{h}_j = \alpha \pmod{\Phi^{\widetilde{G}}(\widetilde{H})}$. As $\Phi^{\widetilde{G}}(\widetilde{H})/[\widetilde{H},\widetilde{H}]= \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$, we obtain $\alpha = f(\beta,1,\dots,1)$ inside $\widehat{\varphi}_1(\ker \widehat{\psi}_2) / \widehat{\varphi}_1(\ker \widehat{\psi}_2^{nr})$.
Note that the conditions of Theorem \[thm:demarch\_wei\_thm\] are always satisfied if $n=2$, so that Theorem \[thm:pollio\] generalizes the main theorem of [@pollio].
Weak approximation for cyclic extensions of prime degree {#sec:eva}
--------------------------------------------------------
In this subsection we extend the result in [@eva Theorem 8.3] to include the weak approximation property for the multinorm one torus of $n$ cyclic extensions of prime degree $p$.
\[thm:eva\] Let $L_1,\dots,L_n$ be non-isomorphic cyclic extensions of $K$ with prime degree $p$. Then, we have $$\operatorname{H}^1(K,\operatorname{Pic}\overline{X})=\begin{cases}
({\mathbb{Z}}/p)^{n-2} \textrm{, if $[L_1\dots L_n:K]=p^2$;}\\
0\textrm{, otherwise.}
\end{cases}$$
The case $n=1$ was proved in [@coll2 Proposition 9.1] and for $n=2$ the result follows from Theorem \[thm:demarch\_wei\_thm\], so assume $n \geq 3$.
Suppose first that $[L_1 \dots L_n : K] > p^2$. Reordering the fields $L_3,\dots,L_n$ if necessary, we can (and do) assume that each one of the fields $L_1,\dots,L_{s-1}$ is contained in $L_1 L_2$ (for some $3 \leq s\leq n$), while none of $L_s,\dots,L_n$ is contained in $L_1 L_2$. We prove two auxiliary claims:
**Claim 1:** $\widetilde{H_i} \subset (\widetilde{H_1} \cap \widetilde{H_i} ).\widetilde{H_s}$ for any $i=1,\dots,s-1$.
**Proof:** Observe that $L_1L_i \cap L_s = K$ as otherwise we would have $L_s \subset L_1 L_i \subset L_1 L_2$, contradicting the assumption on $s$. Therefore $L_i \supset K= L_1 L_i \cap L_s$ and passing to subgroups this implies that ${H_i} \subset ({H_1} \cap {H_i}).{H_s}$, from which the claim follows.
**Claim 2:** $\widetilde{H_i} \subset (\widetilde{H_1} \cap \widetilde{H_i}).\widetilde{H_2}$ for any $i=s,\dots,n$.
**Proof:** Observe that $L_2 \not\subset L_1 L_i$ as otherwise we would have $L_i \subset L_1 L_i = L_1 L_2$, contradicting the assumption on $L_i$. Therefore $L_i \supset K= L_1 L_i \cap L_2$ and passing to subgroups this implies that ${H_i} \subset ({H_1} \cap {H_i}).{H_2}$, from which the claim follows.
Let us now prove that $\operatorname{H}^1(K,\operatorname{Pic}\overline{X})=0$. Since $\bigcap\limits_{i} L_i = K$, by Lemma \[lem:surject\_int\_Galois\]\[lem:surject\_int\_Galois1\] it suffices to show that $$\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr}) \supseteq \{ (h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}') | h_1\dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H}) \}.$$
Let $\alpha=(h_1 \widetilde{H_1}' , \dots , h_n \widetilde{H_n}')$ be such that $h_1 \dots h_n \in \Phi^{\widetilde{G}}(\widetilde{H})$. By Claim 1 above, for $i=3,\dots,s-1$ we can write $h_i = h_{1,i}h_{s,i}$, where $h_{1,i} \in \widetilde{H_1} \cap \widetilde{H_i}$ and $h_{s,i} \in \widetilde{H_s} \cap \widetilde{H_i}$. Using this decomposition, we can apply Lemma \[lem:simpl\_inters\] as done in the proof of Theorem \[thm:demarch\_wei\_thm\] in order to simplify $\alpha$ modulo $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ and assume it has the form $(h_1',h_2,1,\dots,1,h_s',h_{s+1}\dots,h_n)$ for some $h_1' \in \widetilde{H_1},h_s' \in \widetilde{H_s}$. Using Claim 2 and Lemma \[lem:simpl\_inters\] in the same way, we further reduce $\alpha$ modulo $\widetilde{\varphi}_1(\ker \widetilde{\psi}_2^{nr})$ to a vector of the form $(h_1'',h_2',1,\dots,1)$ for some $h_1'' \in \widetilde{H_1},h_2' \in \widetilde{H_2}$ such that $h_1'' h_2' \in \Phi^{\widetilde{G}}(\widetilde{H})$. Finally, since $L_1 \cap L_2 = K$, we have $\widetilde{H}= \widetilde{H_1} \widetilde{H_2}$ and thus $\Phi^{\widetilde{G}}(\widetilde{H}) \subset \Phi^{\widetilde{G}}(\widetilde{H_1})\Phi^{\widetilde{G}}(\widetilde{H_2})$. The result follows by an argument similar to the one given at the end of the proof of Theorem \[thm:demarch\_wei\_thm\].
Now assume that $[L_1\dots L_n:K]=p^2$ (note that this is only possible if $ n \leq p+1$ as a bicyclic field has $p+1$ subfields of degree $p$) and therefore $G=C_p \times C_p$ is abelian. By Proposition \[prop:sq\_free\_mid\_gp\] it suffices to prove that $\ker {\psi}_1 / {\varphi}_1(\ker {\psi}_2^{nr}) \cong ({\mathbb{Z}}/p)^{n-2}$. We first show that ${\varphi}_1(\ker {\psi}_2^{nr})=1$. Let $\alpha \in \ker {\psi}_2^{v}$ for some unramified place $v$ of $N/K$. Write $G_v=\langle g \rangle$ and $\alpha = \bigoplus\limits_{i=1}^{n} \bigoplus\limits_{k=1}^{r_{v,i}} {h}_{i,k} $ for some $g \in G$ and ${h}_{i,k} \in {H_i} \cap {x}_{i,k} \langle g \rangle {x}_{i,k}^{-1}={H_i} \cap \langle g \rangle$. If $g \not\in H_i $ for all $i=1,\dots,n$, then $\alpha$ is the trivial vector and $\varphi_1(\alpha)=(1,\dots,1)$. Otherwise, if $g \in H_{i_0}\cong C_p$ for some index $i_0$, then $g \not\in H_i$ for all $i \neq i_0$ and thus $h_{i,k}=1$ for $i \neq i_0$. In this way, it follows that $1=\psi_2(\alpha)=\prod\limits_{i=1}^{n} \prod\limits_{k=1}^{r_{v,i}} {x}_{i,k}^{-1} {h}_{i,k}{x}_{i,k}= \prod\limits_{k=1}^{r_{v,i_0}} {h}_{i_0,k}$. Therefore, if $\varphi_1(\alpha)=(h_1,\dots,h_n)$, we have $h_i = 1$ if $i \neq i_0$ and $h_{i_0}=\prod\limits_{k=1}^{r_{v,i_0}} {h}_{i_0,k} = 1$. In conclusion, $\varphi_1(\alpha)=(1,\dots,1)$.
On the other hand, we have $\ker \psi_1 = \{(h_1,\dots,h_n) | h_i \in H_i, \prod\limits_{i=1}^{n} h_i = 1\}$. This group is the kernel of the surjective group homomorphism $$\begin{aligned}
f \colon H_1 \times \dots \times H_n &\longrightarrow G \\
(h_1 ,\dots,h_n ) &\longmapsto h_1 \dots h_n \end{aligned}$$
and thus $\ker \psi_1 = \ker f \cong ({\mathbb{Z}}/p)^{n-2}$, as desired.
\[cor:eva\] Let $L=(L_1,\dots,L_n)$ be an $n$-tuple of non-isomorphic cyclic extensions of $K$ with prime degree $p$.
1. \[cor410\_case1\] If $[L_1\dots L_n:K]=p^2$, then weak approximation for the multinorm one torus $T$ holds if and only if the multinorm principle for $L$ fails.
2. Otherwise, both the multinorm principle for $L$ and weak approximation for $T$ hold.
Follows from Voskresenskiĭ’s exact sequence , Theorem \[thm:eva\] and [@eva Theorem 8.3].
\[rem:eva\] In [@eva Proposition 8.5] it is shown that, in the case above, the multinorm principle for $L$ fails if and only if all decomposition groups of the bicyclic extension $L_1 \dots L_n$ are cyclic. We thus have a simple criterion to test the validity of weak approximation for the associated multinorm one torus.
[99]{}
E. Bayer-Fluckiger, T.-Y. Lee, R. Parimala, Hasse principles for multinorm equations. *Advances Math.* **356** (2019).
J.-L. Colliot-Thélène, J.-J. Sansuc, Principal homogeneous spaces under flasque tori: Applications. *J. Algebra* **106** (1987), 148–205.
C. Demarche, D. Wei, Hasse principle and weak approximation for multinorm equations. *Israel J. Math.* **202** (2014), no.1, 275-293.
Y. A. Drakokhrust, On the complete obstruction to the Hasse principle. *Amer. Math. Soc. Transl.*(2) **143** (1989), 29-34.
, The Hasse norm principle for algebraic number fields. *Math. USSR-Izv.* **29** (1987), 299-322.
GAP Group, GAP – Groups, Algorithms, and Programming, Version 4.8.10 (2018). Available at <https://www.gap-system.org>.
W. Hürlimann, On algebraic tori of norm type. *Comment. Math. Helv.* **59** (1984), 539-549.
I. M. Isaacs, *Finite group theory*. Graduate Studies in Math. **92** AMS Providence (2008).
G. Karpilovsky, *The Schur Multiplier*. Clarendon Press, Oxford, 1987.
T.-Y. Lee, The Tate-Shafarevich groups of multinorm-one tori. Preprint (2019), arXiv:1912.09823.
A. Macedo, GAP code (2019). Available at <https://sites.google.com/view/andre-macedo/code>.
A. Macedo, R. Newton, Explicit methods for the Hasse norm principle and applications to $A_n$ and $S_n$ extensions. Preprint (2019), arXiv:1906.03730.
H. Opolka, Zur Aufl[ö]{}sung zahlentheoretischer Knoten. *Math. Z.* **173** (1980) 95–103.
V. P. Platonov, A. S. Rapinchuk, *Algebraic groups and number theory*. Pure and Applied Mathematics **139**, Academic Press, Inc., Boston, MA, 1994.
T. Pollio, On the multinorm principle for finite abelian extensions. *Pure Appl. Math. Q.* **10** (2014), 547-566.
T. Pollio, A. S. Rapinchuk, The multinorm principle for linearly disjoint Galois extensions. *J. Number Theory* **133** (2013), 802-821.
G. Prasad, A. S. Rapinchuk, Local-global principles for embedding of fields with involution into simple algebras with involution. *Comment. Math. Helv.* **85** (2010), 583-645.
J. T. Tate, Global class field theory. pp. 162–203 in: J.W.S. Cassels, A. Fröhlich (eds), *Algebraic number theory*, Second Edition, London Mathematical Society, 2010.
V. E. Voskresenskiĭ, *Algebraic Groups and Their Birational Invariants*. Amer. Math. Soc., Providence, RI, 1998.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'It is shown that there exist nine different ways to describe the flavor mixing, in terms of three rotation angles and one $CP$-violating phase, within the standard electroweak theory of six quarks. For the assignment of the complex phase there essentially exists a continuum of possibilities, if one allows the phase to appear in more than four elements of the mixing matrix. If the phase is restricted to four elements, the phase assignment is uniquely defined. If one imposes the constraint that the phase disappears in a natural way in the chiral limit in which the masses of the $u$ and $d$ quarks are turned off, only three of the nine parametrizations are acceptable. In particular the “standard” parametrization advocated by the Particle Data Group is not permitted. One parametrization, in which the $CP$-violating phase is restricted to the light quark sector, stands up as the most favorable description of the flavor mixing.'
---
= 16.5cm = 24.5cm = -16.5mm = -1.8mm
plus 1pt minus 1pt
[**CERN-TH/97-201**]{}\
[**DPNU-97-39**]{}
[**On the Parametrization of Flavor Mixing\
in the Standard Model**]{}
[**Harald Fritzsch**]{} [^1]\
[*Theory Division, CERN, CH–1211 Geneva 23, Switzerland; and*]{}\
[*Sektion Physik, Universit$\ddot{a}$t M$\ddot{u}$nchen, 80333 M$\ddot{u}$nchen, Germany*]{}
[**Zhi-zhong Xing**]{} [^2]\
[*Department of Physics, Nagoya University, Nagoya 464-01, Japan*]{}
PACS number(s): 12.15.Hh, 11.30.Er, 12.15.Ff
In the standard electroweak theory, the phenomenon of flavor mixing of the quarks is described by a $3\times 3$ unitary matrix, the Cabibbo-Kobayashi-Maskawa (CKM) matrix [@Cabibbo63; @KM73]. This matrix can be expressed in terms of four parameters, which are usually taken as three rotation angles and one phase. A number of different parametrizations have been proposed in the literature [@KM73]–[@FX97]. Of course, adopting a particular parametrization of flavor mixing is arbitrary and not directly a physical issue. Nevertheless it is quite likely that the actual values of flavor mixing parameters (including the strength of $CP$ violation), once they are known with high precision, will give interesting information about the physics beyond the standard model. Probably at this point it will turn out that a particular description of the CKM matrix is more useful and transparent than the others. For this reason, we find it useful to analyze all possible parametrizations and to point out their respective advantages and disadvantages. This is the main purpose of this short note.
In the standard model the quark flavor mixing arises once the up- and down-type mass matrices are diagonalized. The generation of quark masses is intimately related to the phenomenon of flavor mixing. In particular, the flavor mixing parameters do depend on the elements of quark mass matrices. A particular structure of the underlying mass matrices calls for a particular choice of the parametrization of the flavor mixing matrix. For example, in Ref. [@Fritzsch79] it was noticed that a rather special form of the flavor mixing matrix results, if one starts from Hermitian mass matrices in which the (1,3) and (3,1) elements vanish. This has been subsequently observed again in a number of papers [@Hall]. Recently we have studied the exact form of such a description from a general point of view and pointed out some advantages of this type of representation in the discussion of flavor mixing and $CP$-violating phenomena [@FX97]. One of the aims of this work is also to view this parametrization in the context with other ways of describing the flavor mixing.
In the standard model the weak charged currents are given by $$\overline{(u, ~ c, ~ t)}^{~}_L \left ( \matrix{
V_{ud} & V_{us} & V_{ub} \cr
V_{cd} & V_{cs} & V_{cb} \cr
V_{td} & V_{ts} & V_{tb} \cr} \right )
\left ( \matrix{
d \cr s \cr b \cr} \right )_L \; ,
$$ where $u$, $c$, ..., $b$ are the quark mass eigenstates, $L$ denotes the left-handed fields, and $V_{ij}$ are elements of the CKM matrix $V$. In general $V_{ij}$ are complex numbers, but their absolute values are measurable quantities. For example, $|V_{cb}|$ primarily determines the lifetime of $B$ mesons. The phases of $V_{ij}$, however, are not physical, like the phases of quark fields. A phase transformation of the $u$ quark ($u \rightarrow u ~ e^{{\rm
i}\alpha}$), for example, leaves the quark mass term invariant but changes the elements in the first row of $V$ (i.e., $V_{uj} \rightarrow
V_{uj} ~ e^{-{\rm i}\alpha}$). Only a common phase transformation of all quark fields leaves all elements of $V$ invariant, thus there is a five-fold freedom to adjust the phases of $V_{ij}$.
In general the unitary matrix $V$ depends on nine parameters. Note that in the absence of complex phases $V$ would consist of only three independent parameters, corresponding to three (Euler) rotation angles. Hence one can describe the complex matrix $V$ by three angles and six phases. Due to the freedom in redefining the quark field phases, five of the six phases in $V$ can be absorbed; and we arrive at the well-known result that the CKM matrix $V$ can be parametrized in terms of three rotation angles and one $CP$-violating phase. The question about how many different ways to describe $V$ may exist was raised some time ago [@Jarlskog89]. Below we shall reconsider this problem and give a complete analysis.
If the flavor mixing matrix $V$ is first assumed to be a real orthogonal matrix, it can in general be written as a product of three matrices $R_{12}$, $R_{23}$ and $R_{31}$, which describe simple rotations in the (1,2), (2,3) and (3,1) planes: $$\begin{aligned}
R_{12}(\theta) & = & \left ( \matrix{
c^{~}_{\theta} & s^{~}_{\theta} & 0 \cr
- s^{~}_{\theta} & c^{~}_{\theta} & 0 \cr
0 & 0 & 1 \cr} \right ) \; , \nonumber \\ \nonumber \\
R_{23}(\sigma) & = & \left ( \matrix{
1 & 0 & 0 \cr
0 & c_{\sigma} & s_{\sigma} \cr
0 & - s_{\sigma} & c_{\sigma} \cr} \right ) \; , \nonumber \\
\nonumber \\
R_{31}(\tau) & = & \left ( \matrix{
c_{\tau} & 0 & s_{\tau} \cr
0 & 1 & 0 \cr
- s_{\tau} & 0 & c_{\tau} \cr} \right ) \; ,
$$ where $s^{~}_{\theta} \equiv \sin \theta$, $c^{~}_{\theta} \equiv \cos
\theta$, etc. Clearly any two rotation matrices do not commute with each other. There exist twelve different ways to arrange products of these matrices such that the most general orthogonal matrix $R$ can be obtained [@Jarlskog89]. Note that the matrix $R^{-1}_{ij} (\omega) $ plays an equivalent role as $R_{ij} (\omega) $ in constructing $R$, because of $R^{-1}_{ij}(\omega) =
R_{ij}(-\omega)$. Note also that $R_{ij} (\omega) R_{ij}
(\omega^{\prime}) = R_{ij} (\omega + \omega^{\prime})$ holds, thus the product $R_{ij}(\omega) R_{ij}(\omega^{\prime})
R_{kl}(\omega^{\prime\prime})$ or $R_{kl}(\omega^{\prime\prime})
R_{ij}(\omega) R_{ij}(\omega^{\prime})$ cannot cover the whole space of a $3\times 3$ orthogonal matrix and should be excluded. Explicitly the twelve different forms of $R$ read as $$\begin{aligned}
(1) & & R \; =\; R_{12}(\theta) ~ R_{23}(\sigma) ~ R_{12}(\theta^{\prime})
\; , \nonumber \\
(2) & & R \; =\; R_{12}(\theta) ~ R_{31}(\tau) ~ R_{12}(\theta^{\prime})
\; , \nonumber \\
(3) & & R \; =\; R_{23}(\sigma) ~ R_{12}(\theta) ~ R_{23}(\sigma^{\prime})
\; , \nonumber \\
(4) & & R \; =\; R_{23}(\sigma) ~ R_{31}(\tau) ~ R_{23}(\sigma^{\prime})
\; , \nonumber \\
(5) & & R \; =\; R_{31}(\tau) ~ R_{12}(\theta) ~ R_{31}(\tau^{\prime})
\; , \nonumber \\
(6) & & R \; =\; R_{31}(\tau) ~ R_{23}(\sigma) ~ R_{31}(\tau^{\prime})
\; , \nonumber\end{aligned}$$ in which a rotation in the $(i,j)$ plane occurs twice; and $$\begin{aligned}
(7) & & R \; =\; R_{12}(\theta) ~ R_{23}(\sigma) ~ R_{31}(\tau)
\; , \nonumber \\
(8) & & R \; =\; R_{12}(\theta) ~ R_{31}(\tau) ~ R_{23}(\sigma)
\; , \nonumber \\
(9) & & R \; =\; R_{23}(\sigma) ~ R_{12}(\theta) ~ R_{31}(\tau)
\; , \nonumber \\
(10) & & R \; =\; R_{23}(\sigma) ~ R_{31}(\tau) ~ R_{12}(\theta)
\; , \nonumber \\
(11) & & R \; =\; R_{31}(\tau) ~ R_{12}(\theta) ~ R_{23}(\sigma)
\; , \nonumber \\
(12) & & R \; =\; R_{31}(\tau) ~ R_{23}(\sigma) ~ R_{12}(\theta)
\; , \nonumber \end{aligned}$$ where all three $R_{ij}$ are present.
Although all the above twelve combinations represent the most general orthogonal matrices, only nine of them are structurally different. The reason is that the products $R_{ij} R_{kl} R_{ij}$ and $R_{ij} R_{mn} R_{ij}$ (with $ij\neq kl\neq mn$) are correlated with each other, leading essentially to the same form for $R$. Indeed it is straightforward to see the correlation between patterns (1), (3), (5) and (2), (4), (6), respectively, as follows: $$\begin{aligned}
R_{12}(\theta) ~ R_{31}(\tau) ~ R_{12}(\theta^{\prime})
& = & R_{12}(\theta + \pi/2) ~ R_{23}(\sigma = \tau) ~
R_{12}(\theta^{\prime} - \pi/2) \; , \nonumber \\
R_{23}(\sigma) ~ R_{31}(\tau) ~ R_{23}(\sigma^{\prime})
& = & R_{23}(\sigma -\pi/2) ~ R_{12}(\theta = \tau) ~
R_{23}(\sigma^{\prime} + \pi/2) \; , \nonumber \\
R_{31}(\tau) ~ R_{23}(\sigma) ~ R_{31}(\tau^{\prime})
& = & R_{31}(\tau + \pi/2) ~ R_{12}(\theta = \sigma) ~
R_{31}(\tau^{\prime} - \pi/2) \; .
$$ Thus the orthogonal matrices (2), (4) and (6) need not be treated as independent choices. We then draw the conclusion that there exist [*nine*]{} different forms for the orthogonal matrix $R$, i.e., patterns (1), (3) and (5) as well as (7) – (12).
We proceed to include the $CP$-violating phase, denoted by $\varphi$, in the above rotation matrices. The resultant matrices should be unitary such that a unitary flavor mixing matrix can be finally produced. There are several different ways for $\varphi$ to enter $R_{12}$, e.g., $$R_{12} (\theta, \varphi) \; =\; \left ( \matrix{
c^{~}_{\theta} & s^{~}_{\theta} ~ e^{+{\rm i} \varphi} & 0 \cr
- s^{~}_{\theta} ~ e^{-{\rm i} \varphi} & c^{~}_{\theta} & 0 \cr
0 & 0 & 1 \cr} \right ) \; ,
\eqno(4{\rm a})
$$ or $$R_{12} (\theta, \varphi) \; =\; \left ( \matrix{
c^{~}_{\theta} & s^{~}_{\theta} & 0 \cr
- s^{~}_{\theta} & c^{~}_{\theta} & 0 \cr
0 & 0 & e^{-{\rm i} \varphi} \cr} \right ) \; ,
\eqno(4{\rm b})
$$ or $$R_{12} (\theta, \varphi) \; =\; \left ( \matrix{
c^{~}_{\theta} ~ e^{+{\rm i} \varphi} & s^{~}_{\theta} & 0 \cr
- s^{~}_{\theta} & c^{~}_{\theta} ~ e^{-{\rm i} \varphi} & 0 \cr
0 & 0 & 1 \cr} \right ) \; .
\eqno(4{\rm c})
$$ Similarly one may introduce a phase parameter into $R_{23}$ or $R_{31}$. Then the CKM matrix $V$ can be constructed, as a product of three rotation matrices, by use of one complex $R_{ij}$ and two real ones. Note that the location of the $CP$-violating phase in $V$ can be arranged by redefining the quark field phases, thus it does not play an essential role in classifying different parametrizations. We find that it is always possible to locate the phase parameter $\varphi$ in a $2\times 2$ submatrix of $V$, in which each element is a sum of two terms with the relative phase $\varphi$. The remaining five elements of $V$ are real in such a phase assignment. Accordingly we arrive at nine distinctive parametrizations of the CKM matrix $V$, as listed in Table 1, where the complex rotation matrices $R_{12}(\theta, \varphi)$, $R_{23}(\sigma, \varphi)$ and $R_{31}(\tau, \varphi)$ are obtained directly from the real ones in Eq. (2) with the replacement $1
\rightarrow e^{-{\rm i}\varphi}$.
Some instructive relations of each parametrization, as well as the rephasing-invariant measure of $CP$ violation [@Jarlskog85] defined by $\cal J$ through $${\rm Im} \left ( V_{il} V_{jm} V^*_{im} V^*_{jl} \right ) \; =\;
{\cal J} \sum_{k,n=1}^{3} \left ( \epsilon^{~}_{ijk} \epsilon^{~}_{lmn} \right ) \; ,
$$ have also been given in Table 1. One can see that [*P2*]{} and [*P3*]{} correspond to the Kobayashi-Maskawa [@KM73] and Maiani [@Others] representations, although different notations for the $CP$-violating phase and three mixing angles are adopted here. The latter is indeed equivalent to the “standard” parametrization advocated by the Particle Data Group [@Others; @PDG96]. This can be seen clearly if one makes three transformations of quark field phases: $c \rightarrow c ~ e^{-{\rm i} \varphi}$, $t
\rightarrow t ~ e^{-{\rm i} \varphi}$, and $b \rightarrow
b ~ e^{-{\rm i} \varphi}$. In addition, [*P1*]{} is just the one proposed by the present authors in Ref. [@FX97].
[From]{} a mathematical point of view, all nine different parametrizations are equivalent. However this is not the case if we apply our considerations to the quarks and their mass spectrum. It is well known that both the observed quark mass spectrum and the observed values of the flavor mixing parameters exhibit a striking hierarchical structure. The latter can be understood in a natural way as the consequence of a specific pattern of chiral symmetries whose breaking causes the towers of different masses to appear step by step [@Fritzsch87a; @Fritzsch87b; @Hall93]. Such a chiral evolution of the mass matrices leads, as argued in Ref. [@Fritzsch87b], to a specific way to introduce and describe the flavor mixing. In the limit $m_u = m_d =0$, which is close to the real world, since $m_u/m_t \ll 1$ and $m_d/m_b \ll 1$, the flavor mixing is merely a rotation between the $t$–$c$ and $b$–$s$ systems, described by one rotation angle. No complex phase is present; i.e., $CP$ violation is absent. This rotation angle is expected to change very little, once $m_u$ and $m_d$ are introduced as tiny perturbations. A sensible parametrization should make use of this feature. This implies that the rotation matrix $R_{23}$ appears exactly once in the description of the CKM matrix $V$, eliminating [*P2*]{} (in which $R_{23}$ appears twice) and [*P5*]{} (where $R_{23}$ is absent). This leaves us with seven parametrizations of the flavor mixing matrix.
The list can be reduced further by considering the location of the phase $\varphi$. In the limit $m_u = m_d =0$, the phase must disappear in the weak transition elements $V_{tb}$, $V_{ts}$, $V_{cb}$ and $V_{cs}$. In [*P7*]{} and [*P8*]{}, however, $\varphi$ appears particularly in $V_{tb}$. Thus these two parametrizations should be eliminated, leaving us with five parametrizations (i.e., [*P1*]{}, [*P3*]{}, [*P4*]{}, [*P6*]{} and [*P9*]{}). In the same limit, the phase $\varphi$ appears in the $V_{ts}$ element of [*P3*]{} and the $V_{cb}$ element of [*P4*]{}. Hence these two parametrizations should also be eliminated. Then we are left with three parametrizations, [*P1*]{}, [*P6*]{} and [*P9*]{}. As expected, these are the parametrizations containing the complex rotation matrix $R_{23}(\sigma, \varphi)$. We stress that the “standard” parametrization [@PDG96] (equivalent to [*P3*]{}) does not obey the above constraints and should be dismissed.
Among the remaining three parametrizations, [*P1*]{} is singled out by the fact that the $CP$-violating phase $\varphi$ appears only in the $2\times 2$ submatrix of $V$ describing the weak transitions among the light quarks. This is precisely the system where the phase $\varphi$ should appear, not in any of the weak transition elements involving the heavy quarks $t$ and $b$.
In the parametrization [*P6*]{} or [*P9*]{}, the complex phase $\varphi$ appears in $V_{cb}$ or $V_{ts}$, but this phase factor is multiplied by a product of $\sin\theta$ and $\sin\tau$, i.e., it is of second order of the weak mixing angles. Hence the imaginary parts of these elements are not exactly vanishing, but very small in magnitude.
In our view the best possibility to describe the flavor mixing in the standard model is to adopt the parametrization [*P1*]{}. As discussed in Ref. [@FX97], this parametrization has a number of significant advantages in addition to that mentioned above. Especially it is well suited for specific models of quark mass matrices (see, e.g., Refs. [@Fritzsch79; @Hall]).
We conclude: there are nine different ways to describe a real $3\times
3$ flavor mixing matrix in terms of three rotation angles. Introducing a complex phase $\varphi$ does not increase the number of distinct parametrizations, except for the fact that there is a continuum of possibilities for assigning the phase factors. Imposing natural constraints in view of the observed mass hierarchy (i.e., in the limit $m_u = m_d =0$ phases should be absent in the (2,2), (2,3), (3,2) and (3,3) elements of the mixing matrix), we can eliminate six parametrizations, including the original Kobayashi-Maskawa parametrization [@KM73] and the “standard” parametrization proposed in Refs. [@Others; @PDG96]. We propose to use the parametrization [*P1*]{} for the further study of flavor mixing and $CP$-violating phenomena.
[*Acknowledgments:*]{} The work of H.F. was supported in part by GIF contract I-0304-120.07/93 and EEC contract CHRX-CT94-0579 (DG 12 COMA). Z.Z.X. is grateful to A.I. Sanda for his warm hospitality and to the Japan Society for the Promotion of Science for its financial support.
[*Note added:*]{} After completion of this work we received a preprint of A. Rasin [@Rasin97], in which part of the conclusions drawn here was also reached.
[99]{}
N. Cabibbo, Phys. Rev. Lett. [**10**]{}, 531 (1963).
M. Kobayashi and T. Maskawa, Prog. Theor. Phys. [**49**]{}, 652 (1973).
See, e.g., L. Maiani, in: Proc. 1977 Int. Symp. on Lepton and Photon Interactions at High Energies (DESY, Hamburg, 1977), p. 867; L.L. Chau and W.Y. Keung, Phys. Rev. Lett. [**53**]{}, 1802 (1984); H. Fritzsch, Phys. Rev. D [**32**]{}, 3058 (1985); H. Harari and M. Leurer, Phys. Lett. B [**181**]{}, 123 (1986); H. Fritzsch and J. Plankl, Phys. Rev. [**D 35**]{}, 1732 (1987).
F.J. Gilman, Kleinknecht, and B. Renk, Phys. Rev. D [**54**]{}, 94 (1996).
H. Fritzsch and Z.Z. Xing, Report No. YITP-97-31 or hep-ph/9707215.
H. Fritzsch, Nucl. Phys. B [**155**]{}, 189 (1979).
S. Dimopoulos, L.J. Hall, and S. Raby, Phys. Rev. D [**45**]{}, 4192 (1992); L.J. Hall and A. Rasin, Phys. Lett. B [**315**]{}, 164 (1993); R. Barbieri, L.J. Hall, and A. Romanino, Phys. Lett. B [**401**]{}, 47 (1997).
C. Jarlskog, in [*CP Violation*]{}, edited by C. Jarlskog (World Scientific, 1989), p. 3.
C. Jarlskog, Phys. Rev. Lett. [**55**]{}, 1039 (1985); D.D. Wu, Phys. Rev. D [**33**]{}, 860 (1986).
H. Fritzsch, Phys. Lett. B [**184**]{}, 391 (1987).
H. Fritzsch, Phys. Lett. B [**189**]{}, 191 (1987).
L.J. Hall and S. Weinberg, Phys. Rev. D [**48**]{}, 979 (1993).
A. Rasin, Report No. hep-ph/9708216.
[ccc]{}\
Parametrization & & Useful relations\
\
[*P1:*]{} $V \; = \; R_{12}(\theta) ~ R_{23}(\sigma, \varphi)
~ R^{-1}_{12}(\theta^{\prime})$ && ${\cal J} = s^{~}_{\theta} c^{~}_{\theta} s^{~}_{\theta^{\prime}}
c^{~}_{\theta^{\prime}} s^2_{\sigma} c_{\sigma} \sin\varphi$\
$\left ( \matrix{
s^{~}_{\theta} s^{~}_{\theta^{\prime}} c_{\sigma} + c^{~}_{\theta} c^{~}_{\theta^{\prime}} e^{-{\rm i}\varphi} &
s^{~}_{\theta} c^{~}_{\theta^{\prime}} c_{\sigma} - c^{~}_{\theta}
s^{~}_{\theta^{\prime}} e^{-{\rm i}\varphi} & s^{~}_{\theta} s_{\sigma} \cr
c^{~}_{\theta} s^{~}_{\theta^{\prime}} c_{\sigma} - s^{~}_{\theta} c^{~}_{\theta^{\prime}} e^{-{\rm i}\varphi} &
c^{~}_{\theta} c^{~}_{\theta^{\prime}} c_{\sigma} + s^{~}_{\theta}
s^{~}_{\theta^{\prime}} e^{-{\rm i}\varphi} & c^{~}_{\theta} s_{\sigma} \cr
- s^{~}_{\theta^{\prime}} s_{\sigma} & - c^{~}_{\theta^{\prime}}
s_{\sigma} & c_{\sigma} \cr} \right )
$ && $\matrix{
\tan\theta = |V_{ub}/V_{cb}| \cr
\tan\theta^{\prime} = |V_{td}/V_{ts}| \cr
\cos\sigma = |V_{tb}| \cr} $\
\
[*P2:*]{} $V \; = \; R_{23}(\sigma) ~ R_{12}(\theta, \varphi)
~ R^{-1}_{23}(\sigma^{\prime})$ && ${\cal J} = s^2_{\theta} c^{~}_{\theta} s_{\sigma} c_{\sigma} s_{\sigma^{\prime}} c_{\sigma^{\prime}} \sin\varphi$\
$\left ( \matrix{
c^{~}_{\theta} & s^{~}_{\theta} c_{\sigma^{\prime}} & -s^{~}_{\theta} s_{\sigma^{\prime}} \cr
-s^{~}_{\theta} c_{\sigma} & c^{~}_{\theta} c_{\sigma} c_{\sigma^{\prime}} + s_{\sigma} s_{\sigma^{\prime}} e^{-{\rm i}\varphi}
& -c^{~}_{\theta} c_{\sigma} s_{\sigma^{\prime}} + s_{\sigma} c_{\sigma^{\prime}} e^{-{\rm i}\varphi} \cr
s^{~}_{\theta} s_{\sigma} & -c^{~}_{\theta} s_{\sigma} c_{\sigma^{\prime}} + c_{\sigma} s_{\sigma^{\prime}} e^{-{\rm i}\varphi}
& c^{~}_{\theta} s_{\sigma} s_{\sigma^{\prime}} + c_{\sigma} c_{\sigma^{\prime}} e^{-{\rm i}\varphi} \cr} \right )
$ && $\matrix{
\cos\theta = |V_{ud}| \cr
\tan\sigma = |V_{td}/V_{cd}| \cr
\tan\sigma^{\prime} = |V_{ub}/V_{us}| \cr} $\
\
[*P3:*]{} $V \; = \; R_{23}(\sigma) ~ R_{31}(\tau, \varphi)
~ R_{12}(\theta)$ && ${\cal J} = s^{~}_{\theta} c^{~}_{\theta} s_{\sigma} c_{\sigma} s_{\tau} c^2_{\tau} \sin\varphi$\
$\left ( \matrix{
c^{~}_{\theta} c_{\tau} & s^{~}_{\theta} c_{\tau} & s_{\tau} \cr
-c^{~}_{\theta} s_{\sigma} s_{\tau} - s^{~}_{\theta} c_{\sigma} e^{-{\rm i}\varphi}
& -s^{~}_{\theta} s_{\sigma} s_{\tau} + c^{~}_{\theta} c_{\sigma} e^{-{\rm i}\varphi} & s_{\sigma} c_{\tau} \cr
-c^{~}_{\theta} c_{\sigma} s_{\tau} + s^{~}_{\theta} s_{\sigma} e^{-{\rm i}\varphi}
& -s^{~}_{\theta} c_{\sigma} s_{\tau} - c^{~}_{\theta} s_{\sigma} e^{-{\rm i}\varphi} & c_{\sigma} c_{\tau} \cr} \right )
$ && $\matrix{
\tan\theta = |V_{us}/V_{ud}| \cr
\tan\sigma = |V_{cb}/V_{tb}| \cr
\sin\tau = |V_{ub}| \cr} $\
\
[*P4:*]{} $V \; = \; R_{12}(\theta) ~ R_{31}(\tau, \varphi)
~ R^{-1}_{23}(\sigma)$ && ${\cal J} = s^{~}_{\theta} c^{~}_{\theta} s_{\sigma} c_{\sigma} s_{\tau} c^2_{\tau} \sin\varphi$\
$\left ( \matrix{
c^{~}_{\theta} c_{\tau} & c^{~}_{\theta} s_{\sigma} s_{\tau} + s^{~}_{\theta} c_{\sigma} e^{-{\rm i}\varphi}
& c^{~}_{\theta} c_{\sigma} s_{\tau} - s^{~}_{\theta} s_{\sigma} e^{-{\rm i}\varphi} \cr
-s^{~}_{\theta} c_{\tau} & -s^{~}_{\theta} s_{\sigma} s_{\tau} + c^{~}_{\theta} c_{\sigma} e^{-{\rm i}\varphi}
& -s^{~}_{\theta} c_{\sigma} s_{\tau} - c^{~}_{\theta} s_{\sigma} e^{-{\rm i}\varphi} \cr
-s_{\tau} & s_{\sigma} c_{\tau} & c_{\sigma} c_{\tau} \cr} \right )
$ && $\matrix{
\tan\theta = |V_{cd}/V_{ud}| \cr
\tan\sigma = |V_{ts}/V_{tb}| \cr
\sin\tau = |V_{td}| \cr} $\
\
[*P5:*]{} $V \; = \; R_{31}(\tau) ~ R_{12}(\theta, \varphi)
~ R^{-1}_{31}(\tau^{\prime})$ && ${\cal J} = s^2_{\theta} c^{~}_{\theta} s_{\tau} c_{\tau} s_{\tau^{\prime}} c_{\tau^{\prime}} \sin\varphi$\
$\left ( \matrix{
c^{~}_{\theta} c_{\tau} c_{\tau^{\prime}} + s_{\tau} s_{\tau^{\prime}} e^{-{\rm i}\varphi} & s^{~}_{\theta} c_{\tau}
& -c^{~}_{\theta} c_{\tau} s_{\tau^{\prime}} + s_{\tau} c_{\tau^{\prime}} e^{-{\rm i}\varphi} \cr
-s^{~}_{\theta} c_{\tau^{\prime}} & c^{~}_{\theta} & s^{~}_{\theta} s_{\tau^{\prime}} \cr
-c^{~}_{\theta} s_{\tau} c_{\tau^{\prime}} + c_{\tau} s_{\tau^{\prime}} e^{-{\rm i}\varphi} & -s^{~}_{\theta} s_{\tau}
& c^{~}_{\theta} s_{\tau} s_{\tau^{\prime}} + c_{\tau} c_{\tau^{\prime}} e^{-{\rm i}\varphi} \cr} \right )
$ && $\matrix{
\cos\theta = |V_{cs}| \cr
\tan\tau = |V_{ts}/V_{us}| \cr
\tan\tau^{\prime} = |V_{cb}/V_{cd}| \cr} $\
\
[*P6:*]{} $V \; = \; R_{12}(\theta) ~ R_{23}(\sigma, \varphi)
~ R_{31}(\tau)$ && ${\cal J} = s^{~}_{\theta} c^{~}_{\theta} s_{\sigma} c^2_{\sigma} s_{\tau} c_{\tau} \sin\varphi$\
$\left ( \matrix{
-s^{~}_{\theta} s_{\sigma} s_{\tau} + c^{~}_{\theta} c_{\tau} e^{-{\rm i}\varphi} & s^{~}_{\theta} c_{\sigma} &
s^{~}_{\theta} s_{\sigma} c_{\tau} + c^{~}_{\theta} s_{\tau} e^{-{\rm i}\varphi} \cr
-c^{~}_{\theta} s_{\sigma} s_{\tau} - s^{~}_{\theta} c_{\tau} e^{-{\rm i}\varphi} & c^{~}_{\theta} c_{\sigma} &
c^{~}_{\theta} s_{\sigma} c_{\tau} - s^{~}_{\theta} s_{\tau} e^{-{\rm i}\varphi} \cr
-c_{\sigma} s_{\tau} & -s_{\sigma} & c_{\sigma} c_{\tau} \cr} \right )
$ && $\matrix{
\tan\theta = |V_{us}/V_{cs}| \cr
\sin\sigma = |V_{ts}| \cr
\tan\tau = |V_{td}/V_{tb}| \cr} $\
\
[*P7:*]{} $V \; = \; R_{23}(\sigma) ~ R_{12}(\theta, \varphi)
~ R^{-1}_{31}(\tau)$ && ${\cal J} = s^{~}_{\theta} c^2_{\theta} s_{\sigma} c_{\sigma} s_{\tau} c_{\tau} \sin\varphi$\
$\left ( \matrix{
c^{~}_{\theta} c_{\tau} & s^{~}_{\theta} & -c^{~}_{\theta} s_{\tau} \cr
-s^{~}_{\theta} c_{\sigma} c_{\tau} + s_{\sigma} s_{\tau} e^{-{\rm i}\varphi} & c^{~}_{\theta} c_{\sigma}
& s^{~}_{\theta} c_{\sigma} s_{\tau} + s_{\sigma} c_{\tau} e^{-{\rm i}\varphi} \cr
s^{~}_{\theta} s_{\sigma} c_{\tau} + c_{\sigma} s_{\tau} e^{-{\rm i}\varphi} & -c^{~}_{\theta} s_{\sigma}
& -s^{~}_{\theta} s_{\sigma} s_{\tau} + c_{\sigma} c_{\tau} e^{-{\rm i}\varphi} \cr} \right )
$ && $\matrix{
\sin\theta = |V_{us}| \cr
\tan\sigma = |V_{ts}/V_{cs}| \cr
\tan\tau = |V_{ub}/V_{ud}| \cr} $\
\
[*P8:*]{} $V \; = \; R_{31}(\tau) ~ R_{12}(\theta, \varphi)
~ R_{23}(\sigma)$ && ${\cal J} = s^{~}_{\theta} c^2_{\theta} s_{\sigma} c_{\sigma} s_{\tau} c_{\tau} \sin\varphi$\
$\left ( \matrix{
c^{~}_{\theta} c_{\tau} & s^{~}_{\theta} c_{\sigma} c_{\tau} - s_{\sigma} s_{\tau} e^{-{\rm i}\varphi}
& s^{~}_{\theta} s_{\sigma} c_{\tau} + c_{\sigma} s_{\tau} e^{-{\rm i}\varphi} \cr
-s^{~}_{\theta} & c^{~}_{\theta} c_{\sigma} & c^{~}_{\theta} s_{\sigma} \cr
-c^{~}_{\theta} s_{\tau} & -s^{~}_{\theta} c_{\sigma} s_{\tau} - s_{\sigma} c_{\tau} e^{-{\rm i}\varphi}
& -s^{~}_{\theta} s_{\sigma} s_{\tau} + c_{\sigma} c_{\tau} e^{-{\rm i}\varphi} \cr} \right )
$ && $\matrix{
\sin\theta = |V_{cd}| \cr
\tan\sigma = |V_{cb}/V_{cs}| \cr
\tan\tau = |V_{td}/V_{ud}| \cr} $\
\
[*P9:*]{} $V \; = \; R_{31}(\tau) ~ R_{23}(\sigma, \varphi)
~ R^{-1}_{12}(\theta)$ && ${\cal J} = s^{~}_{\theta} c^{~}_{\theta} s_{\sigma} c^2_{\sigma} s_{\tau} c_{\tau} \sin\varphi$\
$\left ( \matrix{
-s^{~}_{\theta} s_{\sigma} s_{\tau} + c^{~}_{\theta} c_{\tau} e^{-{\rm i}\varphi}
& -c^{~}_{\theta} s_{\sigma} s_{\tau} - s^{~}_{\theta} c_{\tau} e^{-{\rm i}\varphi} & c_{\sigma} s_{\tau} \cr
s^{~}_{\theta} c_{\sigma} & c^{~}_{\theta} c_{\sigma} & s_{\sigma} \cr
-s^{~}_{\theta} s_{\sigma} c_{\tau} - c^{~}_{\theta} s_{\tau} e^{-{\rm i}\varphi}
& -c^{~}_{\theta} s_{\sigma} c_{\tau} + s^{~}_{\theta} s_{\tau} e^{-{\rm i}\varphi} & c_{\sigma} c_{\tau} \cr} \right )
$ && $\matrix{
\tan\theta = |V_{cd}/V_{cs}| \cr
\sin\sigma = |V_{cb}| \cr
\tan\tau = |V_{ub}/V_{tb}| \cr} $\
\
[^1]: Electronic address: [email protected]
[^2]: Electronic address: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, an analytical framework is provided to analyze the uplink performance of device-to-device (D2D)-enabled millimeter wave (mmWave) cellular networks with clustered D2D user equipments (UEs). Locations of cellular UEs are modeled as a Poison Point Process (PPP), while locations of potential D2D UEs are modeled as a Poisson Cluster Process (PCP). Signal-to-interference-plus-noise ratio (SINR) outage probabilities are derived for both cellular and D2D links using tools from stochastic geometry. The distinguishing features of mmWave communications such as directional beamforming and having different path loss laws for line-of-sight (LOS) and non-line-of-sight (NLOS) links are incorporated into the outage analysis by employing a flexible mode selection scheme. Also, the effect of beamforming alignment errors on the outage probability is investigated to get insight on the performance in practical scenarios. Moreover, area spectral efficiency (ASE) of the entire network is determined for both underlay and overlay types of sharing. Optimal spectrum partition factor is determined for overlay sharing by considering the optimal weighted proportional fair spectrum partition.'
author:
- 'Esma Turgut and M. Cenk Gursoy [^1]'
title: 'Uplink Performance Analysis in D2D-Enabled mmWave Cellular Networks with Clustered Users'
---
[1.7]{}
Device-to-device (D2D) communication, uplink analysis of mmWave cellular networks, SINR outage probability, Poisson point process, Poisson cluster process, Thomas cluster process, stochastic geometry, mode selection.
Introduction
============
Recent years have witnessed an overwhelming increase in mobile data traffic due to e.g., ever increasing use of smart phones, portable devices, and data-hungry multimedia applications. Since limited available spectrum in microwave ($\mu$Wave) bands does not seem to be capable of meeting this demand in the near future, there has recently been significant interest in moving to new frequency bands. Therefore, the use of large bandwidth at millimeter wave (mmWave) frequency bands to provide much higher data rates and immense capacity has been proposed to be an important part of the fifth generation (5G) cellular networks and has attracted considerable attention recently [@Rappaport1] – [@Ghosh].
Despite the great potential of mmWave bands, they have been considered attractive only for short range-indoor communication due to the increase in free-space path loss with increasing frequency, and poor penetration through solid materials. However, recent channel measurements and recent advances in RF integrated circuit design have motivated the use of these high frequencies for outdoor communication over a transmission range of about 150-200 meters [@Rappaport1], [@Ghosh]. Also, with the employment of highly directional antennas, high propagation loss in the side lobes can be taken advantage of to support simultaneous communication with very limited or almost no interference to achieve lower link outage probabilities, much higher data rates and network capacity than those in $\mu$Wave networks.
Another promising solution to improve the network capacity is to enable device-to-device (D2D) communication in cellular networks. D2D communication allows proximity user equipments (UEs) to establish a direct communication link with each other by bypassing the base station (BS). In other words, conventional two-hop cellular link is replaced by a direct D2D link to enhance the network capacity. Network performance of D2D communication in cellular networks has recently been extensively studied as an important component of fourth generation (4G) cellular networks by using stochastic geometry. In [@Hesham1] and [@Ghosh2], outage and spectrum efficiency of D2D-enabled uplink cellular networks were studied by considering mode selection schemes along with truncated channel inversion power control. In [@Ghosh2], a distance-based mode selection scheme was employed while [@Hesham1] considered a flexible mode selection scheme. Also, effect of spectrum sharing type on the performance was investigated in [@Ghosh2]. In these works, locations of the transmitting potential D2D UEs were modeled using Poisson Point Processes (PPPs) while the receiving D2D UEs were assumed to be distributed within a circle around the transmitting D2D UE. However, in D2D networks, UEs are very likely to form clusters rather than being distributed uniformly in the network. Therefore, a more realistic spatial model has been considered in several recent studies by modeling the locations of the D2D UEs as Poisson Cluster Process (PCP) distributed [@Afshang], [@Afshang2], [@Afshang3]. In [@Afshang], authors obtained expressions for the coverage probability and area spectral efficiency of an out-of-band D2D network. Performance of cluster-centric content placement in a cache-enabled D2D network was studied in [@Afshang2], where the authors have considered a cluster-centric approach which optimizes the performance of the entire cluster rather than the individual D2D UEs. In-band D2D communication where the cellular and D2D networks coexist in the same frequency band was considered in [@Afshang3] by combining PCP with a Poisson Hole Process (PHP). In particular, D2D UE locations are modeled by a Hole Cluster Process (HCP). However, neither of these works on D2D communication has addressed transmission in mmWave frequency bands. Network performance of D2D communication in cellular networks has been gaining even more importance in 5G networks and it is expected to be an essential part of mmWave 5G cellular networks.
Several recent studies have also addressed the mmWave D2D communication. In [@Qiao], authors considered two types of D2D communication schemes in mmWave 5G cellular networks: local D2D and global D2D communications. Local D2D communication is performed by offloading the traffic from the BSs, while global D2D communication is established with multihop wireless transmissions via BSs between two wireless devices associated with different cells. The authors in [@Qiao] also proposed a resource sharing scheme to share network resources among local D2D and global D2D communications by considering the unique features of mmWave transmissions. In [@Guizani], authors have proposed a resource allocation scheme in mmWave frequency bands, which enables underlay D2D communications to improve the system throughput and the spectral efficiency. mmWave D2D multi-hop routing for multimedia applications was studied in [@Eshraghi] to maximize the sum video quality by taking into account the unique characteristics of the mmWave propagation. In [@Turgut], we have studied the uplink performance of D2D-enabled mmWave cellular networks where the locations of both cellular and potential D2D UEs are modeled as a PPP. In other words, correlation among the locations of potential D2D UEs was not taken into account (and also beamsteering errors and area spectral efficient were not addressed in [@Turgut]).
In this work, we consider a single-tier uplink network in which the BSs and cellular UEs coexist with the potential D2D UEs. We model the locations of BSs and cellular UEs as independent homogeneous PPPs. Unlike previous works on mmWave D2D communication systems where the D2D UEs are assumed to be uniformly distributed in the network, we model the locations of potential D2D UEs as a PCP to provide a more appropriate and realistic model. Moreover, potential D2D UEs in the clusters can choose to operate in cellular and D2D mode according to a mode selection scheme. Although there is a higher possibility of operating in D2D mode due to closer distances between the UEs in the clusters, this mode selection strategy provides flexibility and generality in our analysis. Additionally, different from the previous studies on D2D communications, most of which consider only underlay or overlay types of sharing, we take into account both types of sharing strategies to show their impact on the performance of the mmWave D2D networks.
More specifically, our main contributions can be summarized as follows:
- We provide an analytical framework to analyze the uplink performance of D2D-enabled mmWave cellular networks with clustered UEs by using tools from stochastic geometry. In particular, cellular and potential D2D UEs can coexist in the same band, and the cellular UEs are distributed uniformly and potential D2D UEs form clusters in the network.
- An expression for the probability of selecting the D2D mode for a potential D2D UE located in a cluster is derived by considering a flexible mode selection scheme. Laplace transform expressions for both cellular and D2D interference links are obtained. Using these characterizations, we derive SINR outage probability expressions for both cellular and D2D links employing the modified LOS ball model for blockage modeling.
- We investigate the effect of spectrum sharing type on SINR outage probability. The effect of LOS ball model parameters is also identified. Additionally, the impact of alignment errors on the SINR outage probability is investigated to get insight on the performance in practical scenarios.
- Area spectral efficiency (ASE) of the entire network is determined for both underlay and overlay types of sharing. We have shown that an optimal value for the average number of simultaneously active D2D links, maximizing the ASE, exists and this optimal value is independent of cluster center density. Moreover, optimal spectrum partition factor is found for overlay sharing by considering the optimal weighted proportional fair spectrum partition.
The rest of the paper is organized as follows. In Section \[sec:system\_model\], system model is introduced. Characterizations for the transmission strategies and interference models are provided in Section \[Transmission and Interference Characterizations\]. In Section \[sec:Analysis of Uplink SINR Outage Probability\], uplink SINR outage probabilities for both cellular and D2D links are derived initially considering perfect beam alignment, and then in the presence of beamsteering errors. ASE is defined and analyzed in Section \[sec:Analysis of Area Spectral Efficiency\] for the underlay and overlay types of sharing. In Section \[sec:Simulation and Numerical Results\], simulations and numerical results are presented to identify the impact of several system parameters on the performance metrics. Finally, conclusions and suggestions for future work are provided in Section \[sec:Conclusion\]. Proofs are relegated to the Appendix.
System Model {#sec:system_model}
============
In this section, the system model for D2D-communication-enabled mmWave cellular networks with clustered UEs is presented. We consider a single-tier uplink network, where BSs are spatially distributed according to an independent homogeneous PPP $\Phi_{B}$ with density $\lambda_{B}$ on the Euclidean plane. UEs are categorized as cellular UEs and potential D2D UEs. Cellular UEs are distributed according to an independent homogeneous PPP $\Phi_{CU}$ with density $\lambda_{CU}$, while potential D2D UEs are clustered around the cluster centers in which the cluster centers are also distributed according to an independent homogeneous PPP $\Phi_{C}$ with density $\lambda_{C}$. For instance, cellular UEs can be regarded as pedestrians or UEs in transit which are more likely to be uniformly distributed in the network, and therefore homogeneous PPP is a better choice for the modeling of such UEs. On the other hand, potential D2D UEs are located in high UE density areas, i.e. hotspots, and are expected to be closer to each other forming clusters, and thus PCP is a more appropriate and accurate model than a homogeneous PPP.
Cluster members, i.e. potential D2D UEs, are assumed to be symmetrically independently and identically distributed (i.i.d.) around the cluster center. The union of cluster members’ locations form a PCP, denoted by $\Phi_{D}$. In this paper, we model $\Phi_D$ as a Thomas cluster process, where the UEs are scattered around the cluster center $x \in \Phi_C$ according to a Gaussian distribution with variance $\sigma_d^2$ and the probability density function (pdf) of a potential D2D UE’s location is given by [@Haenggi] $$f_{Y}(y)=\frac{1}{2\pi\sigma_d^2} \exp\left( -\frac{\|y\|^2 }{2\sigma_d^2}\right), \quad y \in \mathbb{R}^2.$$ where $y$ is the UE’s location relative to the cluster center and $\|\cdot\|$ is the Euclidean norm. Each potential D2D UE (i.e., each cluster member) in a cluster $x \in \Phi_C$ has the capability of establishing a direct D2D link with the cluster members in the same cluster or they can communicate with a BS in $\Phi_B$. Hence, potential D2D UEs can operate in one of the two modes according to the mode selection scheme: cellular and D2D mode. When operating in D2D mode, a potential D2D UE in the cluster bypasses the BS and communicates directly with its intended receiver in the same cluster. Let $\mathcal{N}^x$ denote the set of all potential D2D UEs in a cluster $x \in \Phi_C$, and $N$ be the total number of potential D2D UEs per cluster, which is assumed to be the same in each cluster. $\mathcal{N}^x$ can be divided into two subsets: set of possible transmitting potential D2D UEs ($\mathcal{N}_t^x$), and set of possible receiving D2D UEs ($\mathcal{N}_r^x$). The set of all simultaneously transmitting potential D2D UEs is denoted by $\mathcal{A}^x \subset \mathcal{N}_t^x$ where $|\mathcal{A}^x|$ is modeled as a Poisson distributed random variable with mean $\bar{n}$. $\mathcal{A}^x$ can also be divided into two subsets: set of simultaneously transmitting potential D2D UEs in D2D mode ($\mathcal{A}_d^x$) and set of simultaneously transmitting potential D2D UEs in cellular mode ($\mathcal{A}_c^x$) which are modeled as Poisson distributed random variables with means $\bar{n}P_{D2D}$ and $\bar{n}(1-P_{D2D})$, respectively. $P_{D2D}$ above is the probability of potential D2D UE selecting the D2D mode, and this probability will be described and characterized in detail later in the paper.
Without loss of generality, a typical receiving node (BS) is assumed to be located at the origin according to Slivnyak’s theorem for cellular UEs and potential D2D UEs transmitting in cellular mode, and these UEs are assumed to be associated with their closest BS. The link between the BSs and cellular UEs/potential D2D UEs transmitting in cellular mode is called the cellular link, and the link between the transmitting and receiving D2D UEs in the same cluster is called the D2D link in the rest of the paper. For the D2D link, we conduct an analysis for a typical D2D UE located at the origin, which is randomly chosen in a randomly chosen cluster which is referred to as the representative cluster centered at $x_0 \in \Phi_C$ throughout the paper.
Transmission Strategies and Interference Characterizations {#Transmission and Interference Characterizations}
==========================================================
In this section, we provide characterizations for the transmission strategies and interference models. In particular, we describe two types of spectrum sharing policies between the cellular and D2D UEs, identify the interference experienced in cellular uplink and D2D links, and characterize the distributions of the link distances. Furthermore, we discuss the mode selection strategy and specify the beamforming assumptions.
Spectrum Sharing
----------------
Cellular spectrum can be shared between the cellular and D2D UEs in two different ways: underlay and overlay. In the underlay type of sharing, D2D UEs can opportunistically access the channel occupied by the cellular UEs. While for the overlay type of sharing, the uplink spectrum is divided into two orthogonal portions, i.e., a fraction $\delta$ of the cellular spectrum is assigned to D2D mode and the remaining $(1-\delta)$ fraction is used for cellular communication, where $\delta$ is the spectrum partition factor [@Ghosh2]. Also, parameter $\beta$ is defined as the spectrum sharing indicator which is equal to one for underlay and zero for overlay type of sharing.
Interference Modeling
---------------------
### Interference in cellular uplink
The total interference in a cellular uplink experienced by a typical receiving node, i.e. the BS located at the origin, emerges from two sources: 1) interference from other cellular UEs/potential D2D UEs transmitting in cellular mode and 2) interference from other potential UEs transmitting in D2D mode (if underlay type of spectrum sharing is employed). Each cellular UE/potential D2D UE transmitting in cellular mode is assigned a unique and orthogonal channel by its associated BS which means that there is no intra-cell interference between UEs transmitting in cellular mode in the same cell. However, we assume universal frequency reuse across the entire cellular network causing inter-cell interference from the other cells’ cellular UEs. Moreover, we consider a congested network scenario in which the total density of cellular UEs and potential D2D UEs in cellular mode is much higher than the density of BSs. In other words, each BS will always have at least one cellular UE to serve in the uplink channel. Different from the downlink communication, where we can model the interfering cellular UEs and potential D2D UEs in cellular mode in different cells as a PPP with density $\lambda_B$, modeling of the cellular interference for uplink is much more complicated [@Dhillon]. For example, an interfering UE in uplink case can be arbitrarily close to the BS, i.e., it can be closer than the UE being served. Therefore, one commonly used approach is to model the uplink other cell interferers as a non-homogeneous PPP $\Phi_c$ with a radially symmetric distance dependent density function given by $$\lambda_u(t)=\lambda_{u,L}(t)+\lambda_{u,N}(t)=\sum_{j \in \{L,N\}} \lambda_B p_{j,c}(t)Q(t^{\alpha_{j,c}}) \label{density_function}$$ where $Q(y)$ is the probability that path loss of a cellular UE to its serving BS is smaller than $y^{-1}$ [@Andrews4]. In the underlay case, we focus on one uplink channel which is shared by the cellular and D2D UEs. Since the potential D2D UEs operating in D2D mode coexist with the cellular UEs in an uplink channel, they cause both intra-cell and inter-cell interference at the BSs. On the other hand, in the overlay case, since the uplink spectrum is divided into two orthogonal portions, there is no cross-mode interference, i.e., no interference from the D2D UEs to the cellular UEs and vice versa.
### Interference in D2D link
The total interference experienced by a typical D2D UE $\in \mathcal{N}_r^{x_0}$ in the representative cluster originates from three different sources: 1) cross-mode interference caused by the other cellular UEs/potential D2D UEs transmitting in cellular mode (if underlay sharing is adopted); 2) intra-cluster interference caused by the simultaneously transmitting D2D UEs in D2D mode inside the representative cluster; and 3) inter-cluster interference caused by the simultaneously transmitting D2D UEs in D2D mode outside the representative cluster. In the overlay case, there is no cross-mode interference, i.e., no interference from the cellular UEs/potential D2D UEs transmitting in cellular mode to the D2D UEs.
Path-loss exponents and link distance modeling {#sec:Path-loss exponents and link distance modeling}
----------------------------------------------
A transmitting UE can either have a line-of-sight (LOS) or non-line-of-sight (NLOS) link to the BS or the receiving UE. In a LOS state, UE should be visible to the receiving nodes, indicating that there is no blockage in the link. On the other hand, in a NLOS state, blockage occurs in the link. Consider an arbitrary link of length $r$, and define the LOS probability function $p(r)$ as the probability that the link is LOS. Using field measurements and stochastic blockage models, $p(r)$ can be modeled as $e^{-\zeta r}$ where decay rate $\zeta$ depends on the building parameter and density [@Bai1]. For analytical tractability, LOS probability function $p(r)$ can be approximated by a step function. In this approach, the irregular geometry of the LOS region is replaced with its equivalent LOS ball model. In this paper, modified LOS ball model is adopted similarly as in [@Andrews3]. According to this model, the LOS probability function of a link $p_L(r)$ is equal to some constant $p_{L}$ if the link distance $r$ is less than ball radius $R_B$ and zero otherwise. The parameters $p_{L}$ and $R_B$ depend on geographical regions. $(p_{L,c}, R_{B,c})$ and $(p_{L,d}, R_{B,d})$ are the LOS ball model parameters for cellular and D2D links, respectively[^2]. Therefore, LOS and NLOS probability function for each link can be expressed as follows: $$\begin{aligned}
\label{LOS_prob_funct}
p_{L,\kappa}(r) &= p_{L,\kappa}\mathds{1}(r \le R_{B,\kappa}) \nonumber \\
p_{N,\kappa}(r) &= (1-p_{L,\kappa})\mathds{1}(r \le R_{B,\kappa})+\mathds{1}(r > R_{B,\kappa})\end{aligned}$$ for $\kappa \in \{c, d\}$ where $\mathds{1}(\cdot)$ is the indicator function. Different path loss laws are applied to LOS and NLOS links, and thus $\alpha_{L,\kappa}$ and $\alpha_{N,\kappa}$ are the LOS and NLOS path-loss exponents for $\kappa \in \{c, d\}$, respectively.
### D2D communication
Regarding the distance modeling for potential D2D UEs which are assumed to be located inside the clusters, there are three types of distances: 1) D2D link distance, i.e., serving distance, 2) intra-cluster interferer distances and 3) inter-cluster interferer distances. Without loss of generality, a typical receiving D2D UE $\in \mathcal{N}_r^{x_0}$ is assumed to be located at the origin, and is associated with another D2D UE $\in \mathcal{A}_d^{x_0}$ located at $y_0$ chosen uniformly at random within the same cluster. Note that the cluster center location is $x_0$ with respect to the origin (where the typical receiving D2D UE is), and transmitting D2D UE location is $y_0$ with respect to the cluster center. Fig. \[Fig\_Rep\_Clu\] illustrates the considered setting where the relative locations are denoted by vectors. Also, let $r_{d0}=\|x_0+y_0\|$ denote the distance between the transmitting and typical receiving D2D UEs. Similarly, let $\{r_{d1}=\|x_0+y\|, \forall y \in \mathcal{A}_d^{x_0} \setminus y_0\}$ denote the set of the distances from simultaneously transmitting D2D UEs in D2D mode inside the representative cluster to a typical receiving D2D UE $\in \mathcal{N}_r^{x_0}$. Distances $r_{d0}$ and $r_{d1}$ are also illustrated in Fig. \[Fig\_Rep\_Clu\]. Note that, $r_{d0}$ is the serving distance, and $\{r_{d1}\}$ is the set of distances from intra-cluster interfering D2D UEs. Actually, these distances are correlated due to the common factor $x_0$. By conditioning on $\omega_0=\|x_0\|$ and using the fact that $y_0$ and $\{y\}$ are i.i.d. zero-mean Gaussian random variables with variance $\sigma_d^2$ in $\mathbb{R}^2$, the serving distance $r_{d0}=\|x_0+y_0\|$ and the the set of distances from intra-cluster interfering D2D UEs $\{r_{d1}=\|x_0+y\|, \forall y \in \mathcal{A}_d^{x_0} \setminus y_0\}$ are conditionally i.i.d. It is shown that conditioning on $\omega_0$ instead of $x_0$ is sufficient [@Afshang]. Therefore, the pdf of each distance is characterized by a Rician distribution [@Afshang]: $$\begin{aligned}
f_{R_{d0}}(r_{d0}|\omega_0)= \text{Ricepdf}(r_{d0},\omega_0;\sigma_d^2) \label{pdfofrd0} \\
f_{R_{d1}}(r_{d1}|\omega_0)= \text{Ricepdf}(r_{d1},\omega_0;\sigma_d^2) \label{pdfofrd1}\end{aligned}$$ where $\text{Ricepdf}(a,b;\sigma_d^2)=\frac{a}{\sigma_d^2}\exp(-\frac{a^2+b^2}{2\sigma_d^2})I_0(\frac{ab}{\sigma_d^2})$ and $I_0(\cdot)$ is the modified Bessel function of the first kind with order zero. Similarly, let $\{r_{d2}=\|x+y\|, \forall y \in \mathcal{A}_d^{x}\}$ denote the set of the distances from simultaneously transmitting D2D UEs in D2D mode in the other clusters to a typical D2D UE $\in \mathcal{N}_r^{x_0}$, i.e., $\{r_{d2}\}$ is the set of distances from inter-cluster interfering D2D UEs. By conditioning on $\omega=\|x\|$, the pdf of each distance is given by $f_{R_{d2}}(r_{d2}|\omega)= \text{Ricepdf}(r_{d2},\omega;\sigma_d^2)$.
### Cellular communication {#sec:cellular link distance modeling}
Recall that cellular UEs and potential D2D UEs transmitting in cellular mode are assumed to be associated with their closest BS, and therefore the pdf of the cellular link distance $r_c$ to the nearest LOS/NLOS BS is given by [@Bai2] $$\label{cellularlinkpdftonearestBS}
f_s(r_c)= 2\pi\lambda_B r_c p_{s,c}(r_c) e^{-2\pi\lambda_B\psi_s(r_c)}/\mathcal{B}_{s,c} \quad \text{for} \; s \in \{L,N\}$$ where $\psi_s(r_c)=\int_0^{r_c} x p_{s,c}(x) dx$, $\mathcal{B}_{s,c}=1-e^{-2\pi\lambda_B\int_0^{\infty} x p_{s,c}(x) dx}$ is the probability that a UE has at least one LOS/NLOS BS, and $p_{s,c}(x)$ is given in (\[LOS\_prob\_funct\]) for $s \in \{L,N\}$. In fact, since potential D2D UEs are distributed according to a PCP around the cluster centers, modeling the distance of potential D2D UEs transmitting in cellular mode to the closest BS by (\[cellularlinkpdftonearestBS\]) is only an approximation. However, we have verified in the simulations that this assumption is quite reasonable especially for small values of the scattering variance $\sigma_d^2$.
Let $\{r_{y_x}=\|x+y\|, \forall x \in \Phi_C, \forall y \in \mathcal{A}_d^{x}\}$ be the set of distances from the cross-mode interferers, i.e. D2D UEs, to a typical BS at the origin. Then, the pdf of each distance is given by $f_{R_{y_x}}(r_{y_x}|\omega)= \text{Ricepdf}(r_{y_x},\omega;\sigma_d^2)$ where $\omega=\|x\|$.
Mode Selection
--------------
In this work, a flexible mode selection scheme similarly as in [@Hesham1] is considered. In this scheme, a potential D2D UE chooses the D2D mode if the biased D2D link quality is at least as good as the cellular uplink quality. In other words, a potential D2D UE will operate in D2D mode if $T_d r_d^{-\alpha_{s,d}} \geq r_c^{-\alpha_{s,c}}$, where $T_d \in [0, \infty)$ is the biasing factor, and $r_c$ and $r_d$ are the cellular and D2D link distances, respectively. Biasing factor $T_d$ has two extremes, $T_d=0$ and $T_d \to \infty$. In the first extreme case, D2D communication is disabled, while in the second case, each potential D2D UE is forced to select the D2D mode. The probability of selecting D2D mode, $P_{D2D}$, is provided in the following Lemma.
*Lemma 1:* Probability of selecting D2D mode for a potential D2D UE located in a cluster $x \in \Phi_C$ is $$\begin{aligned}
P_{D2D}&= \sum_{s \in \{L,N\}} \sum_{s^{\prime} \in \{L,N\}} \int_0^{\infty} \int_0^{\infty} e^{-2\pi\lambda_B \psi_s\left(r_d^{\alpha_{s^{\prime},d}/\alpha_{s,c}}/T_d^{1/\alpha_{s,c}}\right)} f_{R_d}(r_d|\omega) f_{\Omega}(\omega) p_{s^{\prime},d}(r_d) dr_d d\omega\end{aligned}$$ where $\psi_s(a)=\int_0^{a} x p_{s,c}(x) dx$, $p_{s,c}(x)$ and $p_{s^{\prime},d}(r_d)$ are given in (\[LOS\_prob\_funct\]), $f_{R_d}(r_d|\omega)=\text{Ricepdf}(r_{d},\omega;\sigma_d^2)$, and $f_{\Omega}(\omega)=\frac{\omega}{\sigma_d^2}\exp(-\frac{\omega^2}{2\sigma_d^2})$.
*Proof:* See Appendix \[Proof of Lemma 1\].
Directional beamforming {#sec:Directional beamforming}
-----------------------
Antenna arrays at the BSs and UEs are assumed to perform directional beamforming where the main lobe being directed towards the dominant propagation path while smaller side lobes direct energy in other directions. For tractability in the analysis and similar to [@Bai2], [@Bai3], [@Wildman], [@Marco2], antenna arrays are approximated by a sectored antenna model [@Hunter]. The array gains are assumed to be constant $M_{\nu}$ for all angles in the main lobe and another smaller constant $m_{\nu}$ in the side lobe for $\nu \in \{{\text{BS}}, {\text{UE}}\}$. Initially, perfect beam alignment [^3] is assumed in between the transmitting nodes (e.g., cellular or potential D2D UEs) and receiving nodes (e.g., BSs or receiving D2D UEs), leading to an overall antenna gain of $M_{{\text{BS}}}M_{{\text{UE}}}$ for cellular link and $M_{{\text{UE}}}M_{{\text{UE}}}$ for D2D link. In other words, maximum directivity gain can be achieved for the intended link by assuming that the transmitting node and receiving node can adjust their antenna steering orientation using the estimated angles of arrivals. Also, the beam direction of the interfering nodes is modeled as a uniform random variable on $[0,2\pi)$. Therefore, the effective antenna gain is a discrete random variable (RV) described by
$$G=\left\{
\begin{array}{ll}
M_{l}M_{{\text{UE}}} &\text{w. p.} \; p_{M_{l}M_{{\text{UE}}}}=\frac{\theta_{l}}{2\pi} \frac{\theta_{{\text{UE}}}}{2\pi} \\
M_{l}m_{{\text{UE}}} &\text{w. p.} \; p_{M_{l}m_{{\text{UE}}}}=\frac{\theta_{l}}{2\pi} \frac{2\pi-\theta_{{\text{UE}}}}{2\pi} \\
m_{l}M_{{\text{UE}}} &\text{w. p.} \; p_{m_{l}M_{{\text{UE}}}}=\frac{2\pi-\theta_{l}}{2\pi} \frac{\theta_{{\text{UE}}}}{2\pi} \\
m_{l}m_{{\text{UE}}} &\text{w. p.} \; p_{m_{l}m_{{\text{UE}}}}=\frac{2\pi-\theta_{l}}{2\pi} \frac{2\pi-\theta_{{\text{UE}}}}{2\pi}
\end{array}
\right. \label{eq:antennagains}$$
for $l \in \{{\text{BS}}, {\text{UE}}\}$ where $\theta_{\nu}$ is the beam width of the main lobe for $\nu \in \{{\text{BS}}, {\text{UE}}\}$, and $p_{G}$ is the probability of having an antenna gain of $G$.
Analysis of Uplink SINR Outage Probability {#sec:Analysis of Uplink SINR Outage Probability}
==========================================
In this section, we first develop a theoretical framework to analyze the uplink SINR outage probability for a generic UE operating in cellular mode or D2D mode using stochastic geometry. Although a biasing-based mode selection scheme is considered for selecting between D2D and cellular modes, the developed framework can also be applied to different mode selection schemes.
Signal-to-Interference-plus-Noise Ratio (SINR)
----------------------------------------------
Recall that, without loss of generality, we consider a typical receiving node (BS or D2D UE $\in \mathcal{N}_r^{x_0}$ in the representative cluster) located at the origin. Therefore, the SINR experienced at a typical receiving node in cellular and D2D modes, respectively, can be written as $$SINR^{c}=\frac{P_{c}G_0h_0r_c^{-\alpha_{c}(r_c)}}{\sigma_N^2+\underbrace{\sum_{i \in \Phi_c} P_cG_ih_ir_i^{-\alpha_{c}(r_i)}}_{I_{cc}}+\underbrace{\sum_{x \in \Phi_C} \sum_{y \in \mathcal{A}_d^{x}} P_d G_{y_x}h_{y_x}r_{y_x}^{-\alpha_{d}(r_{y_x})}}_{I_{dc}}}$$ $$SINR^{d}=\frac{P_{d}G_0h_0r_{d0}^{-\alpha_{d}(r_{d0})}}{\sigma_N^2+\underbrace{\sum_{i \in \Phi_c} P_cG_ih_ir_i^{-\alpha_{c}(r_i)}}_{I_{cd}}+\underbrace{ \sum_{y \in \mathcal{A}_d^{x_0} \setminus y_0} P_d G_{y_{x_0}}h_{y_{x_0}}r_{d1}^{-\alpha_{d}(r_{d1})}}_{I_{dd_{\text{intra}}}}+\underbrace{\sum_{x \in \Phi_C \setminus x_0} \sum_{y \in \mathcal{A}_d^{x}} P_d G_{y_x}h_{y_x}r_{d2}^{-\alpha_{d}(r_{d2})}}_{I_{dd_{\text{inter}}}}}$$ where $P_{\kappa}$ is the transmit power of the UE operating in mode $\kappa \in \{c,d\}$, $G_0$ is the effective antenna gain of the link which is assumed to be equal to $M_{{\text{BS}}}M_{{\text{UE}}}$ for cellular link and $M_{{\text{UE}}}M_{{\text{UE}}}$ for D2D link, $h_0$ is the small-scale fading gain, $\alpha_{\kappa}(\cdot)$ is the path-loss exponent of the link, which depends on whether the link is LOS on NLOS, $r_c$ and $r_{d0}$ are the cellular and D2D link distances, respectively, $\sigma_N^2$ is the variance of the additive white Gaussian noise component, $I_{c\kappa}$ is the aggregate interference at the receiving node from cellular UEs using the same uplink channel in different cells, which constitute a non-homogeneous PPP $\Phi_c$ with density function given in (\[density\_function\]), and $I_{d\kappa}$ is the aggregate interference at the receiving node from D2D UEs located inside the clusters (hence including both inter-cell and intra-cell D2D UEs). For the D2D link, $I_{dd}$ has two components: intra-cluster interference $I_{dd_{\text{intra}}}$ and inter-cluster interference $I_{dd_{\text{inter}}}$. A similar notation is used for $I_{cc}$, $I_{cd}$, $I_{dc}$, $I_{dd_{\text{intra}}}$ and $I_{dd_{\text{inter}}}$, but note that the effective antenna gains $G_i$, $G_{y_x}$ and $G_{y_{x_0}}$, and path loss exponents $\alpha_{\kappa}(\cdot)$ are different for different interfering links as described in Section \[sec:Directional beamforming\] and Section \[sec:Path-loss exponents and link distance modeling\], respectively. Small-scale fading gains denoted by $h$ are assumed to have an independent exponential distribution in all links. The above description implicitly assumes underlay spectrum sharing between cellular and D2D UEs. Note that since there is no cross-mode interference in the overlay case, the SINR expression in this case reduces to $SINR^{c}=\frac{P_{c}G_0h_0r_c^{-\alpha_{c}(r_c)}}{\sigma_N^2+I_{cc}}$, and $SINR^{d}=\frac{P_{d}G_0h_0r_{d0}^{-\alpha_{d}(r_{d0})}}{\sigma_N^2+I_{dd_{\text{intra}}}+I_{dd_{\text{inter}}}}$, for mode $\kappa \in \{c,d\}$.
Laplace Transform of Interferences
----------------------------------
Before conducting the outage probability analysis, we first provide the Laplace transform expressions for each interference component. The thinning property of Poisson processes can be employed to split the interference component $I_{\chi}$ for $\chi \in \{cc,dc,cd,dd_{\text{intra}},dd_{\text{inter}}\}$ into 8 independent PPPs or PCPs as follows: $$\begin{aligned}
I_{\chi} &= I_{\chi,L} + I_{\chi,N} \nonumber \\
&=\sum_{G \in \big\{ \substack {M_{l}M_{{\text{UE}}},M_{l}m_{{\text{UE}}},\\ m_{l}M_{{\text{UE}}},m_{l}m_{{\text{UE}}}}\big \}} \sum_{j \in \{L,N\}}I_{{\chi,j}}^{G}, \label{eq:6PPP}\end{aligned}$$ for $l \in \{{\text{BS}}, {\text{UE}}\}$ where $I_{\chi,L}$ and $I_{\chi,N}$ are the aggregate LOS and NLOS interferences, and $I_{\chi,j}^{G}$ denotes the interference for $j \in \{L,N\}$ with random antenna gain $G$ defined in (\[eq:antennagains\]). According to the thinning theorem, each independent nonhomogeneous PPP has a density of $\lambda_B p_{j,c}(t)Q(t^{\alpha_{j,c}}) p_{G}$ for $\chi=\{cc,cd\}$, and each independent PCP has a density of $\lambda_C p_G$ for $\chi=\{dc,dd_{\text{inter}}\}$ where $p_{G}$ is given in (\[eq:antennagains\]) for each antenna gain $G$. Similarly, for intracell interference on D2D link, i.e. $I_{dd_{\text{intra}}}$, number of interfering D2D UEs are thinned by multiplying $p_{G}$ with $\bar{n} P_{D2D}-1$ where $\bar{n} P_{D2D}$ is the mean number of simultaneously transmitting potential D2D UEs in D2D mode.
Inserting (\[eq:6PPP\]) into the Laplace transform expression and using the definition of the Laplace transform yield $$\begin{aligned}
\mathcal{L}_{I_{\chi}}(v)&= \mathbb{E}_{I_{\chi}}[e^{-vI_{\chi}}]=\mathbb{E}_{I_{\chi}}[e^{-v(I_{\chi,L}+I_{\chi,N})}] \nonumber \\
&\stackrel{(a)}{=}\mathbb{E}_{I_{\chi,L}}\left[e^{-v \sum_{G} I_{\chi,L}^{G}}\right] \times \mathbb{E}_{I_{\chi,N}}\left[{e^{-v \sum_{G} I_{\chi,N}^{G}}}\right] \nonumber \\
&= \prod_G \prod_j \mathbb{E}_{I_{\chi,j}^G}\left[ e^{-vI_{\chi,j}^G}\right] \nonumber \\
&=\prod_G \prod_j \mathcal{L}_{I_{\chi,j}^G}(v), \label{eq:LT}\end{aligned}$$ where $G \in \{M_{l}M_{{\text{UE}}},M_{l}m_{{\text{UE}}}, m_{l}M_{{\text{UE}}},m_{l}m_{{\text{UE}}}\}$ for $l \in \{{\text{BS}}, {\text{UE}}\}$, $j \in \{L,N\}$, and (a) follows from the fact that $I_{\chi,L}$ and $I_{\chi,N}$ are interferences generated from two independent thinned PPPs or PCPs.
Laplace transform expressions for each interference component are provided in the following Lemmas.
*Lemma 2:* Laplace transform of the aggregate interference at the BS from cellular UEs using the same uplink channel in different cells is given by $$\begin{aligned}
\mathcal{L}_{I_{cc}}(v)=\exp\left(-2\pi\lambda_B\sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_{0}^{\infty} \frac{v P_cG_i t^{-\alpha_{j,c}}}{1+v P_cG_i t^{-\alpha_{j,c}}} Q\left(t^{\alpha_{j,c}}\right) p_{j,c}(t)tdt\right) \label{LT_cc}\end{aligned}$$ where $v=\frac{\Gamma r_c^{\alpha_{s,c}}}{P_{c}G_0}$, $p_{j,c}(t)$ is the LOS/NLOS probability function for cellular link given in (\[LOS\_prob\_funct\]), and $Q(y)$ is defined as [@Andrews4] $$Q(y)=1-\exp\left(-2\pi\lambda_B\left( \int_{0}^{y^{1/\alpha_{L,c}}} x p_{L,c}(x) dx+ \int_{0}^{y^{1/\alpha_{N,c}}} x p_{N,c}(x) dx \right) \right) \label{Q(y)}$$.
*Proof:* See Appendix \[Proof of Lemma 2\].
*Lemma 3:* Laplace transform of the aggregate interference at the BS from both intracell and intercell D2D UEs is given by $$\begin{aligned}
&\mathcal{L}_{I_{dc}}(v)= \nonumber \\
& \exp\left(-2\pi\lambda_C \sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_0^{\infty} \left(1-\exp\left(-\bar{n}P_{D2D} \int_{0}^{\infty}\frac{v P_dG_i u^{-\alpha_{j,d}}}{1+v P_dG_i u^{-\alpha_{j,d}}}f_{U}(u|t)p_{j,d}(u)du\right)\right)tdt\right)\end{aligned}$$ which can be approximated by $$\begin{aligned}
\mathcal{L}_{I_{dc}}(v) \approx \exp\left(-2\pi\lambda_C \bar{n}P_{D2D} \sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_0^{\infty} \frac{v P_dG_i u^{-\alpha_{j,d}}}{1+v P_dG_i u^{-\alpha_{j,d}}}p_{j,d}(u)udu\right)\end{aligned}$$ where $v=\frac{\Gamma r_c^{\alpha_{s,c}}}{P_{c}G_0}$, $p_{j,d}(u)$ is the LOS/NLOS probability function for D2D link given in (\[LOS\_prob\_funct\]), and the approximation follows from the Taylor series expansion of exponential function, i.e. $1-\exp(-x)\approx x$ for small $x$, and the Rician distribution property that $\int_{0}^{\infty}f_{U}(u|t)tdt=u$.
*Proof:* See Appendix \[Proof of Lemma 3\].
*Lemma 4:* Laplace transform of the aggregate interference at the typical UE from cellular UEs using the same uplink channel in different cells is given by $$\begin{aligned}
\mathcal{L}_{I_{cd}}(v)=\exp\left(-2\pi\lambda_B\sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_{0}^{\infty} \frac{v P_cG_i t^{-\alpha_{j,c}}}{1+v P_cG_i t^{-\alpha_{j,c}}}p_{j,c}(t)tdt\right) \label{LT_cd}\end{aligned}$$ where $v=\frac{\Gamma r_{d0}^{\alpha_{s,d}}}{P_{d}G_0}$, $p_{j,c}(t)$ is the LOS/NLOS probability function for cellular link given in (\[LOS\_prob\_funct\]), and $Q(y)$ is given in (\[Q(y)\]).
*Proof:* Proof follows similar steps as in the proof of Lemma 2.
*Lemma 5:* Laplace transform of the intra-cluster interference at the typical UE $\in \mathcal{N}_r^{x_0}$ in the representative cluster is given by $$\begin{aligned}
\mathcal{L}_{I_{dd_{\text{intra}}}}(v|w_0) = \exp \left(-\left(\bar{n}P_{D2D}-1\right) \sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_0^{\infty} \frac{v G_i u^{-\alpha_{j,d}}}{1+v G_i u^{-\alpha_{j,d}}}f_{U}(u|w_0)p_{j,d}(u)du \right)\end{aligned}$$ where $v=\frac{\Gamma r_{d0}^{\alpha_{s,d}}}{P_dG_0}$, and $p_{j,d}(u)$ is the LOS/NLOS probability function for D2D link given in (\[LOS\_prob\_funct\]).
*Proof:* See Appendix \[Proof of Lemma 5\].
*Lemma 6:* Laplace transform of the inter-cluster interference at the typical UE $\in \mathcal{N}_r^{x_0}$ in the representative cluster is given by $$\begin{aligned}
&\mathcal{L}_{I_{dd_{\text{inter}}}}(v) = \nonumber \\
& \exp\left(-2\pi\lambda_C \sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_0^{\infty} \left(1-\exp\left(-\bar{n}P_{D2D} \int_{0}^{\infty}\frac{vP_d G_i u^{-\alpha_{j,d}}}{1+vP_d G_i u^{-\alpha_{j,d}}}f_{U}(u|t)p_{j,d}(u)du\right)\right)tdt\right)\end{aligned}$$ which can be approximated by $$\begin{aligned}
\mathcal{L}_{I_{dd_{\text{inter}}}}(v) \approx \exp\left(-2\pi\lambda_C \bar{n}P_{D2D} \sum_{j \in \{L,N\}} \sum_{i=1}^4 p_{G_i} \int_0^{\infty} \frac{v P_dG_i u^{-\alpha_{j,d}}}{1+vP_d G_i u^{-\alpha_{j,d}}}p_{j,d}(u)udu\right)\end{aligned}$$ where $v=\frac{\Gamma r_{d0}^{\alpha_{s,d}}}{P_dG_0}$.
*Proof:* Proof follows similar steps as in the proof of Lemma 3.
Uplink SINR Outage Probability {#sec:Uplink SINR Outage}
------------------------------
The uplink SINR outage probability ${\text{P}_{\text{out}}}$ is defined as the probability that the received SINR is less than a certain threshold $\Gamma>0$, i.e., ${\text{P}_{\text{out}}}= \mathbb{P}({\text{SINR}}<\Gamma)$. The outage probability for a typical UE in cellular mode is given in the following theorem.
*Theorem 1:* In a single-tier D2D-communication-enabled mmWave cellular network, the outage probability for a typical cellular UE can be expressed as $$\begin{aligned}
&{\text{P}_{\text{out}}}^c(\Gamma)=1-\sum_{s \in \{L,N\}} \int_{0}^{\infty} e^{-\frac{ \Gamma r_c^{\alpha_{s,c}}\sigma_N^2}{P_{c}G_0}} \mathcal{L}_{I_{cc}}\left(\frac{\Gamma r_c^{\alpha_{s,c}}}{P_{c}G_0}\right) \mathcal{L}_{I_{dc}}\left(\frac{\beta \Gamma r_c^{\alpha_{s,c}}}{P_{c}G_0}\right) f_{s}(r_c) \mathcal{B}_{s,c} dr_c \label{eq:Pout_c}\end{aligned}$$ where the Laplace transforms $\mathcal{L}_{I_{cc}}(v)$ and $\mathcal{L}_{I_{dc}}(\beta v)$ are given in Lemma 2 and Lemma 3, respectively, and $f_{s}(r_c)$ is the pdf of the cellular link distance given in (\[cellularlinkpdftonearestBS\]).
*Proof:* See Appendix \[Proof of Theorem 1\].
*Theorem 2:* In a single-tier D2D-communication-enabled mmWave cellular network, the outage probability for a typical D2D UE can be expressed as $$\begin{aligned}
{\text{P}_{\text{out}}}^d(\Gamma)=1-&\sum_{s \in \{L,N\}} \int_{0}^{\infty} e^{-\frac{\Gamma r_{d0}^{\alpha_{s,d}}\sigma_N^2}{P_{d}G_0}} \mathcal{L}_{I_{dd_{\text{intra}}}}\left(\frac{ \Gamma r_{d0}^{\alpha_{s,d}}}{P_{d}G_0}|w_0\right) \mathcal{L}_{I_{dd_{\text{inter}}}}\left(\frac{ \Gamma r_{d0}^{\alpha_{s,d}}}{P_{d}G_0}\right) \mathcal{L}_{I_{cd}}\left(\frac{\beta \Gamma r_{d0}^{\alpha_{s,d}}}{P_{d}G_0}\right) \nonumber \\
&\times p_{s,d}(r_{d0}) f_{R_{d0}}(r_{d0}|w_0)f_{\Omega_0}(w_0) dr_{d0}dw_0 \label{eq:Pout_d}\end{aligned}$$ where the Laplace transforms $\mathcal{L}_{I_{dd_{\text{intra}}}}(v|w_0)$, $\mathcal{L}_{I_{dd_{\text{inter}}}}(v)$ and $\mathcal{L}_{I_{cd}}(\beta v)$ are given in Lemma 4, Lemma 5 and Lemma 6, respectively, $p_{s,d}(r_{d0})$ is the LOS/NLOS probability function for D2D link given in (\[LOS\_prob\_funct\]), $f_{R_{d0}}(r_{d0}|w_0)$ is the pdf of the D2D link distance given in (\[pdfofrd0\]), and $f_{\Omega_0}(\omega_0)=\frac{\omega_0}{\sigma_d^2}\exp(-\frac{\omega_0^2}{2\sigma_d^2})$.
*Proof:* Proof follows similar steps as in the proof of Theorem 1, and the details are omitted for the sake of brevity.
Uplink SINR Outage Probability Analysis In the Presence of Beamsteering Errors
------------------------------------------------------------------------------
In Section \[sec:Uplink SINR Outage\] and the preceding analysis, antenna arrays at the transmitting nodes (cellular or potential D2D UEs) and receiving nodes (BSs or UEs) are assumed to be aligned perfectly and uplink SINR outage probabilities are calculated in the absence of beamsteering errors. However, in practice, it may not be easy to have perfect alignment. Therefore, in this section, we investigate the effect of beamforming alignment errors on the outage probability analysis. We employ an error model similar to that in [@Wildman]. Let $|\epsilon|$ be the random absolute beamsteering error of the transmitting node toward the receiving node with zero-mean and bounded absolute error $|\epsilon|_{\text{max}} \le \pi$. Due to the symmetry in the gain $G_0$, it is appropriate to consider the absolute beamsteering error. The PDF of the effective antenna gain $G_0$ with alignment error can be explicitly written as [@Marco2] $$\begin{aligned}
&f_{G_0}({\text{g}})=F_{|\epsilon|}\left(\frac{\theta_{l}}{2}\right)F_{|\epsilon|}\left(\frac{\theta_{{\text{UE}}}}{2}\right)\delta({\text{g}}-M_{l}M_{{\text{UE}}})+F_{|\epsilon|}\left(\frac{\theta_{l}}{2}\right)\left(1-F_{|\epsilon|}\left(\frac{\theta_{{\text{UE}}}}{2}\right)\right)
\delta({\text{g}}-M_{l}m_{{\text{UE}}}) \nonumber \\
&+\left(1-F_{|\epsilon|}\left(\frac{\theta_{l}}{2}\right)\right)F_{|\epsilon|}\left(\frac{\theta_{{\text{UE}}}}{2}\right) \delta({\text{g}}-m_{l}M_{{\text{UE}}}) +\left(1-F_{|\epsilon|}\left(\frac{\theta_{l}}{2}\right)\right)\left(1-F_{|\epsilon|}\left(\frac{\theta_{{\text{UE}}}}{2}\right)\right) \delta({\text{g}}-m_{l}m_{{\text{UE}}}),
\label{eq:PDFofG}\end{aligned}$$ for $l \in \{{\text{BS}}, {\text{UE}}\}$ where $\delta(\cdot)$ is the Kronecker’s delta function, $F_{|\epsilon|}(x)$ is the CDF of the misalignment error and (\[eq:PDFofG\]) follows from the definition of CDF, i.e., $F_{|\epsilon|}(x)=\mathbb{P}\{|\epsilon|\le x\}$. Assume that the error $\epsilon$ is Gaussian distributed, and therefore the absolute error $|\epsilon|$ follows a half normal distribution with $F_{|\epsilon|}(x)=\text{erf}(x/(\sqrt{2}\sigma_{{\text{BE}}}))$, where $\text{erf}(\cdot)$ denotes the error function and $\sigma_{{\text{BE}}}$ is the standard deviation of the Gaussian error $\epsilon$.
It is clear that the uplink SINR outage probability expressions in Section \[sec:Uplink SINR Outage\] depend on the effective antenna gain $G_0$ between the transmitting and the receiving nodes. Thus, uplink SINR outage probability ${\text{P}_{\text{out}}}^{\kappa}(\Gamma)$ for a typical UE in mode $\kappa \in \{c,d\}$ can be calculated by averaging over the distribution of $G_0$, $f_{G_0}({\text{g}})$, as follows: $$\begin{aligned}
{\text{P}_{\text{out}}}^{\kappa}(\Gamma) &= \int_0^{\infty}{\text{P}_{\text{out}}}^{\kappa}(\Gamma;g)f_{G_0}({\text{g}})d {\text{g}}\nonumber \\
&=F_{|\epsilon|}(\theta_{l}/2)F_{|\epsilon|}(\theta_{{\text{UE}}}/2) {\text{P}_{\text{out}}}^{\kappa}(\Gamma;M_{l}M_{{\text{UE}}})+F_{|\epsilon|}(\theta_{l}/2) \bar{F}_{|\epsilon|}(\theta_{{\text{UE}}}/2) {\text{P}_{\text{out}}}^{\kappa}(\Gamma;M_{l}m_{{\text{UE}}}) \nonumber \\
&+ \bar{F}_{|\epsilon|}(\theta_{\l}/2)F_{|\epsilon|}(\theta_{{\text{UE}}}/2) {\text{P}_{\text{out}}}^{\kappa}(\Gamma;m_{l}M_{{\text{UE}}})+\bar{F}_{|\epsilon|}(\theta_{\l}/2) \bar{F}_{|\epsilon|}(\theta_{{\text{UE}}}/2) {\text{P}_{\text{out}}}^{\kappa}(\Gamma;m_{l}m_{{\text{UE}}}),\end{aligned}$$ for $l \in \{{\text{BS}}, {\text{UE}}\}$ where we define $\bar{F}_{|\epsilon|}(\theta/2)=1-F_{|\epsilon|}(\theta/2)$.
Analysis of Area Spectral Efficiency {#sec:Analysis of Area Spectral Efficiency}
====================================
In Section \[sec:Analysis of Uplink SINR Outage Probability\], we have analyzed the uplink outage probability and obtained outage probability expressions for a typical cellular and D2D link. In this section, we consider another performance metric, namely area spectral efficiency (ASE), to measure the network capacity. ASE is defined as the average number of bits transmitted per unit time per unit bandwidth per unit area. It can be mathematically defined as follows: $$\text{ASE}=\left(\lambda_B(1-{\text{P}_{\text{out}}}^c(\Gamma))+\bar{n}P_{D2D} \lambda_C (1-{\text{P}_{\text{out}}}^d(\Gamma))\right)\log_2(1+\Gamma) \label{eq:ASE_underlay}$$ where ${\text{P}_{\text{out}}}^c(\Gamma)$ and ${\text{P}_{\text{out}}}^d(\Gamma)$ are given in (\[eq:Pout\_c\]) and (\[eq:Pout\_d\]), respectively, $\bar{n}P_{D2D} \lambda_C$ and $\lambda_B$ are the average number of simultaneously active D2D links and cellular links per unit area, respectively. Note that ASE defined in (\[eq:ASE\_underlay\]) is valid for a saturated network scenario, i.e., each BS has at least one cellular UE to serve in the uplink channel. If the network is not saturated, the presence of inactive BSs will lead to increased SINR for both cellular and D2D links (due to lower interference), and outage probability will decrease. However, ASE may be lower as a result of fewer number of active cellular links per unit area. The ASE expression in (\[eq:ASE\_underlay\]) is given for underlay type of spectrum sharing. For overlay type of sharing, the uplink spectrum is divided into two orthogonal portions. Therefore, ASE can be redefined as follows: $$\text{ASE}=\left((1-\delta)\lambda_B(1-{\text{P}_{\text{out}}}^c(\Gamma))+\delta \bar{n}P_{D2D} \lambda_C (1-{\text{P}_{\text{out}}}^d(\Gamma))\right)\log_2(1+\Gamma)$$ where $\delta$ is the spectrum partition factor. In the case of overlay spectrum sharing, the following optimization problem can be formulated in order to determine the optimal value of $\delta$ that maximizes the ASE: $$\delta^{\ast}=\arg \underset{\delta \in [0,1]}{\max} \; \text{ASE}$$ The solution of this optimization problem is given as follows: if $\lambda_B(1-{\text{P}_{\text{out}}}^c(\Gamma)) > \bar{n}P_{D2D} \lambda_C (1-{\text{P}_{\text{out}}}^d(\Gamma))$, $\delta^{\ast}=0$; otherwise, $\delta^{\ast}=1$. In other words, all bandwidth is assigned to the cellular or D2D link depending on which one is performing better. Therefore, this is a greedy approach that does not address any fairness considerations. In numerical results, we have shown that if cellular communication is disabled, i.e. $\delta=1$, ASE is maximized. To overcome this unfairness in the spectrum allocation between D2D and cellular communication, we consider the optimal weighted proportional fair spectrum partition which is formulated as follows: $$\delta^{\ast}=\arg \underset{\delta \in [0,1]}{\max} \; w_c\log\big((1-\delta)\lambda_B(1-{\text{P}_{\text{out}}}^c(\Gamma))\log_2(1+\Gamma)\big)+w_d\log\left(\delta \bar{n}P_{D2D} \lambda_C (1-{\text{P}_{\text{out}}}^d(\Gamma))\log_2(1+\Gamma)\right) \label{eq:objection_func}$$ where $w_c$ and $w_d$ are the introduced weights. If we take the derivative of the objective function in (\[eq:objection\_func\]) with respect to $\delta$ and make it equal to zero, the optimal spectrum partition factor is obtained as $\delta^{\ast}=\frac{w_d}{w_c+w_d}=w_d$ which is simply equal to the weight we assign to the potential D2D UEs. In other words, $w_d$ portion of the spectrum should be assigned to D2D communication to achieve proportional fairness.
Simulation and Numerical Results {#sec:Simulation and Numerical Results}
================================
In this section, theoretical expressions are evaluated numerically. We also provide simulation results to validate the accuracy of the proposed model for the D2D-enabled uplink mmWave cellular network with clustered UEs as well as to confirm the accuracy of the analytical characterizations. In the numerical evaluations and simulations, unless stated otherwise, the parameter values listed in Table \[Table\] are used.
**Parameters** **Values**
--------------------------------------------------------------------------------- --------------------------------
$\alpha_{L,c}$, $\alpha_{N,c}$; $\alpha_{L,d}$, $\alpha_{N,d}$ 2, 4; 2, 4
$M_{\nu}$, $m_{\nu}$, $\theta_{\nu}$ for $\nu \in \{{\text{BS}}, {\text{UE}}\}$ 20dB, -10dB, $30^o$
$\lambda_B$, $\lambda_C$ $10^{-5}$, $10^{-4}$ $(1/m^2)$
$N$, $\bar{n}$ 40, 3
$(p_{L,c}, R_{B,c})$, $(p_{L,d}, R_{B,d})$ (1, 100), (1, 50)
$\beta$, $\delta$, $T_d$ 1, 0.2, 1
$\Gamma$, $\sigma_N^2$, $\sigma_d^2$ 0dB, -74dBm, 25
$P_c$, $P_d$ 200mW, 200mW
: System Parameters
\[Table\]
First, we investigate the effect of UE distribution’s standard deviation $\sigma_d$ on the probability of selecting D2D mode for different values of the LOS probability function $p_{L,c}$ for cellular link and $p_{L,d}$ for D2D link in Fig. \[Fig\_PD2D\]. As the standard deviation increases, the distance between the transmitting and receiving potential D2D UEs also increases. As a result, probability of selecting the D2D mode decreases. Also, since the number of LOS BSs increases with the increase in $p_{L,c}$, probability of selecting D2D mode decreases with increasing $p_{L,c}$. On the other hand, probability of selecting D2D mode increases when we increase $p_{L,d}$ as a result of increasing number of LOS potential D2D UEs in the cluster. As we have discussed in Section \[sec:cellular link distance modeling\], although the cluster centers are distributed according to a PPP, modeling the pdf of the distance between the nearest LOS/NLOS BS and potential D2D UE using Eq. (\[cellularlinkpdftonearestBS\]) is only an approximation because the potential D2D UEs are distributed according to a PCP around the cluster centers. However, as shown in Fig. \[Fig\_PD2D\], this pdf assumption agrees well with the simulation results especially for small values of $\sigma_d$. On the other hand, there is a minor deviation between the analysis and simulation results for larger values of $\sigma_d$. This is because potential D2D UEs are located farther from the cluster center for larger $\sigma_d$.
![Probability of selecting D2D mode as a function of UE distribution’s standard deviation $\sigma_d$ for different values of $p_{L,c}$ and $p_{L,d}$. Simulation results are also plotted with markers. []{data-label="Fig_PD2D"}](PD2D.eps){width="\figsize\textwidth"}
In Fig. \[Fig\_Navg\_lambdaC\], we plot the SINR outage probability of cellular and D2D links as a function of average number of simultaneously transmitting potential D2D UEs $\bar{n}$ in each cluster for different values of cluster center density $\lambda_C$. Moreover, the effect of spectrum sharing type is investigated. As described in Section \[sec:system\_model\], $\beta$ indicates the type of spectrum sharing; i.e., it is equal to one for underlay and zero for the overlay scheme. For the underlay type of spectrum sharing, when the average number of simultaneously transmitting potential D2D UEs gets higher, both intra-cluster and inter-cluster interferences increase and as a result SINR outage probabilities for both cellular and D2D links increase. Similarly, inter-cluster interference increases with the increase in cluster center density. Therefore, outage probabilities for both cellular and D2D links increase. For the overlay type of spectrum sharing, outage probability is smaller for cellular UEs compared to underlay and it is independent of $\bar{n}$ since cross-mode interference from D2D UEs becomes zero in the case of overlay spectrum sharing. On the other hand, outage probability of D2D UEs remains the same with both overlay and underlay sharing, showing that the effect of cross-mode interference from cellular UEs is negligible even under the congested network scenario assumption.
![SINR outage probability as a function of average number of simultaneously active D2D links $\bar{n}$ for different values of cluster center density $\lambda_C$ ($\Gamma=40dB$). Simulation results are also plotted with markers. []{data-label="Fig_Navg_lambdaC"}](Navg_lambdaC.eps){width="\figsize\textwidth"}
In Fig. \[Fig\_standard\_deviation\_RBd\], we investigate the effect of UE distribution’s standard deviation $\sigma_d$ on SINR outage probability of D2D links for different values of LOS the ball radius $R_{B,d}$. We have two different observations depending on the value of $\sigma_d$. For small values of $\sigma_d$, i.e. when the potential D2D UEs in the cluster are distributed closer to each other, outage probability is less for small LOS ball radius $R_{B,d}$. On the other hand, outage probability with smaller LOS ball radius $R_{B,d}$ becomes greater for bigger values of $\sigma_d$. For small $\sigma_d$, main link is more likely be a LOS link and effect of interference is small if the LOS ball radius is small, hence outage probability is low. However, the main link becomes more likely to be a NLOS link and the effect of interference becomes relatively more dominant with the increasing $\sigma_d$.
![SINR outage probability as a function of UE distribution’s standard deviation $\sigma_d$ for different values of LOS ball radius $R_{B,d}$ ($\Gamma=20dB$). []{data-label="Fig_standard_deviation_RBd"}](standard_deviation_RBd.eps){width="\figsize\textwidth"}
Next, we compare the SINR outage probabilities for different values of the antenna main lobe gain $M_{\nu}$ and beam width of the main lobe $\theta_{\nu}$ for $\nu \in \{{\text{BS}}, {\text{UE}}\}$ in Fig. \[Fig\_antenna\_pattern\]. Outage probability improves with the increase in the main lobe gain $M_{\nu}$ for the same value of $\theta_{\nu}$ for $\nu \in \{{\text{BS}}, {\text{UE}}\}$. On the other hand, since we assume perfect beam alignment for serving links, outage probability increases with the increase in the beam width of the main lobe due to the growing impact of the interference. Finally, we notice that for given SINR threshold, the outage probabilities for D2D links are smaller than those for cellular links, owing to generally smaller communication distances in D2D links.
![SINR outage probability as a function of the threshold in dB for different antenna parameters. Simulation results are also plotted with markers. []{data-label="Fig_antenna_pattern"}](antenna_pattern.eps){width="\figsize\textwidth"}
Effect of beam steering errors between the transmitting nodes (cellular or potential D2D UEs) and receiving nodes (BSs or UEs) on the SINR outage probability of cellular and D2D links is shown in Fig. \[Fig\_BSE\]. As shown in the figure, outage probability becomes worse for both cellular and D2D links with the increase in the standard deviation of the alignment error. Although the interference from interfering nodes remains unchanged, its effect grows with the increase in alignment error on the main link. This proves the importance of having perfect beam alignment to achieve improved performance.
![SINR outage probability as a function of the threshold in dB for different alignment errors $\sigma_{BE}$. Simulation results are also plotted with markers. []{data-label="Fig_BSE"}](BSE.eps){width="\figsize\textwidth"}
In the numerical analysis, we also investigate the ASE for underlay type of sharing. In Fig. \[Fig\_ASE\_underlay\], we plot the ASE as a function of the average number of simultaneously active D2D links $\bar{n}$ in each cluster for different values of the cluster center density $\lambda_C$. With the increase in the average number of simultaneously active D2D links, ASE first increases and then decreases. Therefore, an optimal value that maximizes ASE exists. Below this optimal value, increasing the average number of simultaneously active D2D links helps in improving spatial frequency reuse. Once the optimal value is exceeded, however, the effect of intra-cluster interference offsets the benefit of having larger average number of simultaneously active D2D links. Moreover, increasing cluster center density for the same average number of simultaneously active D2D links in each cluster improves ASE. Although the inter-cluster interference increases with larger cluster center density, spatial frequency reuse improves with larger cluster center density, i.e. $\bar{n}\lambda_C$ increases. Since inter-cluster interference does not have a dominant impact on outage probability and intra-cluster interference remains the same, ASE increases for the same average number of simultaneously active D2D links. Interestingly, the optimal number of simultaneously active D2D links is independent of the cluster center density because intra-cluster interference is more dominant than inter-cluster interference.
![Area Spectral Efficiency (ASE) for underlay type of sharing as a function of average number of simultaneously active D2D links $\bar{n}$ for different values of cluster center density $\lambda_C$ ($\Gamma=40dB$). []{data-label="Fig_ASE_underlay"}](ASE_underlay.eps){width="\figsize\textwidth"}
Finally, ASE is investigated for overlay type of spectrum sharing. In Fig. \[Fig\_ASE\_overlay\], ASE is plotted as a function of the average number of simultaneously active D2D links in each cluster $\bar{n}$ for different values of spectrum partition factor $\delta$. As mentioned in Section \[sec:Analysis of Area Spectral Efficiency\], ASE is maximized if all bandwidth is assigned to D2D links, i.e. $\delta=1$. We also consider the optimal weighted proportional fair spectrum partition in Section \[sec:Analysis of Area Spectral Efficiency\], and plot the objective function in (\[eq:objection\_func\]) as a function of $\delta$. As shown in Fig. \[Fig\_obj\_function\], optimal spectrum partition factor is equal to $\delta^{\ast}=0.4=w_d$ which validates our result in Section \[sec:Analysis of Area Spectral Efficiency\].
![Area Spectral Efficiency (ASE) for overlay type of sharing as a function of average number of simultaneously active D2D links $\bar{n}$ for different values of spectrum partition factor $\delta$ ($\Gamma=40dB$). []{data-label="Fig_ASE_overlay"}](ASE_overlay.eps){width="\figsize\textwidth"}
![Objective function as a function of spectrum partition factor $\delta$ ($\Gamma=40dB$). []{data-label="Fig_obj_function"}](obj_function.eps){width="\figsize\textwidth"}
Conclusion {#sec:Conclusion}
==========
In this paper, we have provided an analytical framework to compute the SINR outage probabilities for both cellular and D2D links in a D2D-enabled mmWave cellular network with clustered UEs. Distinguishing features of mmwave communications, such as directional beamforming with sectored antenna model and modified LOS ball model for blockage modeling, have been considered in the analysis. BSs and cellular UEs are assumed to be distributed according to independent PPPs, while potential D2D UEs locations’ are modeled as a PCP. Potential D2D UEs in the clusters are allowed to choose cellular or D2D mode according to a flexible mode selection scheme. Under these assumptions, we have analyzed the interference experienced in cellular uplink and D2D links, and characterized the SINR outage probabilities.
Numerical results show that probability of selecting D2D mode decreases with increasing UE distribution’s standard deviation $\sigma_d$ and increasing $p_{L,c}$, while increase in $p_{L,d}$ leads to higher D2D mode selection probability. We have also shown that more simultaneously transmitting potential D2D UEs and/or higher cluster center density result in higher outage probabilities for both cellular and D2D links due to the growing impact of interference. Moreover, the type of spectrum sharing plays a crucial role in the SINR outage performance of cellular UEs. Another interesting observation is that smaller LOS ball radius is preferred for small values of $\sigma_d$ while the opposite is advantageous for large values of $\sigma_d$. Moreover, increasing the main lobe gain and decreasing the beam width of the main lobe result in lower SINR outage. Effect of alignment error on outage probability is also quantified and importance of beam alignment in improving the performance is noted. Finally, ASE of the whole network is analyzed for both underlay and overlay types of sharing. We have shown that there is an optimal number of simultaneously active D2D links maximizing ASE. This optimal number is independent of cluster center density and spectrum partition factor. For overlay, there exists an optimal spectrum partition factor if the optimal weighted proportional fair spectrum partition is considered. Analyzing the outage performance of cellular and D2D UEs for a different PCP such as uniformly distributed potential D2D UEs, and investigating the effect of using different mode selection schemes remains as future work.
Proof of Lemma 1
----------------
Probability of selecting the D2D mode for a potential D2D UE located in a cluster $x \in \Phi_C$ can be computed as \[Proof of Lemma 1\] $$\begin{aligned}
P_{D2D}&=\sum_{s \in \{L,N\}} \sum_{s^{\prime} \in \{L,N\}} \mathbb{P}\left (T_d r_d^{-\alpha_{s^{\prime},d}} \geq r_c^{-\alpha_{s,c}}\right) p_{s^{\prime},d}(r_d) \mathcal{B}_{s,c} \nonumber \\
&= \sum_{s \in \{L,N\}} \sum_{s^{\prime} \in \{L,N\}}\mathbb{P}\left (r_c \geq r_d^{\alpha_{s^{\prime} ,d} /\alpha_{s,c}}T_d^{-1/\alpha_{s,c}}\right ) p_{s^{\prime},d}(r_d) \mathcal{B}_{s,c} \nonumber \\
&=\sum_{s \in \{L,N\}} \sum_{s^{\prime} \in \{L,N\}} \int_0^{\infty} \int_0^{\infty} \bar{F}_s\left(\frac{r_d^{\alpha_{s^{\prime},d}/\alpha_{s,c}}}{T_d^{1/\alpha_{s,c}}}\right) f_{R_d}(r_d|\omega) f_{\Omega}(\omega) p_{s^{\prime},d}(r_d)\mathcal{B}_{s,c} dr_d d\omega \nonumber \\
&\stackrel{(a)}{=}\sum_{s \in \{L,N\}} \sum_{s^{\prime} \in \{L,N\}} \int_0^{\infty} \int_0^{\infty} e^{-2\pi\lambda_B \psi_s\left(r_d^{\alpha_{s^{\prime},d}/\alpha_{s,c}}/T_d^{1/\alpha_{s,c}}\right)} f_{R_d}(r_d|\omega) f_{\Omega}(\omega) p_{s^{\prime},d}(r_d) dr_d d\omega\end{aligned}$$ where $\bar{F}_s(r_c)=e^{-2\pi\lambda_B\psi_s(r_c)}/\mathcal{B}_{s,c}$ is the complementary cumulative distribution function (ccdf) of the cellular link distance $r_c$ to the nearest LOS/NLOS BS, $p_{s^{\prime},d}(r_d)$ is the LOS/NLOS probability function for the D2D link given in (\[LOS\_prob\_funct\]), and (a) follows by substituting the cdf of $r_c$ into the expression.
Proof of Lemma 2 {#Proof of Lemma 2}
----------------
Laplace transform of the aggregate interference at the BS from cellular UEs transmitting in the same uplink channel in different cells can be calculated using (\[eq:LT\]) as follows: $$\begin{aligned}
\mathcal{L}_{I_{cc}}(v)= \prod_{G} \prod_j \mathcal{L}_{I_{cc,j}^G}(v), \label{eq:LT_cc6}\end{aligned}$$ where the Laplace transform for $I_{cc,j}^G$ can be computed using stochastic geometry as follows: $$\begin{aligned}
\mathcal{L}_{I_{cc,j}^G}(v)&\stackrel{(a)}{=} \exp\left(-2\pi\lambda_{B}p_{G} \int_{0}^{\infty}\left(1-\mathbb{E}_h \left[ e^{-v P_{c}G h t^{-\alpha_{j,c}} }\right]\right)Q(t^{\alpha_{j,c}}) p_{j,c}(t) t dt\right) \nonumber \\
&\stackrel{(b)}{=} \exp\left(-2\pi\lambda_B p_{G} \int_{0}^{\infty} \frac{v P_cG t^{-\alpha_{j,c}}}{1+v P_cG t^{-\alpha_{j,c}}}Q(t^{\alpha_{j,c}})p_{j,c}(t)tdt\right), \label{eq:LT_cc_j}\end{aligned}$$ where (a) follows from computing the probability generating functional (PGFL) of PPP, the lower limit of integration is determined using the fact that the minimum distance of interfering UEs in cellular mode to the typical BS is equal to $r_c$, (b) is obtained by computing the moment generating function (MGF) of the exponentially distributed random variable $h$, and $Q(y)$ is given in (\[Q(y)\]). By inserting (\[eq:LT\_cc\_j\]) into (\[eq:LT\_cc6\]) for $j \in \{L,N\}$ and $G \in \{M_{{\text{BS}}}M_{{\text{UE}}},M_{{\text{BS}}}m_{{\text{UE}}}, m_{{\text{BS}}}M_{{\text{UE}}},m_{{\text{BS}}}m_{{\text{UE}}}\}$, the Laplace transform expression in Lemma 2 can be obtained.
Proof of Lemma 3 {#Proof of Lemma 3}
----------------
Laplace transform of the aggregate interference at the BS from both intracell and intercell D2D UEs can be calculated using (\[eq:LT\]) $$\begin{aligned}
\mathcal{L}_{I_{dc}}(v)= \prod_G \prod_j \mathcal{L}_{I_{dc,j}^G}(v), \label{eq:LT_dc6}\end{aligned}$$ where the Laplace transform for $I_{dc,j}^G$ can be computed using stochastic geometry and following the similar steps as in [@Afshang]: $$\begin{aligned}
\label{eq:LT_dc_j}
&\mathcal{L}_{I_{dc,j}^G}(v) \nonumber \\
&\stackrel{(a)}{=}\mathbb{E}_{\Phi_C}\left[\prod_{x \in \Phi_C} \mathbb{E}_{\mathcal{A}_d^{x}} \left[\prod_{y \in \mathcal{A}_d^{x}} \mathbb{E}_{h_{y_x}}\left[ e^{-vP_dGh_{y_x}r_{y_x}^{-\alpha_{j,d}}} \right] \right] \right] \nonumber \\
&\stackrel{(b)}{=}\mathbb{E}_{\Phi_C}\left[\prod_{x \in \Phi_C} \mathbb{E}_{\mathcal{A}_d^{x}} \left[\prod_{y \in \mathcal{A}_d^{x}} \frac{1}{1+vP_dGr_{y_x}^{-\alpha_{j,d}}}\right] \right] \nonumber \\
&\stackrel{(c)}{=}\mathbb{E}_{\Phi_C}\left[\prod_{x \in \Phi_C} \sum_{k=0}^{N/2}\left(\int_{\mathbb{R}^2} \frac{1}{1+vP_dGr_{y_x}^{-\alpha_{j,d}}} p_{j,d}(r_{y_x})f_{Y}(y)dy \right)^k \mathbb{P}(K=k|K<N/2)\right] \nonumber \\
&\stackrel{(d)}{=}\exp\left( -\lambda_C \int_{\mathbb{R}^2}\left(1-\sum_{k=0}^{N/2}\left(\int_{\mathbb{R}^2} \frac{1}{1+vP_dGr_{y_x}^{-\alpha_{j,d}}} p_{j,d}(r_{y_x})f_{Y}(y)dy \right)^k \frac{(\bar{n}P_{D2D})^ke^{-(\bar{n}P_{D2D})}}{k!\eta} \right)dx \right) \nonumber \\
&\stackrel{(e)}{=} \exp \left( -2\pi\lambda_C \int_0^{\infty}\left(1 -\sum_{k=0}^{N/2} \left(\int_{0}^{\infty} \frac{1}{1+vP_dGu^{-\alpha_{j,d}}}p_{j,d}(u)f_U(u|t)du\right)^k \frac{(\bar{n}P_{D2D})^ke^{-(\bar{n}P_{D2D})}}{k!\eta} \right) tdt \right) \nonumber \\
&\stackrel{(f)}{=} \exp \left( -2\pi\lambda_C \int_0^{\infty}\left(1 -\exp\left(-\bar{n}P_{D2D} \int_{0}^{\infty} \frac{vP_dGu^{-\alpha_{j,d}}}{1+vP_dGu^{-\alpha_{j,d}}}p_{j,d}(u)f_U(u|t)du \right) \right) tdt \right)\end{aligned}$$ where $\{r_{y_x}=\|x+y\|, \forall x \in \Phi_C, \forall y \in \mathcal{A}_d^{x}\}$, (a) follows from the assumption of independent fading gains across all interfering links, (b) is obtained by computing the moment generating function (MGF) of the exponentially distributed random variable $h_{y_x}$, (c) follows from the fact that the locations of the cluster members in each cluster are independent when conditioned on $x \in \Phi_C$ and expectation over the number of interfering devices which are Poisson distributed conditioned on the total being less than $N/2$ where $\eta=\sum_{l=0}^{N/2}\frac{(\bar{n}P_{D2D})^le^{-(\bar{n}P_{D2D})}}{l!}$, (d) is determined by computing the probability generating functional (PGFL) of PPP, (e) follows by converting the coordinates from Cartesian to polar, (f) follows from the assumption that $\bar{n}P_{D2D} \ll N/2$. By inserting (\[eq:LT\_dc\_j\]) into (\[eq:LT\_dc6\]) for $j \in \{L,N\}$ and $G \in \{M_{{\text{BS}}}M_{{\text{UE}}},M_{{\text{BS}}}m_{{\text{UE}}}, m_{{\text{BS}}}M_{{\text{UE}}},m_{{\text{BS}}}m_{{\text{UE}}}\}$, we obtain the Laplace transform expression in Lemma 3.
Proof of Lemma 5 {#Proof of Lemma 5}
----------------
Laplace transform of the intra-cluster interference at the typical UE $\in \mathcal{N}_r^{x_0}$ in the representative cluster can be calculated using (\[eq:LT\]) as follows: $$\begin{aligned}
\mathcal{L}_{I_{dd_{\text{intra}}}}(v|w_0))= \prod_G \prod_j \mathcal{L}_{I_{dd_{\text{intra}},j}^G}(v), \label{eq:LT_ddintra6}\end{aligned}$$ where the Laplace transform for $I_{dd_{\text{intra}},j}^G$ conditioned on $w_0$ can be computed following similar steps as in the proof of Lemma 3: $$\begin{aligned}
\label{eq:LT_ddintra_j}
\mathcal{L}_{I_{dd_{\text{intra}},j}^G}(v) &= \mathbb{E}_{\mathcal{A}_d^{x_0}} \left[\prod_{y \in \mathcal{A}_d^{x_0} \setminus y_0} \mathbb{E}_{h_{y_{x_0}}}\left[ e^{-vP_dGh_{y_{x_0}}r_{d1}^{-\alpha_{j,d}}} \right] \right] \nonumber \\
&\stackrel{(a)}{=} \mathbb{E}_{\mathcal{A}_d^{x_0}} \left[\prod_{y \in \mathcal{A}_d^{x_0} \setminus y_0} \frac{1}{1+vP_dGr_{d1}^{-\alpha_{j,d}}}\right] \nonumber \\
&\stackrel{(b)}{=} \sum_{k=0}^{N/2-1}\left(\int_{\mathbb{R}^2} \frac{1}{1+vP_dGr_{d1}^{-\alpha_{j,d}}} p_{j,d}(r_{d1})f_{Y}(y)dy \right)^k \mathbb{P}(K=k|K<N/2-1) \nonumber \\
&=\sum_{k=0}^{N/2-1}\left(\int_{\mathbb{R}^2} \frac{1}{1+vP_dGr_{d1}^{-\alpha_{j,d}}} p_{j,d}(r_{d1})f_{Y}(y)dy \right)^k \frac{(\bar{n}P_{D2D}-1)^ke^{-(\bar{n}P_{D2D}-1)}}{k!\eta} \nonumber \\
&\stackrel{(c)}{=} \exp \left(-\left(\bar{n}P_{D2D}-1\right) \int_0^{\infty} \frac{vP_d G u^{-\alpha_{j,d}}}{1+vP_dG u^{-\alpha_{j,d}}}f_{U}(u|w_0)p_{j,d}(u)du \right)\end{aligned}$$ where $\{r_{d1}=\|x_0+y\|, \forall y \in \mathcal{A}_d^{x_0} \setminus y_0\}$, $\eta=\sum_{l=0}^{N/2-1}\frac{(\bar{n}P_{D2D}-1)^le^{-(\bar{n}P_{D2D}-1)}}{l!}$, (a) is obtained by computing the moment generating function (MGF) of the exponentially distributed random variable $h_{y_{x_0}}$, (b) follows from the fact that the locations of the intra-cluster D2D UEs simultaneously transmitting in D2D mode are independent when conditioned on $x_0 \in \Phi_C$ and expectation over the number of interfering devices which are Poisson distributed conditioned on the total being less than $N/2-1$, (c) follows by converting the coordinates from Cartesian to polar and using the assumption that $\bar{n}P_{D2D} \ll N/2$. By inserting (\[eq:LT\_ddintra\_j\]) into (\[eq:LT\_ddintra6\]) for $j \in \{L,N\}$ and $G \in \{M_{{\text{UE}}}M_{{\text{UE}}},M_{{\text{UE}}}m_{{\text{UE}}}, m_{{\text{UE}}}M_{{\text{UE}}},m_{{\text{UE}}}m_{{\text{UE}}}\}$, we readily obtain the Laplace transform expression in Lemma 4.
Proof of Theorem 1
------------------
The outage probability for a typical UE in cellular mode can be calculated as follows: \[Proof of Theorem 1\] $$\begin{aligned}
&{\text{P}_{\text{out}}}^c(\Gamma)= {\text{P}_{\text{out},L}}^c(\Gamma)\mathcal{B}_{L,c}+{\text{P}_{\text{out},N}}^c(\Gamma)\mathcal{B}_{N,c} \nonumber \\
&{\text{P}_{\text{out}}}^c(\Gamma)=\sum_{s \in \{L,N\}}\mathbb{P}\left( \frac{P_c G_0h_0r_c^{-\alpha_{s,c}}}{\sigma^2+I_{cc}+I_{dc}} \leq \Gamma \right) \mathcal{B}_{s,c} \nonumber \\
&=\sum_{s \in \{L,N\}} \int_0^{\infty}\hspace{-0.2cm} \mathbb{P}\left (h_0 \leq \frac{\Gamma r_c^{\alpha_{s,c}}}{P_cG_0}\left(\sigma^2+I_{cc}+I_{dc}\right) |r_c \right)f_{s}(r_c) \mathcal{B}_{s,c} dr_c \nonumber \\
&= 1- \sum_{s \in \{L,N\}} \int_0^{\infty} e^{-v\sigma^2} \mathcal{L}_{I_{cc}}(v) \mathcal{L}_{I_{dc}}(\beta v) f_{s}(r_c) \mathcal{B}_{s,c} dr_c \label{OutageProbability}\end{aligned}$$ where $v=\frac{\Gamma r_c^{\alpha_{s,c}}}{P_{c}G_0}$, and (\[OutageProbability\]) follows from $h_0$ $\sim$ $\exp(1)$, and by noting that Laplace transforms of interference at the BS from cellular UEs and D2D UEs are independent.
[1.5]{}
[99]{} T. Rappaport, S. Sun, R. Mayzus, H. Zhao, Y. Azar, K. Wang, G. N. Wong, J. K. Schulz, M. Samimi, and F. Gutierrez, “Millimeter wave mobile communications for 5G cellular: It will work!," *IEEE Access*, vol. 1, pp. 335-349, May 2013.
J. G. Andrews, S. Buzzi, W. Choi, S. V. Hanly, A. Lozano, A. C. K. Soong, and J. C. Zhang, “What will 5G be?," *IEEE J. Sel. Areas Commun.*, vol. 32, no. 6, pp. 1065-1082, Jun. 2014.
S. Rangan, T.S. Rappaport, and E. Erkip, “Millimeter-wave cellular wireless networks: Potentials and challenges," *Proc. of the IEEE*, vol. 102, no. 3, pp. 366-385, Mar. 2014.
A. Ghosh, T. A. Thomas, M. C. Cudak, R. Ratasuk, P. Moorut, F. W. Vook, T. S. Rappaport, G. R. MacCartney, S. Sun, and S. Nie, “Millimeter wave enhanced local area systems: A high data rate approach for future wireless networks," *IEEE Journal on Sel. Areas in Comm.*, Special Issue on 5G, Jul., 2014.
H. ElSawy and E. Hossain, “Analytical modeling of mode selection and power control for underlay D2D communication in cellular networks," *IEEE Trans. Commun.*, vol. 62, no. 11, pp. 4147-4161, Nov. 2014.
X. Lin, J. G. Andrews, and A. Ghosh, “Spectrum sharing for device-to-device communication in cellular networks," *IEEE Trans. Wireless Commun.*, vol. 13, no. 12, pp. 6727-6740, Dec. 2014.
J. G. Andrews, A. K. Gupta, and H. S. Dhillon, “A primer on cellular network analysis using stochastic geometr," 2016. Available at: https://arxiv.org/abs/1604.03183.
M. Afshang, H. S. Dhillon, and P. H. J. Chong, “Modeling and performance analysis of clustered device-to-device networks," *IEEE Trans. Wireless Commun.*, vol. 15, no. 7, pp. 4957-4972, Jul. 2016.
M. Afshang, H. S. Dhillon, and P. H. J. Chong, “Fundamentals of Cluster-Centric Content Placement in Cache-Enabled Device-to-Device Networks," *IEEE Trans. Commun.*, vol. 64, no. 6, pp. 2511-2526, Jun. 2016.
M. Afshang and H. S. Dhillon, “Spatial modeling of device-to-device networks: Poisson cluster process meets Poisson hole process," *Proc. Annual Asilomar Conf. on Signals, Syst., Comp.*, pp. 317–321, Nov. 2015.
J. Qiao, X. Shen, J. Mark, Q. Shen, Y. He, and L. Lei, “Enabling deviceto-device communications in millimeter-wave (5G) cellular networks," *IEEE Commun. Mag.*, vol. 53, no. 1, pp. 209-215, Jan. 2015.
Z. Guizani and N. Hamdi, “mmWave E-Band D2D communications for 5G-Underlay networks: Effect of power allocation on D2D and cellular users throughputs," *IEEE Symposium on Computers and Communication*, June 2015.
N. Eshraghi, B. Maham, and V. Shah-Mansouri, “Millimeter-wave device-to-device multi-hop routing for multimedia applications," *IEEE International Conference on Communications (ICC)*, May 2016.
E. Turgut and M. C. Gursoy,“Uplink performance analysis in D2D-enabled mmwave cellular networks," *Proc. of the IEEE Vehicular Technology Conference (VTC)-Fall*, Sep. 2017.
M. Haenggi, *Stochastic Geometry for Wireless Networks*. Cambridge, U.K.: Cambridge Univ. Press, 2012.
T. Bai and R. Vaze and R. W. Heath Jr., “Analysis of blockage effects on urban cellular networks," *IEEE Trans. Wireless Commun.*, vol. 13, no. 9, pp. 5070-5083, Sept. 2014.
S. Singh, M. Kulkarni, A. Ghosh, and J. Andrews, “Tractable model for rate in self-backhauled millimeter wave cellular networks," *IEEE J. Sel. Areas Commun.*, vol. 33, no. 10, pp. 2196-2211, Oct. 2015.
T. Bai, and R. W. Heath, “Coverage and rate analysis for millimeter wave cellular networks," *IEEE Trans. Wireless Commun.*, vol. 14, no. 2, pp. 1100-1114, Feb. 2015.
A. Thornburg, T. Bai, and R. W. Heath, “Performance analysis of mmWave ad hoc networks," *IEEE Trans. Signal Process.*, *IEEE Trans. Signal Processing.*, vol. 64, no. 15, pp. 4065-4079. Aug. 2016.
J. Wildman, P. H. J. Nardelli, M. Latva-aho, and Steven Weber, “On the joint impact of beamwidth and orientation error on throughput in directional wireless Poisson networks," *IEEE Trans. Wireless Commun.*, vol. 13, no. 12, pp. 7072-7085, Dec. 2014.
M. Di Renzo, “Stochastic geometry modeling and analysis of multi-tier millimeter wave cellular networks," *IEEE Trans. Commun.*, vol. 14, no. 9, pp. 5038-5057, Sep. 2015.
A. Hunter, J. Andrews, and S. Weber, “Transmission capacity of ad hoc networks with spatial diversity," *IEEE Trans. Wireless Commun.*, vol. 7, no. 12, pp. 5058-5071, Dec. 2008.
J. G. Andrews, T. Bai, M. N. Kulkarni, A. Alkhateeb, A. K. Gupta, and R. W. Heath, “Modeling and analyzing millimeter wave cellular systems," *IEEE Trans. on Commun.*, vol. 65, no. 1, pp. 403-430, Jan 2017.
[^1]: The authors are with the Department of Electrical Engineering and Computer Science, Syracuse University, Syracuse, NY, 13244 (e-mail: [email protected], [email protected]).
[^2]: Throughout the paper, subscripts $c$ and $d$ denote associations with cellular and D2D links, respectively.
[^3]: Subsequently, beamsteering errors are also addressed.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We extend the notion of Rokhlin dimension from topological dynamical systems to $C^*$-correspondences. We show that in the presence of finite Rokhlin dimension and a mild quasidiagonal-like condition (which, for example, is automatic for finitely generated projective correspondences), finite nuclear dimension passes from the scalar algebra to the associated Toeplitz–Pimsner and (hence) Cuntz–Pimsner algebras. As a consequence we provide new examples of classifiable $C^*$-algebras: if $A$ is simple, unital, has finite nuclear dimension and satisfies the UCT, then for every finitely generated projective $\mathcal{H}$ with finite Rokhlin dimension, the associated Cuntz–Pimsner algebra $\mathcal{O} (\mathcal{H})$ is classifiable in the sense of Elliott’s Program.'
address:
- 'N.B. and A.Z. can be reached at Department of Mathematics, Penn State University, State College, PA 16802.'
- 'A.T. can be reached at Department of Mathematical Sciences, Fraser Noble Building, Aberdeen, AB24 3UE. '
author:
- 'Nathanial P. Brown, Aaron Tikuisis, and Aleksey M. Zelenberg'
title: 'Rokhlin dimension for $C^*$-correspondences'
---
[^1]
Introduction
============
The topological notion of covering dimension was extended to the noncommutative context by Winter and Zacharias in [@Winter-Zacharias1]. Their *nuclear dimension* has contributed to the most important and broadly applicable advances in the theory of nuclear $C^*$-algebras in the last 40 years. For example, the Toms–Winter conjecture asserts that finite nuclear dimension is often equivalent to structural properties analogous to those exploited by Connes in his proof of uniqueness of the injective II$_1$-factor ([@Connes]). In a remarkable breakthrough, this audacious conjecture was confirmed in the unique-trace case in [@Sato-White-Winter], and has now been confirmed for much broader classes (see [@Brown-Bosa-Sato-Tikuisis-White-Winter]). As another stunning example, in 2015 finite nuclear dimension led to the completion of Elliott’s Classification Program for the cases of most interest (cf. [@Elliott-Gong-Lin-Niu], [@Tikuisis-White-Winter]). In short, nuclear dimension has revolutionized the field.
It is thus important to know which examples have finite nuclear dimension. In the influential paper [@Matui-Sato] it was shown that all Kirchberg algebras have nuclear dimension at most three. (In fact, they have dimension one; see [@Brown-Bosa-Sato-Tikuisis-White-Winter] and [@Ruiz-Sims-Sorensen].) In [@Szabo] Szabo proved that if $\mathbb{Z}^n$ acts freely on a compact metric space $X$ of finite covering dimension, then the associated crossed product $C(X) \rtimes \mathbb{Z}^n$ has finite nuclear dimension. His work was inspired by and depended upon [@Hirshberg-Winter-Zacharias], where the classical measure-theoretic Rokhlin property was exported to the realm of topology. This so-called *Rokhlin dimension* makes sense in the noncommutative context too, so we can study it for $C^*$-dynamical systems. There is mounting evidence that Rokhlin dimension is for $C^*$-dynamical systems what nuclear dimension is for $C^*$-algebras: ubiquitous and fundamental.
For example, it is intimately connected with nuclear dimension via the crossed product construction. One of the main results of [@Hirshberg-Winter-Zacharias] was that $A\rtimes_{\alpha} \mathbb{Z}$ has finite nuclear dimension whenever $A$ does and the automorphism $\alpha$ has finite Rokhlin dimension. Since $A\rtimes_{\alpha} \mathbb{Z}$ can be realized as a Cuntz–Pimsner algebra, it is natural to seek an extension of Rokhlin dimension to the context of $C^*$-correspondences, then ask whether the associated Toeplitz–Pimnser and/or Cuntz–Pimsner algebras have finite nuclear dimension?
That is the subject of this paper. Indeed, we generalize Rokhlin dimension to $C^*$-correspondences in Definition \[Rokhlin-Dimension-Def\], then prove the following.
\[Main-Theorem\] Suppose that $\mathcal{H}$ is a countably generated $C^*$-correspondence over a separable unital $C^*$-algebra $A$, satisfying a technical quasidiagonal-like condition (see Theorem \[main\_thm\]). Then $$(\dim_{\mathrm{nuc}}(\mathcal{T}(\mathcal{H})) + 1) \le 2(\dim_{\mathrm{nuc}}(A)+1)(\dim_{\mathrm{Rok}}(\mathcal{H})+1).$$
The quasidiagonal-like condition is probably unnecessary, and would be if a certain algebra could be shown directly to have finite nuclear dimension (cf. Lemma \[quasi\_hypothesis\]). In any case, it is satisfied in many examples, including all finitely generated projective correspondences (see Example \[QD-examples\]).
As an application, we provide new examples of classifiable $C^*$-algebras in the sense of Elliott’s Program. Thanks to [@Elliott-Gong-Lin-Niu] and [@Tikuisis-White-Winter], this amounts to verifying simplicity, finite nuclear dimension and the Universal Coefficient Theorem (UCT) of [@Rosenberg-Schochet].
(Corollary \[application2\]) \[application\] Assume $A$ is simple, unital, satisfies the UCT and has finite nuclear dimension. For every finitely generated projective $\mathcal{H}$ with finite Rokhlin dimension, the associated Cuntz–Pimsner algebra $\mathcal{O}(\mathcal{H})$ is also simple, unital, satisfies the UCT and has finite nuclear dimension.
Note that we’ve substantially generalized the $C^*$-dynamical system case from [@Hirshberg-Winter-Zacharias Theorem 4.1] because $A\rtimes_{\alpha} \mathbb{Z}$ is the Cuntz–Pimsner algebra over a singly generated projective correspondence. Also, this corollary holds whenever Theorem \[Main-Theorem\] does, so it’s likely true for arbitrary correspondences of finite Rokhlin dimension (and definitely true when the quasidiagonal-like hypothesis is satisfied).
The proof of Theorem \[Main-Theorem\] follows [@Hirshberg-Winter-Zacharias] very closely, at least in spirit. The main technical innovation is finding suitable replacements for the outgoing and incoming maps used in the proof of Theorem 4.1 in [@Hirshberg-Winter-Zacharias]. But we also have to worry about the nuclear dimension of the range of the outgoing maps, an easy task in the crossed-product case. The majority of the paper is devoted to laying out these issues, and resolving them. Once that is done, the proof of finite nuclear dimension is very similar to [@Hirshberg-Winter-Zacharias].
Here is an outline of what follows. In Section \[Preliminaries\], we establish notation and give relevant background information on nuclearity, nuclear dimension, Hilbert $C^*$-modules, $C^*$-correspondences, and Cuntz–Pimsner algebras. Section \[Rokhlin dimension for correspondences\] is about the definition of Rokhlin dimension for $C^*$-correspondences. The heavy lifting is contained in Section \[Heavy lifting\], culminating with the proof of our main result (Theorem \[main\_thm\]). In Section \[Amalgamated\], we observe a cute application: in certain circumstances, reduced amalgamated free products have finite nuclear dimension. Finally, in Section \[classifiability\], we show how to use work of Schweizer ([@Schweizer]) to deduce Corollary \[application\] from Theorem \[Main-Theorem\].
Preliminaries {#Preliminaries}
=============
Notation and terminology
------------------------
Throughout all that follows, we will use the following conventions, assumptions, and notation. When we refer to an *ideal* of a $C^*$-algebra, we mean a closed, two-sided ideal. If $A$ is a $C^*$-algebra, let $A_+$, $\text{Ball}_1(A)$, $\mathcal{M}(A)$, and $\mathcal{Z}(A)$ denote the positive cone, unit ball, multiplier algebra, and center, respectively, of $A$. If $x,y\in A$ are self-adjoint, let $x\approx_{\epsilon}y$, $x\perp y$, and $x\perp_{\epsilon}y$ mean $\lVert x - y\rVert < \epsilon$, $xy = 0$, and $xy \approx_{\epsilon} 0$, respectively. If $C,D\subseteq A$, let $C'$ denote the commutant of $C$ in $A$ and let $C\subset_{\scriptscriptstyle{\epsilon}}D$ mean that for every contraction $c\in C$, there is a $d\in D$ satisfying $c\approx_{\epsilon} d$. If $\varphi:A \rightarrow B$ is a linear map to a $C^*$-algebra $B$, we will say that $\varphi$ is $\epsilon$-contractive if $\lVert \varphi\rVert < 1 + \epsilon$. We will use the abbreviation c.p.(c.) to mean completely positive (and contractive). We write $\mathbb{K}$ for the compact operators on $\ell^2(\mathbb{N})$. For a $C^*$-algebra $A$, we define $$A_\infty = \ell^2(\mathbb N,A) / c_0(\mathbb N,A),$$ and view $A$ as a subalgebra of $A_\infty$ in the canonical way (consisting of elements represented by constant sequences).
Order zero maps and nuclear dimension
-------------------------------------
Throughout this section let $A$ and $K$ be $C^*$-algebras.
\[Order Zero Maps\] A c.p. map $\psi:K\rightarrow A$ is *order zero* if it preserves orthogonality: for every pair of positive elements $x_1,x_2\in K$, $$x_1\perp x_2 \Rightarrow \psi(x_1) \perp \psi(x_1).$$
$A$ has *nuclear dimension* at most $n$, written $\dim_{\mathrm{nuc}}(A) \le n$, if for every finite subset $F\subset A$ and $\epsilon > 0$, there is a finite-dimensional $C^*$-algebra $K = K^{(0)}\oplus\cdots\oplus K^{(n)}$, a c.p.c. map $\varphi: A\rightarrow K$, and a c.p. map $\psi: K \rightarrow A$ satisfying
1. $\lVert \psi\circ\varphi(a) - a\rVert < \epsilon$ for every $a\in F$, and
2. for each $i = 0,\ldots,n$, the restriction of $\psi$ to $K^{(i)}$ is contractive and order zero.
The following is well-known and underpins many nuclear dimension computations in the literature.
\[nucdimtool\] Fix $m,n \in \mathbb N$. Let $A$ be a $C^*$-algebra. Then $\dim_{\mathrm{nuc}}(A)\leq m(n+1)-1$ if, for every finite set $F \subset A$ and $\epsilon>0$, there exists a $C^*$-algebra $B$ of nuclear dimension at most $n$ and c.p. maps $$A \stackrel{\scriptstyle{\psi}}\longrightarrow B \stackrel{\scriptstyle{\phi}}\longrightarrow A_\infty$$ such that $\psi$ is c.p.c., $\phi$ is a sum of $m$ c.p.c. order zero maps, and $\|\psi\phi(a)-a\|<\epsilon$ for $a \in F$.
By [@Tikuisis-Winter Proposition 2.5], we only need to show that the inclusion of $A$ into $A_\infty$ has nuclear dimension at most $(m+1)n-1$ To this end, let $F \subset A$ be finite and let $\epsilon>0$. Find $B,\psi$, and $\phi$ as in the hypotheses, for the finite set $F$ and with $\epsilon/2$ in place of $\epsilon$. Since $B$ has nuclear dimension at most $m$, we can factor the identity map on $B$, up to $\frac\epsilon{2n}$ on $\psi(F)$, as $$B \stackrel{\scriptstyle{\beta}}\longrightarrow K \stackrel{\scriptstyle{\alpha}}\longrightarrow B$$ such that $\beta$ is c.p.c. and $\alpha$ is a sum of $(m+1)$ c.p.c. order zero maps. Then the inclusion $A \to A_\infty$ factors, up to $\epsilon$ on $F$, as $$A \stackrel{\scriptstyle{\beta\psi}}\longrightarrow K \stackrel{\scriptstyle{\phi\alpha}}\longrightarrow A_\infty,$$ where $\beta\psi$ is c.p.c., and $\phi\alpha$ decomposes as a sum of $n(m+1)$ c.p.c. order zero maps.
If $J\unlhd A$ is an ideal, then by [@Winter-Zacharias2 Proposition 2.9] we have $$\dim_{\mathrm{nuc}}(A) \le \dim_{\mathrm{nuc}}(J) + \dim_{\mathrm{nuc}}(A/J) + 1.$$ However a slight modification of the proof of [@Kirchberg-Winter Proposition 5.1] implies the following statement.
\[propextension\] If $J \unlhd A$ has a quasicentral approximate unit consisting of projections, then $\dim_{\mathrm{nuc}}(A) = \max\{\dim_{\mathrm{nuc}}(J),\dim_{\mathrm{nuc}}(A/J)\}$.
Correspondences
---------------
Throughout this section let $A$ be a unital $C^*$-algebra. We give a brief overview of $C^*$-modules (all of which are assumed to be over $A$), $C^*$-correspondences, and Cuntz–Pimsner algebras. For comprehensive treatments, see [@Brown-Ozawa; @Lance; @Pimsner].
Let $\mathcal{H}$ be a Hilbert $C^*$-module. We say $\mathcal{H}$ is *full* if the set $\langle \mathcal{H},\mathcal{H}\rangle : = \{\langle x,y\rangle \ : \ x,y\in \mathcal{H}\}$ is dense in $A$.We say $\mathcal{H}$ is *free* if it has an orthonormal set of generators; it is *finitely generated projective* if it is an orthogonal direct summand in a finitely generated free module.
If $\mathcal{K}$ is another $C^*$-module, we denote by $\mathbb{B}(\mathcal{H,K})$ and $\mathbb{K}(\mathcal{H,K})$ the adjointable and compact operators from $\mathcal{H}$ to $\mathcal{K}$, respectively. $\mathbb{K}(\mathcal{H,K})$ is the closed span of operators $e_{x,y}$ (over $x \in \mathcal K, y \in \mathcal H$), where $$\label{RankOneFormula}
e_{x,y}(z) = x\ldotp \langle y,z\rangle_{\mathcal H}, \quad z \in \mathcal H.$$ In the case $\mathcal{H = K}$, we write $\mathbb{B}(\mathcal{H})$ and $\mathbb{K}(\mathcal{H})$. Besides the operator norm topology, there is another natural topology on $\mathbb{B}(\mathcal{H,K})$: a sequence $T_n$ converges *strictly* to $T$ if $T_n(x) \rightarrow T(x)$ and $T_n^*(y) \rightarrow T^*(y)$ for every $x\in \mathcal{H}$ and $y\in \mathcal{K}$. If $\{\mathcal{H}_i\}_{i\in I}$ is a collection of $C^*$-modules, their direct sum $\bigoplus \mathcal{H}_i$, defined as $$\{(x_i)_{i \in I} \in \prod_i \mathcal H_i \mid \sum \langle x_i, x_i\rangle \text{ converges in norm}\},$$ is also a $C^*$-module. If $\mathcal{H}$ is a free $C^*$-module whose orthonormal generators are indexed by a set $I$, then $\mathcal{H}\cong \bigoplus_IA$.
We will need the following important theorem of Kasparov.
\[Kasparov’s Stabilization Theorem\] If $\mathcal{H}$ is a countably generated, then there is a countably generated free $C^*$-module $\mathcal{H}'$ satisfying $$\mathcal{H}\oplus \mathcal{H}'\cong \mathcal{H}'.$$
Kasparov’s stabilization theorem implies $\mathbb{K}(\mathcal{H})$ is isomorphic to a hereditary subalgebra of $A\otimes \mathbb{K}$. By [@Winter-Zacharias2 Proposition 2.5], this implies $$\label{Compact-Dimension}
\dim_{\mathrm{nuc}}(\mathbb{K}(\mathcal{H})) \le \dim_{\mathrm{nuc}}(A).$$
We say $\mathcal{H}$ is a *$C^*$-correspondence* (or simply a correspondence) if there is a unital injective \*-homomorphism $\omega: A\rightarrow \mathbb{B}(\mathcal{H})$. The map $\omega$ is called the left action of $A$ on $\mathcal{H}$ and we will write $\omega_a(x)$ as $a\ldotp x$. The simplest example is the *identity correspondence*, which is the $C^*$-module $A$ (over $A$) with the left action given by left multiplication. If $\{\mathcal{H}_i\}_{i\in I}$ is a collection of correspondences, their direct sum $\bigoplus \mathcal{H}_i$ is also a correspondence with the left action given by $\omega_{\bigoplus \mathcal{H}_i} = \bigoplus\omega_{\mathcal{H}_i}$.
We say two correspondences $\mathcal{H}$ and $\mathcal{K}$ are *unitarily equivalent* (and write $\mathcal{H}\approx \mathcal{K}$) if there exists an adjointable map $U:\mathcal{H} \rightarrow \mathcal{K}$ such that
1. $U(a\ldotp x) = a\ldotp U(x)$ for every $x\in \mathcal{H}$ and $a\in A$, and
2. $U^*U = \text{id}_\mathcal{H}$, and $UU^* = \text{id}_{\mathcal{K}}$.
The algebraic tensor product $\mathcal{H}\odot\mathcal{K}$ of two correspondences $\mathcal{H}$ and $\mathcal{K}$ is naturally a right $A$-module with an $A$-valued semi-inner product given by $$\label{TensorIP}
\langle x_1\otimes y_1,x_2\otimes y_2\rangle = \langle y_1, \langle x_1,x_2\rangle\ldotp y_2\rangle.$$ Denote by $\mathcal{H}\otimes \mathcal{K}$ the $C^*$-module obtained from $\mathcal{H}\odot \mathcal{K}$ by separation and completion. In $\mathcal H \otimes \mathcal K$, the following identity holds $$\label{TensorBalance}
x \otimes (a\ldotp y) = (x\ldotp a) \otimes y, \quad x \in \mathcal H, y \in \mathcal K, a \in A.$$ There is an injective \*-homomorphism $\mathbb{B}(\mathcal{H}) \rightarrow \mathbb{B}(\mathcal{H}\otimes \mathcal{K})$ given by $T\mapsto T\otimes 1$, where $(T\otimes 1)(x\otimes y) = Tx\otimes y$ (and a \*-homomorphism $A'\cap \mathbb{B}(\mathcal{K}) \rightarrow \mathbb{B}(\mathcal{H}\otimes \mathcal{K})$ given by $T\mapsto 1\otimes T$). In particular $\mathcal{H}\otimes \mathcal{K}$ is a correspondence, called the *interior tensor product* of $\mathcal{H}$ and $\mathcal{K}$. Elements of the form $h\otimes k \in \mathcal{H}\otimes \mathcal{K}$ are called *elementary tensors*.
If $\mathcal H$ is a correspondence over a $C^*$-algebra $A$, define $$\mathcal H_\infty = \ell^\infty(\mathbb N, \mathcal H)/c_0(\mathbb N,\mathcal H),$$ and one can easily check that this is a correspondence over the sequence algebra $A_\infty$ in the obvious way. It contains $\mathcal H$ as the subset consisting of elements with constant sequence representatives. In particular, the product (in either order) of an element of $A_\infty$ and an element of $\mathcal H$ makes sense as an element of this $\mathcal H_\infty$.
For a correspondence $\mathcal H$ over a $C^*$-algebra $A$, a *representation* of $\mathcal{H}$ on a $C^*$-algebra $B$ is a pair $(\pi,\tau)$ consisting of a \*-homomorphism $\pi: A\rightarrow B$ and a linear map $\tau: \mathcal{H} \rightarrow B$ satisfying $\tau(a\ldotp x\ldotp b) = \pi(a)\tau(x)\pi(b)$ and $\tau(x)^*\tau(y) = \pi(\langle x,y\rangle)$ for every $x,y\in \mathcal{H}$ and $a,b\in A$. Denote by $C^*(\pi,\tau)$ the $C^*$-subalgebra of $B$ generated by $\pi(A)$ and $\tau(\mathcal{H})$. We say a representation $(\pi,\tau)$ *admits a gauge action* if there is an action $\beta$ of $\mathbb{T} = \{z\in \mathbb{C} \ : \ \lvert z\rvert = 1\}$ on $C^*(\pi,\tau)$ such that $\beta_z(\pi(a)) = \pi(a)$ and $\beta_z(\tau(\xi)) = z\tau(\xi)$ for all $a\in A$ and $\xi\in \mathcal{H}$.
Cuntz–Pimsner algebras
----------------------
We now briefly review the construction of Cuntz–Pimsner algebras. These were first defined by Pimsner in [@Pimsner]. Note that [@Pimsner] contains the proofs of Theorems \[Toeplitz-Properties\] and \[Cuntz-Universal\], and Proposition \[useful\_cuntz\]. See also [@Brown-Ozawa Section 4.6]
For a single correspondence $\mathcal{H}$, set $\mathcal{H}^{\otimes 0} = A$ and $\mathcal{H}^{\otimes k} = \mathcal{H}\otimes \cdots \otimes \mathcal{H}$. The *full Fock space* and *$p^{\text{th}}$-cutoff Fock space* over $\mathcal{H}$ are the correspondences defined by $$\mathcal{F}(\mathcal{H}) = \bigoplus_{k = 0}^{\infty}\mathcal{H}^{\otimes k} \ \ \ \text{and} \ \ \ \mathcal{F}_p(\mathcal{H}) = \bigoplus_{k=0}^{p-1}\mathcal{H}^{\otimes k}.$$ We say an elementary tensor $x = x_1\otimes\cdots\otimes x_k \in \mathcal{H}^{\otimes k}\subset\mathcal{F}(\mathcal{H})$ has *length $k$* and write $\lvert x\rvert = k$. When $\mathcal H$ is a free module with orthonormal basis $\{\xi_i\}_{i \in I}$, we set $W$ equal to the set of all elementary tensors in the $\xi_i$, and $$\label{W_kDef}
W_k = \{\mu \in W \mid |\mu| = k\}, \quad W_{<p} = \{\mu \in W \mid |\mu| < p\}.$$ The left action of $A$ on $\mathcal{F}(\mathcal{H})$ is given by $a\ldotp (x_1\otimes\cdots\otimes x_k) = (a\ldotp x_1)\otimes \cdots \otimes x_k$. Moreover for each $x \in \mathcal{H}$ we define the *creation* operator $T_x\in \mathbb{B}(\mathcal{F}(\mathcal{H}))$ by $T_x(a) = x\ldotp a$ and $T_x(x_1\otimes \cdots\otimes x_n) = x\otimes x_1\otimes\cdots\otimes x_n$. If $a\in\mathcal{H}^{\otimes 0}$ set $T_a = a$, and if $x = x_1\otimes\cdots\otimes x_k\in \mathcal{H}^{\otimes k}$ set $T_x = T_{x_1}\cdots T_{x_k}$. Note that for an elementary tensor $y$, $$T_x^*(y) = \begin{cases} \langle x,y'\rangle. y'', \quad &y=y' \otimes y'', |y|=|x|; \\ 0, \quad &|y|<|x|. \end{cases}$$ Using rank-one operators $e_{x,y}\in \mathbb{K}(\mathcal{H}^{\otimes \lvert x\rvert}, \mathcal{H}^{\otimes \lvert y\rvert})$ associated to elementary tensors in $\mathcal{F}(\mathcal{H})$, we have the identity $$\label{Strict-Expansion}
T_xT_y^* = \sum_{k = 0}^{\infty}e_{x,y}\otimes 1_{\mathcal{H}^{\otimes k}},$$ where convergence is understood to be strict.
\[Toeplitz-Pimsner Algebra\] Let $\mathcal{H}$ be a correspondence over $A$. The *Toeplitz–Pimsner algebra* $\mathcal{T}(\mathcal{H})$ is the $C^*$-subalgebra of $\mathbb{B}(\mathcal{F}(\mathcal{H}))$ generated by $A$ and $\{T_x \ | \ x\in \mathcal{H}\}$.
\[Toeplitz-Properties\] Let $\mathcal{T}(\mathcal{H})$ be the Toeplitz–Pimsner algebra of a correspondence $\mathcal{H}$ over $A$.
1. For every $\alpha\in \mathbb{C}$, $x,y\in \mathcal{H}$, and $a,b\in A$, the creation operators satisfy $$\label{Tequations}
T_{\alpha x + y} = \alpha T_x + T_y, \ \ \ T_{a\ldotp x\ldotp b} = aT_xb, \ \ \ T_x^*T_y = \langle x,y \rangle.$$ In particular, $$\label{Tspan}
\mathcal{T}(\mathcal{H}) = \overline{\mathrm{span}\{T_xT_y^* \mid x,y \in \mathcal F(\mathcal H) \text{ elementary tensors}\}}.$$
2. ([@Fowler-Muhly-Raeburn]) The Toeplitz–Pimsner algebra $\mathcal T(\mathcal H)$ is the unique $C^*$-algebra (up to isomorphism) generated by a representation $(\pi,\tau)$ of $\mathcal H$, such that $$\pi(A) \cap \overline{\mathrm{span}\{\tau(x)\tau(y)^* \mid x,y \in \mathcal H\}} = 0,$$ and which admits a gauge action.
The Toeplitz–Pimsner algebra is too large for many purposes, so we define the Cuntz–Pimsner algebra $\mathcal{O}(\mathcal{H})$ to be a natural quotient of $\mathcal{T}(\mathcal{H})$. Denote by $J(\mathcal{H})$ the $C^*$-subalgebra of $\mathbb{B}(\mathcal{F}(\mathcal{H}))$ generated by $$\bigcup_{n=0}^\infty \mathbb{B}(\bigoplus_{k=0}^n \mathcal{H}^{\otimes k}).$$ The multiplier algebra $\mathcal{M}(J(\mathcal{H}))$ can be identified with all $T\in \mathbb{B}(\mathcal{F}(\mathcal{H}))$ satisfying both $TJ(\mathcal{H}) \subset J(\mathcal{H})$ and $J(\mathcal{H})T\subset J(\mathcal{H})$. In particular, there is there is an inclusion $\mathcal{T}(\mathcal{H}) \subset \mathcal{M}(J(\mathcal{H}))$. (However, note that $\mathcal{T}(\mathcal{H})$ may not contain $J(\mathcal H)$.)
\[CPalgebra\] The *Cuntz–Pimsner algebra* $\mathcal{O}(\mathcal{H})$ is the $C^*$-algebra $Q(\mathcal{T}(\mathcal{H}))$, where $Q:\mathcal{M}(J(\mathcal{H})) \rightarrow \mathcal{M}(J(\mathcal{H}))/J(\mathcal{H})$ is the quotient map. We denote by $S_x$ the image of the creation operator $T_x$ under $Q$.
Here is another description of $\mathcal{O}(\mathcal{H})$. Let $I_{\mathcal{H}} = A\cap \mathbb{K}(\mathcal{H}) \subset \mathbb{B}(\mathcal{H})$. Since $I_{\mathcal{H}}$ is an ideal in $A$ and $\mathcal{F}(\mathcal{H})I_{\mathcal{H}}$ is a $\mathbb{B}(\mathcal{F}(\mathcal{H}))$-invariant subcorrespondence of $\mathcal{F}(\mathcal{H})$, we can conclude $\mathbb{K}(\mathcal{F}(\mathcal{H})I_{\mathcal{H}}) = \overline{\mathrm{span}}\{e_{x,y} \ | \ x,y\in \mathcal{F (\mathcal{H})I_{\mathcal{H}}\}}$ is an ideal in $\mathbb{B}(\mathcal{F}(\mathcal{H}))$.
\[useful\_cuntz\] $\mathbb{K}(\mathcal{F}(\mathcal{H})I_{\mathcal{H}}) \subset\mathcal{T}(\mathcal{H})$, and in particular $\mathbb{K}(\mathcal{F}(\mathcal{H})I_{\mathcal{H}}) = \ker Q|_{\mathcal{T}(\mathcal{H})}$. In other words, $$\mathcal{O}(\mathcal{H}) \cong \mathcal{T}(\mathcal{H})/\mathbb{K}(\mathcal{F}(\mathcal{H})I_{\mathcal{H}}).$$
\[Cuntz-Universal\] Let $\mathcal{O}(\mathcal{H})$ be the Cuntz–Pimsner algebra of a correspondence $\mathcal{H}$ over $A$. For every $\alpha \in \mathbb{C},x,y\in \mathcal{H}$, and $a,b\in A$, the following identity holds: $$S_{\alpha x + y} = \alpha S_x + S_y, \ \ \ S_{a\ldotp x\ldotp b} = aS_xb, \ \ \ S_x^*S_y = \langle x,y\rangle.$$
\[Cuntz=Pimsner\] If $\mathcal{H}$ is a finitely generated projective correspondence, then $\mathbb{K}(\mathcal{H}) = \mathbb{B}(\mathcal{H})$ (see [@Wegge-Olsen]) and hence $A\cap \mathbb{K}(\mathcal{H}) = A$. Since $A$ is unital, $\mathbb{K}(\mathcal{F}(\mathcal{H})I_{\mathcal{H}}) = \mathbb{K}(\mathcal{F}(\mathcal{H}))$. This shows $\mathcal{T}(\mathcal{H})$ contains all of $\mathbb{K}(\mathcal{F}(\mathcal{H}))$ and $\mathcal O(\mathcal H) \cong \mathcal T(\mathcal H)/\mathbb K(\mathcal F(\mathcal H))$. At the other extreme, if $A\cap\mathbb{K}(\mathcal{H}) = \{0\}$, the kernel of $Q$ is trivial and hence there is a \*-isomorphism $\mathcal{O}(\mathcal{H}) \rightarrow \mathcal{T}(\mathcal{H})$ sending $a$ to $a$ and $S_x$ to $T_x$.
Rokhlin dimension for $C^*$-correspondences {#Rokhlin dimension for correspondences}
===========================================
Here is our definition of Rokhlin dimension for $C^*$-correspondences.
\[Rokhlin-Dimension-Def\] Let $A$ be a separable $C^*$-algebra and let $\mathcal{H}$ be a countably generated correspondence over $A$. We say that $\mathcal{H}$ has *Rokhlin dimension at most $d$*, $\dim_{\mathrm{Rok}}(\mathcal{H}) \le d$, if any $p\in \mathbb{N}$, there exist positive contractions $$\{f_k^l\}_{l=0,\dots,d;\, k \in \mathbb Z/p} \subset A_\infty \cap A'$$ satisfying
1. $f_k^lf_{k'}^l=0$ for all $l$ and all $k \ne k'$,
2. $\sum_{k,l}f_k^l = 1$, and
3. $z\ldotp f_k^l = f_{k+1}^l\ldotp z$ in $\mathcal H_\infty$, for all $k$, $l$, and $z \in \mathcal H$.
\[Rokhlin-Def-Rem\] (i) One can of course reformulate this definition without using the sequence algebra $A_\infty$ and the sequence correspondence $\mathcal H_\infty$. Namely, we have $\dim_{\mathrm{Rok}}(\mathcal H) \leq d$ if and only if, for any $\epsilon > 0$, any $p\in \mathbb{N}$, any finite set $F\subset A$, and any finite set $\mathcal{V} \subset \mathcal{H}$, there exist positive contractions $$\{f_k^l\}_{l=0,\dots,d;\, k \in \mathbb Z/p} \subset A$$ satisfying
1. $\lVert f_k^lf_{k'}^l\rVert < \epsilon$ when $k \ne k'$ and all $l$.
2. $\lVert \sum_{k,l}f_k^l - 1\rVert < \epsilon$.
3. $\lVert z\ldotp f_k^l - f_{k+1}^l\ldotp z\rVert < \epsilon$ for all $k$, $l$, and $z \in \mathcal{V}$.
4. $\lVert [f_k^l,a]\rVert < \epsilon$ for all $k,l$ and $a\in F$.
\(ii) If $A$ is a $C^*$-algebra and $\alpha \in \text{Aut}(A)$, one may define a correspondence $A^\alpha$ over $A$ as the singly generated $C^*$-module $A$ with with the left action given by $a\ldotp b = \alpha(a)b$. There is a canonical isomorphism $\mathcal O(A^\alpha) \cong A \rtimes_\alpha \mathbb Z$ (this isomorphism fixes $A$ and sends $S_1$ to the canonical unitary of the crossed product). Our definition of Rokhlin dimension of $C^*$-correspondences is designed to (almost) coincide with the Rokhlin dimension of $\alpha$ (with single towers) as defined by Hirshberg, Winter, and Zacharias in [@Hirshberg-Winter-Zacharias Definition 2.3(c)]. Specifically, we ask for single towers all of height $p+1$ (whereas their definition of $\dim_{\mathrm{Rok}}^s(A,\alpha)\leq d$ asks that each colour has a single tower of height either $p$ or $p+1$). In [@Hirshberg-Winter-Zacharias Proposition 2.8 and Remark 2.9], they show that, up to a possible factor of $2$, their Rokhlin dimension coincides with the version with single towers all of height $p+1$. The same argument applies to variants on the definition of Rokhlin dimension for correspondences, and as such, we have chosen to work with the simplest version of Rokhlin dimension.
\(iii) We can simultaneously express condition (3) and the requirement that the Rokhlin contractions commute with $A$ by asking that, for any elementary tensor $z \in \mathcal F(\mathcal H)$, $$z\ldotp f_k^l = f_{k+\lvert z\rvert}^l\ldotp z.$$
Nuclear dimension of Toeplitz–Pimsner and Cuntz–Pimsner algebras {#Heavy lifting}
================================================================
In [@Hirshberg-Winter-Zacharias Theorem 4.1], it is shown that for a $C^*$-algebra $A$ with finite nuclear dimension and an automorphism $\alpha:A \to A$ of finite Rokhlin dimension, the crossed product $A \rtimes_\alpha \mathbb Z$ has finite nuclear dimension. In this section we generalize this result to correspondences of finite Rokhlin dimension (this is truly a generalization, see Remark \[Rokhlin-Def-Rem\] (i)), subject to a technical condition which is satisfied, for example, by correspondences which are finitely generated and projective as Hilbert $C^*$-modules.
One can recast the argument used to prove [@Hirshberg-Winter-Zacharias Theorem 4.1] in terms of the Fock-space representation of $\mathcal{T}(A^\alpha)$, where $A^\alpha$ is as in Remark \[Rokhlin-Def-Rem\] (i). Specifically, the argument makes use of outgoing maps from $\mathcal T(A^\alpha)$ to a compression of the Fock space representation; the range of these maps land in a subalgebra of $\mathbb B(\mathcal F(A^\alpha))$ which, in this case, is isomorphic to some $M_n(A)$. It is possible to then define incoming maps $M_n(A) \to \mathcal T(A^\alpha)$, using compressions by row vectors corresponding to Rokhlin towers. Our argument for general correspondences is based on this outline; however, there are two significant technical differences. First, the codomain of the outgoing map will generally not be a matrix algebra over $A$, and so further input is needed to get a (uniform) bound on its nuclear dimension. Second, there need to be suitable replacements for the row vector compressions used to construct the incoming maps. In this section, we deal with these technicalities.
A compressed Fock space representation and nuclear dimension
------------------------------------------------------------
Let $\mathcal{H}$ be a countably generated correspondence over $A$. For each $p \in \mathbb{N}$, form $$\mathcal{F}_p(\mathcal{H}) = \bigoplus_{k=0}^{p-1}\mathcal{H}^{\otimes k}.$$ Set $$\begin{aligned}
D_p(\mathcal{H}) = \mathrm{span}\{e_{x,y} \otimes 1_{\mathcal H}^{\otimes k} \mid & x,y \in \mathcal F_p(\mathcal H) \text{ elementary tensors}, \\ &\quad \max\{|x|,|y|\}+k<p\}.\end{aligned}$$ This is a $C^*$-subalgebra of $\mathbb B(\mathcal F_p(\mathcal H))$. Evidently, $\mathbb{K}(\mathcal{F}_p(\mathcal{H})) $ is an ideal of $D_p(\mathcal{H})$.
\[quasi\_hypothesis\] Suppose that for every $p\in \mathbb{N}$, there is an approximate unit consisting of projections in $\mathbb{K}(\mathcal{F}_p(\mathcal{H}))$ that are quasicentral in $D_p(\mathcal{H})$. Then for every $p\in \mathbb{N}$ we have $$\dim_{\mathrm{nuc}}(D_p(\mathcal{H})) \le \dim_{\mathrm{nuc}}(A).$$
By induction in $p$. It is not hard to see that $D_1(\mathcal{H}) \cong A$. Suppose the result holds for $D_{i}$ for $i = 1,\ldots,p$. Set $$\tilde{D}_p = \overline{\mathrm{span}}\{(e_{x,y}\otimes 1_{\mathcal{H}^{\otimes k}})\otimes 1_{\mathcal{H}} \ : 0 \le \max\{\lvert x\rvert ,\lvert y\rvert\} + k < p\} \subset D_{p+1}(\mathcal{H}).$$ It’s clear that $\tilde{D}_p\cong D_p(\mathcal{H})$ and that $$D_{p+1}(\mathcal{H}) = C\text{*}(\mathbb{K}(\mathcal{F}_{p+1}(\mathcal{H})) \cup \tilde{D}_p) \ \ \ \text{and} \ \ \ \mathbb{K}(\mathcal{F}_{p+1}(\mathcal{H}))\cap\tilde{D}_p = \{0\}.$$ This shows that $D_{p+1}(\mathcal{H})$ is an extension of (an algebra isomorphic to) $D_p(\mathcal{H})$ by the compacts $\mathbb{K}(\mathcal{F}_{p+1}(\mathcal{H}))$. The result follows by induction, Proposition \[propextension\], and By the inductive hypothesis and , both of these algebras have nuclear dimension at most $\dim_{\mathrm{nuc}}(A)$. By the hypotheses of this lemma, we may apply Proposition \[propextension\] to conclude that the nuclear dimension of $D_{p+1}(\mathcal H)$ is also at most $\dim_{\mathrm{nuc}}(A)$.
\[QD-examples\] (i) If $\mathcal{H}$ is a finitely generated projective correspondence, then so is $\mathcal{H}^{\otimes k}$ for any $k \ge 0$ (see [@Lance Proposition 4.7]). Hence, $\mathbb{K}(\mathcal{F}_p(\mathcal{H}))$ is unital for any $p\in \mathbb{N}$ and we get $$\mathbb{K}(\mathcal{F}_p(\mathcal{H})) = D_p(\mathcal{H}) = \mathbb{B}(\mathcal{F}_p(\mathcal{H})).$$ Thus, such $\mathcal H$ does satisfy the condition of Lemma \[quasi\_hypothesis\] (i.e., for every $p$, there is an approximate unit of projections in $\mathbb K(\mathcal F_p(\mathcal H))$ which is quasicentral in $D_p(\mathcal H)$).
\(ii) Let $\mathcal{K}$ be countably generated free Hilbert $A$-module with orthonormal basis $(\xi_i)_{i=1}^\infty$. Let $\{\alpha_i\}_{i=1}^\infty \subset \mathrm{Aut}(A)$ be a sequence of automorphisms of $A$, and define a correspondence by the left action $a\ldotp \xi_i = \xi_i\ldotp \alpha_i(a)$ for $i \in \mathbb N$. This correspondence also satisfies the hypothesis of Lemma \[quasi\_hypothesis\].
For $\mu = \xi_1\cdots \xi_k \in W_k$, set $$\alpha_{\mu} = \alpha_{k} \circ \cdots \circ \alpha_{1} \in \mathrm{Aut}(A),$$ and note that $a.\mu = \mu.\alpha_\mu(a)$.
Let $W_{<p}^n$ denote the finite subset of $W_{<p}$ consisting of elementary tensors of length $<p$ in $\xi_1,\dots,\xi_n$. Define the finite rank projection $q_n = \sum_{\zeta \in W_{<p}^n}e_{\zeta,\zeta}$. This projection is quasicentral in $D_p(\mathcal{K})$, since for $\mu,\nu \in W$ and $k \in \mathbb N$, for sufficiently large $n$ (namely, when $n$ is greater than all indices of basis elements appearing in $\mu$ and $\nu$), we have $$\begin{aligned}
q_n(e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}}) & =
\sum_{\zeta \in W_{<p}^n}e_{\zeta,\zeta}\sum_{\eta \in W_k} e_{\mu.a\otimes \eta, \nu\otimes \eta} \\
&= \sum_{\zeta \in W_{<p}^n}e_{\zeta,\zeta}\sum_{\eta\in W_k}e_{\mu\otimes \eta\ldotp \alpha_{\eta}(a),\nu\otimes \eta} \\
& = \sum_{\zeta\in W_{<p}^n}\sum_{\eta\in W_k}e_{\zeta\ldotp \langle\zeta, \mu\otimes \eta\rangle\alpha_{\eta}(a), \nu\otimes \eta}\\
& = \sum_{\eta\in W_k^n}e_{\mu\otimes \eta\ldotp \alpha_{\eta}(a),\nu\otimes \eta}\\
& = \sum_{\eta\in W_k}\sum_{\zeta\in W_{<p}^n}e_{\mu\otimes \eta\ldotp \alpha_{\eta}(a), \zeta\ldotp \langle\zeta, \nu\otimes \eta\rangle}\\
& = \sum_{\eta\in W_k}e_{\mu\otimes \eta\ldotp \alpha_{\eta}(a), \nu\otimes \eta}\sum_{\zeta\in W_{<p}^n}e_{\zeta,\zeta}\\
& = e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}}q_n. \end{aligned}$$
Incoming maps
-------------
This section works extensively with countably generated free correspondences: first we use the structure of such a correspondence $\mathcal K$ to define incoming maps $D_p(\mathcal K) \to \mathcal T(\mathcal K)$, and then we use Kasparov’s stabilization theorem to apply such incoming maps to general countably generated correspondences. Throughout, $\mathcal K$ will generally be a countably generated free correspondence, while $\mathcal H$ will be an arbitrary (countably generated) Hilbert module or correspondence.
For a countably generated free correspondence $\mathcal K$ over $A$ with orthonormal basis $(\xi_i)_{i=1}^\infty$, recall that $W \subset \mathcal F(\mathcal K)$ denotes the set of elementary tensors in this generating set, while (from ) $W_k$ and $W_{<p}$ denote the subsets of elementary tensors of length $k$ and $<p$ respectively. As Hilbert $A$-modules, we have $$\mathcal{K}^{\otimes k}\cong \bigoplus_{W_k}A, \ \ \ \mathcal{F}_p(\mathcal{K}) \cong \bigoplus_{W_{<p}}A, \ \ \ \mathcal{F}(\mathcal{K}) \cong \bigoplus_WA.$$
### Free correspondences {#technical0}
\[i-love-lamp\] Let $A$ be a $C^*$-algebra, let $\mathscr{H}$ be a Hilbert $A$-module, and let $I$ be an index set. There is an inclusion $\mathbb{B}(\bigoplus_IA) \hookrightarrow \mathbb{B}(\bigoplus_I \mathscr{H})$. More specifically, an operator $S\in \mathbb{B}(\bigoplus_IA)$ acts adjointably on $\bigoplus_I\mathscr{H}$ via $$\label{i-love-lamp-eq}
(x_i)_{i \in I} \mapsto (\sum_{i\in I} (\langle \xi_j,S\xi_i\rangle \ldotp x_i)_{j \in J},$$ where $(\xi_i)_{i\in I}$ is the canonical orthonormal basis for $\bigoplus_I A$.
The map $a\otimes x \mapsto a\ldotp x$ implements an isomorphism between $A\otimes \mathscr{H}$ and $\mathscr{H}$. Let us show that the map $m: (a_i)_{i\in I} \otimes x\mapsto (a_i \ldotp x)_{j \in I}$ extends to an isomorphism between $(\bigoplus_IA)\otimes \mathscr{H}$ and $\bigoplus_I\mathscr{H}$. We have $$\begin{aligned}
\langle (a_i)_i \otimes x, (b_i)_i \otimes y\rangle
&\stackrel{\eqref{TensorIP}}=& \langle x, \langle (a_i)_i,(b_i)_i\rangle\ldotp y\rangle \\
&=& \langle x, \sum_i a_i^*b_i \ldotp y \rangle \\
&=& \sum_i \langle a_i\ldotp x, b_i\ldotp y \rangle \\
&=& \langle (a_i\ldotp x)_i, (b_i\ldotp y)_i \rangle \\
&=& \langle m((a_i)_i \otimes x), m((b_i)_i \otimes y) \rangle\end{aligned}$$ so that $m$ can be extended, as a $\langle \cdot,\cdot\rangle$-preserving linear functional, to the tensor product $(\bigoplus_IA)\otimes \mathscr{H}$. Moreover, the image of $m$ contains the dense set consisting of finitely supported elements in $\bigoplus_I\mathscr{H}$, so that $m$ is a unitary operator.
The result follows since $T\mapsto mTm^{-1}$ is a \*-isomorphism between $\mathbb{B}((\bigoplus_IA)\otimes \mathscr{H})$ and $\mathbb{B}(\bigoplus_I\mathscr{H})$, and composing this with the natural embedding $\mathbb{B}(\bigoplus_IA) \hookrightarrow\mathbb{B}((\bigoplus_IA)\otimes \mathscr{H})$ yields a map satisfying .
\[techcortwo\] Let $A$ be a $C^*$-algebra, let $\mathcal K$ be a countably generated free correspondence with orthonormal basis $\{\xi_i\}_{i\in I}$, and define $W_{<p}$ by . For each $p \in \mathbb{N}$, there is an inclusion $\mathbb{B}(\mathcal{F}_p(\mathcal{K})) \hookrightarrow \mathbb{B}(\bigoplus_{W_{<p}}\mathcal{F}(\mathcal{K}))$ that sends $e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}}$ (where $\mu,\nu \in W$, $a \in A$, and $\max\{|\mu|,|\nu|\} + k < p$) $$\label{techcortwo-eq1}
(x_{\zeta})_{\zeta \in W_{<p}} \mapsto \sum_{\eta,\eta'\in W_k} (\delta_{\zeta',\mu\otimes \eta'} \langle \eta', a\ldotp \eta\rangle \ldotp x_{\nu\otimes \eta})_{\zeta'' \in W_{<p}},$$ and $a$ (as an operator in $\mathbb{B}(\mathcal{F}_p(\mathcal{K}))$) to the operator $$\label{techcortwo-eq2}
(x_{\zeta})_{\zeta \in W_{<p}} \mapsto \sum_{\zeta \in W_{<p}} (\langle \zeta',a\ldotp \zeta \rangle\ldotp x_{\zeta})_{\zeta' \in W_{<p}}.$$
Applying Lemma \[i-love-lamp\] with $\mathcal H=\mathcal F(\mathcal K)$ and $I=W_{<p}$ yields an embedding $\iota:\mathbb B(\bigoplus_{W_{<p}} A) \to \mathbb B(\bigoplus_{W_{<p}} \mathcal F(\mathcal K))$ such that for $S \in \mathbb B(\bigoplus_{W_{<p}} A)$ and $(x_{\zeta})_{\zeta \in W_{<p}}$, $$\iota(S)((x_\zeta)_{\zeta\in W_{<p}}) = \sum_{\zeta \in W_{<p}} (\langle \xi_{\zeta'},S\xi_\zeta \rangle\ldotp x_\zeta)_{\zeta' \in W_{<p}}.$$ We use the canonical identification of $\mathcal F_p(\mathcal K)$ with $\bigoplus_{W_{<p}} A$ to view $\iota$ as an inclusion of $\mathbb(F_p(\mathcal K))$ to $\mathbb B(\bigoplus_{W_{<p}} \mathcal F(\mathcal K))$. This identification takes an elementary tensor $\zeta \in W_{<p}$ to the orthonormal basis element $\xi_\zeta$. Since $a \in A$ acts on $\mathcal F_p(\mathcal K)$ by sending $\zeta \in W_{<p}$ to $a.\zeta$, follows immediately. For , using and , first note that for $\zeta \in W_{<p}$, $$(e_{\mu\ldotp a,\nu} \otimes 1_{\mathcal H^{\otimes k}})(\zeta) = \begin{cases} \mu \otimes a\ldotp \eta, \quad &\zeta=\nu \otimes \eta, |\eta|=k; \\ 0, \quad &\text{otherwise}. \end{cases}$$ For $\zeta = \mu \otimes \eta \in W_{<p}$ where $\eta \in W_k$ and for $\zeta' \in W_{<p}$, $$\begin{aligned}
\langle \zeta',(e_{\mu\ldotp a,\nu} \otimes 1_{\mathcal H^{\otimes k}})\zeta \rangle
&= \langle \zeta',\mu \otimes a\ldotp \eta \rangle \\
&= \begin{cases} \langle \eta', a\ldotp \eta \rangle, \quad &\zeta' = \mu \otimes \eta', |\eta'| = k; \\ 0, \quad &\text{otherwise}. \end{cases}\end{aligned}$$ Putting this together, we have $$\begin{aligned}
\iota(e_{\mu\ldotp a,\nu} \otimes 1_{\mathcal H^{\otimes k}})((x_\zeta)_{\zeta \in W_{<p}})
&= \sum_{\zeta \in W_{<p}} (\langle \zeta',(e_{\mu\ldotp a,\nu} \otimes 1_{\mathcal H^{\otimes k}})\zeta \rangle\ldotp x_\zeta)_{\zeta' \in W_{<p}} \\
&= \sum_{\eta,\eta' \in W_k} (\delta_{\zeta',\mu \otimes \eta'} \langle \eta', a\ldotp \eta\rangle \ldotp x_{\mu \otimes \eta})_{\zeta' \in W_{<p}},\end{aligned}$$ as required.
\[Rows-Vectors\] Let $A$ be a $C^*$-algebra, let $\mathscr{H}$ be a Hilbert $A$-module, and let $I$ be a countable index set. Let $\{T_i\}_{i\in I}$ be a collection of isometries in $\mathbb{B}(\mathscr{H})$ with orthogonal ranges such that $\sum_i T_iT_i^*$ converges strictly in $\mathbb{B}(\mathscr{H})$. Then the map $[T_i]_I:\bigoplus_I\mathscr{H} \rightarrow \mathscr{H}$ given by $(x_i)_{i\in I}\mapsto \sum_{i \in I} T_ix_i$ is an adjointable operator with adjoint given by $$\label{TiAdjoint}
[T_i]_I^*(x) = (T_i^*x)_{i \in I}.$$ Moreover, $[T_i]_I$ is an isometry.
The hypothesis that $\sum_i T_iT_i^*$ converges strictly implies that the formula for $[T_i]_I^*(x)$ does define an element of $\bigoplus_I \mathcal H$, and therefore this formula produces a well-defined map $\mathcal H \to \bigoplus_I \mathcal H$. For an indexed family $(x_i)_{i \in I}$ of pairwise orthogonal elements of $\mathcal H$, note that $\sum_i x_i$ converges to an element $x \in \mathcal H$ if and only if $(x_i)_{i \in I}$ represents an element $y$ of $\bigoplus_I \mathcal H$, and in this case, $\langle x,x\rangle_{\mathcal H} = \langle y,y\rangle_{\bigoplus_I \mathcal H}$. Since the $T_i$ have orthogonal ranges, it follows from these facts that $[T_i]_I$ is a well-defined isometry. It is an easy calculation to see that the formula for $[T_i]_I^*$ does indeed provide an adjoint to $[T_i]_I$.
The assumption that $\sum_i T_iT_i^*$ converges strictly is not automatic; here is an example. Let $A=l^\infty(\mathbb C)$ and let $\mathcal H = \bigoplus_{\mathbb N} A$; write an element of $\mathcal H$ as $(x_i^j)$ where for each $j$, $\sum_i |x_i^j|^2$ converges (and is uniformly bounded in $j$). For each $j \in \mathbb N$, pick an injective map $\theta_j:\mathbb N \times \mathbb N \to \mathbb N$ such that $\theta_j(1,j)=1$. For each $k \in \mathbb N$, define $T_k:\mathcal H \to \mathcal H$ by $T_k((x_i^j)) = (y_i^j)$ where $$y_i^j = \begin{cases} x_{i'}^j, \quad &i=\theta^j(i',k); \\ 0,\quad &\text{otherwise}. \end{cases}$$ Using the fact that the $\theta_j$ are injective, one easily computes $$\langle (x_i^j), (x_i^j) \rangle = (\sum_i |x_i^j|^2)_j = \langle T_k((x_i^j)), T_k(x_i^j)) \rangle,$$ so that each $T_k$ is isometric. Injectivity of the $\theta_j$ also implies that the $T_k$ have pairwise orthogonal ranges.
To see that $\sum_k T_kT_k^*$ does not converge strictly, let us check that $\sum_k T_kT_k^* \xi_1$ does not converge (in $\mathcal H$). Note that $\xi_1 = (\delta_{i,1})_{i,j}$, and that for $(x_i^j) \in \mathcal H$, $$T_kT_k^*(x_i^j) = (\chi_{A_{j,k}}(i)y_i^j),$$ where $$A_{j,k} = \{\theta_j(n,k) \mid n \in \mathbb N\}.$$ By our choice of $\theta_j$, we have $1 \in A_{j,k}$ if and only if $j=k$. Therefore, $$\sum_{k=1}^n T_kT_k^*\xi_1 = \xi_1\ldotp \chi_{\{1,\dots,n\}}$$ (viewing $\chi_{\{1,\dots,n\}}$ as an element of $l^\infty(\mathbb N)=A$). The sequence $(\chi_{\{1,\dots,n\}})_n$ does not converge (in norm) in $A$, so that $\sum_k T_kT_k^* \xi_1$ does not converge in $\mathcal H$.
\[techcorone\] Let $A$ be a $C^*$-algebra, let $\mathcal K$ be a countably generated free correspondence with orthonormal basis $\{\xi_i\}_{i\in I}$, and define $W_{k}$ by . For any $k \ge 0$, there is an isometry $[T_{\eta}]_{W_k} \in \mathbb{B}(\bigoplus_{W_k}\mathcal{F}(\mathcal{K}), \mathcal{F}(\mathcal{K}))$ given by $$(x_{\eta})_{\eta \in W_k} \mapsto \sum_{\eta\in W_k}T_{\eta}( x_{\eta}).$$
Note that $(T_\eta)_{\eta \in W_k}$ is a family of isometries with pairwise orthogonal images, and that $\sum_{\eta\in W_k}T_{\eta}T_{\eta}^*$ converges strictly to $1_{\mathcal{F}(\mathcal{K})} - 1_{\mathcal{F}_k(\mathcal{K})}$ in $\mathbb{B}(\mathcal{F}(\mathcal{K}))$. Thus setting $I=W_k$ and $\mathscr{H} =\mathcal{F}(\mathcal{K})$, the hypotheses of Lemma \[Rows-Vectors\] are satisfied; this lemma shows that $[T_\eta]_{W_k}$ is a well-defined isometry.
For each $p \in \mathbb{N}$ and $0 \le k < p$, we can regard $[T_{\eta}]_{W_k}$ as being an element in $\mathbb{B}(\bigoplus_{W_{<p}}\mathcal{F}(\mathcal{K}), \mathcal{F}(\mathcal{K}))$ by identifying $\bigoplus_{W_{<p}}\mathcal{F}(\mathcal{K})$ with $\bigoplus_{W_0}\mathcal{F}(\mathcal{K})\oplus\cdots\oplus \bigoplus_{W_{p-1}}\mathcal{F}(\mathcal{K})$, and defining $[T_{\eta}]_{W_k}$ to be zero on $\bigoplus_{W_{k'}}\mathcal{F}(\mathcal{K})$ for $k' \ne k$.
Recall from Definition \[CPalgebra\] that, for $x \in \mathcal K$, $S_x$ denotes the image of $T_x$ in the Cuntz–Pimsner algebra $\mathcal O(\mathcal K)$.
\[maps\] Let $A$ be a $C^*$-algebra, let $\mathcal K$ be a countably generated free correspondence with orthonormal basis $\{\xi_i\}_{i\in I}$, and define $W_{<p}$ by . Let $p \in \mathbb{N}$ and let $\mathcal{G} = (g_0,\ldots,g_{p-1})$ be an tuple of positive contractions in $A$. Set $$\label{Rdef}
R_{\mathcal G} = \sum_{k=0}^{p-1}g_k[T_{\eta}]_{k} \in \mathbb{B}(\bigoplus_{W_{<p}}\mathcal{F}(\mathcal{K}), \mathcal{F}(\mathcal{K})).$$ Define $\sigma_{\mathcal G}:D_p(\mathcal K) \to \mathbb B(\mathcal F(\mathcal K))$ to be the following composition: $$D_p(\mathcal K) \subset \mathbb B(\mathcal F_p(\mathcal K)) \to \mathbb B(\bigoplus_{W_{<p}}\mathcal F(\mathcal K)) \stackrel{R_{\mathcal G}\cdot R_{\mathcal G}^*}{\longrightarrow} \mathbb B(\mathcal F(\mathcal K)),$$ where the first map is the inclusion given by Corollary \[techcortwo\]. Let $Q:\mathcal M(J(\mathcal K)) \to \mathcal M(J(\mathcal K))/J(\mathcal K)$ denote the quotient map. Then $\sigma_{\mathcal G}(D_p(\mathcal K)) \subset Q^{-1}(\mathcal O(\mathcal K))$, and so there exists a c.p. map $$\rho_{\scriptscriptstyle{\mathcal{G}}} = Q \circ \sigma_{\mathcal G}: D_p(\mathcal{K}) \rightarrow \mathcal{O}(\mathcal{K}).$$ This map satisfies $$\label{rhoMapDef}
\rho_{\scriptscriptstyle{\mathcal G}}(e_{x,y}\otimes 1_{\mathcal{K}^{\otimes k}}) = g_{k +\lvert x\rvert}S_xS_y^*g_{k + \lvert y\rvert}$$ for elementary tensors $x,y$ in $\mathcal{F}_p(\mathcal{K})$ and $k \in \mathbb N$ such that $\max\{|x|,|y|\}+k<p$.
For $a\in A$, $\mu,\nu \in W_{<p}$, and $k \ge 0$ such that $e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}} \in D_p(\mathcal{K})$, and $z\in \mathcal{F}(\mathcal{K})$, we have $$\begin{aligned}
(R_{\mathcal G}e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}}R_{\mathcal G}^*)(z)
&\stackrel{\eqref{TiAdjoint}}=& R_{\mathcal G}(e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}})(T_{\zeta}^*g_{\lvert \zeta\rvert}\ldotp z)_{\zeta \in W_{<p}} \\
&\stackrel{\eqref{techcortwo-eq1}}=& R_{\mathcal G}(\sum_{\eta,\eta'\in W_k} \delta_{\zeta',\mu\otimes \eta'} \langle \eta',a\ldotp \eta\rangle \ldotp T_{\nu \otimes \eta}^* g_{\lvert \nu\otimes \eta\rvert}\ldotp z)_{\zeta'\in W_{<p}} \\
&=& \sum_{\eta,\eta'\in W_{k}} g_{\lvert \mu\otimes\eta'\rvert} T_{\mu\otimes\eta'} \langle \eta',a\ldotp \eta\rangle\ldotp T_{\nu\otimes\eta}^* g_{\lvert \nu\otimes\eta \rvert}\ldotp z \\
&=& g_{\lvert \mu\rvert + k} T_\mu \left(\sum_{\eta,\eta'\in W_{k}} T_{\eta'}\langle \eta',a\ldotp \eta\rangle T_\eta^*\right) T_\nu^* g_{\lvert \nu\rvert+k}\ldotp z.\end{aligned}$$ Note that for a basis element $\xi \in W$, $$\begin{aligned}
\sum_{\eta,\eta' \in W_k} T_{\eta'}\langle \eta',a\ldotp \eta\rangle T_\eta^* \xi
&= \begin{cases} \sum_{\eta'\in W_k} T_{\eta'} \langle \eta',a\ldotp \eta\rangle \xi', \quad &\xi=\eta\otimes \xi', \eta\in W_k; \\ 0,\quad &|\xi|<k \end{cases} \\
&= \begin{cases} a\ldotp \eta\otimes \xi', \quad &\xi=\eta\otimes \xi', \eta\in W_k; \\ 0,\quad &|\xi|<k \end{cases} \\
&= (1_{\mathcal F(\mathcal K)}-1_{\mathcal F_k(\mathcal K)})\xi,\end{aligned}$$ and by linearity, this formula continues to hold for all $\xi \in \mathcal F(\mathcal K)$. Putting these together, we find $$(R_{\mathcal G}e_{\mu\ldotp a,\nu}\otimes 1_{\mathcal{K}^{\otimes k}}R_{\mathcal G}^*)(z) = (g_{\lvert\mu\rvert+k}T_{\mu}a(1_{\mathcal{F}(\mathcal{K})} - 1_{\mathcal{F}_k(\mathcal{K})})T_{\nu}^*g_{\lvert \nu\rvert}+k)(z).$$ Define $\rho_{\mathcal G}=Q\circ\sigma_{\mathcal G}:D_p(\mathcal K) \to \mathcal{M}(J(\mathcal{K}))$, and we see that holds for $x=\mu.a, y=\nu$ where $\mu,\nu \in W_{<p}$. By density, holds for all elementary tensors $x,y$, and the image of $\rho_{\mathcal G}$ is contained in $\mathcal O(\mathcal K)$ (i.e., the image of $\sigma_{\mathcal G}$ is contained in $Q^{-1}(\mathcal O(\mathcal K))$).
\[moremaps\] Let $A$ be a $C^*$-algebra, let $\mathcal K$ be a countably generated free correspondence with orthonormal basis $\{\xi_i\}_{i\in I}$, and define $W_{<p}$ by . Let $\mathcal{G} = (g_0,\ldots,g_{p-1})$, let $\delta$ be the maximum value of $\lVert g_ig_j\rVert$, $i\ne j$, and define $R_{\mathcal G}$ by .
1. Regarding $g_k^2|_{\mathcal{K}^{\otimes k}}$ as an element of $\mathbb{B}(\bigoplus_{W_{<p}}\mathcal{F}(\mathcal{K}))$, we have $$R_{\mathcal G}^*R_{\mathcal G} \approx_{p^2\delta} g_0^2|_{\mathcal{K}^{\otimes 0}} + \cdots + g_{p-1}^2|_{\mathcal{K}^{\otimes p-1}}.$$
2. Let $w$ and $z$ be elementary tensors in $\mathcal{F}_p(\mathcal{K})$ satisfying $w\ldotp g_l \approx_{\delta} g_{l + \lvert w\rvert}\ldotp w$ and $z\ldotp g_l\approx_{\delta} g_{l + \lvert z\rvert}\ldotp z$. If $k \in \mathbb N$ is such that $\max\{\lvert w\rvert,\lvert z\rvert\} + k < p$, then $$\lVert [R_{\mathcal G}^*R_{\mathcal G}, e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}}]\rVert < 2(p^2 + 2)\delta.$$
(1): For $(x_\zeta)_{\zeta\in W_{<p}} \in \bigoplus_{W_{<p}} \mathcal F(\mathcal K)$, $$\begin{aligned}
[T_\eta]_{W_k}^*g_k^2[T_\eta]_{W_k}(x_\zeta)_{\zeta\in W_{<p}}
&= [T_\eta]_{W_k}^*\sum_{\zeta \in W_k} g_k\cdot \zeta \otimes x_\zeta \\
&= (\chi_{W_k}(\zeta) g_k^2\cdot x_\zeta)_{\zeta \in W_{<p}},\end{aligned}$$ and therefore, $$\begin{aligned}
g_0^2|_{\mathcal{K}^{\otimes 0}} + \cdots + g_{p-1}^2|_{\mathcal{K}^{\otimes p-1}}
&= \sum_k [T_\eta]_{W_k}^*g_k^2[T_\eta]_{W_k} \\
&\approx_{p^2\delta} \sum_{k,k'} [T_\eta]_{W_k}^*g_kg_{k'}[T_\eta]_{W_{k'}} \\
&= R_{\mathcal G}^*R_{\mathcal G}.\end{aligned}$$ where $\zeta\in W_k$ and $x\in \mathcal{F}(\mathcal{K})$.
(2): Using (1) we have $$\begin{aligned}
(e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}})(R_{\mathcal G}^*R_{\mathcal G}) & \approx_{p^2\delta}(e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}})(g_0^2|_{\mathcal{K}^{\otimes 0}} + \cdots + g_{p-1}^2|_{\mathcal{K}^{\otimes p-1}}) \\
& = (e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}})(g^2_{k + \lvert z\rvert}|_{\mathcal{K}^{\otimes k + \lvert z\rvert}}) \\
& = e_{w,g^2_{k + \lvert z\rvert}\ldotp z}\otimes 1_{\mathcal{K}^{\otimes k}} \\
& \approx_{2\delta} e_{w,z\ldotp g_k^2}\otimes 1_{\mathcal{K}^{\otimes k}} \\
& = e_{w\ldotp g^2_k,z}\otimes 1_{\mathcal{K}^{\otimes k}} \\
& \approx_{2\delta} e_{g^2_{k + \lvert w\rvert}\ldotp w,z}\otimes 1_{\mathcal{K}^{\otimes k}} \\
& = (g^2_{k + \lvert w\rvert}|_{\mathcal{K}^{\otimes k + \lvert w\rvert}})(e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}}) \\
& = (g^2_0|_{\mathcal{K}^{\otimes 0}} + \cdots + g^2_{p-1}|_{\mathcal{K}^{\otimes p-1}})e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}} \\
& \approx_{p^2\delta} (R_{\mathcal G}^*R_{\mathcal G})(e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}}). \end{aligned}$$
### Countably generated correspondences {#technical}
Let $\mathcal{H}$ be a countably generated correspondence over $A$. By Kasparov’s Stabilization Theorem (Theorem \[Kasparov’s Stabilization Theorem\]), there is a countably generated free Hilbert $A$-module $\mathcal{H}'$ such that their direct sum $\mathcal{K} = \mathcal{H}\oplus\mathcal{H}'$ is free. Choose a left action of $A$ on $\mathcal{H}'$ such that $A\cap\mathbb{K}(\mathcal{H}') = \{0\}$ (one may take, for example, the canonical left action of $A$ on $\bigoplus_{\mathbb N} A$). The diagonal action of $A$ turns $\mathcal{K}$ into a correspondence. The orthogonal projection $P_{\mathcal{H}} \in \mathbb{B}(\mathcal{K})$ onto $\mathcal{H}\oplus 0$ commutes with the image of the left action of $A$ on $\mathcal{K}$, so for $k > 0$ the map $$P_{\mathcal{H}}^{\otimes k} : x_1\otimes\cdots\otimes x_k \mapsto P_{\mathcal{H}}x_1\otimes \cdots \otimes P_{\mathcal{H}}x_k$$ is an orthogonal projection in $\mathbb{B}(\mathcal{K}^{\otimes k})$.
Let $A$ be a $C^*$-algebra, let $\mathcal H$ be a countably generated correspondence over $A$, and let $\mathcal K$ and $P_{\mathcal H}$ be as above. Allowing $P_{\mathcal{H}}^{\otimes 0}$ to mean the identity map on $A$, we have
1. $\mathcal{H}^{\otimes k} = P_{\mathcal{H}}^{\otimes k}(\mathcal{K}^{\otimes k})$,
2. $\mathcal{F}_p(\mathcal{H}) = (\sum_{k=0}^{p-1}P_{\mathcal{H}}^{\otimes k})(\mathcal{F}_p(\mathcal{K}))$, and
3. $\mathcal{F}(\mathcal{H}) = (\sum_{k\ge 0}P_{\mathcal{H}}^{\otimes k})(\mathcal{F}(\mathcal{K}))$.
For the first part, the map $V_k: x_1\otimes \cdots \otimes x_k \mapsto (x_1, 0)\otimes \cdots\otimes (x_k, 0)$ implements an isomorphism between the $k$-fold tensor product $\mathcal{H}^{\otimes k}$ and the image of $P_{\mathcal{H}}^{\otimes k}$ in $\mathcal{K}^{\otimes k}$. The second part follows easily from the first since the sums are finite. Lastly, it is clear that the series $\sum_{k\ge 0}P_{\mathcal{H}}^{\otimes k}$ converges strictly to a projection in $\mathbb{B}(\mathcal{F}(\mathcal{K}))$ with the desired property.
Denote the projections $\sum_{k=0}^{p-1}P_{\mathcal{H}}^{\otimes k}$ and $\sum_{k\ge 0}P_{\mathcal{H}}^{\otimes k}$ given above by $P_{\mathcal{F}_p(\mathcal{H})}$ and $P_{\mathcal{F}(\mathcal{H})}$, respectively. If $z = (x_1,y_1)\otimes\cdots\otimes (x_k,y_k)\in \mathcal{K}^{\otimes k}$ is an elementary tensor, we say that $z$ is an *$\mathcal{H}$-elementary tensor* if $y_1 = \cdots = y_k = 0$. Equivalently, $z$ is $\mathcal{H}$-elementary if $P_{\mathcal{H}}^{\otimes k}(z) = z$.
\[sumuniversecuntz\] Let $(\pi,\tau)$ be the representation of $\mathcal{H}$ in $\mathcal{O}(\mathcal{K})$ given by $\pi: a\mapsto a$ and $\tau: x\mapsto S_{(x,0)}$. Then there is a \*-isomorphism $\theta: C^*(\pi,\tau) \rightarrow \mathcal{T}(\mathcal{H})$ sending $a$ to $a$ and $S_{(x,0)}$ to $T_x$.
Consider first the representation $(\tilde{\pi},\tilde{\tau})$ of $\mathcal{H}$ in $\mathcal{T}(\mathcal{K})$ given by $\tilde{\pi}: a\mapsto a$ and $\tilde{\tau}: x\mapsto T_{(x,0)}$. It is clear that $$\tilde{\pi}(A)\cap \overline{\mathrm{span}}\{\tilde{\tau}(x)\tilde{\tau}(y)^* \ : \ x,y\in \mathcal{H}\} = \{0\}.$$ If we restrict the gauge action on $\mathcal{T}(\mathcal{K})$ to $C^*(\pi,\tau)$, Theorem \[Toeplitz-Properties\] (2) gives a \*-isomorphism $\tilde{\theta}: C^*(\tilde{\pi},\tilde{\tau}) \rightarrow \mathcal{T}(\mathcal{H})$ sending $a$ to $a$ and $T_{(x,0)}$ to $T_x$. Now, since the left action of $A$ on $\mathcal{H}'$ was defined so that $A\cap \mathbb{K}(\mathcal{H}') = \{0\}$, each $a\in A\subset \mathbb{B}(\mathcal{K})$ differs from any finite sum of rank-one operators by at least $\lVert a\rVert$. Therefore, $A\cap \mathbb{K}(\mathcal{K}) = \{0\}$ and so by Remark \[Cuntz=Pimsner\] there is a \*-isomorphism $\bar{\theta}: \mathcal{O}(\mathcal{K}) \rightarrow \mathcal{T}(\mathcal{K})$ sending $a$ to $a$ and $S_{(x,y)}$ to $T_{(x,y)}$. Moreover, it’s clear that $\bar{\theta}(C^*(\pi,\tau)) = C^*(\tilde{\pi},\tilde{\tau})$ so taking $\theta = \tilde{\theta}\circ \bar{\theta}|_{\scriptscriptstyle{C^*(\pi,\tau)}}$ completes the proof. Here is a diagram illustrating the maps involved.
\(m) \[matrix of math nodes, row sep=4em, column sep=9em\] [ (A,) & C\^\*(,) ()\
() & C\^\*(,)()\
]{}; (m-1-1) edge node\[auto\] [$ (\pi,\tau)$]{} (m-1-2) (m-1-1) edge node\[auto,swap\] [$(a\mapsto a,x\mapsto T_x)$]{} (m-2-1) (m-1-2) edge node\[auto\] [$\bar{\theta}|_{\scriptscriptstyle{C^*(\pi,\tau)}}$]{} (m-2-2) (m-2-2) edge node\[auto\] [$\tilde{\theta}$]{} (m-2-1) (m-1-1) edge node\[descr\] [$ (\tilde{\pi},\tilde{\tau}) $]{} (m-2-2);
\[embed\] Let $A$ be a $C^*$-algebra, let $\mathcal H$ be a countably generated correspondence over $A$, and let $\mathcal K$ be as above. Let $D_p^0(\mathcal{K})$ be the $C^*$-subalgebra of $D_p(\mathcal{K})$ generated by elements of the form $e_{x,y}\otimes 1_{\mathcal{K}^{\otimes k}}$ where $x$ and $y$ are $\mathcal{H}$-elementary tensors. Then the map $$\gamma: e_{x_1\otimes\cdots\otimes x_k,y_1\otimes\cdots\otimes y_l}\mapsto e_{(x_1,0)\otimes\cdots\otimes (x_k,0), (y_1,0)\otimes\cdots\otimes(y_l,0)}$$ extends to an isomorphism $\gamma: D_p(\mathcal{H}) \rightarrow D_p^0(\mathcal{K})$. In particular if for each $p$, there is an approximate unit of projections in $\mathbb K(\mathcal F_p(\mathcal H))$ which is quasicentral in $D_p(\mathcal H)$, then $\dim_{\mathrm{nuc}}(D_p^0(\mathcal{K})) \le \dim_{\mathrm{nuc}}(A)$.
The map $\bar{V} = \sum_{k=0}^{p-1}V_k$ sending $a$ to $a$ and $x_1\otimes\cdots\otimes x_k$ to $(x_1,0)\otimes \cdots\otimes (x_k,0)$ extends to a unitary from $\mathcal{F}_p(\mathcal{H})$ to $P_{\mathcal{F}_p(\mathcal{H})}(\mathcal{F}_p(\mathcal{K}))$. Let $l,m,k \ge 0$ satisfy $0 \le \max\{l,m\} + k < p$ and let $x,y,w,z \in\mathcal{F}_p(\mathcal{H})$ be elementary tensors of lengths $l,m,m,$ and $k$, respectively. We have $$\begin{aligned}
\bar{V}^*e_{\bar{V}x,\bar{V}y}\otimes 1_{\mathcal{K}^{\otimes k}}\bar{V}(w\otimes z) & = \bar{V}^*(\bar{V}x\ldotp \langle \bar{V}y,\bar{V}w \rangle\otimes \bar{V}z)\\
& = \bar{V}^*(\bar{V}x\ldotp \langle y,w\rangle \otimes \bar{V}z) \\
& = e_{x,y}\otimes 1_{\mathcal{H}^{\otimes k}}(w\otimes z),\end{aligned}$$ so $\gamma$ is implemented by unitary conjugation and the result follows.
\[spock\_cor\] Let $A$ be a $C^*$-algebra, let $\mathcal H$ be a countably generated correspondence over $A$, and let $\mathcal K$ be as above. Let $\mathcal{G} = (g_0,\ldots,g_{p-1})$ be a tuple of positive contractions in $A$. There is a c.p. map $\bar\rho_{\mathcal G}:D_p(\mathcal{H}) \rightarrow \mathcal{T}(\mathcal{H})$ such that, for elementary tensors $x,y$ in $\mathcal{F}_p(\mathcal{H})$ and $k$ such that $\max\{\lvert x\rvert,\lvert y\rvert\}+k<p$, $$\label{barrhoeq}
\bar\rho_{\mathcal G}(e_{x,y}\otimes 1_{\mathcal{H}^{\otimes k}}) = g_{k + \lvert x\rvert}T_xT_y^*g_{k+\lvert y\rvert},$$ and the following commutes
\(m) \[matrix of math nodes, row sep=2em,column sep=8em, text height=1.5ex, text depth=0.25ex\] [ D\_p() & D\_p\^0()D\_p()\
() & C\^\*(,)()\
]{}; (m-1-1) edgenode\[auto\] [$\gamma$ (Lemma \[embed\])]{} (m-1-2) (m-1-1) edgenode\[auto\] [$\bar{\rho}_{\scriptscriptstyle{\mathcal{G}}}$]{} (m-2-1) (m-1-2) edgenode\[auto\] [$\rho_{\scriptscriptstyle{\mathcal{G}}}$ (Lemma \[maps\])]{} (m-2-2) (m-2-2) edgenode\[auto\] [$\theta$ (Prop. \[sumuniversecuntz\])]{} (m-2-1);
Define $\rho_{\mathcal G}:D_p(\mathcal K) \to \mathcal O(\mathcal K)$ by Lemma \[maps\] and $\gamma:D_p(\mathcal H) \to D_p^0(\mathcal K)$ from Lemma \[embed\]. Let $x=x_1 \otimes \cdots \otimes x_l$ and $y=y_1\otimes \cdots \otimes y_m$. We compute $$\begin{aligned}
\rho_{\mathcal G}(\gamma(e_{x,y} \otimes 1_{\mathcal H^{\otimes k}}))
&= \rho_{\mathcal G}(\gamma(e_{x_1\otimes\cdots\otimes x_l,y_1\otimes\cdots\otimes y_m} \otimes 1_{\mathcal H^{\otimes k}})) \\
&= \rho_{\mathcal G}(e_{(x_1,0)\otimes\cdots\otimes (x_l,0),(y_1,0)\otimes\cdots\otimes (y_m,0)} \otimes 1_{\mathcal K^{\otimes k}}) \\
&= g_{k + l}S_{(x_1,0)\otimes\cdots\otimes (x_l,0)}S_{(y_1,0)\otimes\cdots\otimes(y_m,0)}^*g_{k + m},\end{aligned}$$ and we see that the result is in $C^*(\pi,\tau)$ where $(\pi,\tau)$ is as in Proposition \[sumuniversecuntz\]. Hence, the image of $\rho_{\mathcal G} \circ\gamma$ is contained in $C^*(\pi,\tau)$ and using $\theta:C^*(\pi,\tau) \to \mathcal T(\mathcal H)$ defined by Proposition \[sumuniversecuntz\], we may define $$\bar\rho_{\mathcal G} = \theta \circ \rho_{\mathcal G} \circ \gamma:D_p{\mathcal H} \to \mathcal T(\mathcal H).$$ We further compute $$\begin{aligned}
\bar\rho_{\mathcal G}(\gamma(e_{x,y} \otimes 1_{\mathcal H^{\otimes k}}))
&= \theta(g_{k + l}S_{(x_1,0)\otimes\cdots\otimes (x_l,0)}S_{(y_1,0)\otimes\cdots\otimes(y_m,0)}^*g_{k + m}) \\
&= g_{k + l}T_{x_1\otimes \cdots\otimes x_l}T_{y_1\otimes\cdots\otimes y_m}^*g_{k + m},\end{aligned}$$ as required.
Let $A$ be a $C^*$-algebra, let $\mathcal H$ be a countably generated correspondence over $A$. Consider a tuple $\mathcal G=(g_0,\dots,g_{p-1})$ of positive contractions from $A_\infty$. We may lift this to a sequence $(\mathcal G_i)$ of tuples of positive contractions from $A$. Using this lift and Corollary \[spock\_cor\], we define $$\label{barrhoDef}
\bar\rho_{\mathcal G} = (\rho_{\mathcal G_i}):D_p(\mathcal H) \to \mathcal O(\mathcal H)_\infty.$$ Note that this is independent of the choice of the lift $(\mathcal G_i)$.
\[lulu\] Let $A$ be a separable $C^*$-algebra, let $\mathcal H$ be a countably generated correspondence over $A$, and let $p \in \mathbb N$. If $\mathcal{G} = (g_0,\ldots,g_{p-1})$ is a tuple of orthogonal positive contractions in $A_\infty$ satisfying $z \ldotp g_k = g_{k + \lvert z\rvert}\ldotp z$ for all $k$ and all elementary tensors $z\in \mathcal F_p(\mathcal H)$ then the c.p. map $\bar{\rho}_{\scriptscriptstyle{\mathcal{G}}}$ from is contractive and order zero.
Analogously to , define $\rho_{\mathcal G} = (\rho_{\mathcal G_i}):D_p(\mathcal K) \to \mathcal O(\mathcal K)_\infty$ and $\sigma_{\mathcal G} = (\sigma_{\mathcal G_i}):D_p(\mathcal K) \to \mathbb B(\mathcal F(\mathcal K))_\infty$. We have $$\begin{aligned}
\bar\rho_{\mathcal G} &= \theta_\infty \circ \rho_{\mathcal G} \circ \gamma \text{ and} \\
\rho_{\mathcal G} &= Q_\infty \circ \sigma_{\mathcal G}.\end{aligned}$$ By Lemma \[moremaps\] (1) and the hypothesis, $$\lVert [R_{\mathcal G_i}^*R_{\mathcal G_i}, e_{w,z}\otimes 1_{\mathcal{K}^{\otimes k}}]\rVert \to 0$$ for all $\mathcal H$-elementary tensors $w,z\in \mathcal F(\mathcal{K})$. By the definition of $\gamma$ (from Lemma \[embed\]), this implies that $$\lVert [R_{\mathcal G_i}^*R_{\mathcal G_i}, \gamma(x)] \rVert \to 0$$ for all $x \in D_p(\mathcal H)$. By the formula for $\sigma_{\mathcal G_i}$ (from Lemma \[maps\]), it follows that $\sigma_{\mathcal G}$ is order zero. Consequently, $\rho_{\mathcal G} \circ \gamma$, and therefore also $\bar\rho_{\mathcal G}$, is order zero.
To see that $\bar\rho_{\mathcal G}$ is contractive, we have by Lemma \[moremaps\] (2) that $$\|R_{\mathcal G_i}^*R_{\mathcal G_i} - \sum_k g_{k,i}^2|_{\mathcal K^{\otimes k}}\| \to 0,$$ where $\mathcal G_i = (g_{0,i},\dots,g_{p-1,i})$. Since the $g_i$ are orthogonal and contractive, $\sum_i g_i^2$ is contractive. It follows from the formula for $\sigma_{\mathcal G_i}$ that $\sigma_{\mathcal G}$ is contractive. Hence, so is $\bar\rho_{\mathcal G}$.
Proof of the main theorem {#Proof}
-------------------------
We have one technical lemma before the main result. For $p \in \mathbb N$ and $k =0,\dots,p-1$, set $$\label{d_kDef}
d_p(k) = 1 - \frac{\lvert p-1-2k\rvert}{p-1} \in [0,1]$$\
.
\[bumplemma\] Let $p$ be an odd integer. If $f_0,\ldots f_{p-1}$ are positive contractions in a $C^*$-algebra $A$, then for any $0 \le N\le p-1$, $$\label{bump}
\sum_{k = N}^{p-1}d_p(k).\left(f_k + f_{\frac{p-1}{2} + k\ (\mathrm{mod}\ p)}\right) \ \approx_{\frac{4N^2}{p-1}} \sum_{k=0}^{p-1}f_k.$$
It’s easy to verify that $$d_p(k) + d_p(\tfrac{p-1}{2} + k\ (\mathrm{mod}\ p)) = 1 \ \text{for}\ k =0,\dots,p-1$$ so that $$\sum_{k=0}^{p-1} d_p(k)\left(f_k + f_{\frac{p-1}{2} + k (\mathrm{mod}\ p)}\right)
= \sum_{k=0}^{p-1}f_k.$$ Moreover, $$\lVert\sum_{k=0}^{N-1}d_p(k)\left(f_k + f_{\frac{p-1}{2} + k}\right)\rVert \le 2\sum_{k=0}^{N-1}d_p(k) \le 2N\left(1 - \frac{p-1-2N}{p-1}\right) = \frac{4N^2}{p-1},$$ so the result follows.
We now have the ingredients to prove the main result.
\[main\_thm\] Suppose that $\mathcal{H}$ is a countably generated $C^*$-correspondence over a separable unital $C^*$-algebra $A$. Assume further that for every $p\in \mathbb{N}$, there is an approximate unit consisting of projections in $\mathbb{K}(\mathcal{F}_p(\mathcal{H}))$ that are quasicentral in $D_p(\mathcal{H})$. Then $$(\dim_{\mathrm{nuc}}(\mathcal{T}(\mathcal{H})) + 1) \le 2(\dim_{\mathrm{nuc}}(A)+1)(\dim_{\mathrm{Rok}}(\mathcal{H})+1).$$
Example \[QD-examples\] gives two important examples of classes of correspondences which satisfy the technical hypothesis of this theorem, namely, that for every $p$, there is an approximate unit of projections in $\mathbb K(\mathcal F_p(\mathcal H))$ which is quasicentral in $D_p(\mathcal H)$.
If either $\dim_{\mathrm{nuc}}(A)$ or $\dim_{\mathrm{Rok}}(\mathcal{H},A) = \infty$ there is nothing to show. Otherwise, let $\dim_{\mathrm{nuc}}(A) = n$ and $\dim_{\mathrm{Rok}}(\mathcal{H}) = d$. We will use Lemma \[nucdimtool\] with $m=d+1$. Therefore let $F$ be a finite subset of $\mathcal{T}(\mathcal{H})$ and fix $\epsilon > 0$.
Noting , we may assume without loss of generality that $F$ consists of elements of the form $T_xT_y^*$ where $x,y$ are elementary tensors. Pick $N$ such that every element of $F$ is $T_xT_y^*$ for some elementary tensors $x,y \in \mathcal F_{N+1}(\mathcal H)$. Let $p \in \mathbb{N}$ be large enough so that $$\label{pestimate}
(d+1)\left(2\sqrt{\frac{2N}{p-1}} + \frac{4N^2}{p-1}\right) < \epsilon$$ and let $P\in \mathbb{B}(\mathcal{F}(\mathcal{H}))$ be the projection onto the $p^{\text{th}}$-cutoff Fock space $\mathcal{F}_p(\mathcal{H})$.
For $k=0,\dots,p-1$, use $d_p(k)$ as defined in ; for convenience, we set $d_p(k)=0$ for $k \geq p$. Using these, set $\Delta = \text{diag}(d_p(0),\ldots,d_p(p-1))$ and let $\phi: \mathcal{T}(\mathcal{H}) \rightarrow \mathbb B(\mathcal F_p(\mathcal H))$ be compression by $\sqrt{\Delta} P$; that is, for elementary tensors $\zeta,x,y$, $$\phi(T_xT_y^*)(\zeta) =
\begin{cases}
d_p(|\zeta|)d_p(|\zeta|+|x|-|y|) x \otimes \langle y,\zeta_1\rangle \ldotp \zeta_2, \quad &\zeta=\zeta_1\otimes \zeta_2, |\zeta_1|=|y|; \\ 0, \quad &|\zeta|<|y| \end{cases}$$ (here it is convenient that $d_p(k)=0$ for $k \geq p$). It is not hard to see that $\phi(\mathcal T(\mathcal H))$ is contained in $D_p(\mathcal H)$.
Find Rokhlin contractions $\{f_k^l\}_{l = 0,\ldots,d;\ k \in \mathbb Z/p}$ in $A_\infty$ satisfying
1. $f_k^lf_{k'}^l=0$ for all $l$ and $k\ne k'$,
2. $\sum_{k,l}f_k^l =1$,
3. $z\ldotp f_k^l = f_{k + \lvert z\rvert}^l\ldotp z$ for all $k,l$, and all elementary tensors $z \in \mathcal F(\mathcal H)$.
Note that (3) implies that $z\ldotp (f_k^l)^m = (f_{k + \lvert z\rvert}^l)^m\ldotp z$ for all $m \in \mathbb N$, which in turn implies that $z\ldotp g(f_k^l) = g(f_{k + \lvert z\rvert}^l)\ldotp z$ for any continuous function $g \in C([0,1])$. In particular, $$\label{Rokhlinsqrt}
z\ldotp (f_k^l)^{1/2} = (f_{k + \lvert z\rvert}^l)^{1/2}\ldotp z$$ For $l=0,\dots,d$, set $$\mathcal{G}^l = ((f_0^l)^{1/2},\ldots,(f_{p-1}^l)^{1/2}) \ \ \ \text{and} \ \ \ \hat{\mathcal{G}}^l = ((f_{\frac{p-1}{2} + 0}^l)^{1/2},\ldots,(f_{\frac{p-1}{2} + p-1}^l)^{1/2}),$$ two tuples of orthogonal positive elements in $A_\infty$. Define $\bar{\rho}_{\scriptscriptstyle{\mathcal{G}^l}}$ and $\bar{\rho}_{\scriptscriptstyle{\hat{\mathcal{G}}^l}}$ as in Corollary \[spock\_cor\]; using , Corollary \[lulu\] tells us that these maps are c.p.c. and order zero.
To simplify notation, write $\rho^l$ and $\hat{\rho}^l$ for $\bar{\rho}_{\scriptscriptstyle{\mathcal{G}^l}}$ and $\bar{\rho}_{\scriptscriptstyle{\hat{\mathcal{G}}^l}}$, respectively. Thus we have the following diagram.
\(m) \[matrix of math nodes, row sep=2em,column sep=8em, text height=1.5ex, text depth=0.25ex\] [ T(H) & & T(H)\_\
& D\_p(H) &\
]{}; (m-1-1) edgenode\[auto\] (m-1-3) (m-1-1) edgenode\[auto,swap\] [$\phi$]{} (m-2-2) (m-2-2) edgenode\[auto,swap\] [$\sum_l (\rho_l + \hat\rho_l)$]{} (m-1-3);
By Lemma \[quasi\_hypothesis\], $D_p(\mathcal H)$ has nuclear dimension at most $n$. We shall show that this diagram commutes, up to $\epsilon$ on $F$. As $\phi$ is c.p.c. and each $\sum_l (\rho_l+\hat\rho_l)$ is overtly a sum of $2(d+1)$ c.p.c. order zero maps, this will finish verifying the hypotheses of Lemma \[nucdimtool\], and thus finish the proof.
Consider an element of $F$, necessarily of the form $T_xT_y^*$, $x,y \in \mathcal F_{N+1}(\mathcal H)$. For the moment, assume that $\lvert x\rvert \ge \lvert y\rvert$. Applying $\phi$ to $T_xT_y^*$, using , and noting that $\sqrt{d_p(k +\lvert x\rvert)}$ is within $\sqrt{\frac{2N}{p-1}}$ of $\sqrt{d_p(k + \lvert y\rvert)}$ for $k \in \mathbb Z/p$, we have $$\begin{aligned}
\phi(T_xT_y^*) &= \sum_{k = 0}^{p-1- \lvert x\rvert}\sqrt{d_p(k + \lvert x\rvert)d_p(k + \lvert y\rvert)}e_{x,y}\otimes 1_{\mathcal{H}^{\otimes k}}\notag\\
\label{equationcutdown}
& \approx_{\sqrt{\frac{2N}{p-1}}} \sum_{k=0}^{p-1-\lvert x\rvert}d_p(k+\lvert x\rvert)e_{x,y}\otimes 1_{\mathcal{H}^{\otimes k}},\end{aligned}$$ where the error estimate in the last line is a maximum (rather than a sum) because of orthogonality.
Next, applying $\rho^l$ to $\phi(T_xT_y^*)$, we get $$\begin{aligned}
\notag
\rho^{l}\circ\phi(T_xT_y^*)
&\stackrel{\eqref{equationcutdown}}\approx_{\sqrt{\frac{2N}{p-1}}} \sum_{k=0}^{p-1-\lvert x\rvert} d_p(k + \lvert x\rvert) \rho^l(e_{x,y} \otimes 1_{\mathcal H^{\otimes k}}) \\
\notag
&\stackrel{\eqref{barrhoeq}}= \sum_{k=0}^{p-1-\lvert x\rvert}d_p(k + \lvert x\rvert) (f_{k + \lvert x\rvert}^l)^{1/2}T_xT_y^*(f_{k + \lvert y\rvert}^l)^{1/2} \\
\notag
&= \sum_{k=\lvert x\rvert}^{p-1}d_p(k) (f_{k}^l)^{1/2}T_xT_y^*(f_{k - \lvert x\rvert+ \lvert y\rvert}^l)^{1/2} \\
\notag
&= \sum_{k=\lvert x\rvert}^{p-1}d_p(k) (f_{k}^l)^{1/2}T_x(f_{k - \lvert x\rvert}^l)^{1/2}T_y^* \\
\label{rhol-eq1}
&= (\sum_{k=\lvert x\rvert}^{p-1}d_p(k)f_k^l) T_xT_y^*,\end{aligned}$$ and likewise, $$\label{rhol-eq2}
\hat{\rho}^l\circ\phi(T_xT_y^*) \approx_{\sqrt{\frac{2N}{p-1}}} (\sum_{k=\lvert x\rvert}^{p-1}d_p(k)f_{\frac{p-1}{2} + k}^l) T_xT_y^*.$$ Summing these terms, we obtain $$\begin{aligned}
\sum_{l=0}^d (\rho^l + \hat{\rho}^l)\circ \phi(T_xT_y^*) \hspace*{-3em}\,
&\stackrel{\eqref{rhol-eq1},\eqref{rhol-eq2}}{_{\phantom{2(d+1)\sqrt{\frac{2N}{p-1}}}}\approx_{2(d+1)\sqrt{\frac{2N}{p-1}}}}& \sum_{l=0}^d \sum_{k=|x|}^{p-1} d_p(k) (f_k^l + f_{\frac{p-1}2+k}^l)T_xT_y^* \\
&\stackrel{\text{Lemma \ref{bumplemma}}}{_{\phantom{{(d+1)\frac{4N^2}{p-1}}}}\approx_{(d+1)\frac{4N^2}{p-1}}}& \sum_{l=0}^d \sum_{k=0}^{p-1} f_k^l T_xT_y^* \\
&=& T_xT_y^*.\end{aligned}$$ By , this yields $$\sum_{l=0}^d (\rho^l + \hat{\rho}^l)\circ \phi(T_xT_y^*) \approx_\epsilon T_xT_y^*,$$ as required. A nearly identical argument shows the same estimate in the case $\lvert x\rvert < \lvert y\rvert$.
\[Main-Theorem\_Cor\] Under the same hypotheses as the previous theorem, $$(\dim_{\mathrm{nuc}}(\mathcal{O}(\mathcal{H})) + 1) \le 2(\dim_{\mathrm{nuc}}(A)+1)(\dim_{\mathrm{Rok}}(\mathcal{H})+1).$$
This follows from the fact that $\mathcal{O}(\mathcal{H})$ is a quotient of $\mathcal{T}(\mathcal{H})$ by the ideal $\mathbb{K}(\mathcal{F}(\mathcal{H})I_{\mathcal{H}})$, and by [@Winter-Zacharias2 Proposition 2.3].
Nuclear dimension of certain free products {#Amalgamated}
==========================================
We now use the results given in the previous sections to deduce finite nuclear dimension of a certain class of reduced amalgamated free products. For a more comprehensive treatment of free products, see the monograph [@Voiculescu-Dykema-Nica]. We start with a well-known example.
\[free-group\] There is a \*-isomorphism $$C(\mathbb{T})\text{\large{\textasteriskcentered}} C(\mathbb{T}) \cong C^*_r(\mathbb{F}_2).$$ Here the free product is being taken with respect to the state $f\mapsto \int_{\mathbb{T}}f(z)dz$ on $C(\mathbb{T})$.
The free group $\mathbb{F}_2$ on two generators is not amenable, so the reduced group $C^*$-algebra is not nuclear and in particular has infinite nuclear dimension. What this shows is that unlike other canonical $C^*$-constructions, finite nuclear dimension is not in general preserved under the reduced amalgamated free product construction, even in the abelian case. However, the next result by Speicher from [@Speicher] shows that there are exceptions.
\[Speicher\] Let $\mathcal{H}_i$ be correspondences over $A$. Then $$(\mathcal{T}(\bigoplus_I\mathcal{H}_i), E_{\bigoplus \mathcal{H}_i}) \cong \text{\Large{\textasteriskcentered}}_A(\mathcal{T}(\mathcal{H}_i), E_{\mathcal{H}_i}).$$
\[2-Main-Theorem\] Let $\mathcal{H}$ be a correspondence over $A$ and let $I$ be an arbitrary (countable) set. If $\dim_{\mathrm{Rok}}(\mathcal{H}) < \infty$, then $\dim_{\mathrm{Rok}}(\bigoplus_I\mathcal{H}) < \infty$.
The left action of $A$ on $\bigoplus_I\mathcal{H}$ is given by $a\ldotp (x_i)_{i\in I} = (a\ldotp x_i)_{i\in I}$. Let $\epsilon > 0$, $p \in \mathbb{N}$, $F$ a finite subset of $A$ and $\mathcal{V}$ a finite subset of $\bigoplus_I\mathcal{H}$. Since the finitely supported elements of $\bigoplus_I \mathcal H$ are dense, we may assume that $\mathcal V$ consists only of finitely supported elements; let $N\in \mathbb N$ be such that each is supported on at most $N$ elements. Set $\mathcal V'$ equal to the set of all elements in $\mathcal H$ which appear in some component of some element of $\mathcal V$; this is a finite set. Find Rokhlin contractions $(f_k^l)_{k\in \mathbb Z/p}^{l=0,\ldots,d} \in A$ for the correspondence $\mathcal H$, with respect to $(\epsilon/N,p,F,\mathcal{V}')$. For $x=(x_i)_{i\in I} \in \mathcal V$, $$f_k^l\ldotp (x_i) = (f_k^l\ldotp x_i) \approx_{N\epsilon/N} (x_i\ldotp f_{k+1}^l) = (x_i) \ldotp f_{k+1}^l.$$ Thus, the $f_k^l$ are Rokhlin contractions for $\bigoplus_I \mathcal H$, with respect to $(\epsilon,p,F,\mathcal V)$.
For countably generated Hilbert modules $\mathcal H_1$ and $\mathcal H_2$, is it the case that $$\dim_{\mathrm{Rok}}(\mathcal H_1 \oplus \mathcal H_2) = \max\{ \dim_{\mathrm{Rok}}(\mathcal H_1), \dim_{\mathrm{Rok}}(\mathcal H_2)\}?$$
In light of Theorem \[Main-Theorem\] and Proposition \[2-Main-Theorem\], we obtain the following statement.
\[Second-Main-Theorem\] Let $\mathcal{H}$ be a finitely generated projective correspondence over $A$. If $\dim_{\mathrm{nuc}}(A)$ and $\dim_{\mathrm{Rok}}(\mathcal{H})$ are both finite, then the amalgamated free product of any finite number of copies of $\mathcal{T}(\mathcal{H})$ (with respect to the canonical expectation $E_{\mathcal{H}}$) has finite nuclear dimension.
In comparison with Example \[free-group\], the following example demonstrates an interesting consequence of Theorem \[Second-Main-Theorem\].
If $\varphi$ is a minimal homeomorphism of $\mathbb{T}$, then $\dim_{\mathrm{Rok}}(\varphi^*)$ is finite. By Theorem \[Second-Main-Theorem\], the nuclear dimension of $$\mathcal{T}(C(\mathbb{T})^{\varphi^*})\text{\large{\textasteriskcentered}}\mathcal{T}(C(\mathbb{T})^{\varphi^*}) \cong \mathcal{T}(C(\mathbb{T})\oplus C(\mathbb{T}))$$ is also finite. Here the free product is being taken with respect to the usual conditional expectation $E_{C(\mathbb{T})\oplus C(\mathbb{T})}$.
Classifiability of certain Cuntz–Pimsner algebras {#classifiability}
=================================================
In this section we show that in the presence of finite Rokhlin dimension, the algebras $\mathcal{O}(\mathcal{H})$ are often *classifiable*, by which we mean unital and simple with finite nuclear dimension and satisfying the UCT (cf. [@Tikuisis-White-Winter]). “Satisfying the UCT” is a $KK$-theoretic property for separable $C^*$-algebras, equivalent to being $KK$-equivalent to an abelian $C^*$-algebra, introduced by Rosenberg and Schochet in [@Rosenberg-Schochet]. The class of algebras which satisfy the UCT (sometimes called the *bootstrap class*) is closed under a number of natural operations (see [@Blackadar:KBook §22.3 and §23]), although it is unknown whether every separable nuclear $C^*$-algebra is in this class.
\[classif\_lem\_1\] For simple $C^*$-algebras, if $A$ satisfies the UCT, so does $\mathcal{O}(\mathcal{H})$.
By [@Pimsner Theorem 4.4], the Toeplitz Pimsner algebra $\mathcal{T}(\mathcal{H})$ is $KK$-equivalent to $A$. Consequently, $\mathcal{T}(\mathcal{H})$ satisfies the UCT whenever $A$ does. Since $\mathcal{O}(\mathcal{H})$ is a quotient of $\mathcal{T}(\mathcal{H})$ by the ideal $\mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}})$, it suffices to prove (by the two-out-of-three principle) that $\mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}})$ satisfies the UCT. To this end observe that $\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}}$ is countably generated, so Kasparov’s stabilization result implies $$\mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}}\oplus\ell^2(A))\cong \mathbb{K}(\ell^2(A)) \cong A\otimes\mathbb{K}\sim_{\mathrm{Morita}} A,$$ where $\sim_{\mathrm{Morita}}$ denotes Morita equivalence of $C^*$-algebras. Since $A$ is simple, so is $A\otimes \mathbb{K}\cong \mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}}\oplus\ell^2(A))$. This means every hereditary $C^*$ subalgebra $B\subseteq \mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}}\oplus\ell^2(A))$ is full, and hence satisfies $$B\sim_{\mathrm{Morita}} \mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}}\oplus\ell^2(A))\sim_{\mathrm{Morita}} A.$$ The result follows because $\mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}})$ is hereditary in $\mathbb{K}(\mathcal{\mathcal{F}(\mathcal{H})I_{\mathcal{H}}}\oplus\ell^2(A))$, and because the UCT is preserved under Morita equivalence.
Next, following Schweizer, we deal with simplicity of $\mathcal{O}(\mathcal{H})$.
$\mathcal{H}$ is called *minimal* if there are no nontrivial ideals $J \subset A$ such that $\{\langle x,j\ldotp y\rangle \ : \ x,y\in \mathcal{H},\ j\in J\}\subseteq J$. $\mathcal{H}$ is *nonperiodic* if $\mathcal{H}^{\otimes k}\approx A \Rightarrow k = 0$, where we regard $A$ as the identity correspondence over itself.
Generalizing the well-known fact that $C(X)\rtimes_{\sigma}\mathbb{Z}$ is simple if and only if the homeomorphism $\sigma$ is minimal, Schweizer proved the following beautiful result.
\[thm:schweizer\] $\mathcal{O}(\mathcal{H})$ is simple if and only if $\mathcal{H}$ is minimal and nonperiodic.
Note that minimality is automatic whenever the scalar algebra $A$ is simple, so in this case, we are left to worry about periodicity.
\[classif\_lem\_2\] If $\mathcal{H}$ has finite Rokhlin dimension, it is nonperiodic.
Assume $\dim_{\text{Rok}}(\mathcal{H}) \le d$ and that there is some $k > 0$ such that $\mathcal{H}^{\otimes k}\approx A$. This means there is an adjointable unitary bimodule map $U:\mathcal{H}^{\otimes k} \rightarrow A$. Let $v = U^{-1}(1_A)$ and fix $\epsilon > 0$. It is straightforward to check that $\dim_{\text{Rok}}(\mathcal{H}) \le d$ implies $\dim_{\text{Rok}}(\mathcal{H}^{\otimes k}) \le d$, so we can find a set of positive contractions $\{f_1^l,f_2^l\}_{l = 0,\ldots,d}\subset A$ satisfying the following three estimates:
1. $\lVert f_1^lf_2^l\rVert < \epsilon$ for every $l$,
2. $\lVert\sum_{l=0}^d(f_1^l + f_2^l) - 1_A\rVert < \epsilon$, and
3. $\lVert v\ldotp f_1^l - f_2^l\ldotp v\rVert < \epsilon$ for every $l$.
Applying $U$ to the third estimate gives $\lVert f_1^l - f_2^l\rVert < \epsilon$ for every $l$. Combining this with the first estimate gives $\lVert f_i^l\rVert < \sqrt{2\epsilon}$ for $i = 1,2$. But this implies $\lVert \sum_{l=0}^d f_1^l + f_2^l\rVert < 2(d+1)\sqrt{2\epsilon}$, which contradicts the second estimate when $\epsilon$ is sufficiently small. Thus, $\mathcal{H}$ must be nonperiodic.
Combining Theorem \[thm:schweizer\] and the previous lemma we see that for $C^*$-correspondences with finite Rokhlin dimension, $\mathcal{O}(\mathcal{H})$ is simple and unital whenever $A$ is. For simple $A$, by Lemma \[classif\_lem\_1\], $\mathcal{O}(\mathcal{H})$ satisfies the UCT whenever $A$ does. Thus applying Theorem \[main\_thm\], we obtain the following corollary.
\[application2\] If $A$ is classifiable, $\dim_{\text{Rok}}(\mathcal{H}) <\infty$ and for every $p\in \mathbb{N}$ there is an approximate unit consisting of projections in $\mathbb{K}(\mathcal{F}_p(\mathcal{H}))$ that are quasicentral in $D_p(\mathcal{H})$ (e.g., if $\mathcal{H}$ is finitely generated projective), then $\mathcal{O}(\mathcal{H})$ is classifiable.
Suppose $A$ is a unital Kirchberg algebra and $\alpha:A\rightarrow A$ is an automorphism with finite Rokhlin dimension (such automorphisms are generic by [@Hirshberg-Winter-Zacharias Theorem 3.4]). If $\mathcal{H}$ is a countably generated free $A$-module with basis $(\xi_i)_{i=1}^{\infty}$, we can define a correspondence with the left action $a\ldotp \xi_i = \xi_i\ldotp \alpha(a)$. Note that $A\cap \mathbb{K}(\mathcal{H}) = \{0\}$, so Remark \[Cuntz=Pimsner\] and the $KK$-equivalence of $A$ and $\mathcal{T}(\mathcal{H})$ imply $A$ is $KK$-equivalent to $\mathcal{O}(\mathcal{H})$. Moreover, $\dim_{\text{Rok}}(\mathcal{H}) < \infty$, so $\mathcal{O}(\mathcal{H})$ is a simple, separable $C^*$-algebra with finite nuclear dimension. Since the creation operator associated to $\xi_1$ is a proper isometry, $\mathcal{O}(\mathcal{H})$ must be a Kirchberg algebra. Invoking Kirchberg–Phillips classification, we conclude that $\mathcal{O}(\mathcal{H})\cong A$.
We end this paper by pointing out that our work does not recapture the interesting examples of classifiable Cuntz–Pimsner algebras constructed by Kumjian in [@Kumjian]. Indeed, it can be shown that they don’t have finite Rokhlin dimension.
[10]{}
B. Blackadar. *[$K$]{}-theory for operator algebras*, second edition, Mathematical Sciences Research Institute Publications, **5**, Cambridge University Press, Cambridge (1998).
L.G. Brown, *Stable isomorphism of hereditary subalgebras of $C^*$-algebras*, Pacific Journal of Mathematics, Vol. 71, No. 2, 1977.
N.P. Brown and N. Ozawa. *$C^*$-algebras and finite-dimensional approximations*, Graduate Studies in Mathematics, **88**, American Mathematical Society, Providence, RI, 2008.
N.P. Brown, J. Bosa, Y. Sato, A. Tikuisis, S. White and W. Winter, *Covering dimension of $C^*$-algebras and $2$-coloured classification*, to appear in Mem. Amer. Math. Soc. arXiv:1506.03974.
A. Connes, *Classification of injective factors: cases II$_1$, II$_{\infty}$, III$_{\lambda}, \ \lambda \ne 1$*, Ann. of Math. (2) **104** (1976), 73-115.
K.J. Dykema and D. Shlyakhtenko, *Exactness of Cuntz–Pimsner $C^*$-algebras*, Proc. Edinb. Math. Soc. (2) **44** (2001), no. 2, 425-444.
G.A. Elliott, G. Gong, H. Lin and Z. Niu, *On the classification of simple amenable $C^*$-algebras with finite decomposition rank, II*. arXiv:1507.03437.
N.J. Fowler, P.S. Muhly and I. Raeburn, *Representations of Cuntz–Pimsner algebras*, Indiana Univ. Math J. **52**(2003), no. 3, 569-605.
P. Green, *The local structure of twisted covariance algebras*, Acta Math. **140** (1978), no. 3-4, 191-250.
I. Hirshberg, W. Winter and J. Zacharias. *Rokhlin dimension and $C^*$-dynamics*, Comm. Math. Phys. **335** (2015), no. 2, 637-570.
E. Kirchberg and W. Winter. *Covering dimension and quasidiagonality*, Internat. J. Math. **15** (2004), no. 1, 63-85.
A. Kishimoto *The Rohlin property for shifts on UHF algebras and automorphisms of Cuntz algebras*, J. Funct. Anal. **140** (1996), no. 1, 100-123.
A. Kumjian, *On certain Cuntz–Pimsner algebras*, Pacific Journal of Mathematics, Vol. 217, No. 2, 2004.
H. Lin, *Stable approximately unitary equivalence of homomorphisms*, J. Operator Theory, **47** (2002), 343-378.
H. Matui and Y. Sato, *Decomposition rank of UHF-absorbing $C^*$-algebras*, Duke Math. J. **163** (2014), no. 14, 2687-2708.
M.V. Pimsner, *A class of $C^*$-algebras generalizing both Cuntz–Krieger algebras and crossed products by $\mathbb{Z}$*, Free Probability Theory (Waterloo, ON, 1995), Fields Inst, Commun., **12**, American Mathematical Society, Providence, RI, (1997), 189-212.
J. Rosenberg and C. Schochet, *The Künneth theorem and the universal coefficient theorem for Kasparov’s generalized K-functor*, Duke Math. J. **55** (1987), 431-474.
E. Ruiz, A. Sims, and A. Sørensen, *UCT-Kirchberg algebras have nuclear dimension one*, Adv. Math. **279** (2015), 1-28.
Y. Sato, S. White and W. Winter, *Nuclear dimension and $\mathcal{Z}$-stability*, Invent. Math **580** (2015), DOI: 10.1007/s00222-015-0580-1
J. Schweizer, *Dilations of $C^*$-correspondences and the simplicity of Cuntz–Pimsner algebras*, Journal of Functional Analysis *180*, 404-425 (2001).
R. Speicher, *Combinatorial theory of the free product with amalgamation and operator-valued free probability*, Mem. Amer. Math. Soc. **132** (1998), no. 627.
G. Szabó: *The Rokhlin dimension of topological $\mathbb{Z}^m$-actions*, Proc. Lond. Math. Soc. (3) **110** (2015), no. 3, 673-694.
A. Tikuisis, S. White, and W.Winter, *Quasidiagonality of nuclear $C^*$-algebras*, to appear in Ann. of Math. (2). arXiv:1509.08318. A. Tikuisis and W. Winter, *Decomposition rank of $\mathcal Z$-stable $\mathrm C^*$-algebras.* Anal. PDE **7** (2014), no. 3, pp. 673-700.
D.V. Voiculescu, K.J. Dykema and A. Nica, *Free random variables, A noncommutative probability approach to free products with applications to random matrices, operator algebras and harmonic analysis on free groups*. CRM Monograph Series, **1**, American Mathematical Society, Providence, RI, (1992).
N.E. Wegge-Olsen, *$K$-theory and $C^*$-algebras. A friendly approach*, Oxford Science Publications, The Clarendon Press, Oxford University Press, New York, (1993).
W. Winter and J. Zacharias. *Completely positive maps of order zero*, Münster J. Math **2** (2009), 311-324.
W. Winter and J. Zacharias. *The nuclear dimension of $C^*$-algebras*, Adv. Math. **224** (2010), no. 2, 461-498.
[^1]: N.B. and A.Z. were partially supported by NSF grant DMS-1201385. A.T. was partially supported by an NSERC Postdoctoral Fellowship and EPSRC grant EP/N00874X/1.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This article investigates the signature of the seventeen multi-connected flat spaces in cosmic microwave background (CMB) maps. For each such space it recalls a fundamental domain and a set of generating matrices, and then goes on to find an orthonormal basis for the set of eigenmodes of the Laplace operator on that space. The basis eigenmodes are expressed as linear combinations of eigenmodes of the simply connected Euclidean space. A preceding work, which provides a general method for implementing multi-connected topologies in standard CMB codes, is then applied to simulate CMB maps and angular power spectra for each space. Unlike in the 3-torus, the results in most multi-connected flat spaces depend on the location of the observer. This effect is discussed in detail. In particular, it is shown that the correlated circles on a CMB map are generically not back-to-back, so that negative search of back-to-back circles in the WMAP data does not exclude a vast majority of flat or nearly flat topologies.'
author:
- Alain Riazuelo
- Jeffrey Weeks
- 'Jean-Philippe Uzan'
- Roland Lehoucq
- 'Jean-Pierre Luminet'
date: 13 November 2003
title: 'Cosmic microwave background anisotropies in multi-connected flat spaces'
---
Introduction
============
Among all multi-connected three-dimensional spaces, “flat spaces”[^1] have been studied the most extensively in the cosmological context. This is due to the computational simplicity of the simplest compact flat three-manifold, the 3-torus, which has been used extensively in numerical simulations. The main goal of this article is to provide tools to compute the CMB properties and produce high resolution CMB maps for all seventeen multi-connected flat spaces [^2], following the general method introduced in our preceding work [@rulw02].
Recent measurements show that the density parameter $\Omega_0$ is close to unity and the observable universe is approximately flat. CMB data obtained by the Archeops balloon experiments [@archeops] and more recently by WMAP [@wmap] place strong constraints on the curvature. In addition, WMAP [@map] and later the Planck satellite [@planck] do and will provide full sky maps of CMB anisotropies, offering an opportunity to probe the topological properties of our universe. This observational constraint on the curvature radius of the universe motivates the detailed study of flat spaces even though spherical spaces are also promising candidates [@wlu02; @glluw; @lwugl; @luw02; @prl; @nat].
At present, the status of the constraints on the topology of flat spaces is evolving rapidly driven by the release of the WMAP data. Previous analysis, based on the COBE data, mainly constrained the topology of a 3-torus (see Refs. [@sokolov93; @staro93; @stevens93; @costa95; @levin98; @levin99; @pf; @inoue00; @inoue01; @inoue2; @roukema] and Refs. [@lachieze95; @uzan99; @levin02] for reviews of different methods for searching for the topology).
The WMAP data [@wmap] possess some anomalies on large angular scales that may be explained by a topological structure. In particular, the quadrupole is abnormally low, the octopole is very planar and the alignment between the quadrupole and octopole is also anomalous [@teg1]. Besides many other potential explanations [@quad], it was suggested that a toroidal universe with a smaller dimension on the order of half the horizon scale may explain all these anomalies [@teg1] but it was latter shown, on the basis of a finer statistical analysis, that it did not [@teg2]. Another topology was recently proposed to explain some of this anomaly in the case of a slightly positively curved space, namely the Poincaré dodecahedral space [@nat].
The first results of the search for the topology through the “circles in the sky” method [@cornish98] gave negative results for back-to-back or almost back-to-back circles [@teg2; @cssk]. While the first applies only to back-to-back circle with no twist, the second study includes an arbitrary twist and conclude that “it rules out the possibility that we live in a universe with topology smaller than $24 {\,{\mbox{Gpc}^{}}}$”. As will be discussed in this paper, back-to-back circles are generic only for homogeneous topologies such as e.g. 3-tori and a subclass of lens spaces. In non-homogeneous spaces the relative position of the circles depends on the position of the observer in the fundamental polyhedron.
In conclusion, as demonstrated by these preliminary results, only the toroidal spaces have been really constrained [@teg2; @cssk]. Besides, a series of studies have pointed out a departure of the WMAP data from statistical isotropy. Copi [*et al.*]{} [@copi] recently argued in particular that they are inconsistent with an isotropic Gaussian distribution at 98.8% confidence level. Previous studies pointed toward a possible North-South asymmetry of the data [@eriksen; @park]. Spaces with non-trivial topology are a class of models in which global isotropy (and possibly global homogeneity) is broken. Simulated CMB maps of these spaces may help to construct estimators for quantifying the departure of the temperature distribution from isotropy, and also give a deeper understanding of recent results.
Let us emphasize that in the case where the topological scale is slightly larger than the size of the observable universe, no matching circles will be observed. This might also happen for a configuration where the circles would all lie in the direction of the galactic disk where the signal-to-noise ratio might be too low. Contrary to the simply connected case, the correlation matrix, $C_{\ell m}^{\ell' m'}
\equiv \left< a_{\ell m} a_{\ell' m'} \right>$, of the coefficients of the development of the temperature fluctuations on spherical harmonics, will not be proportional to $\delta_{\ell \ell'} \delta_{m
m'}$. The study of this correlation matrix could offer the possibility to probe topology (slightly) beyond the horizon. Computing the correlation matrix $C_{\ell m}^{\ell' m'}$ for different multi-connected spaces will help design the best strategy to constrain the deviation from the simply connected case, and gives a concrete example of cosmological models in which the global homogeneity and isotropy are broken.
As described in detail in our preceding work [@rulw02], what is needed for any CMB computation are the eigenmodes of the Laplacian $$\label{Helmotz1}
\Delta {\Upsilon^{[{\Gamma}]}_{{k}\,{}}} = - k^2 {\Upsilon^{[{\Gamma}]}_{{k}\,{}}} ,$$ with boundary conditions compatible with the given topology. These eigenmodes can be developed on the basis ${\cal Y}_{k \ell m}$ of the (spherically symmetric) eigenmodes of the universal covering space as $$\label{eq:1}
{\Upsilon^{[{\Gamma}]}_{{k}\,{s}}}
= \sum_{\ell = 0}^\infty
\sum_{m = -\ell}^{\ell}
{\xi^{[{\Gamma}]\,{s}}_{{k}\,{\ell}\,{m}}} {\cal Y}_{k \ell m} ,$$ so that all the topological information is encoded in the coefficients ${\xi^{[{\Gamma}]\,{s}}_{{k}\,{\ell}\,{m}}}$, where $s$ labels the various eigenmodes sharing the same eigenvalue $- k^2$. Ref. [@rulw02] computes these coefficients for the torus and lens spaces and Refs. [@glluw; @cornish99] discuss more general cases.
To summarize, this article aims at several goals. First, it will give the complete classification of flat spaces and the exact form of the eigenmodes of the Laplacian for each of them. It will also provide a set of simulated CMB maps for most of these spaces. Among other effects, it will illustrate the effect of non-compact directions and discuss the influence of the position of the observer in the case of non-homogeneous spaces, which has never been discussed before. It also explains the structure of the observed CMB spectrum in the case of a very anisotropic (i.e., flattened or elongated in one direction) fundamental domain.
This article is organized as follows. We start by recalling the properties of the eighteen flat spaces (Section \[sec02\]) as well as the eigenmodes of the simply connected three-dimensional Euclidean space ${{{{\bf {E}}}}^3}$ (Section \[sec2\]), and in particular how to convert planar waves, which suit the description of topology, to spherical waves, which are more convenient for CMB computation. Then, in Section \[SectionEigenmodesOfMulticonnected\], we explain how to extract the modes of a given multi-connected space from the modes of ${{{{\bf {E}}}}^3}$. This method is then applied to give the eigenmodes of the ten compact flat spaces (Section \[SectionCompactSpaces\]), the five multi-connected flat spaces with two compact directions (“chimney spaces”, Section \[SectionDoublyPeriodic\]) and the two multi-connected flat spaces with only one compact direction (“slab spaces”, Section \[SectionSinglyPeriodic\]). Applying the general formalism developed in our previous work [@rulw02], we produce CMB maps for some of these spaces. With three exceptions the manifolds are not homogeneous, in the sense that a given manifold does not look the same from all points. To discuss the implication of the observed CMB and the genericity of the maps, we detail in Section \[SectionLocationOfObserver\] the influence of the position of the observer on the form of the eigenmodes and we study its consequences on the observed CMB maps. We show in particular that the matched circles are generically not back-to-back, but their relative position depends on the topology, the precise shape of the fundamental domain, and the position of the observer.
Notation {#notation .unnumbered}
========
The local geometry of the universe is described by a locally Euclidean Friedmann-Lemaître metric $$\label{fl_metric} {{\rm d}}s^2
= - c^2 {{\rm d}}t^2
+ a^2 (t) \left[{{\rm d}}{\chi}^2 + {\chi}^2{{\rm d}}\omega^2 \right] ,$$ where $a(t)$ is the scale factor, $t$ the cosmic time, ${{\rm d}}\omega^2 \equiv {{\rm d}}\theta^2 + \sin^2 \theta \, {{\rm d}}\varphi^2$ the infinitesimal solid angle.
The eighteen flat spaces {#sec02}
========================
Let us start by recalling the list of flat spaces. They are obtained as the quotient ${{{{\bf {E}}}}^3}/ \Gamma$ of three-dimensional Euclidean space ${{{{\bf {E}}}}^3}$ by a group $\Gamma$ of symmetries of ${{{{\bf {E}}}}^3}$ that is discrete and fixed point free. The classification of such spaces has long been known [@feodoroff85; @bierbach11], motivated by the study of crystallography and completed in 1934 [@novacki34]. The ten compact flat spaces are quotients of the 3-torus; six are orientable and four are non-orientable. Fig. \[FigureCompactSpaces\] shows fundamental polyhedra. The non-compact spaces form two families, the chimney space and its quotients having two compact directions (Fig. \[FigureChimneySpaces\]) and the slab space and its quotient having only one compact direction (Fig. \[FigureSlabSpaces\]). The terms [*slab space*]{} and [*chimney space*]{} were coined by Adams and Shapiro in their beautiful exposition of the flat three-dimensional topologies [@toponames]. Table \[tab1\] summarizes the properties of the whole family of flat spaces.
0.25cm
$\quad$ Symbol $\quad$ Name $\quad$ \# Compact Directions $\quad$ $\quad$ Orientable $\quad$
------------------------ --------------------------------------- --------------------------------------- ----------------------------
$E_{1}$ 3-torus 3 Yes
$E_{2}$ half turn space 3 Yes
$E_{3}$ quarter turn space 3 Yes
$E_{4}$ third turn space 3 Yes
$E_{5}$ sixth turn space 3 Yes
$E_{6}$ Hantzsche-Wendt space 3 Yes
$E_{7}$ Klein space 3 No
$E_{8}$ Klein space with horizontal flip 3 No
$E_{9}$ Klein space with vertical flip 3 No
$E_{10}$ Klein space with half turn 3 No
$E_{11}$ chimney space 2 Yes
$E_{12}$ chimney space with half turn 2 Yes
$E_{13}$ chimney space with vertical flip 2 No
$E_{14}$ chimney space with horizontal flip 2 No
$E_{15}$ chimney space with half turn and flip 2 No
$E_{16}$ slab space 1 Yes
$E_{17}$ slab space with flip 1 No
$E_{18}$ Euclidean space 0 Yes
: Classification of the 18 three-dimensional flat spaces.[]{data-label="tab1"}
Eigenmodes of ${{{{\bf {E}}}}^3}$ {#sec2}
=================================
The eigenmodes of Euclidean space ${{{{\bf {E}}}}^3}$ admit two different bases: a basis of planar waves and a basis of spherical waves. The former are more convenient when seeking eigenbases for multi-connected spaces, while the latter are more convenient for simulating CMB maps. This section considers both bases and the conversion between them, as also detailed in the particular case of the torus in Ref. [@rulw02].
Planar waves {#SectionPlanarWaves}
------------
Each vector ${{\bf k}}$ defines a planar wave $$\label{PlanarWave}
\Upsilon_{{\bf k}}({{\bf x}}) = {{\rm e}^{i {{\bf k}}\cdot {{\bf x}}}} .$$ The defining vector ${{\bf k}}$, called the [*wave vector*]{}, lives in the dual space, so the dot product ${{\bf k}}\cdot {{\bf x}}$ is always dimensionless. These modes are indeed not square integrable and are normalized as $$\label{norm1}
\int_{{{{{\bf {R}}}}^3}} \Upsilon_{{\bf k}}({{\bf x}}) \Upsilon_{{{\bf k}}'}^* ({{\bf x}})
\frac{{{\rm d}}^3 {{\bf x}}}{(2 \pi)^3}
= {\delta^{{\rm D}}({{{\bf k}}- {{\bf k}}'})} .$$
Spherical waves
---------------
Each spherical wave factors into a radial part and an angular part, $$\label{SphericalWave}
{\cal Y}_{k \ell m} (\chi, \theta, \varphi)
= \sqrt{\frac{2}{\pi}} \,
(2 \pi)^{3 / 2} \,
j_\ell (k \chi) \,
Y_\ell^m (\theta, \varphi) ,$$ where $(\chi, \theta, \varphi)$ are the usual spherical coordinates $$\begin{aligned}
\label{SphericalCoordinates}
x & = & \chi \; \sin{\theta} \; \cos{\varphi} \nonumber \\
y & = & \chi \; \sin{\theta} \; \sin{\varphi} \nonumber \\
z & = & \chi \; \cos{\theta} .\end{aligned}$$ The radial factor $j_\ell (k \chi)$ is the spherical Bessel function of index $\ell$, and the angular factor $Y_\ell^m (\theta, \varphi)$ is the standard spherical harmonic. The mode ${\cal Y}_{k \ell m}$ is not square integrable and is normalized according to $$\int_{{{{{\bf {R}}}}^3}} {\cal Y}_{k \ell m} {\cal Y}_{k' \ell' m'}^*
\frac{\chi^2 {{\rm d}}\chi {{\rm d}}\cos \theta {{\rm d}}\varphi}{(2 \pi)^3}
= \frac{1}{k^2} {\delta^{{\rm D}}({k - k'})} {\delta_{{\ell}\,{\ell'}}} {\delta_{{m}\,{m'}}} ,$$ which is analogous to the normalization (\[norm1\]) and which determines the numerical coefficient $\sqrt{2 / \pi}$.
Conversion
----------
Subsequent sections will find explicit bases for the eigenmodes of multi-connected flat three-manifolds as linear combinations of planar waves. The planar waves may easily be converted to spherical waves using Eqns 5.17.3.14 and 5.17.2.9 of Ref. [@vmk]: $$\begin{aligned}
\label{Conversion}
\Upsilon_{{\bf k}}({{\bf x}})
& = & {{\rm e}^{i {{\bf k}}\cdot {{\bf x}}}} \nonumber \\
& = & \sum_{\ell = 0}^\infty \;
i^\ell \; j_\ell (k \, |{{\bf x}}|) \; (2\ell + 1) \;
P_\ell (\cos{\theta_{{{\bf k}}, {{\bf x}}}}) \nonumber \\
& = & \sum_{\ell = 0}^\infty \;
i^\ell \; j_\ell (k \, |{{\bf x}}|) \;
\left(4 \pi \sum_{m = - \ell}^{\ell}
Y_\ell^m ({\bf \hat x}) Y_\ell^{m*} ({\bf \hat k})
\right) \nonumber \\
& = & \sum_{\ell = 0}^\infty \,
\sum_{m = - \ell}^{\ell} \;
i^\ell \; Y_\ell^{m*} ({\bf \hat k}) \;
\left[ \; 4 \pi \; j_\ell (k \, |{{\bf x}}|) \;
Y_\ell^m ({\bf \hat x}) \; \right] \nonumber \\
& = & \sum_{\ell = 0}^\infty \,
\sum_{m = - \ell}^{\ell} \;
\left( \; i^\ell \; Y_\ell^{m*} ({\bf \hat k}) \right)
{\cal Y}_{k l m} ({{\bf x}}) ,\end{aligned}$$ where $k = |{{\bf k}}|$, ${\bf\hat k} \equiv {{\bf k}}/ |{{\bf k}}|$, and ${\bf\hat x}
\equiv {{\bf x}}/ |{{\bf x}}|$.
In particular, the conversion formula (\[Conversion\]) lets one easily translate a planar wave $\Upsilon_{{\bf k}}$ to the framework we developed in Ref. [@rulw02], which expresses each basis eigenmode as a sum of spherical waves $$\Upsilon_{{\bf k}}= \Upsilon_{k, s}
= \sum_{\ell = 0}^\infty
\sum_{m = - \ell}^{\ell}
\xi^s_{k \ell m} {\cal Y}_{k \ell m} ,$$ where $s$ indexes the different $\Upsilon_{{\bf k}}$ whose wave vectors ${{\bf k}}$ share the same modulus $k$. In the Euclidean case the index may be chosen to be $s = \hat{{\bf k}}$. Comparison with (\[Conversion\]) immediately gives the coefficients $$\xi^{\hat {{\bf k}}}_{k \ell m} = i^\ell \; Y_\ell^{m*} ({\bf \hat k}) .$$
Eigenmodes of Multi-Connected Spaces {#SectionEigenmodesOfMulticonnected}
====================================
A multi-connected flat three-manifold is the quotient ${{{{\bf {E}}}}^3}/ \Gamma$ of Euclidean space ${{{{\bf {E}}}}^3}$ under the action of a group $\Gamma$ of isometries. The group $\Gamma$ is called the [*holonomy group*]{} and is always discrete and fixed point free. Each eigenmode $\hat\Upsilon$ of the multi-connected space ${{{{\bf {E}}}}^3}/ \Gamma$ lifts to a $\Gamma$-periodic eigenmode $\Upsilon$ of ${{{{\bf {E}}}}^3}$, that is, to an eigenmode of ${{{{\bf {E}}}}^3}$ that is invariant under the action of the holonomy group $\Gamma$. Common practice blurs the distinction between eigenmodes of ${{{{\bf {E}}}}^3}/ \Gamma$ and $\Gamma$-periodic eigenmodes of ${{{{\bf {E}}}}^3}$, and we follow that practice here. Thus the task of finding the eigenmodes of the multi-connected space ${{{{\bf {E}}}}^3}/ \Gamma$ becomes the task of finding the $\Gamma$-periodic eigenmodes of ${{{{\bf {E}}}}^3}$. In this section we investigate how an isometry $\gamma \in \Gamma$ acts on the space of eigenmodes. The two lemmas we obtain will make it easy to determine the eigenmodes of specific multi-connected spaces in subsequent sections.
Every isometry $\gamma$ of Euclidean space ${{{{\bf {E}}}}^3}$ factors as a rotation/reflection followed by a translation. If we write the rotation/reflection as a $3 \times 3$ matrix $M$ in the orthogonal group ${\rm O} (3)$ and write the translation as a vector ${{\bf T}}$, then $\gamma$ acts on ${{{{\bf {E}}}}^3}$ as $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\quad \mapsto \quad
{\left(\begin{array}{rrr}
{M_{x x}} & {M_{x y}} & {M_{x z}} \\
{M_{y x}} & {M_{y y}} & {M_{y z}} \\
{M_{z x}} & {M_{z y}} & {M_{z z}} \end{array} \right)}
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
+ {\left(\begin{array}{c}
{T_x} \\
{T_y} \\
{T_z} \end{array} \right)} .$$ This isometry of ${{{{\bf {E}}}}^3}$ induces a natural action $\Upsilon_{{\bf k}}({{\bf x}})
\mapsto \Upsilon_{{\bf k}}(M {{\bf x}}+ {{\bf T}})$ on the space of eigenmodes.\
[**Lemma 1 (Action Lemma).**]{} [*The natural action of an isometry $\gamma$ of ${{{{\bf {E}}}}^3}$ takes a planar eigenmode $\Upsilon_{{\bf k}}$ to another planar eigenmode ${{\rm e}^{i {{\bf k}}\cdot {{\bf T}}}} \Upsilon_{{{\bf k}}M}$.*]{}\
[*Proof.*]{} Keeping in mind that ${{\bf k}}$ is a row vector while ${{\bf x}}$ is a column vector, the proof is an easy computation: $$\begin{aligned}
\label{Lemma1Proof}
\Upsilon_{{\bf k}}({{\bf x}})
& = & {{\rm e}^{i {{\bf k}}\cdot {{\bf x}}}} \nonumber \\
& \mapsto & {{\rm e}^{i {{\bf k}}\cdot (M{{\bf x}}+ {{\bf T}})}} \nonumber \\
& = & {{\rm e}^{i {{\bf k}}\cdot {{\bf T}}}} \; {{\rm e}^{i {{\bf k}}M {{\bf x}}}} \nonumber \\
& = & {{\rm e}^{i {{\bf k}}\cdot {{\bf T}}}} \; \Upsilon_{{{\bf k}}M}({{\bf x}}) .\end{aligned}$$ [[*Q.E.D.*]{}]{}\
[**Lemma 2 (Invariance Lemma).**]{}
*If $\gamma$ is an isometry of ${{{{\bf {E}}}}^3}$ with matrix part $M$ and translational part $T$, the mode $\Upsilon_{{\bf k}}$ is a planar wave, and $n$ is the smallest positive integer such that ${{\bf k}}= {{\bf k}}M^n$ (typically $n$ is simply the order of the matrix $M$), then the action of $\gamma$*
1. preserves the $n$-dimensional space of eigenmodes spanned by $\lbrace \Upsilon_{{\bf k}}, \Upsilon_{{{\bf k}}M}, \ldots, \Upsilon_{{{\bf k}}M^{n-1}} \rbrace$ as a set, and
2. fixes a specific element $$a_0 \Upsilon_{{\bf k}}+ a_1 \Upsilon_{{{\bf k}}M}
+ \ldots
+ a_{n - 1} \Upsilon_{{{\bf k}}M^{n - 1}} ,$$ if and only if for each $j {\mbox{ (mod }{n})}$ $$a_{j + 1} = {{\rm e}^{i {{\bf k}}M^j {{\bf T}}}} a_j .$$
[*Proof.*]{} Both parts are immediate corollaries of Lemma 1. Specifically, the action of $\gamma$ takes a linear combination $$\label{Lemma1ProofBefore}
a_0 \Upsilon_{{\bf k}}+ a_1 \Upsilon_{{{\bf k}}M}
+ \ldots
+ a_{n - 2} \Upsilon_{{{\bf k}}M^{n - 2}}
+ a_{n - 1} \Upsilon_{{{\bf k}}M^{n - 1}} ,$$ to $$\label{Lemma1ProofAfter}
a_0 {{\rm e}^{i {{\bf k}}{{\bf T}}}} \Upsilon_{{{\bf k}}M}
+ a_1 {{\rm e}^{i {{\bf k}}M {{\bf T}}}} \Upsilon_{{{\bf k}}M^2}
+ \ldots
+ a_{n - 2} {{\rm e}^{i {{\bf k}}M^{n - 2} {{\bf T}}}} \Upsilon_{{{\bf k}}M^{n - 1}}
+ a_{n - 1} {{\rm e}^{i {{\bf k}}M^{n - 1} {{\bf T}}}} \Upsilon_{{{\bf k}}} ,$$ so it’s clear that the $n$-dimensional subspace spanned by $\lbrace
\Upsilon_{{\bf k}}, \Upsilon_{{{\bf k}}M}, \ldots, \Upsilon_{{{\bf k}}M^{n-1}}
\rbrace$ is preserved as a set. Equating (\[Lemma1ProofBefore\]) to (\[Lemma1ProofAfter\]) and comparing coefficients proves the second part. [[*Q.E.D.*]{}]{}\
Compact Flat Three-Manifolds {#SectionCompactSpaces}
============================
We will first find the eigenmodes of the 3-torus, and then use them to find the eigenmodes of the remaining compact flat three-manifolds.
3-Torus {#SubsectionThreeTorus}
-------
The [*3-torus*]{} is the quotient of Euclidean space ${{{{\bf {E}}}}^3}$ under the action of three linearly independent translations ${{\bf T}}_1$, ${{\bf T}}_2$ and ${{\bf T}}_3$. Its fundamental domain is a parallelepiped. Its eigenmodes are the eigenmodes of ${{{{\bf {E}}}}^3}$ invariant under the translations ${{\bf T}}_1$, ${{\bf T}}_2$ and ${{\bf T}}_3$ (recall from Section \[SectionEigenmodesOfMulticonnected\] the convention that eigenmodes of the quotient are represented as periodic eigenmodes of ${{{{\bf {E}}}}^3}$). The Invariance Lemma (with $n = 1$) implies that an eigenmode $\Upsilon_{{\bf k}}$ of ${{{{\bf {E}}}}^3}$ is invariant under the translation ${{\bf T}}_1$ if and only if ${{\rm e}^{i {{\bf k}}\cdot {{\bf T}}_1}} = 1$ which occurs precisely when ${{\bf k}}\cdot {{\bf T}}_1$ is an integer multiple of $2 \pi$. Thus geometrically the allowed values of the wave vector ${{\bf k}}$ form a family of parallel planes orthogonal to ${{\bf T}}_1$. Similarly, the eigenmode $\Upsilon_{{\bf k}}$ is invariant under the translation ${{\bf T}}_2$ (resp. ${{\bf T}}_3$) if and only if ${{\bf k}}$ lies on a family of parallel planes orthogonal to ${{\bf T}}_2$ (resp. ${{\bf T}}_3$), defined by ${{\bf k}}\cdot
{{\bf T}}_2 \in 2 \pi Z$ (resp. ${{\bf k}}\cdot {{\bf T}}_3 \in 2 \pi Z$). The eigenmode $\Upsilon_{{\bf k}}$ is invariant under all three translations ${{\bf T}}_1$, ${{\bf T}}_2$ and ${{\bf T}}_3$ if and only if it lies on all three families of parallel planes simultaneously. The intersection of the three families forms a lattice of discrete points. Fig. \[FigureTorusLattice\] illustrates the construction for the 2-torus; the construction for the 3-torus is analogous. This lattice of points defines the standard basis for the eigenspace of a torus.\
[**Definition.**]{} The [*standard basis*]{} for a 3-torus ${{{{\bf {T}}}}^3}$ generated by three linearly independent translations ${{\bf T}}_1$, ${{\bf T}}_2$ and ${{\bf T}}_3$ is the set $B = \lbrace \Upsilon_{{\bf k}}\; | \; {{\bf k}}\cdot {{\bf T}}_i \in 2 \pi Z \mbox{ for } i = 1, 2, 3 \rbrace$.\
The most important special case of a 3-torus is the [*rectangular 3-torus*]{} generated by three mutually orthogonal translations $$\begin{aligned}
{{\bf T}}_1 & = & (L_x, 0, 0) \nonumber \\
{{\bf T}}_2 & = & (0, L_y, 0) \nonumber \\
{{\bf T}}_3 & = & (0, 0, L_z) ,\end{aligned}$$ in which case the allowed wave vectors ${{\bf k}}$ take the form $$\label{RectangularTorusBasis}
{{\bf k}}= 2 \pi \, \left(\frac{n_x}{L_x}, \,
\frac{n_y}{L_y}, \,
\frac{n_z}{L_z} \right) ,$$ for integer values of $n_x$, $n_y$ and $n_z$, thus forming a rectangular lattice (Fig. \[FigureHoneycomb\] left).
The second most important special case is the [*hexagonal 3-torus*]{} generated by $$\begin{aligned}
\label{HexagonalGenerators}
{{\bf T}}_1 & = & \left(L, 0, 0 \right) \nonumber \\
{{\bf T}}_2 & = & \left(- \frac{1}{2} L,
+ \frac{\sqrt{3}}{2} L,
0 \right) \nonumber \\
{{\bf T}}_3 & = & \left(- \frac{1}{2} L,
- \frac{\sqrt{3}}{2} L,
0 \right) \nonumber \\
{{\bf T}}_4 & = & \left(0, 0, L_z\right) .\end{aligned}$$ These four generators and their inverses define a fundamental domain as a hexagonal prism. The first three generators are linearly dependent (${{\bf T}}_1 + {{\bf T}}_2 + {{\bf T}}_3 = 0$); eliminating any one of them suggests an alternative fundamental domain as a prism with a rhombic base. Even though the hexagonal and rhombic prisms look different, they define the same manifold. Either way, the allowed wave vectors ${{\bf k}}$ form a hexagonal lattice (Fig. \[FigureHoneycomb\] right) and may be parameterized as $$\label{HexagonalTorusBasis}
{{\bf k}}\quad = \quad 2 \pi \, \left(- \frac{n_1}{L},
\frac{2 n_1 - n_2}{\sqrt{3} L},
\frac{n_3}{L_z} \right) ,$$ for integer values of $n_1$, $n_2$ and $n_3$.
In the case of a general 3-torus, one writes the translations ${{\bf T}}_1$, ${{\bf T}}_2$ and ${{\bf T}}_3$ as the columns of a $3 \times 3$ matrix $T$ and solves the equation ${{\bf k}}T = 2 \pi (n_1, n_2, n_3)$ to find the allowable wave vectors ${{\bf k}}= 2 \pi (n_1, n_2, n_3) T^{-1}$ for integers $n_1$, $n_2$ and $n_3$.\
When one wants to simulate CMB maps, one needs to know not only the modes themselves but also how the modes are paired under complex conjugation. The reason is that the cosmological fields are in fact real-valued stochastic variables. Any such field can be decomposed into Fourier modes as $$\phi({{\bf x}}, t) = \int \frac{{{\rm d}}^3 {{\bf k}}}{(2 \pi)^{3 / 2}} \phi_{{\bf k}}(t)
{{\rm e}^{i {{\bf k}}. {{\bf x}}}} \hat e_{{\bf k}},$$ where $\hat e_{{\bf k}}$ is a complex, usually Gaussian, random variable satisfying $$\left<\hat e_{{\bf k}}\hat e_{{{\bf k}}'}^* \right> = {\delta^{{\rm D}}({{{\bf k}}- {{\bf k}}'})} .$$ The evolution equations of the cosmological perturbations involve time derivatives and a Laplacian so that the coefficient $\phi_{{\bf k}}(t)$ can be decomposed as $$\phi_{{\bf k}}(t) = \phi_k (t) {{\rm e}^{i \theta_{{\bf k}}}} ,$$ where $\theta_{{\bf k}}$ is a phase that is constant throughout the evolution. By absorbing the phase into the random variable, we can always choose $\phi_{{\bf k}}$ to be a real function of $k$ only, i.e., $\phi_k(t)$. Since ${{\rm e}^{i {{\bf k}}\cdot {{\bf x}}}}$ and ${{\rm e}^{- i {{\bf k}}\cdot
{{\bf x}}}}$ are conjugate in the preceding decomposition, the fact that $\phi ({{\bf x}}, t)$ is real implies that $$\label{random_relation}
\hat e_{{\bf k}}^* = \hat e_{- {{\bf k}}} .$$ This latter constraint may not hold for all the other spaces studied in this article and we will need to give its analog for each case.
Quotients of the 3-Torus
------------------------
For ease of illustration we first explain our general method for the two-dimensional Klein bottle. Figure \[FigureKleinBottle\] shows a portion of the Klein bottle’s universal cover, in which alternate images of the fundamental domain appear mirror reversed. Half the holonomies are pure translations while the other half are glide reflections. In other words, the holonomy group $\Gamma$ contains an index 2 subgroup $\Gamma' \subset \Gamma$ comprising the pure translations. While ${{{{\bf {E}}}}^2}/ \Gamma$ gives the original Klein bottle, ${{{{\bf {E}}}}^2}/ \Gamma'$ gives a torus whose fundamental domain is the square formed by the solid lines in Figure \[FigureKleinBottle\] (ignoring the dotted lines). According to the convention introduced in Section \[SectionEigenmodesOfMulticonnected\], the Klein bottle’s eigenmodes are represented as $\Gamma$-periodic functions on ${{{{\bf {E}}}}^2}$. But every $\Gamma$-periodic function is automatically a $\Gamma'$-periodic function as well, because $\Gamma'$ is a subgroup of $\Gamma$. Thus every eigenmode of the Klein bottle is [*a priori*]{} an eigenmode of the 2-torus. The task in finding the eigenspace of the Klein bottle is to start with the eigenspace of the torus and find the subspace that is invariant under the glide reflection (the one taking the lower half of a square to the upper half). In practice this is quite simple. A rectangular torus has holonomy group $\Gamma'$ generated by the two translations $${\left(\begin{array}{c}
{x} \\
{y} \end{array} \right)} \; \mapsto \; {\left(\begin{array}{c}
{x} \\
{y} \end{array} \right)} + {\left(\begin{array}{c}
{L_x} \\
{0} \end{array} \right)}
\qquad \mbox{and} \qquad
{\left(\begin{array}{c}
{x} \\
{y} \end{array} \right)} \; \mapsto \; {\left(\begin{array}{c}
{x} \\
{y} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{L_y} \end{array} \right)} .$$ The standard eigenbasis for this torus takes the form (\[RectangularTorusBasis\]), namely $B = \lbrace \Upsilon_{{\bf k}}\; | \; {{\bf k}}= 2 \pi (\frac{n_x}{L_x}, \frac{n_y}{L_y}) \mbox{ for }
n_x, n_y \in Z \rbrace$. To extend this $\Gamma'$ to the full holonomy group $\Gamma$ of the Klein bottle, we add the glide reflection $${\left(\begin{array}{c}
{x} \\
{y} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} \\
{0} & {1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{L_y / 2} \end{array} \right)} ,$$ and ask which elements of the basis $B$ it preserves. The Invariance Lemma provides a ready answer: when $k_x \neq 0$ the two-dimensional subspace $\lbrace \Upsilon_{k_x, k_y}, \Upsilon_{- k_x, k_y} \rbrace$ is preserved as a set (Part 1 of the Invariance Lemma) while the mode $\Upsilon_{k_x, k_y} + (-1)^{n_y} \Upsilon_{- k_x, k_y}$ is fixed exactly (Part 2 of the Invariance Lemma). In the exceptional case that $k_x = 0$, the one-dimensional subspace $\lbrace \Upsilon_{0, k_y}
\rbrace$ is preserved as a set, while $\Upsilon_{0, k_y}$ is fixed if and only if $n_y$ is even (because when $n = 1$, Part 2 of the Invariance Lemma requires $a_0 = {{\rm e}^{i {{\bf k}}\cdot {{\bf T}}}} a_0$ which implies ${{\bf k}}\cdot {{\bf T}}/ (2 \pi) = n_z / 2 \in Z$). In summary, an orthonormal basis for the space of eigenmodes of the Klein bottle is the union of $$\begin{array}{lcl}
\left[ \Upsilon_{2 \pi (\frac{n_x}{L_x}, \frac{n_y}{L_y})}
+ (- 1)^{n_y} \Upsilon_{2 \pi (- \frac{n_x}{L_x}, \frac{n_y}{L_y})}
\right] / \sqrt{2}
& \quad \mbox{for} \quad
& n_x \in Z^+, n_y \in Z, \\
\Upsilon_{2 \pi (0, \frac{n_y}{L_y})}
& \quad \mbox{for} \quad
& n_y \in 2 Z . \\
\end{array}$$\
Let us now apply this same method to each of the nine quotients of the 3-torus.
### Half turn space
The analysis of the half turn space closely follows that of the Klein bottle given immediately above. The only difference is that the Klein bottle’s holonomy group contained translations and glide reflections, while the half turn space’s holonomy group contains translations and corkscrew motions. Specifically, we begin with the generators for the holonomy group $\Gamma'$ of a rectangular 3-torus $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} \; \mapsto \; {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{L_x} \\
{0} \\
{0} \end{array} \right)}
, \qquad
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} \; \mapsto \; {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{L_y} \\
{0} \end{array} \right)}
, \qquad
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} \; \mapsto \; {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z} \end{array} \right)} .$$ and add a generator for the half-turn corkscrew motion $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 2} \end{array} \right)} ,$$ to get the full holonomy group $\Gamma$ of the half turn space. The Invariance Lemma shows that when $(k_x, k_y) \neq (0, 0)$ the two-dimensional subspace $\lbrace \Upsilon_{k_x, k_y, k_z},
\Upsilon_{- k_x, - k_y, k_z} \rbrace$ is preserved as a set (Part 1 of the Invariance Lemma) while the mode $\Upsilon_{k_x, k_y, k_z} +
(-1)^{n_z} \Upsilon_{- k_x, - k_y, k_z}$ is fixed exactly (Part 2 of the Invariance Lemma). In the exceptional case that $(k_x, k_y) = (0,
0)$, the one-dimensional subspace $\lbrace \Upsilon_{0, 0, k_z}
\rbrace$ is preserved as a set, while $\Upsilon_{0, 0, k_z}$ is fixed if and only if $n_z$ is even. In summary, an orthonormal basis for the space of eigenmodes of the half turn space, $\Upsilon_{k_x, k_y,
k_z}^{[E_2]}$, is the union of $$\begin{array}{lcl}
\frac{1}{\sqrt{2}}
\left[\Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},\frac{n_z}{L_z})}
+ (-1)^{n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x},-\frac{n_y}{L_y},\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& (n_x \in Z^+, n_y, n_z \in Z) \mbox{ or } (n_x = 0, n_y \in Z^+, n_z \in Z), \\
\Upsilon_{2\pi(0,0,\frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& n_z \in 2Z . \\
\end{array}$$ In terms of the notations used in Ref. [@rulw02], it leads to the coefficients $$\xi_{k\ell m}^{\hat{{\bf k}}}=\left\lbrace
\begin{array}{lcl}
\frac{i^\ell}{\sqrt{2}}
\left[Y_\ell^{m*}(\hat{{\bf k}})+ (-1)^{n_z}Y_\ell^{m*}(\hat{{\bf k}}M)
\right]
& \quad \mbox{for} \quad
& (n_x \in Z^+, n_y, n_z \in Z) \mbox{ or } (n_x = 0, n_y \in Z^+, n_z \in Z), \\
i^\ell Y_\ell^{m*}(\hat{{\bf k}})
& \quad \mbox{for} \quad
& (n_x,n_y)=(0,0),\quad n_z \in 2Z . \\
\end{array}\right.,$$ ${{\bf k}}$ being given by Eq. (\[RectangularTorusBasis\]). Passing from the expression of the modes to the coefficients ${\xi^{{\hat
{{\bf k}}}}_{{k}\,{\ell}\,{m}}}$ is straightforward and in the following we will give only the expressions of the modes.
If desired, one could construct a more general half turn space from a right prism with a parallelogram base, instead of a rectangular box.
To find the analog of Eq. (\[random\_relation\]), one simply needs to check that $$\Upsilon_{k_x, k_y, k_z}^{[E_2]*}
= (- 1)^{n_z} \Upsilon_{k_x, k_y, - k_z}^{[E_2]} ,$$ so that it follows that
1. when $k_z \not = 0$, $\hat e_{{\bf k}}$ is a complex random variable satisfying $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_z} \hat e_{k_x, k_y, - k_z} ,$$
2. when $k_z = 0$, $\hat e_{{\bf k}}$ is a real random variable.
### Quarter turn space
The quarter turn space is similar to the half turn space, but with with a quarter turn corkscrew motion $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{0} & {- 1} & {0} \\
{1} & {0} & {0} \\
{0} & {0} & {1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_x / 4} \end{array} \right)} .$$ In particular this implies that $L_x = L_y$. The Invariance Lemma shows that when $(k_x, k_y) \neq (0, 0)$ the four-dimensional subspace $\lbrace \Upsilon_{k_x, k_y, k_z}$, $\Upsilon_{k_y, - k_x, k_z}$, $\Upsilon_{- k_x, - k_y, k_z}$, $\Upsilon_{- k_y, k_x, k_z} \rbrace$ is preserved as a set, while the mode $\Upsilon_{k_x, k_y, k_z} +
i^{n_z} \Upsilon_{k_y, - k_x, k_z} + (-1)^{n_z} \Upsilon_{- k_x, -
k_y, k_z} + (-i)^{n_z} \Upsilon_{- k_y, k_x, k_z}$ is fixed exactly. In the exceptional case that $(k_x, k_y) = (0, 0)$, the one-dimensional subspace $\lbrace \Upsilon_{0, 0, k_z} \rbrace$ is preserved as a set, while $\Upsilon_{0, 0, k_z}$ is fixed if and only if $n_z$ is a multiple of 4. In summary, an orthonormal basis for the space of eigenmodes of the quarter turn space, $\Upsilon_{k_x, k_y, -
k_z}^{[E_3]}$, is the union of $$\begin{aligned}
& & \frac{1}{2}
\left[ \Upsilon_{2\pi ( \frac{n_x}{L_x}, \frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ i^{ n_z} \Upsilon_{2\pi ( \frac{n_y}{L_y},-\frac{n_x}{L_x}, \frac{n_z}{L_z})}
+ i^{2 n_z} \Upsilon_{2\pi (-\frac{n_x}{L_x},-\frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ i^{3 n_z} \Upsilon_{2\pi (-\frac{n_y}{L_y}, \frac{n_x}{L_x}, \frac{n_z}{L_z})}
\right] \nonumber\\
& & \qquad\qquad \mbox{for} \quad
n_x \in Z^+, n_y \in Z^+ \cup \lbrace 0 \rbrace, n_z \in Z, \nonumber\\
& & \Upsilon_{2\pi (0,0,\frac{n_z}{L_z})}
\quad \mbox{for} \quad n_z \in 4Z.\end{aligned}$$
As in the case of the half turn space, one can easily check that $$\Upsilon_{k_x, k_y, k_z}^{[E_3]*}
= (- 1)^{n_z} \Upsilon_{k_x, k_y, - k_z}^{[E_3]} ,$$ so that the analog of Eq. (\[random\_relation\]) is given by
1. when $k_z \not = 0$, $\hat e_{{\bf k}}$ is a complex random variable satisfying $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_z} \hat e_{k_x, k_y, - k_z} ,$$
2. when $k_z = 0$, $\hat e_{{\bf k}}$ is a real random variable.
### Third turn space
The third turn space is a three-fold quotient of a hexagonal 3-torus, not a rectangular one. To the generators (\[HexagonalGenerators\]) of the hexagonal 3-torus we add a one-third turn corkscrew motion $$\label{ThirdTurnGenerator}
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} \; \mapsto \;
{\left(\begin{array}{rrr}
{- \frac{1}{2}} & {- \frac{\sqrt{3}}{2}} & {0} \\
{\frac{\sqrt{3}}{2}} & {- \frac{1}{2}} & {0} \\
{0} & {0} & {1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 3} \end{array} \right)} .$$ The eigenmodes $\Upsilon_{{\bf k}}$ of the hexagonal 3-torus are already known from Eqn. (\[HexagonalTorusBasis\]) (and illustrated in Figure \[FigureHoneycomb\]). Applying the Invariance Lemma to them with the additional generator (\[ThirdTurnGenerator\]) yields the eigenbasis, $\Upsilon_{k_x, k_y, k_z}^{[E_4]}$, $$\begin{array}{lcl}
\frac{1}{\sqrt{3}}
\left[ \Upsilon_{{{\bf k}}}
+ \zeta^{n_3} \Upsilon_{{{\bf k}}M}
+ \zeta^{2 n_3} \Upsilon_{{{\bf k}}M^2}
\right]
& \quad \mbox{for} \quad
& n_1 \in Z^+,\; n_2 \in Z^+ \cup \lbrace 0 \rbrace, n_3 \in Z, \\
\Upsilon_{2\pi (0,0,\frac{n_3}{L_z})}
& \quad \mbox{for} \quad
& n_3 \in 3Z, \\
\end{array}$$ where $\zeta = {{\rm e}^{2 i \pi / 3}}$ is a cube root of unity and it is easily checked that $$\begin{aligned}
{{\bf k}}& = & 2 \pi \left(\frac{- n_2}{L}, \;
\frac{2 n_1 - n_2}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M & = & 2 \pi \left(\frac{n_1}{L}, \;
\frac{2 n_2 - n_1}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M^2 & = & 2 \pi \left(\frac{n_2 - n_1}{L}, \;
\frac{- n_1 - n_2}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) .\end{aligned}$$
One can check that $$\Upsilon_{n_1, n_2, n_3}^*
= \left\lbrace \begin{array}{ll}
\zeta^{2 n_3} \Upsilon_{n_2, n_2 - n_1, - n_3} &
\hbox{when} \, n_2 > n_1 \\
\zeta^{n_3} \Upsilon_{n_1 - n_2, n_1, - n_3} &
\hbox{when}\, n_1 \geq n_2
\end{array}\right. .$$ It follows that the analog of Eq. (\[random\_relation\]) is given by
1. $$\hat{e}_{n_1, n_2, n_3}^*
= \left\lbrace \begin{array}{ll}
\zeta^{2 n_3} \hat{e}_{n_2, n_2 - n_1, - n_3} &
\hbox{when} \, n_2 > n_1 \\
\zeta^{n_3} \hat{e}_{n_1 - n_2, n_1, - n_3} &
\hbox{when}\, n_1 \geq n_2
\end{array}\right. .$$
### Sixth turn space
The sixth turn space is like the third turn space, but with a one-sixth turn corkscrew motion $$\label{SixthTurnGenerator}
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{\frac{1}{2}} & {- \frac{\sqrt{3}}{2}} & {0} \\
{\frac{\sqrt{3}}{2}} & {\frac{1}{2}} & {0} \\
{0} & {0} & {1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 6} \end{array} \right)} .$$ The same reasoning as before shows the eigenbasis, $\Upsilon_{k_x,
k_y, k_z}^{[E_5]}$, to be $$\begin{array}{lcl}
\frac{1}{\sqrt{6}}
\left[ \Upsilon_{{{\bf k}}}
+ \zeta^{n_3} \Upsilon_{{{\bf k}}M}
+ \zeta^{2 n_3} \Upsilon_{{{\bf k}}M^2}
+ \zeta^{3 n_3} \Upsilon_{{{\bf k}}M^3}
+ \zeta^{4 n_3} \Upsilon_{{{\bf k}}M^4}
+ \zeta^{5 n_3} \Upsilon_{{{\bf k}}M^5}
\right]
& \mbox{ for } \;
& n_1 \in Z^+,\; n_2 \in Z^+ \cup \lbrace 0 \rbrace,\;n_2 < n_1, n_3 \in Z, \\
\Upsilon_{2\pi (0,0,\frac{n_3}{L_z})}
& \mbox{ for }
& n_3 \in 6Z, \\
\end{array}$$ where $\zeta = {{\rm e}^{2i \pi / 6}}$ is a sixth root of unity and it is easily checked that $$\begin{aligned}
{{\bf k}}& = & 2 \pi \left(\frac{- n_2}{L}, \;
\frac{2 n_1 - n_2}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M & = & 2 \pi \left(\frac{n_1 - n_2}{L}, \;
\frac{n_1 + n_2}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M^2 & = & 2 \pi \left(\frac{n_1}{L}, \;
\frac{2 n_2 - n_1}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M^3 & = & 2 \pi \left(\frac{n_2}{L}, \;
\frac{n_2 - 2 n_1}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M^4 & = & 2 \pi \left(\frac{n_2 - n_1}{L}, \;
\frac{- n_1 - n_2}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right) \nonumber\\
{{\bf k}}M^5 & = & 2 \pi \left(\frac{- n_1}{L}, \;
\frac{n_1 - 2 n_2}{\sqrt{3} L}, \;
\frac{n_3}{L_z} \right).\end{aligned}$$ One can check that $$\Upsilon_{n_1, n_2, n_3}^* = (- 1)^{n_3} \Upsilon_{n_1, n_2, - n_3} ,$$ so that when $n_3 \neq 0$ an analog relation exists with the corresponding complex random variables, and when $n_3 = 0$, the random variable $\hat{e}_{n_1, n_2, 0}$ is real.
### Hantzsche-Wendt space
The fundamental polyhedron of the Hantzsche-Wendt space is a rhombic dodecahedron circumscribed about a rectangular box of size $(L_x / 2,
L_y / 2, L_z / 2)$. The holonomy group is generated by the three half-turn corkscrew motions $$\begin{aligned}
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; & \mapsto & \;
{\left(\begin{array}{rrr}
{\;\;\; 1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {- 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
+ {\left(\begin{array}{c}
{L_x / 2} \\
{L_y / 2} \\
{0} \end{array} \right)} \nonumber \\
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; & \mapsto & \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {- 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
+ {\left(\begin{array}{c}
{0} \\
{L_y / 2} \\
{L_z / 2} \end{array} \right)} \nonumber \\
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; & \mapsto & \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
+ {\left(\begin{array}{c}
{L_x / 2} \\
{0} \\
{L_z / 2} \end{array} \right)} .
\label{HantzscheWendtGenerators}\end{aligned}$$ The composition of these three generators is the identity, so any two suffice to generate the group. Each element of the Hantzsche-Wendt group has a rotational component $M \in \lbrace {{\rm diag}}(1, 1, 1)$, ${{\rm diag}}(1, -1, -1)$, ${{\rm diag}}(-1, 1, -1)$, ${{\rm diag}}(-1, -1, 1)
\rbrace$. The pure translations (elements with $M = {{\rm diag}}(1, 1, 1)$) form a subgroup of index 4; the corresponding four-fold cover of the Hantzsche-Wendt space is a rectangular 3-torus of size $(L_x, L_y,
L_z)$, whose holonomy is generated by the squares of the above corkscrew motions. Thus we may begin with the eigenspace for a rectangular 3-torus (\[RectangularTorusBasis\]) and ask what subspace remains fixed by the three Hantzsche-Wendt generators (\[HantzscheWendtGenerators\]). The Invariance Lemma shows that all three generators preserve the (typically four-dimensional) subspace $\lbrace \Upsilon_{k_x, k_y, k_z}, \Upsilon_{k_x, - k_y, - k_z},
\Upsilon_{- k_x, k_y, - k_z}, \Upsilon_{- k_x, - k_y, k_z} \rbrace$ as a set, and fix the linear combination $$\Upsilon_{k_x,k_y,k_z}
\; + \; (- 1)^{n_x - n_y} \Upsilon_{k_x, - k_y, - k_z}
\; + \; (- 1)^{n_y - n_z} \Upsilon_{- k_x, k_y, - k_z}
\; + \; (- 1)^{n_z - n_x} \Upsilon_{- k_x, - k_y, k_z} .
\label{HantzscheWendtFixedPoint}$$ Visualizing the four wave vectors in the subscripts of (\[HantzscheWendtFixedPoint\]) as alternate corners of the cube $(\pm k_x, \, \pm k_y, \, \pm k_z) = 2 \pi (\frac{\pm n_x}{L_x}, \,
\frac{\pm n_y}{L_y}, \, \frac{\pm n_z}{L_z})$, one sees that subspace will be degenerate if and only if at least two of the indices $\lbrace
n_x, n_y, n_z \rbrace$ are zero. In the degenerate case a two-dimensional subspace like $\lbrace \Upsilon_{k_x, 0, 0},
\Upsilon_{- k_x, 0, 0} \rbrace$ is preserved as a set, while the mode $\Upsilon_{k_x, 0, 0} + \Upsilon_{- k_x, 0, 0}$ is preserved by all three generators if and only if $n_x$ is even. Thus the Hantzsche-Wendt’s eigenspace has basis, $\Upsilon_{k_x, k_y,
k_z}^{[E_6]}$, $$\begin{aligned}
& & \frac{1}{2}
\left[ \Upsilon_{2\pi ( \frac{n_x}{L_x}, \frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ (-1)^{n_x - n_y} \Upsilon_{2\pi ( \frac{n_x}{L_x},-\frac{n_y}{L_y},-\frac{n_z}{L_z})}
+ (-1)^{n_y - n_z} \Upsilon_{2\pi (-\frac{n_x}{L_x}, \frac{n_y}{L_y},-\frac{n_z}{L_z})}
+ (-1)^{n_z - n_x} \Upsilon_{2\pi (-\frac{n_x}{L_x},-\frac{n_y}{L_y}, \frac{n_z}{L_z})}
\right] \nonumber\\
& & \qquad\qquad \mbox{for} \quad
(n_x, n_y \in Z^+, n_z \in Z) \mbox{ or }
(n_x = 0, n_y, n_z \in Z^+) \mbox{ or }
(n_y = 0, n_x, n_z \in Z^+),\nonumber\\
& & \frac{1}{\sqrt{2}} \left[
\Upsilon_{2\pi (\frac{n_x}{L_x},0,0)}
+ \Upsilon_{2\pi (-\frac{n_x}{L_x},0,0)}
\right]
\quad \mbox{for} \quad n_x \in 2Z^+,\nonumber\\
& & \frac{1}{\sqrt{2}} \left[
\Upsilon_{2\pi (0,\frac{n_y}{L_y},0)}
+ \Upsilon_{2\pi (0,-\frac{n_y}{L_y},0)}
\right]
\quad \mbox{for} \quad n_y \in 2Z^+,\nonumber\\
& & \frac{1}{\sqrt{2}} \left[
\Upsilon_{2\pi (0,0,\frac{n_z}{L_z})}
+ \Upsilon_{2\pi (0,0,-\frac{n_z}{L_z})}
\right]
\quad \mbox{for} \quad n_z \in 2Z^+.\end{aligned}$$
One can easily check that $$\Upsilon_{k_x, k_y, k_z}^{[E_6]*}
= (- 1)^{n_x - n_z} \Upsilon_{k_x, k_y, - k_z}^{[E_6]} ,$$ when $(n_x, n_y \in Z^+, n_z \in Z)$ or $(n_x = 0, n_y, n_z \in Z^+)$ or $(n_y = 0, n_x, n_z \in Z^+)$ and that otherwise $\Upsilon_{k_x,
k_y, k_z}^{[E_6]}$ is real. It follows that the analog of Eq. (\[random\_relation\]) is given by
1. when $n_x, n_y \in Z^+$ and $n_z \in Z$, $\hat e_{{\bf k}}$ is a complex random variable satisfying $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_x - n_z} \hat e_{k_x, k_y, - k_z} .$$ It is thus a real random variable if $k_z = 0$ and $n_x \in 2 Z$ and a purely imaginary random variable if $k_z = 0$ and $n_x \notin 2 Z$
2. when $n_x = 0$ and $n_y, n_z \in Z^+$, $\hat e_{{\bf k}}$ is a random variable satisfying $$\hat e_{0, k_y, k_z}^* = (- 1)^{n_y} \hat e_{0, k_y, k_z} ,$$ so that it is a real random variable when $n_y \in 2 Z^+$ and purely imaginary otherwise,
3. when $n_y = 0$ and $n_x, n_z \in Z^+$, $\hat e_{{\bf k}}$ is a random variable satisfying $$\hat e_{k_x, 0, k_z}^* = (- 1)^{n_z} \hat e_{k_x, 0, k_z} ,$$ so that it is a real random variable when $n_z \in 2 Z^+$ and purely imaginary otherwise,
4. when $n_x \in Z^+$, $n_y \in Z^+$ or $n_z \in Z^+$, $\hat
e_{k_x, 0, 0}$, $\hat e_{0, k_y, 0}$ and $\hat e_{0, 0, k_z}$ are real random variables.
### Klein space
Klein space is generated by two glide reflections $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
+ {\left(\begin{array}{c}
{L_x / 2} \\
{L_y / 2} \\
{0} \end{array} \right)} , \qquad
{\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
+ {\left(\begin{array}{c}
{L_x / 2} \\
{- L_y / 2} \\
{0} \end{array} \right)} , \qquad
\label{KleinGlideReflections}$$ along with a simple translation $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} \; \mapsto \; {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z} \end{array} \right)} .
\label{KleinSimpleTranslation}$$ The first (resp. second) glide reflection corresponds to the two upper (resp. lower) faces of the hexagonal prism in Figure \[FigureCompactSpaces\], taking one to the other so that the small dark-colored (resp. light-colored) windows match. The simple translation takes the front hexagonal face to the back hexagonal face so that the doors match.
The square of either glide reflection is a horizontal translation ${{\bf T}}= (L_x, 0, 0)$ (taking the unmarked left wall to the unmarked right wall in Figure \[FigureCompactSpaces\]), while the composition of one glide reflection with the inverse of the other is a vertical translation ${{\bf T}}= (0, L_y, 0)$. Thus the Klein space is the two-fold quotient of a rectangular 3-torus of size $(L_x, L_y, L_z)$.
To find the Klein space’s eigenmodes, we begin with the modes (\[RectangularTorusBasis\]) of the rectangular 3-torus and ask which remain invariant under the glide reflections (\[KleinGlideReflections\]). The Invariance Lemma shows the Klein space’s eigenmodes have the orthonormal basis $$\begin{array}{lcl}
\frac{1}{\sqrt{2}}
\left[\Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},\frac{n_z}{L_z})}
+ (-1)^{n_x + n_y} \Upsilon_{2\pi(\frac{n_x}{L_x},-\frac{n_y}{L_y},\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_y \in Z^+,\;n_x, n_z \in Z, \\
\Upsilon_{2\pi(\frac{n_x}{L_x},0,\frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& n_x \in 2Z,\;n_z \in Z. \\
\end{array}
\label{PlainKleinBasis}$$
One can easily check that $$\Upsilon_{k_x, k_y, k_z}^{[E_7]*}
= (- 1)^{n_x + n_y} \Upsilon_{- k_x, k_y, - k_z}^{[E_7]} ,$$ when $n_y \in Z^+, \; n_x, n_z \in Z$ and that $\Upsilon_{k_x, k_y,
k_z}^{[E_7]}$ is real otherwise. It follows that the analog of Eq. (\[random\_relation\]) is given by
1. when $n_y \in Z^+, \; n_x, n_z \in Z$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_x + n_y} \hat e_{- k_x, k_y, - k_z} .$$ It is thus a real random variable when $k_x = k_z = 0$ and $n_y \in 2
Z$ and a purely imaginary random variable when when $k_x = k_z = 0$ and $n_y \notin 2 Z$,
2. when $n_x \in 2 Z$, $n_y = 0$ and $n_z \in Z$, $\hat
e_{k_x, 0, k_z}$ is a real random variable.
### Klein space with horizontal flip
The Klein space with horizontal flip is a two-fold quotient of the plain Klein space. It includes the same two glide reflections (\[KleinGlideReflections\]) as the Klein space, but adds a square root $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 2} \end{array} \right)} ,
\label{KleinHorizontalFlip}$$ of the Klein space’s $L_z$-translation (\[KleinSimpleTranslation\]). Because the Klein space with horizontal flip is a quotient of the plain Klein space, every eigenmode of the former is automatically an eigenmode of the latter (recall the reasoning of the first paragraph of Section \[SectionEigenmodesOfMulticonnected\]). Thus our task is to decide which of the Klein space’s eigenmodes (\[PlainKleinBasis\]) are preserved by the new generator (\[KleinHorizontalFlip\]). The Invariance Lemma shows the orthonormal basis to be $$\begin{array}{lcl}
\frac{1}{2}
\left[ \Upsilon_{2\pi( \frac{n_x}{L_x}, \frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ (-1)^{n_x + n_y} \Upsilon_{2\pi( \frac{n_x}{L_x},-\frac{n_y}{L_y}, \frac{n_z}{L_z})}
\right.& & \\
\quad\left. + (-1)^{n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x}, \frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ (-1)^{n_x + n_y + n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x},-\frac{n_y}{L_y}, \frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_x, n_y \in Z^+,\;n_z \in Z, \\
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi(0,\frac{n_y}{L_y},\frac{n_z}{L_z})}
+ (-1)^{n_y} \Upsilon_{2\pi(0,-\frac{n_y}{L_y},\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_y \in Z^+,\; n_z \in 2Z, \\
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi(\frac{n_x}{L_x},0,\frac{n_z}{L_z})}
+ (-1)^{n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x},0,\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_x \in 2Z^+,\; n_z \in Z, \\
\Upsilon_{2\pi(0,0,\frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& n_z \in 2Z. \\
\end{array}
\label{KleinHorizontalFlipModes}$$
Following the same procedure as before, we obtain that the analog of Eq. (\[random\_relation\]) is given by
1. when $n_x, n_y \in Z^+, \; n_z \in Z$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_x + n_y + n_z} \hat e_{k_x, k_y, - k_z} .$$ It is thus a real random variable when $n_z = 0$ and $n_x + n_y \in 2
Z$ and a purely imaginary random variable when $n_z = 0$ and $n_x +
n_y \notin 2 Z$,
2. when $n_x = 0$, $n_y \in Z^+$ and $n_z \in Z$, $\hat e_{0, k_y,
k_z}$ must satisfy $$\hat e_{0, k_y, k_z}^* = (- 1)^{n_y} \hat e_{0, k_y, - k_z} ,$$ so that it is a real random variable when $k_z = 0$ and $n_y \in 2 Z$ and a purely imaginary random variable when $n_z = 0$ and $n_y \notin
2 Z$,
3. when $n_x \in Z^+$, $n_y = 0$ and $n_z \in Z$, $\hat e_{k_x, 0,
k_z}$ must satisfy $$\hat e_{k_x, 0, k_z}^* = (- 1)^{n_z} \hat e_{k_x, 0, - k_z} ,$$ so that it is a real random variable when $k_z = 0$,
4. when $n_z \in 2 Z$ it has to satisfy $$\hat e_{0, 0, k_z}^* = \hat e_{0, 0, - k_z} .$$
### Klein space with vertical flip
The Klein space with vertical flip replaces (\[KleinHorizontalFlip\]) with $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 2} \end{array} \right)} ,
\label{KleinVerticalFlip}$$ whose action interchanges the modes $\Upsilon_{(k_x, k_y, k_z)}
\leftrightarrow (-1)^{n_z} \Upsilon_{(k_x, - k_y, k_z)}$. Consistency with the glide reflections (\[KleinGlideReflections\]), whose action interchanges $\Upsilon_{(k_x, k_y, k_z)} \leftrightarrow (-1)^{n_x +
n_y} \Upsilon_{(k_x, - k_y, k_z)}$, requires $n_x + n_y \equiv n_z
{\mbox{ (mod }{2})}$. Thus the orthonormal basis is $$\begin{array}{lcl}
\frac{1}{\sqrt{2}}
\left[\Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},\frac{n_z}{L_z})}
+ (-1)^{n_x + n_y} \Upsilon_{2\pi(\frac{n_x}{L_x},-\frac{n_y}{L_y},\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_y \in Z^+,\;n_x, n_z \in Z,\;n_x + n_y \equiv n_z {\mbox{ (mod }{2})}, \\
\Upsilon_{2\pi(\frac{n_x}{L_x},0,\frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& n_x,n_z \in 2Z. \\
\end{array}$$
Following the same procedure as before, we obtain that the analog of Eq. (\[random\_relation\]) is given by
1. when $n_y \in Z^+$ and $n_x, n_z \in Z$ with $n_x + n_y \equiv
n_z {\mbox{ (mod }{2})}$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_z} \hat e_{- k_x, k_y, - k_z} ,$$ so that it is a real random variable when $k_x = k_z = 0$,
2. when $n_x, n_z \in 2 Z$, it has to be such that $$\hat e_{k_x, 0, k_z}^* = \hat e_{- k_x, 0, - k_z} .$$
### Klein space with half turn
The Klein space with half turn replaces (\[KleinHorizontalFlip\]) or (\[KleinVerticalFlip\]) with $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {- 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 2} \end{array} \right)} ,
\label{KleinHalfTurn}$$ The orthonormal eigenbasis, which differs only slightly from that of the Klein space with horizontal flip (\[KleinHorizontalFlipModes\]), is $$\begin{array}{lcl}
\frac{1}{2}
\left[ \Upsilon_{2\pi( \frac{n_x}{L_x}, \frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ (-1)^{n_x + n_y} \Upsilon_{2\pi( \frac{n_x}{L_x},-\frac{n_y}{L_y}, \frac{n_z}{L_z})}\right. & & \\
\quad \left. + (-1)^{n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x},-\frac{n_y}{L_y}, \frac{n_z}{L_z})}
+ (-1)^{n_x + n_y + n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x}, \frac{n_y}{L_y}, \frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_x, n_y \in Z^+,\;n_z \in Z, \\
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi(0,\frac{n_y}{L_y},\frac{n_z}{L_z})}
+ (-1)^{n_y} \Upsilon_{2\pi(0,-\frac{n_y}{L_y},\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_y \in Z^+,\; n_z \in Z,\; n_y \equiv n_z {\mbox{ (mod }{2})}, \\
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi(\frac{n_x}{L_x},0,\frac{n_z}{L_z})}
+ (-1)^{n_z} \Upsilon_{2\pi(-\frac{n_x}{L_x},0,\frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& n_x \in 2Z^+,\; n_z \in Z, \\
\Upsilon_{2\pi(0,0,\frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& n_z \in 2Z. \\
\end{array}$$
Following the same procedure as before, we obtain that the analog of Eq. (\[random\_relation\]) is given by
1. when $n_x, n_y \in Z^+$ and $n_z \in Z$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_z} \hat e_{k_x, k_y, - k_z} ,$$ so that it is a real random variable when $k_z = 0$,
2. when $n_y \in Z^+$, $n_z \in Z$ and $n_y \equiv n_z {\mbox{ (mod }{2})}$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{0, k_y, k_z}^* = (- 1)^{n_y} \hat e_{0, k_y, - k_z} ,$$ so that it is a real random variable when $k_z = 0$ and $n_y \in 2 Z$ and a purely imaginary random variable when $k_z = 0$ and $n_y \notin
2 Z$.
3. when $n_x \in 2 Z^+$ and $n_z \in Z$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, 0, k_z}^* = (- 1)^{n_z} \hat e_{k_x, 0, - k_z} ,$$ so that it is a real random variable when $k_z = 0$
4. when $n_z \in 2 Z$, it has to be such that $$\hat e_{0, 0, k_z}^* = \hat e_{0, 0, - k_z} .$$
Doubly Periodic Spaces {#SectionDoublyPeriodic}
======================
We will first find the eigenmodes of the chimney space, and then use them to find the eigenmodes of its quotients.
Chimney Space {#SubsectionChimneySpace}
-------------
Just as a 3-torus is the quotient of Euclidean space ${{{{\bf {E}}}}^3}$ under the action of three linearly independent translations ${{\bf T}}_1$, ${{\bf T}}_2$ and ${{\bf T}}_3$, a [*chimney space*]{} is the quotient of ${{{{\bf {E}}}}^3}$ by only two linearly independent translations ${{\bf T}}_1$ and ${{\bf T}}_2$. Its fundamental domain is an infinitely tall chimney whose cross section is a parallelogram (Figure \[FigureChimneySpaces\]). And just as the allowable wave vectors ${{\bf k}}$ for an eigenmode $\Upsilon_{{\bf k}}$ of a 3-torus were defined by the intersection of three families of parallel planes (Section \[SubsectionThreeTorus\]), the allowable wave vectors ${{\bf k}}$ for an eigenmode $\Upsilon_{{\bf k}}$ of a chimney space are defined by the intersection of two families of parallel planes. Thus the allowable wave vectors form a latticework of parallel lines.
The most important special case is the [*rectangular chimney space*]{} generated by two orthogonal translations $$\begin{aligned}
{{\bf T}}_1 & = (L_x, 0, 0) , \nonumber \\
{{\bf T}}_2 & = (0, L_y, 0) ,
\label{ChimneySpaceTranslations}\end{aligned}$$ in which case the allowed wave vectors ${{\bf k}}$ take the form $$\label{RectangularChimneyBasis}
{{\bf k}}= 2\pi\,\left(\frac{n_x}{L_x},\, \frac{n_y}{L_y},\, r_z\right)$$ for integer values of $n_x$ and $n_y$ and real values of $r_z$. The corresponding orthonormal basis $$\begin{array}{lcl}
\Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},r_z)}
& \quad \mbox{for} \quad
& n_x, n_y \in Z,\;r_z \in R
\end{array}$$ is continuous, not discrete as for the 3-torus. Nevertheless, restricting to a fixed modulus $k = |{{\bf k}}|$ recovers a finite-dimensional basis.
The next four spaces are quotients of the chimney space, so their eigenmodes will form subspaces of those of the chimney space itself.
As in the case of the torus, the random variable $\hat e_{{\bf k}}$ is a complex random variable satisfying $$\hat e_{{\bf k}}^* = \hat e_{- {{\bf k}}} .$$
Quotients of the Chimney Space {#SubsectionChimneySpaceQuotients}
------------------------------
### Chimney space with half turn
The chimney space with half turn (Figure \[FigureChimneySpaces\]) is generated by the rectangular chimney space’s $x$-translation ${{\bf T}}_1 = (L_x, 0, 0)$ along with $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {- 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_y / 2} \end{array} \right)} .
\label{ChimneyHalfTurn}$$ Even though the eigenmodes are not discrete, the Invariance Lemma applies exactly as in the compact case, giving the eigenbasis $$\begin{array}{lcl}
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},r_z)}
+ (-1)^{n_y} \Upsilon_{2\pi(-\frac{n_x}{L_x},\frac{n_y}{L_y},-r_z)}
\right]
& \quad \mbox{for} \quad
& (n_x \in Z^+, r_z \in R) \mbox{ or } (n_x = 0, r_z \in R^+), \\
\Upsilon_{2\pi(0,\frac{n_y}{L_y},0)}
& \quad \mbox{for} \quad
& n_y \in 2Z. \\
\end{array}$$
This case is analogous to the case of the half turn space so the analog of Eq. (\[random\_relation\]) is given by
1. when $(n_x \in Z^+, r_z \in R)$ or $(n_x = 0, r_z \in R^+)$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_y} \hat e_{k_x, - k_y, k_z} .$$ It is thus a real random variable when $n_y = 0$ and complex otherwise,
2. when $n_y \in 2 Z$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{0, k_y, 0}^* = \hat e_{0, - k_y, 0} .$$
### Chimney space with vertical flip
The chimney space with vertical flip (Figure \[FigureChimneySpaces\]) is generated by the translation ${{\bf T}}_1 = (L_x, 0, 0)$ along with $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {- 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{L_y / 2} \\
{0} \end{array} \right)} .
\label{ChimneyVerticalFlip}$$ The Invariance Lemma shows the orthonormal eigenbasis to be $$\begin{array}{lcl}
\frac{1}{2}
\left[ \Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},r_z)}
+ (-1)^{n_y} \Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},-r_z)}
\right]
& \quad \mbox{for} \quad
& n_x, n_y \in Z,\; r_z \in R^+, \\
\Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},0)}
& \quad \mbox{for} \quad
& n_x \in Z,\; n_y \in 2Z. \\
\end{array}$$
One might also consider a chimney space with a vertical flip in the $x$-direction as well as the $y$-direction. Surprisingly, such a space turns out to be equivalent to a chimney space with a single flip, but with cross section a parallelogram rather than a rectangle. In other words, the chimney space with two flips has the same topology as the chimney space with one flip, even though they may differ geometrically.
Concerning the properties of the random variable, the analog of Eq. (\[random\_relation\]) is given by
1. when $n_x, n_y \in Z$ and $r_z \in R^+$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_y} \hat e_{- k_x, - k_y, k_z} ,$$ it is thus a real random variable when $k_x = k_y = 0$,
2. when $n_x \in Z$ and $n_y \in 2 Z$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{k_x, k_y, 0}^* = \hat e_{- k_x, - k_y, 0} .$$
### Chimney space with horizontal flip
The chimney space with horizontal flip (Figure \[FigureChimneySpaces\]) is generated by the translation ${{\bf T}}_1 = (L_x, 0, 0)$ along with $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{L_y / 2} \\
{0} \end{array} \right)} .
\label{ChimneyHorizontalFlip}$$ The Invariance Lemma shows the orthonormal eigenbasis to be $$\begin{array}{lcl}
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi(\frac{n_x}{L_x},\frac{n_y}{L_y},r_z)}
+ (-1)^{n_y} \Upsilon_{2\pi(-\frac{n_x}{L_x},\frac{n_y}{L_y},r_z)}
\right]
& \quad \mbox{for} \quad
& n_x \in Z^+,\; n_y \in Z,\; r_z \in R, \\
\Upsilon_{2\pi(0,\frac{n_y}{L_y},r_z)}
& \quad \mbox{for} \quad
& n_y \in 2Z,\; r_z \in R. \\
\end{array}$$
Concerning the properties of the random variable, the analog of Eq. (\[random\_relation\]) is given by
1. when $n_x \in Z$, $n_y \in Z$ and $r_z \in R$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_y} \hat e_{k_x, - k_y, - k_z} ,$$ it is thus a real random variable when $k_y = k_z = 0$,
2. when $n_y \in 2 Z$ and $r_z \in R$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{0, k_y, 0}^* = \hat e_{0, - k_y, 0} .$$
### Chimney space with half turn and flip
The chimney space with half turn and flip (Figure \[FigureChimneySpaces\]) is generated by $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {- 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{L_y / 2} \\
{0} \end{array} \right)} ,
\label{ChimneyHalfTurnFlip1}$$ and $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {- 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{L_x / 2} \\
{0} \\
{0} \end{array} \right)} .
\label{ChimneyHalfTurnFlip2}$$ It is a four-fold quotient of the plain chimney space, unlike the preceding examples, which were two-fold quotients. The Invariance Lemma gives an orthonormal basis for its eigenmodes $$\begin{array}{lcl}
\frac{1}{2}
\left[ \Upsilon_{2\pi( \frac{n_x}{L_x}, \frac{n_y}{L_y}, r_z)}
+ (-1)^{n_y} \Upsilon_{2\pi(-\frac{n_x}{L_x}, \frac{n_y}{L_y},-r_z)}
\right.& & \\
\quad \left. + (-1)^{n_x} \Upsilon_{2\pi( \frac{n_x}{L_x}, \frac{n_y}{L_y},-r_z)}
+ (-1)^{n_x + n_y} \Upsilon_{2\pi(-\frac{n_x}{L_x}, \frac{n_y}{L_y}, r_z)}
\right]
& \quad \mbox{for} \quad
& n_x \in Z^+,\; n_y \in Z,\; r_z \in R^+, \\
\frac{1}{\sqrt{2}}
\left[\Upsilon_{2\pi( 0, \frac{n_y}{L_y}, r_z)}
+ \Upsilon_{2\pi( 0, \frac{n_y}{L_y}, -r_z)}
\right]
& \quad \mbox{for} \quad
& n_y \in 2Z,\; r_z \in R^+, \\
\frac{1}{\sqrt{2}}
\left[ \Upsilon_{2\pi( \frac{n_x}{L_x}, \frac{n_y}{L_y}, 0)}
+ (-1)^{n_y} \Upsilon_{2\pi( -\frac{n_x}{L_x}, \frac{n_y}{L_y}, 0)}
\right]
& \quad \mbox{for} \quad
& n_x \in 2Z^+,\; n_y \in Z, \\
\Upsilon_{2\pi( 0, \frac{n_y}{L_y}, 0)}
& \quad \mbox{for} \quad
& n_y \in 2Z. \\
\end{array}$$
Following the same procedure as before, we obtain that the analog of Eq. (\[random\_relation\]) is given by
1. when $n_x \in Z^+$, $n_y \in Z$ and $r_z \in R^+$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_y} \hat e_{k_x, - k_y, k_z} ,$$ so that it is a real random variable when $k_y = 0$,
2. when $n_y \in 2 Z$ and $r_z \in R^+$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{0, k_y, k_z}^* = \hat e_{0, - k_y, k_z} ,$$ so that it is a real random variable when $k_y = 0$,
3. when $n_x \in 2 Z^+$ and $n_y \in Z$, $\hat e_{{\bf k}}$ must satisfy $$\hat e_{k_x, 0, k_z}^* = (- 1)^{n_y} \hat e_{k_x , - k_y, 0} ,$$ so that it is a real random variable when $k_y = 0$,
4. when $n_y \in 2 Z$, it has to be such that $$\hat e_{0, k_y, 0}^* = \hat e_{0, - k_y, 0} .$$
Singly Periodic Spaces {#SectionSinglyPeriodic}
======================
We will first find the eigenmodes of the slab space, and then use them to find the eigenmodes of the slab space with flip.
Slab Space {#SubsectionSlabSpace}
----------
Just as a 3-torus is the quotient of Euclidean space under the action of three linearly independent translations and a chimney space is the quotient by two translations, a [*slab space*]{} is the quotient of ${{{{\bf {E}}}}^3}$ by a single translation. Its fundamental domain is an infinitely tall and wide slab (Figure \[FigureSlabSpaces\]), with opposite faces identified straight across. The allowable wave vectors ${{\bf k}}$ for an eigenmode $\Upsilon_{{\bf k}}$ of a slab space define a family of parallel planes.
If we choose coordinates so that the translation takes the form $${{\bf T}}= (0, 0, L_z) ,
\label{SlabSpaceTranslation}$$ then the allowed wave vectors ${{\bf k}}$ are $$\label{SlabSpaceBasis}
{{\bf k}}= 2 \pi \, \left(r_x, \, r_y, \, \frac{n_z}{L_z} \right) ,$$ for real values of $r_x$ and $r_y$ and integer values of $n_z$. The corresponding orthonormal basis is $$\begin{array}{lcl}
\Upsilon_{2\pi\,(r_x,\, r_y,\, \frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& r_x, r_y \in R,\;n_z \in Z.
\end{array}$$ Even if we restrict to a fixed modulus $k = |{{\bf k}}|$, the eigenmodes of slab space remain continuous, not discrete.
If desired, one could construct a more general slab space by identifying opposite faces with a rotation. Such a space would be topologically the same as a standard slab space, but geometrically different.
As in the case of the torus and of the chimney space, the random variable $\hat e_{{\bf k}}$ is a complex random variable satisfying $$\hat e_{{\bf k}}^* = \hat e_{- {{\bf k}}} .$$
Slab Space with Flip {#SubsectionSlabSpaceWithFlip}
--------------------
A [*slab space with flip*]{} is generated by $${\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)}
\; \mapsto \;
{\left(\begin{array}{rrr}
{- 1} & {0} & {0} \\
{0} & {\;\;\; 1} & {0} \\
{0} & {0} & {\;\;\; 1} \end{array} \right)} {\left(\begin{array}{c}
{x} \\
{y} \\
{z} \end{array} \right)} + {\left(\begin{array}{c}
{0} \\
{0} \\
{L_z / 2} \end{array} \right)} .
\label{SlabSpaceWithFlip}$$ The Invariance Lemma provides the orthonormal basis for its eigenmodes $$\begin{array}{lcl}
\left[
\Upsilon_{2\pi\,( r_x,\, r_y,\, \frac{n_z}{L_z})}
+ (-1)^{n_z} \Upsilon_{2\pi\,(-r_x,\, r_y,\, \frac{n_z}{L_z})}
\right]
& \quad \mbox{for} \quad
& r_x \in R^+,\; r_y \in R,\; n_z \in Z, \\
\Upsilon_{2\pi\,( 0,\, r_y,\, \frac{n_z}{L_z})}
& \quad \mbox{for} \quad
& r_y \in R,\;n_z \in 2Z.
\end{array}$$
Concerning the properties of the random variable, the analog of Eq. (\[random\_relation\]) is given by
1. when $r_x \in R^+$, $r_y\in R$ and $n_z \in Z$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{k_x, k_y, k_z}^* = (- 1)^{n_z} \hat e_{k_x, - k_y, - k_z} .$$ It is thus a real random variable when $k_y = k_z = 0$,
2. when $r_y \in R$ and $n_z \in 2 Z$, $\hat e_{{\bf k}}$ satisfies $$\hat e_{0, k_y, k_z}^* = \hat e_{0, - k_y, - k_z} .$$
Numerical simulations {#SectionCMBMaps}
=====================
We can now compute the correlation matrix and simulate CMB maps for the 17 multi-connected spaces described in the previous sections.
In all the simulations, we have considered a flat $\Lambda$CDM model with $\Omega_\Lambda = 0.7$, a Hubble parameter $H_0 \equiv 100 h
{\,{\mbox{km}^{}}} {\,{\mbox{s}^{-1}}} {\,{\mbox{Mpc}^{-1}}}$ with $h = 0.62$, a baryon density $\omega_{{\rm b}}\equiv \Omega_{{\rm b}}h^2 = 0.019$ and a spectral index $n_{{\scriptscriptstyle {\rm S}}}= 1$. With this the radius of the last scattering surface is $R_{{\rm LSS}}= 15.0 {\,{\mbox{Gpc}^{}}}$.
We present a series of CMB maps with a resolution of $\ell = 120$ for the different spaces. These maps are represented on a sphere portraying the last scattering surface seen from outside in the universal cover. Images of the last scattering surface under the action of a holonomy and its inverse are shown and their intersection gives a pair of matched circles. All the plots presented here contain only the Sachs-Wolfe contribution and omit both the Doppler and integrated Sachs-Wolfe contributions.
For the compact spaces, the characteristics of the fundamental polyhedra are
- $L_x = L_y = 0.64$ for all spaces,
- $L_z = 1.28$ for the half- and quarter-turn spaces (Figs. \[plot1\], \[plot2\])
- $L_z = 1.92$ for the third- and sixth-turn spaces (Figs. \[plot3\] and \[plot4\])
- $L_z = 0.64$ for the Hantzsche-Wendt and Klein spaces (Figs. \[plot5\] through \[plot9\]).
Turning to the $C_\ell$, let us examine the effects of the topology and of the volume of the fundamental domain.
To understand the properties of the angular power spectrum on large scales, let us develop a simple geometrical argument based on the properties of the eigenmodes of the Laplacian operator (see Ref. [@inoue2] for an analogous discussion and Ref. [@prl] for the spherical case). In the simply connected Euclidean space ${{{{\bf {E}}}}^3}$, the number $N_{{\scriptscriptstyle {\rm SC}}}$ of modes between $k$ and $k + \Delta k$ is simply given by $N_{{\scriptscriptstyle {\rm SC}}}(k) = 4 \pi k^2 \Delta k$, whatever the scale. Now, due to the topology, most modes will disappear from the spectrum and we are left with wavenumbers of modulus $$k = 2 \pi \sqrt{\left( \frac{n_x}{L_x}\right)^2
+ \left(\frac{n_y}{L_y}\right)^2
+ \left(\frac{n_z}{L_z}\right)^2 } .$$ On very small scales (large $k$), the Weyl formula [@lwugl] allows us to determine the number $N_{{\scriptscriptstyle {\rm MC}}}$ of modes remaining in the spectrum: asymptotically, $N_{{\scriptscriptstyle {\rm MC}}}(< k) \sim V k^3 / 6 \pi^2$ (see, e.g., Fig. 2 of Ref. [@rulw02]). It follows that the number of modes between $k$ and $k + \Delta k$ is now given by $N_{{{\scriptscriptstyle {\rm MC}}}, \infty} (k) \sim V k^2
\Delta k / 2 \pi^2 \sim V N_{{\scriptscriptstyle {\rm SC}}}/ (2 \pi)^3$. Thus we may set the overall normalization on small scales where the effect of the topology reduces to an overall rescaling. But this has implications concerning the large scales.
Consider a rectangular torus with a square cross section of size $L_x = L_y$ and with height $L_z$, and let the relative proportions of $L_x$ and $L_z$ vary.
When $L_x \gg L_z $, the space looks like a slab space and the modes on large scales (i.e., such that $2 \pi / L_x \ll k \ll 2 \pi / L_z$) have a modulus $k \sim \frac{2 \pi}{L_x} \sqrt{n_x^2 + n_y^2}$, so that they approach a two-dimensional distribution. Since the number of modes with $k < k_0$ is given by $N (< k_0) = \sum_{n = 0}^{k_0^2}
r_2(n) = \pi k_0^2 + {\cal O} (k_0)$, where $r_n (p)$ is the number of representations of $p$ by $n$ squares, allowing zeros and distinguishing signs and order (e.g., $r_2 (5) = 8$ and $r_3 (4) =
6$), we obtain that the number of modes between $k$ and $k + \Delta k$ is now given by $N_{{{\scriptscriptstyle {\rm MC}}}, 0} (k) \sim L_x^2 k \Delta k / 2
\pi$. Defining the relative weight as $$w (k) \equiv \frac{(2 \pi)^3}{V} \frac{N_{{{\scriptscriptstyle {\rm MC}}}, 0} (k)}{N_{{\scriptscriptstyle {\rm SC}}}(k)} ,$$ we obtain that $w \sim (\pi / k L_z) \gg 1$ so that the large scale modes are boosted compared with the mode distribution of the simply connected space exactly as if the spectral index $n_{{\scriptscriptstyle {\rm S}}}$ were lowered by 1. In the hypothesis of a scale invariant spectrum $n_{{\scriptscriptstyle {\rm S}}}= 1$, one therefore expects that the $\ell (\ell + 1) C_\ell$ spectrum will behave as $\ell^{- 1}$ for the relevant scales
When $L_x \ll L_z$, the space looks like a chimney space and the modes on large scales (i.e., such that $2 \pi / L_z \ll k \ll 2 \pi / L_x$) have a modulus $k \sim 2 \pi n_z / L_z$ so that they approach a one-dimensional distribution. It follows that the number of modes between $k$ and $k + \Delta k$ is now given by $N_{{{\scriptscriptstyle {\rm MC}}}, 0} (k)
\sim L_z \Delta k / \pi$, so that $w(k) \sim (2 \pi / k^2 L_x^2) \gg
1$. Again, this will imply a relative boost of the spectrum on large scales as if the spectral index were lowered by 2.
When $L_x \sim L_z$, as long as we are above the mode cut-off, one has a three-dimensional distribution of modes so that the relative weight of large scale modes is $w \sim 1$, as in a simply connected space. The signature of the topology in the $C_\ell$ exists at sufficiently large scales in the form of small spikes around the expected value in a simply connected space due to the discrete nature of the $k$-spectrum.
These results are summarized in Fig. \[oprol\].
Location of the Observer {#SectionLocationOfObserver}
========================
The 3-torus, chimney space, and slab space are exceptional because they are globally homogeneous. A globally homogeneous space looks the same to all observers within it; that is, a global isometry will take any point to any other point. The remaining multi-connected flat spaces, by contrast, are not globally homogeneous and may look different to different observers. For ease of illustration, consider the two-dimensional Klein bottle: the self-intersections of the “last scattering circle” are different for an observer sitting on an axis of glide symmetry (Figure \[KleinBottleBasepoints\] left) than for an observer sitting elsewhere (Figure \[KleinBottleBasepoints\] right). Analogously in three dimensions, the lattice of images of the last scattering surface may differ tremendously for observers sitting at different locations within the same space. The power spectrum, the statistical anisotropies, and the matching circles may all differ.
Moving the observer to a new basepoint would needlessly complicate existing computer software for simulating CMB maps. It is much easier to move the whole universe, leaving the observer fixed! In technical terms, we want to replace an eigenmode $\Upsilon_{{\bf k}}({{\bf x}})$ with the translated mode $\Upsilon_{{\bf k}}({{\bf x}}+ {{\bf x}}_{{\rm obs}})$, where ${{\bf x}}_{{\rm obs}}$ is the desired location for the observer. The translated mode is quite easy to compute: $$\begin{aligned}
\label{TranslatedObserver}
\Upsilon_{{\bf k}}({{\bf x}})
& \mapsto & \Upsilon_{{\bf k}}({{\bf x}}+ {{\bf x}}_{{\rm obs}}) \nonumber \\
& = & {{\rm e}^{i {{\bf k}}\cdot ({{\bf x}}+ {{\bf x}}_{{\rm obs}})}} \nonumber \\
& = & {{\rm e}^{i {{\bf k}}\cdot {{\bf x}}_{{\rm obs}}}} \; {{\rm e}^{i {{\bf k}}{{\bf x}}}} \nonumber \\
& = & {{\rm e}^{i {{\bf k}}\cdot {{\bf x}}_{{\rm obs}}}} \; \Upsilon_{{{\bf k}}} ({{\bf x}}) .\end{aligned}$$ For a simple mode $\Upsilon_{{\bf k}}({{\bf x}})$, the translation produces a phase shift (by a factor of ${{\rm e}^{i {{\bf k}}\cdot {{\bf x}}_{{\rm obs}}}}$) and nothing more. The full effect is seen when one considers linear combinations of simple modes: $$\begin{aligned}
\label{psh}
a_1 \Upsilon_{{{\bf k}}_1} ({{\bf x}})
+ a_2 \Upsilon_{{{\bf k}}_2} ({{\bf x}})
\quad\mapsto\quad
a_1 {{\rm e}^{i {{\bf k}}_1 \cdot {{\bf x}}_{{\rm obs}}}} \Upsilon_{{{\bf k}}_1} ({{\bf x}})
+ a_2 {{\rm e}^{i {{\bf k}}_2 \cdot {{\bf x}}_{{\rm obs}}}} \Upsilon_{{{\bf k}}_2} ({{\bf x}}) .\end{aligned}$$ Each term undergoes a different phase shift, so the final sum may be qualitatively different from the original.
Note that the phase shift (\[psh\]) induced by the change of the position of the observer does not influence the properties of the statistical variable $\hat e_{{\bf k}}$, but does influence the way a given mode contributes to a given angular scale. This is depicted in Figure \[figsw\], where the angular power spectrum is shown in a half-turn space for various positions of the observer. Corresponding examples of maps are shown in Fig. \[genobs1\].
Conclusions {#SectionConclusions}
===========
This article has presented the tools required to compute CMB maps for all multi-connected flat spaces. We gave for each space
- the polyhedron and holonomy group,
- the eigenmodes of the Laplacian.
We then presented simulated maps for all of the nine compact non homogeneous spaces. On the basis of the angular power spectra we compared the effect of different topologies and different configurations for a given topology. We also implemented the effect of an arbitrary position of the observer which yields significant effects for non-homogeneous spaces. We investigated this effect both on simulated maps and angular power spectra. In particular, it shows that generically matched circles are not back-to-back and that their relative position depends on the position of the observer.
All these tools and simulations will be of great help for extending the conclusions reached on the torus and to investigate their genericity as well as for providing test maps for any method wishing to detect (an interpret) the breakdown of global isotropy.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank Adam Weeks Marano for drawing the figures of the fundamental polyhedra. We also thank François Bouchet and Simon Prunet for discussions, and Neil Cornish, David Spergel, Glenn Starkman and Max Tegmark for fruitful exchanges. J.W. thanks the MacArthur Foundation for its support.
[100]{}
A. Riazuelo, J.-P. Uzan, R. Lehoucq, and J.W. Weeks, to appear in ; J.-P. Uzan and A. Riazuelo, C. R.Acad. Sci. (Paris), to appear.
A. Benoît [[*et al.*]{}]{}, A. & A. [**399**]{}, L19 (2003); A. Benoît [[*et al.*]{}]{}, A. & A. [**399**]{}, L25 (2003).
D. Spergel [[*et al.*]{}]{}, [Astrophys. J. Suppl. ]{}[**148**]{}, 175 (2003).
WMAP homepage: \[[http://map.gsfc.nasa.gov/ ]{}\].
Planck homepage: \[[http://astro.estec.esa.nl/Planck/ ]{}\].
J. Weeks, R. Lehoucq, and J.-P. Uzan, [Class. Quant. Grav. ]{}[**20**]{}, 1529 (2003).
E. Gausmann, R. Lehoucq, J.-P. Luminet, J.-P. Uzan, and J. Weeks, [Class. Quant. Grav. ]{}[**18**]{}, 5155 (2001).
R. Lehoucq, J. Weeks, J.-P. Uzan, E. Gausmann, and J.-P. Luminet, [Class. Quant. Grav. ]{}[**19**]{} 4683, (2002).
R. Lehoucq, J.-P. Uzan, and J. Weeks, Kodai Math. Journal [**26**]{} (2003) 119.
J.-P. Uzan, A. Riazuelo, R. Lehoucq, and J. Weeks, to appear in .
J.-P. Luminet, J. Weeks, A. Riazuelo, R. Lehoucq, and J.-P. Uzan, Nature (London) [**425**]{} (2003) 593.
I.Y. Sokolov, [JETP ]{}Lett. [**57**]{}, 617 (1993).
A.A. Starobinsky, [JETP ]{}Lett. [**57**]{}, 622 (1993).
D. Stevens, D. Scott, and J. Silk, , 20 (1993).
A. de Oliveira-Costa and G.F. Smoot, , 447 (1995).
J. Levin, E. Scannapieco, G. de Gasperis, and J. Silk, , 123006 (1998).
E. Scannapieco, J. Levin, and J. Silk, Month. Not. R. Astron.Soc. [**303**]{}, 797 (1999).
R. Bowen and P. Ferreira, , 04132 (2002).
K.T. Inoue, , 103001 (2000).
K.T. Inoue, [Class. Quant. Grav. ]{}[**18**]{}, 1967 (2001).
K.T. Inoue and N. Sugiyama, (2003) 043003.
B. Roukema, Month. Not. R. Astron. Soc. [**312**]{}, 712 (2000); [*ibid.*]{} [Class. Quant. Grav. ]{}[**17**]{}, 3951 (2000).
M. Lachièze-Rey and J.-P. Luminet, Phys. Rept. [**254**]{} 135, (1995).
J.-P. Uzan, R. Lehoucq, and J.-P. Luminet, in [*XIX$^{\it th}$ Texas Symposium on Relativistic Astrophysics and Cosmology*]{}, edited by E. Aubourg [[*et al.*]{}]{} (Tellig, Châtillon, France), CD-ROM file [04/25]{}.
J. Levin, Phys. Rept. [**365**]{}, 251 (2002).
M. Tegmark, A. de Oliveira-Costa, and A. Hamilton, \[[arXiv:astro-ph/0302496]{}\].
J.-P. Uzan, U. Kirchner, and G.F.R. Ellis, Month. Not. R. Astron.Soc. [**344**]{}, L65 (2003); G. Efstathiou, Month. Not. R. Astron.Soc. [**343**]{}, L95 (2003); C.R. Contaldi, M. Peloso, L. Kofman, and A. Linde, JCAP [**0307**]{}, 002 (2003); J.M. Cline, P. Crotty, and J. Lesgourgues, JCAP [**0309**]{}, 010 (2003).
A. de Oliveira-Costa, M. Tegmark, M. Zaldarriaga, and A. Hamilton, \[[arXiv:astro-ph/0307282]{}\].
N.J. Cornish, D. Spergel, and G. Starkmann, [Class. Quant. Grav. ]{}[**15**]{}, 2657 (1998).
N. Cornish, D. Spergel, G. Starkman, E. Komatsu, \[[arXiv:astro-ph/0310233]{}\].
C.J. Copi, D. Huterer, and G.D. Starkman, \[[arXiv:astro-ph/0310511]{}\].
H.K. Eriksen, F.K. Hansen, A.J. Banday, K.M. Gorski, and P.B. Lilje, \[[arXiv:astro-ph/0307507]{}\].
C.-G. Park, \[[arXiv:astro-ph/0307469]{}\].
N.J. Cornish and D.N. Spergel, \[[arXiv:math.DG/9906017]{}\].
E. Feodoroff, Russian Journ. for Crystallography and mineralogy [**21**]{} (1885) 1.
L. Bierberbach, Mathematische Annalen [**70**]{}, (1911) 297; [*Ibid*]{}, [**72**]{} (1912) 400.
W. Novacki, Commentarii Mathematici Helvetici [**7**]{} (1934) 81.
C. Adams and J. Shapiro, American Scientist [**89**]{}, 443 (2001).
Barry Cipra, [*What’s Happening in the Mathematical Sciences*]{}, Amer. Math. Soc. (2002).
D.A. Varshalovich, A.N. Moskalev, and V.K. Khersonskii, [*Quantum theory of angular momentum*]{} (World Scientific, Singapore, 1988).
[^1]: In this article, we follow the cosmological use and we call “flat spaces” the eighteen types of Euclidean spaces, and “Euclidean space” the simply connected universal cover ${{{{\bf {E}}}}^3}$.
[^2]: Test maps for these spaces are available on demand.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Time evolution of wave packets built from the eigenstates of the Dirac equation for a hydrogenic system is considered. We investigate the space and spin motion of wave packets which, in the non-relativistic limit, are stationary states with a probability density distributed uniformly along the classical, elliptical orbit (elliptic WP). We show that the precession of such a WP, due to relativistic corrections to the energy eigenvalues, is strongly correlated with the spin motion. We show also that the motion is universal for all hydrogenic systems with an arbitrary value of the atomic number $Z$.'
address:
- |
Institute of Physics, University of Zielona Góra, 65-246 Zielona Góra, Poland\
[ [email protected]]{}
- |
Institute of Physics, Maria Curie-Skłodowska University, 20-031 Lublin, Poland\
[ [email protected]]{}
- |
$^\spadesuit$ Institut des Sciences Nucléaires, F 38026 Grenoble Cedex, France\
[ [email protected]]{}
- |
$^\clubsuit$ The Weizmann Institute of Science 76100 Rehovot, Israel\
[ [email protected]]{}
author:
- 'P Rozmej, M Turek, R Arvieu$^\spadesuit$, and I S Averbukh$^\clubsuit$'
title: Relativistic precession and spin dynamics of an elliptic Rydberg wave packet
---
The detailed study of the time evolution of quantum wave packets (WPs) in simple atomic or molecular systems has been the object of growing attention for more than ten years [@r1]. Most of the previous theoretical studies were done in non-relativistic framework. In the field of relativistic quantum mechanics most of the efforts have been focused on the problem of the interaction between the atoms and a mixture of static fields with, most of the time, intense laser fields [@r2; @r3; @r4; @r5; @r6; @r7; @r8; @r9; @r10; @r11]. Under these conditions the use of a relativistic theory is fully justified since the external field is then able to bring considerable energy to the WP. For isolated atoms, however, the use of relativistic dynamics is more questionable, if the WP is followed or observed only during a short period of time. In ref. [@r12] relativistic wave packets, corresponding to circular orbits, have been constructed for hydrogenic atoms with a large $Z$, and propagated over time according to the Dirac equation. Particular attention was paid to the spin collapse event, [*i.e.*]{} to the maximum entanglement, in the course of time, of the spin degree of freedom with the spatial ones. This phenomenon was first shown to take place for a WP in a harmonic oscillator with a spin-orbit force [@r13], where it is periodic. For this reason it has been called the [*spin-orbit pendulum*]{}. In the Dirac equation with a Coulomb potential, it is produced by the built-in spin-orbit force and it is not periodic. The time scale where this effect can manifest itself was discussed in ref. [@r12], as a function of the charge $Z$ of the atom and the average principal quantum number $N$ of the WP. This phenomenon is expected to take place on a longer time scale like the other time dependent quantum effects of spreading and of fractional revivals [@r14]. We intend below to complement this work by studying the possible relativistic precession of elliptic WPs and by comparing this precession to the spin motion. A preliminary version of the present work was already presented at the ECAMP VII conference [@ecamp].
There are two possible ways to build up a WP ’on top of a classical elliptic orbit’ in non-relativistic mechanics. One of them is by kicking properly a well designed WP, for example a Gaussian WP, that is then evolved in time by the free Hamiltonian until it spreads above an average classical ellipsis. This method is not very easy to apply and its main inconvenience is to produce an internal motion within the ellipsis that is able to blur the precession. Therefore, we have preferred a second method, much simpler and even more elegant, which consists of using the coherent WPs of ref. [@r15], which are stationary states of the non-relativistic Coulomb problem. The space probability density of these states was indeed shown to be highly concentrated around a classical Bohr-Sommerfeld ellipsis. If these states can be adapted to relativistic dynamics their time evolution will doubtlessly be due to relativistic effects, [*i.e.*]{} due to the fine structure that will act as a perturbation. Let us first show how to adapt these states to a relativistic theory. We want the spatial part of the large components of the new WP to tend (in the non-relativistic limit) toward the state $|n \gamma\rangle $ of [@r15] which is defined as $$\begin{aligned}
\label{e1}
\fl
|n \gamma\rangle = \sum_{l,m} (-1)^{(l+m)/2} \frac{2^{n-l-1}(n-1)!}
{[(l-m)/2]![(l+m)/2]!} \left[ \frac{(l+m)!(l-m)!(2l+1)}{(n-l-1)!(n+l)!}
\right]^{1/2} \nonumber \\
\times \left(\sin \frac{\gamma}{2}\right)^{n-m-1}
\left(\cos \frac{\gamma}{2}\right)^{n+m-1}\, |n,l,m\rangle
\;\; = \;\; \sum_{lm} w_{lm}^{(n)}\,|n,l,m\rangle \; . \end{aligned}$$
The probability density $\langle n\gamma |n \gamma\rangle$ is fairly localized onto a Bohr orbit with eccentricity $\epsilon=\sin \gamma$ and the average angular momentum is
$$\label{e2}
l_{av} = (n-1)\, \cos{\gamma} \;,$$
where $n$ is the principal quantum number of the orbitals $|n,l,m\rangle$ which are admixed in (\[e1\]). The sum contains $n(n+1)/2$ values of $m$ with $l+m$ even. The contribution of states with $m<l$ decreases very rapidly with $m$ (more than one order of magnitude for each 2 units of $m$ as shown in fig. 1). The dominant weights are those with $m=l$ and their distribution is nearly Gaussian. The relative contribution of states with $m<l$ increases, however, for larger values of the eccentricity parameter $\epsilon$. The larger admixture of these states causes much faster daphasing of the WP. Therefore for illustration of typical precession phenomena he have chosen the case $\epsilon=0.4$.
In order to study the entanglement of the spin degrees of freedom with the orbital ones, it is natural in a non-relativistic theory to start from a product state of an arbitrary spinor $\left( \begin{array}{c}a\\ b \end{array}\right)$ with the state $|n \gamma\rangle$. $$\label{e3}
|\Psi_{nr}\rangle = |n \gamma\rangle \,\left( \begin{array}{c}
a\\ b \end{array}\right)
= \sum_{lm} w_{lm}^{(n)}\,|n,l,m\rangle \,\left( \begin{array}{c}
a\\ b \end{array}\right) \,.$$ It may be expanded in eigenstates of total angular momentum $|n,l,j=l+s,m_j\rangle$ with $m_j=m+1/2$ or $m-1/2$ and $s=+1/2$ or $-1/2$.
$$\begin{aligned}
\label{e5}
\fl
|\Psi_{nr}\rangle = \sum_{lm} \, w_{lm}^{(n)} \,
\left\{ \,
a \, \left( \,
\sqrt{\frac{l+1+m}{2l+1}} \, | n, l, j_>, m+1/2 \rangle
- \sqrt{\frac{l- m}{2l+1}} \, | n, l, j_<, m+1/2 \rangle \,
\right) \right. \\
+\left.
b \, \left( \,
\sqrt{\frac{l+1-m}{2l+1}} \, | n, l, j_>, m-1/2 \rangle
+ \sqrt{\frac{l+ m}{2l+1}} \, | n, l, j_<, m-1/2 \rangle \,
\right) \right\} \nonumber \,. \end{aligned}$$
In a similar way as in ref. [@r12], for a circular WP, the state (\[e5\]) is transformed into a four component relativistic state $\Psi_r$ by replacing in (\[e5\]) the non-relativistic states $|n,l,j,m_j\rangle$ by the eigenstates of the Dirac equation with the same quantum numbers. In this manner the WP gets small components in the most natural way. The radial parts of the large and of the small components are taken obviously as different ones. The probability density of the relativistic WP built in this way is represented in fig. 2.
The time evolution of the WP is produced by introducing the phase factors $\exp{(-iE_l^+\,t)}$ and $\exp{(-iE_l^-\,t)}$ as coefficients of states with $j=l+1/2$ and $j=l-1/2$ with their corresponding eigenvalues. The four components $c_i(t)$ $i=1,\ldots, 4$ of the WP at time $t$ are given in the \[aA\] with $\Psi_r$ defined as $$\label{e6}
|\Psi_r(t)\rangle = \left( \begin{array}{c} |c_1(t)\rangle \\
|c_2(t)\rangle \\|c_3(t)\rangle \\|c_4(t)\rangle \end{array}
\right) \;.$$
It should be stressed that $\Psi_r$ for $t=0$ is not any more a product state of the form of eq. (\[e3\]) due to the built-in entanglement contained in the solutions of the Dirac equation. However, since the small components of $\Psi_r$ are indeed very small, this defect has no important effect on the magnitude of the initial spin. Since the spin-orbit coupling effect, [*i.e.*]{} spin-orbit pendulum [@r13], manifests itself more efficiently if $s$ and $l$ are perpendicular to each other, we choose for our discussion: $$\label{e7}
a = b =\frac{1}{\sqrt{2}} \quad\quad \mbox{\em i.e.}
\quad\quad \langle s_x \rangle_{nr} = \frac{1}{2} \;.$$
Precession of quantum elliptical states in the Coulomb field has already been considered in ref. [@crawford], starting also from the same coherent state as done here. However, the precession was studied only in non-relativistic quantum dynamics as a perturbation effect and no treatment of the spin was attempted.
Let us finally discuss the time units relevant to our problem. To the lowest order in terms of the fine structure constant, the energy of an eigenstate $|n,l,j=l+s\rangle$ of the Dirac equation can be written (in a.u.) for a hydrogenic atom of charge $Z$ as $$\label{e8}
E_{nlj} = \overline{E}_n-\frac{1}{2} \frac{Z^4\alpha^2}{n^3(l+s+1/2)}$$ with $s=+1/2$ or $-1/2$. The constant energy $\overline{E}_n$ produces no effect on WP, since it depends only on $n$. Let us define an average time unit $T_p$ ($p$ for precession): $$\begin{aligned}
\label{e9}
T_p & = & \frac{2\pi}{\langle \frac{dE_{nlj}}{dl} \rangle} \\
& = & \frac{4\pi n^3}{Z^4\alpha^2} \langle (l+s+\frac{1}{2})^2\rangle
\approx \frac{4\pi n^3\,l_{av}^2}{Z^4\alpha^2} \label{e10}\\
& = & \frac{2 \,l_{av}^2\,T_K}{(Z\alpha)^2}\;.\label{e10a}\end{aligned}$$ Brackets $\langle\rangle$ in (\[e9\])-(\[e10\]) denote average values, $l_{av}$ is given by (\[e2\]). For $n=50$, $\epsilon=0.4$ precession time $T_p$ ranges from $1.96\cdot 10^{-11}$s for $Z=92$ to $1.4\cdot 10^{-3}$s for $Z=1$. $T_K$ in (\[e10a\]) denotes the Kepler period.
It is necessary to compare $T_p$ to the radiative lifetime $T_{n,l}^{\mbox{\tiny rad}}$ of hydrogenic levels. We will use the estimation of $T_{n,l}^{\mbox{\tiny rad}}$ for a single $n,l$ state given in [@chang]. In atomic units it is $$\label{trad}
T_{n,l}^{\mbox{\tiny rad}} = \frac{3}{2\,\alpha^3\,Z^4}\,n^3
\left(l+\frac{1}{2}\right)^2\;,
% \frac{3\hbar^3}{2m\,\alpha^3\,Z^4\,e^4}\,
% n^3\left(l+\frac{1}{2}\right)^2$$ which was found to agree within 10% with experimental data. One obtains the universal ratio: $$\label{tptr}
\frac{T_p}{T_{n,l}^{\mbox{\tiny rad}}} = \frac{8\pi\,\alpha}{3} \approx 0.061$$ which guarantees that the precession of the wave packet takes place long enough time before even a single photon is emitted. The occurrence of $\alpha$ in this ratio is understandable, since $T_p$ is a classical-like quantity, while $T_{n,l}^{\mbox{\tiny rad}}$ is proportional to $1/\hbar$, because it can be expressed as a ratio of a typical energy of the emitted photon to the classical radiation rate corresponding to the orbital motion.
When $t=T_p$ the linear terms in the expansion of $E_{nlj}$ in the power of $l$ contribute on the average to $2\pi$ in the exponential factors and the WP is expected to restore. However the terms of the higher order dephase differently the various partial waves, and this leads to a spreading of the WP near the initial shape [@r14]. See the discussion on the magnitude of these terms in \[aB\]. Expression (10) or (11) (with the Kepler period $T_K=2\pi n^3/Z^2$) is also recognized as the classical precession time in the relativistic Coulomb problem [@r16].
Let us note that $dE_{nlj}/dl$ is also (to the first order) the energy difference between two spin-orbit partners. Therefore the precession time can also be interpreted as the recurrence time of the spin. Hence we should observe a strong correlation between the spatial motion of the density: the precession, and the spin motion.
Let us discuss now the dynamics dependence on the atomic number, $Z$. Formula (\[e10\]) suggests that the relativistic effects under consideration in this article, depend crucially on $Z$. For sure the highest possible $Z$ are indeed required to lower $T_p$ as much as possible. It is interesting to stress nevertheless, that within a very good approximation, a scaling of $Z$ is possible which leads to the universal behaviour of the wave packet (\[e1\]).
First of all since $E$, approximated by eq. (\[e8\]), scales as $Z^4$, [*i.e.*]{} as $T_p^{-1}$, the variable $Z$ disappears from $Et$ if we use the reduced time $t/T_p$. The autocorrelation function (\[ac\]) which is expressed only in terms of $Et$ is the simplest quantity which has a universal behaviour, provided the same $n$ and $\epsilon$ are used for all values of $Z$.
The other quantities, like the probability density and the the spin expectation values depend on the radial wave functions and radial integrals. In a non-relativistic theory scaling of the radial wave function is elementary, it is obtained by dividing the radial wave fuction by $Z^{3/2}$ and multiplying the radial variable by $Z$. This is an exact property. The ratio small/large components of the relativistic solutions scale as $\sqrt{(1-E)/(1+E)}$, [*i.e.*]{} roughly like $Z^2$ and the radial wave functions depend also in a more complicated manner on $Z$. Nevertheless, on the whole as seen in fig. 5 below, the small components contribute a very small part of the probability density even for $Z=92$ (see also fig. 1 of ref. [@r12]). The scaling of the probability density is therefore entirely governed by the large components [*i.e.*]{} by the non-relativistic theory.
In a similar way we have checked that the radial integrals which contribute to the spin expectation values have the same properties: the integrals $G_+, G_-$, and $G_{+-}$ are equal to 1, within less than 10$^{-5}$ and the other $F_+ ,F_- \dots$ contribute in a very small manner, also of the order of 10$^{-5}$.
Therefore, the universal behaviour of the wave packet is justified and except for fig. 2 no value of $Z$ is given. For longer times, the energy factors omitted in eq. (\[e8\]) which involve higher powers of $Z$, play a role and produce a genuine $Z$ dependence. Those effects will not be discussed here.
The probability density of the wave packet with $n=50, \epsilon=0.4$ and $a=b=1/\sqrt{2}$, with $Z=92$, is shown for a set of times up to $t=T_p$ in fig. 3 and fig. 4. The part of the density coming from the small components, shown in fig. 5, is also concentrated on top of an ellipsis but its magnitude is a thousands times smaller than the total density. For $t<T_p/4$ the density precesses as described classically with a small dispersion. However the spreading takes place very rapidly for larger $t$ and is quite extended for $t=T_p$.
The precession of the ellipsis, the recurrence and spreading can be seen in a more condensed manner in the autocorrelation function represented in fig. 6 for three different spin preparations (spin up, spin down and $a=b=1/\sqrt{2}$). For small $t/T_p$ the WP becomes almost orthogonal to its initial parent, for $t=T_p$ a recurrence occurs but the overlap is only 0.7. Another peak occurs at $t=2T_p$ but higher frequencies become important and spread the recurrence. For $t>3T_p$ these higher frequencies play a dominant role. An example of an approximate revival for $t=22.6\,T_p$ is displayed in fig. 7. Fractional revivals [@r14] can also be seen to some extent. Two examples of 1/3 and 1/2 revivals are presented in fig. 8.
The rough independence of the autocorrelation function on the spin preparation requires some explanation. We can approximate this function by $$\label{ac1}
\langle \Psi_{r}(0)|\Psi_{r}(t) \rangle \approx a^2 \sum_l
w_{ll}^2 \exp{(-iE^+_l t)}+b^2 \sum_l w_{ll}^2 \exp{(-iE^-_l t)} \;, % (12)$$ where we have neglected the terms with very small weights $w^2_{l,m\neq l}$. The contribution of the two sums above is almost the same because $w_{ll}$ has a smooth variation with $l$ on the one hand, and because of the exact equality $ E_l^-=E_{l-1}^+ $ on the other hand. Finally the time evolution of the spin averages is presented in fig. 9. This figure exhibits what we can call the relativistic spin-orbit pendulum. Although the wave packet is not prepared initially in a pure state of spin, the impurity is very small (for $a=b=1/\sqrt{2}$ one has $\langle \sigma_x\rangle \approx 0.9997$). As time goes on the spin stays very nearly in the $Oxy$ plane and rotates around $Oz$ with period $T_p$. Its magnitude slowly decreases, however, and for $5<t/T_p<10$ the spin is amost totally entangled with the orbital motion, since the average of its three projections are almost zero. During a period of time that last for about $5T_p$ the angular momentum of the spin is transferred to the orbital motion and therefore the mean trajectory is not planar anymore [@r13]. Since the orbital angular momentum of $40\hbar$ is much higher than $\hbar/2$ this geometrical effect can hardly be represented graphically. From $t\approx 15T_p$ a revival of spin occurs. The spin rotates and increases its magnitude at the same time. This event stays even longer than initially. However the recurrence is only partial.
In conclusion we can say, that for not too long time, the precession and spin motion of the WPs are fairly well described by the following approximation: non-relativistic wave functions of the form (\[e3\]) and relativistic energy eigenvalues. From this point of view the full, computationally very demanding, relativistic approach is unnecessary. However, this conclusion may be formulated not a priori but only a posteriori.
We would like to stress the richness of the dynamics just described. Indeed, in addition to the relativistic precession of the ellipsis we have obtained for longer times its fractional revivals. During the evolution the spin of the electron is entangled with its orbital motion to various degrees. All this agree completely with previous results [@r14] and [@r13]. However, it was obtained here in a full relativistic calculation. Since we have been able to scale the atomic number $Z$ we have given a universal behaviour to our WP. It is, however, clear that this scaling is destroyed in real atoms in a more realistic theory which would take into account quantum defects. Their inclusion would also distort the dynamics, for example it would change the precession time, in a way that is out of reach of our simple theory. As far as purely relativistic effects are concerned, like the importance of the small components or the zitterbewegung, we have found them negligible in the Coulomb problem, in contrast to the Dirac oscillator [@r19] in which they play a major role.
The authors would like to express their gratitude to the two referees who attracted their attention to the interplay between relativistic precession and radiative processes (eq. (12)) in the Coulomb problem. I.S. Averbukh thanks also C.R. Stroud, Jr. for useful discussions.
Details of the calculations {#aA}
===========================
Replacing a non-relativistic wave function in (\[e5\]) by the eigenstates of the Dirac equation, one obtains for $\Psi_r(t)$: $$\begin{aligned}
\label{a1}
\fl
\Psi_r(t)=\sum_{lm} \, w_{lm}^{(n)} & &
\left\{ a \sqrt{\frac{l+1+m}{2l+1}}
\pmatrix{ i g_{n'_{+}} \Omega_{l, j_>, m+1/2} \cr
- f_{n'_{+}} \Omega_{l+1, j_>, m+1/2} }
\exp{(-i E^+_lt)}\right.\nonumber
\\
& & \left. + a \sqrt{\frac{l-m}{2l+1}}
\pmatrix{ i g_{n'_{-}} \Omega_{l, j_<, m+1/2} \cr
- f_{n'_{-}} \Omega_{l-1, j_<, m+1/2} }
\exp{(-i E^-_lt)}\right.
\\
& &\left.
+ b \sqrt{\frac{l+1-m}{2l+1}}
\pmatrix{ i g_{n'_{+}} \Omega_{l, j_>, m-1/2} \cr
- f_{n'_{+}} \Omega_{l+1, j_>, m-1/2} }
\exp{(-i E^+_lt)}\right.\nonumber
\\
& & \left. - b \sqrt{\frac{l+m}{2l+1}}
\pmatrix{ i g_{n'_{-}} \Omega_{l, j_<, m-1/2} \cr
- f_{n'_{-}} \Omega_{l-1, j_<, m-1/2} }
\exp{(-i E^-_lt)}\nonumber\right\} \;.\end{aligned}$$ We have used the notations of [@r12]: $g(r)$ and $f(r)$ are the radial parts of the large and small components associated with the quantum numbers $n'=n-(j+1/2)$. These functions are multiplied by the spherical tensors $\Omega_{l, j, m_j}$ which are defined by eq. 4a and 4b of [@r12]. The energy of the spin orbit partners of a given value of $l$ is denoted by $E_l^+$ if $j=l+1/2$ and $E_l^-$ if $j=l-1/2$, respectively.
For numerical calculations it is convenient to rewrite components of (\[a1\]) in the following form: $$\begin{aligned}
\fl
|c_1(t)\rangle = i \sum_{l} \left\{ g_{n'_+} \exp{(-iE^+_lt)} \sum_{m} \,
w_{lm}^{(n)} \left( a \frac{l+1+m}{2l+1}|l,m\rangle
\right.\right. \nonumber
\\ \left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
+b \frac{\sqrt{(l+1-m)(l+m)}}{2l+1} |l,m-1\rangle \right)\\
+ \, g_{n'_-} \exp{(-iE^-_lt)}\sum_{m} \, w_{lm}^{(n)} \left(
a \frac{l-m}{2l+1}|l,m\rangle
\right.\nonumber
\\ \left.\left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
-b \frac{\sqrt{(l+1-m)(l+m)}}{2l+1} |l,m-1\rangle \right)\right\}
\nonumber \;,\end{aligned}$$ $$\begin{aligned}
\fl
|c_2(t)\rangle = i \sum_{l} \left\{ g_{n'_+} \exp{(-iE^+_lt)} \sum_{m} \,
w_{lm}^{(n)} \left( b \frac{l+1-m}{2l+1}| l,m \rangle
\right.\right. \nonumber
\\ \left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
+a \frac{\sqrt{(l+1+m)(l-m)}}{2l+1} |l,m+1\rangle \right) \\
+ \, g_{n'_-} \exp{(-iE^-_lt)}\sum_{m} \, w_{lm}^{(n)} \left(
b \frac{l+m}{2l+1}| l,m \rangle
\right.\nonumber
\\ \left.\left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
-a \frac{\sqrt{(l+1+m)(l-m)}}{2l+1} |l,m+1\rangle \right)\right\}
\nonumber \;,\end{aligned}$$
$$\begin{aligned}
\fl
|c_3(t)\rangle = \sum_{l} \left\{ f_{n'_+} \exp{(-iE^+_lt)} \sum_{m} \,
w_{lm}^{(n)} \left(
a \sqrt{\frac{(l+1+m)(l+1-m)}{(2l+1)(2l+3)}}|l+1, m \rangle
\right.\right. \nonumber
\\ \left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
+ b \sqrt{\frac{(l+1-m)(l+2-m)}{(2l+1)(2l+3)}}|l+1, m-1\rangle
\right) \\
+ \,f_{n'_-} \,\exp{(-iE^-_lt)} \sum_{m} \, w_{lm}^{(n)} \left(
a \sqrt{\frac{(l+m)(l-m)}{(2l+1)(2l-1)}}|l-1,m\rangle
\right.\nonumber
\\ \left.\left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
-b \sqrt{\frac{(l+m)(l-1+m)}{(2l+1)(2l-1)}} |l-1,m-1\rangle
\right)\right\} \nonumber \;,\end{aligned}$$
$$\begin{aligned}
\fl
|c_4(t)\rangle = \sum_{l} \left\{ f_{n'_+} \exp{(-iE^+_lt)} \sum_{m} \,
w_{lm}^{(n)} \left(
-b \sqrt{\frac{(l+1+m)(l+1-m)}{(2l+1)(2l+3)}}|l+1, m \rangle
\right.\right. \nonumber
\\ \left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
- a \sqrt{\frac{(l+1+m)(l+2+m)}{(2l+1)(2l+3)}}|l+1, m+1\rangle
\right) \\
+ \,f_{n'_-} \,\exp{(-iE^-_lt)} \sum_{m} \, w_{lm}^{(n)} \left(
-b \sqrt{\frac{(l+m)(l-m)}{(2l+1)(2l-1)}}|l-1,m\rangle
\right.\nonumber
\\ \left.\left.\quad\quad\quad\quad\quad\quad\quad\quad\quad
+a \sqrt{\frac{(l-m)(l-1-m)}{(2l+1)(2l-1)}} |l-1,m+1\rangle
\right)\right\} \nonumber \;.\end{aligned}$$
Using the explicit form of the WP, one obtains for the average values of the spin operators:
$$\begin{aligned}
\fl
\langle \sigma_x \rangle_t =
2 a b \sum_{lm} \left\{ w_{l,m}^2
\left[ G_{+} \frac{(l+1)^2-m^2}{(2l+1)^2}
+G_{-}\frac{l^2-m^2}{(2l+1)^2}
-F_{+}\frac{(l+1)^2-m^2}{(2l+1)(2l+3)} \right.\right. \nonumber \\
\hspace{5ex} \left. -F_{-}\frac{l^2-m^2}{(2l+1)(2l-1)}
+2G_{+-}\frac{l(l+1)+m^2}{(2l+1)^2} \cos{(\omega_l t)} \right]
\nonumber \\
\fl \hspace{5ex}
+ w_{l,m}w_{l-2,m-2} \left[ F_{-+}
\sqrt{\frac{(l+m)(l-1+m)(l-2+m)(l-3+m)}{(2l-1)^2(2l+1)(2l+3)}}
\cos{(\omega''_lt)} \right] \nonumber \\
\fl \hspace{5ex}
+ w_{l,m}w_{l,m-2} \frac{\sqrt{(l+m)(l-1+m)}}{(2l+1)} \hspace{-1ex}\\
\times \left[
\frac{\sqrt{(l+1-m)(l+2-m)}}{(2l+1)}
[G_{+}+G_{-}-2G_{-+}\cos{(\omega_l t)}] \right. \nonumber \\
\left.
\hspace{3ex}
-\frac{\sqrt{(l+2-m)(l-3-m)}}{(2l-1)} F_{-}
-\frac{\sqrt{(l+2-m)(l+1-m)}}{(2l+3)} F_{+} \right] \nonumber \\
\fl \hspace{5ex}
- w_{l,m}w_{l-2,m} \left[2\, F_{-+}
\sqrt{\frac{(l^2-m^2)((l-1)^2-m^2)}{(2l-1)^2(2l+1)(2l-3)}}
\cos{(\omega''_l t)} \right] \nonumber \\
\fl \hspace{5ex} \left.
+ w_{l,m}w_{l+2,m-2} \left[ F'_{-+}
\sqrt{\frac{(l+1-m)(l+2-m)(l+3-m)(l+4-m)}{(2l+1)(2l+3)^2(2l+5)}}
\cos{(\omega'_l t)} \right] \right\}
\nonumber ,\end{aligned}$$
$$\begin{aligned}
\fl
\langle \sigma_y \rangle_t =
2 a b \sum_{lm} \left\{ w_{l,m}^2 \left[
\frac{2m}{2l+1} G_{+-} \sin{(\omega_l t)}\right] \right.\\
\fl \hspace{5ex}
- w_{l,m}w_{l-2,m-2} \left[ F_{-+}
\sqrt{\frac{(l+m)(l-1+m)(l-2+m)(l-3+m)}{(2l-1)^2(2l+1)(2l+3)}}
\sin{(\omega''_l t)} \right] \nonumber \\
\fl \hspace{5ex}
+ w_{l,m}w_{l-2,m} \left[2\, F_{-+}
\sqrt{\frac{(l^2-m^2)((l-1)^2-m^2)}{(2l-1)^2(2l+1)(2l-3)}}
\sin{(\omega''_l t)} \right] \nonumber \\
\fl \hspace{5ex} \left.
- w_{l,m}w_{l+2,m-2} \left[ F'_{-+}
\sqrt{\frac{(l+1+m)(l+2-m)(l+3-m)(l+4-m)}{(2l-1)(2l+3)^2(2l+5)}}
\sin{(\omega'_l t)} \right] \right\}
\nonumber ,\end{aligned}$$
and
$$\begin{aligned}
\fl
\langle \sigma_z \rangle_t = \sum_{lm} \left\{ w_{l,m}^2 \left[
a^2 \frac{2m+1}{2l+1} \left(
G_{+} \frac{l+1+m}{2l+1} - G_{-} \frac{l-m}{2l+1}
-F_{+} \frac{l+1+m}{2l+3} + F_{-} \frac{l-m}{2l-1} \right) \right. \right.
\nonumber \\
\hspace{2ex} \left. + b^2 \frac{2m-1}{2l+1} \left(
G_{+} \frac{l+1-m}{2l+1} - G_{-} \frac{l+m}{2l+1}
-F_{+} \frac{l+1-m}{2l+3} + F_{-} \frac{l+m}{2l-1} \right) \right.
\nonumber \\
\hspace{2ex} \left. + 4 \,G_{+-}\, \cos{(\omega_l t)} \left(
a^2 \frac{(l-m)(l+1+m)}{(2l+1)^2} - b^2 \frac{(l+m)(l+1-m)}{(2l+1)^2}
\right) \right]
\nonumber \\
\fl \hspace{5ex} \left.
+ 4(a^2-b^2)\, w_{l,m}w_{l+2,m}\, F'_{-+}
\sqrt{\frac{((l+1)^2-m^2)((l+2)^2-m^2)}{(2l+1)(2l+3)^2(2l+5)}}
\cos{(\omega'_l t)} \right\} \end{aligned}$$
In the above formulas, the following notations have been introduced: $$\begin{aligned}
\omega_l &=& (E^+_l-E^-_l) \;, \\
\omega'_l &=& (E^-_{l+2}-E^+_l) \;, \\
\omega''_l &=& (E^-_{l}-E^+_{l-2}) \end{aligned}$$ Note that $\omega'_l=\omega''_{l+2}$. Radial integrals are denoted as follows: $$\begin{aligned}
G_{+} &=& \int_0^{\infty} \left( g^+_l(r) \right)^2 r^2 dr \;, \\
G_{-} &=& \int_0^{\infty} \left( g^-_l(r) \right)^2 r^2 dr \;, \\
F_{+} &=& \int_0^{\infty} \left( f^+_l(r) \right)^2 r^2 dr \;, \\
F_{-} &=& \int_0^{\infty} \left( f^-_l(r) \right)^2 r^2 dr \;, \\
G_{+-} &=& \int_0^{\infty} g^+_l(r)g^-_l(r) r^2 dr \;, \\
F_{+-} &=& \int_0^{\infty} f^+_{l+2}(r) f^-_l(r) r^2 dr \;, \\
F_{-+} &=& \int_0^{\infty} f^+_{l-2}(r) f^-_l(r) r^2 dr \;, \\
F'_{-+} &=& \int_0^{\infty} f^+_{l}(r) f^-_{l+2}(r) r^2 dr \;.\end{aligned}$$ Apart from the case of $G_{+}, G_{-}, F_{+}, F_{-}$ for $l=n-1$, which are relatively easily obtained analytically, all other radial integrals have been calculated numerically (using quadruple precision).
The autocorrelation function can be calculated from (\[a1\]) in a straightforward way: $$\begin{aligned}
\label{ac}
\fl \hspace{4ex}
\langle \Psi_{r}(0)|\Psi_{r}(t) \rangle
&=& \sum_l \left\{ \exp{(-iE_l^+t)}\left[ \sum_m w_{l,m}^2 \left(
a^2\frac{l+1+m}{2l+1}+b^2\frac{l+1-m}{2l+1} \right)\right]\right.
\nonumber \\
&& \hspace{2ex} +\left. \exp{(-iE_l^-t)}\left[ \sum_m w_{l,m}^2 \left(
a^2\frac{l-m}{2l+1}+b^2\frac{l+m}{2l+1} \right)\right]\right\} \end{aligned}$$
{#aB}
Let us use the following notation: $$\label{b1}
k = j+1/2 \quad,\quad {\cal E}_k= \frac{E_{nlj}}{m_0c^2} \quad,\quad
x = (Z\alpha)^2 \,.$$ The exact eigenergies ${\cal E}_k$ (in units ${m_0c^2}$) are given by $$\label{b2}
{\cal E}_k= \left[ 1+\frac{x^2}{(n-k+\sqrt{k^2-x^2})^2}\right]^{-1/2} \,.$$ Expanding this expression in Taylor series with respect to $x$ one obtains $$\label{b3}
\fl
{\cal E}_k= 1 - \frac{x^2}{2\,n^2} - \frac{x^4}{4\,n^3}\left(\frac{2}{k}
- \frac{3}{2\,n}\right) - \frac{x^6}{4\,n^3}\left( \frac{1}{2\,k^3}
+ \frac{3}{2\,n\,k^2} - \frac{3}{n^2\,k}+\frac{5}{4\, n^3} \right)
+ [O(x)]^8\,.$$ In eq. \[e8\] only the lowest term depending on $k$ in $x^4$ [*i.e.*]{} $ \delta_4{\cal E}_k= - \frac{x^4}{4\,n^3}\frac{2}{k}$ has been included. The higher order term $ \delta_6{\cal E}_k$ contributes very little, because the ratio $\delta_6{\cal E}_k/\delta_4{\cal E}_k= x^2(\frac{1}{k^2}+\frac{3}{4nk}
-\frac{3}{2n^2}+\frac{5k}{8n^3})$ reaches the maximum value about 0.0005 for $\epsilon=0.4$ and $Z=92$ and stays much smaller for lower $Z$. Then the time evolution for not too long period is mainly determined by the lowest order contribution (\[e8\]).
The precession time is determined by the derivative $$\label{b4}
\frac{\partial {\cal E}_k}{\partial k}\left|_{k=l_{av}}\right. =
\frac{x^4}{2\,n^3\,k^2}\left[ 1+ \frac{3x^2}{2}\left( \frac{1}{2k^2} +
\frac{1}{nk} - \frac{1}{2n^2}\right)\right] \;.$$ Again in eqs. (\[e9\])-(\[e10\]) only term of the order of $x^4$ has been used. The $x^6$-order term contributes at most about 0.00022 of the $x^4$-order term for $\epsilon=0.4$ and $Z=92$. Therefore we conclude that the $x^6$-order term can be safely neglected in estimation of the precession time $T_p$.
[99]{} For an introduction and a list of references see Schleich W P 2001 [*Quantum Optics in Phase Space*]{} (Berlin, Viley-Vch. ) Bosanac S D 1993 [*J. Phys. A: Math. Gen.* ]{} [**25**]{} 5523 Klystra N J, Ermolayev A M and Joachin C J 1997 [*J. Phys. B: At. Mol. Opt. Phys.* ]{} [**30**]{} L449 Rathe U W, Keitel C H, Protopapas M and Knight P L 1997 [*J. Phys. B: At. Mol. Opt. Phys.* ]{} [**30**]{} L531 Rutherford G H and Grobe R 1998 [*J. Phys. A: Math. Gen.* ]{} [**31**]{} 9931 Su Q, Smetanko B A and Grobe R 1998 [*Laser Phys.* ]{} [**8**]{} 93 Braun J W, Su Q and Grobe R 1999 [*Phys. Rev. A*]{} [**59**]{} 604 Wagner R E, Peverly P J, Su Q and Grobe R 1999 [*Phys. Rev. A*]{} [**61**]{} 035402 Krekora P, Wagner R E, Su Q and Grobe R 2001 [*Phys. Rev. A*]{} [**63**]{} 025404 Krekora P, Su Q and Grobe R 2001 [*J. Phys. B: At. Mol. Opt. Phys.* ]{} [**34**]{} 2795 Nilsen H M, Masden L B and Hansen J P 2001 [*J. Phys. B: At. Mol. Opt. Phys.* ]{} [**34**]{} L39 Arvieu R, Rozmej P and Turek M 2000 [*Phys. Rev. A*]{} [**62**]{} 022514 Arvieu R and Rozmej P 1994 [*Phys. Rev. A*]{} [**50**]{} 4376 ;\
1996 [*J. Phys. B: At. Mol. Opt. Phys.* ]{} [**29**]{} 1339 Averbukh I S and Perelman N F 1989 [*Phys. Lett. A*]{} [**139**]{} 449 ; 1989 [*Sov. Phys. JETP*]{} [**69**]{} 464 Rozmej P, Arvieu R, Averbukh I S and Turek M 2001 ECAMP VII, The Seventh European Conference on Atomic and Molecular Physics, Berlin, 2-6 April 2001, Europhysics Converence Abstracts, Vol. 25B, p.99 Gay J C, Delande D and Bommier A 1989 [*Phys. Rev. A*]{} [**39**]{} 6587 Crawford M C 2001 [*Am. J. Phys.*]{} [**69**]{} 1182 Chang E C 1985 [*Phys. Rev. A*]{} [**31**]{} 495 Sommerfeld A 1921 [*Atombau und Spektrallinien*]{} (F. Fievag und Sohn, Braunschweig) Rozmej P, Arvieu R 1999 [*J. Phys. A: Math. Gen.* ]{} [**32**]{} 5637
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We study the stochastic stability in the zero-noise limit from a quantitative point of view.
We consider smooth expanding maps of the circle, perturbed by additive noise. We show that in this case the zero-noise limit has a quadratic speed of convergence, as conjectured by Lin, in 2005, after numerical experiments (see [@Lin]). This is obtained by providing an explicit formula for the first and second term in the Taylor’s expansion of the response of the stationary measure to the small noise perturbation. These terms depend on important features of the dynamics and of the noise which is perturbing it, as its average and variance.
We also consider the zero-noise limit from a quantitative point of view for piecewise expanding maps showing estimates for the speed of convergence in this case.
address:
- 'Dipartimento di Matematica, Università di Pisa, Largo Bruno Pontecorvo 5, 56127 Pisa, Italy'
- '—'
author:
- Stefano Galatolo
- Hugo Marsan
title: 'Quadratic response and speed of convergence of invariant measures in the zero-noise limit.'
---
Introduction
============
Deterministic dynamical systems are often used as models of physical and natural phenomena despite the ubiquitous presence in nature of small random perturbations or fluctuations. It is natural to study the robustness of the deterministic model to such random perturbations and which of the aspects of the deterministic dynamics are stable under small random perturbations. In this paper we consider the many important aspects of the statistical behavior of the system which are encoded in its invariant measures. We study hence quantitatively how these measures change when the system is perturbed by the adding of a small quantity of noise, in the so called zero-noise limit. More precisely, we study this limit and its speed of convergence from a quantitative point of view, also considering first and second order terms in this convergence. We will see that these terms depend on important features of the dynamics and of the noise which is perturbing it, as its average and variance.
Let $S_{0}:=(X,T)$ be a discrete time deterministic dynamical system where $X\ $is a metric space and$\ \ T:X\rightarrow X$ is a Borel measurable map. It is well known that $(X,T)$ can have several invariant measures, let us consider one of these measures and denote it by $\mu _{0}$.
Suppose now we perturb the system at each iteration by the adding of a small quantity of noise whose amplitude is expressed by a certain parameter $\delta $ $\in \lbrack 0,\overline{\delta })$ obtaining a family of random systems $\{S_{\delta }\}_{\delta \in \lbrack 0,\overline{\delta })}$ (these systems will formally be defined as suitable random dynamical systems, a precise definition will be given in Section \[QR2\]). Suppose $\{\mu
_{\delta }\}_{\delta \in \lbrack 0,\overline{\delta })}$ are stationary measures for $S_{\delta }.$ It is natural to investigate under which assumptions one may have$$\lim_{\delta \rightarrow 0}\mu _{\delta }= \mu _{0}.$$In this case the system and the measure $\mu _{0}$ are said to be *statistically stable* under small noise perturbations (or in the zero-noise limit). An invariant measure of a deterministic model which is stable under the small random perturbations which are present in nature is a measure that can be observed in the real phenomenon behind the model. For this reason this zero-noise limit was proposed by A. N. Kolmogorov as a tool to select the *physically meaningful* measures among the a priori many invariant measures of a deterministic system (see e.g. [@ER], [@Y]).
The statistical stability for the zero-noise limit (also called stochastic stability) was proved for several classes of systems, starting from uniformly hyperbolic ones to many interesting cases of non-uniform hyperbolic behavior ([@Ki],[@Y2], [@BK1], [@BK2], [@BY], [@met], [@Sh], [@AV], [@LSV], [@AT], [@AA], [@A], [@AK], [@S], [@BV]).
The mere existence of the zero-noise limit gives a qualitative information on the behavior of the system under perturbation. In practice it can be useful to have quantitative information on this convergence, both on the speed of the convergence and on the “direction” of change of the invariant measure after the perturbation. In [@Lin] several numerical experiments have been done to estimate the speed of convergence in the limit, conjecturing a quadratic speed in the case of smooth expanding maps and linear speed for the piecewise expanding and hyperbolic case. Other exponents have been conjectured in cases of weakly chaotic, non-uniformly hyperbolic systems.
In this paper we will consider these kinds of questions, investigating both quantitative estimates for the speed of the convergence and the direction of change of the invariant measure of the system under perturbation. This is strongly related to the linear response theory, although in this case we will not be only interested in the linear term in the response of the system to the perturbation, but also in the higher order terms, and in particular to the quadratic one.
The *Linear Response* means to quantify the response of the system when submitted to a certain infinitesimal perturbation as a derivative. For example, if one is interested in the linear response of the stationary measure of the system, we will consider the derivative of the invariant measure of interest with respect to the perturbation.
More precisely, let $\{S_{\delta }\}_{\delta \in \lbrack 0,\overline{\delta })}$ as above be the family of systems arising by some small perturbation of the initial system $S_{0}$ with stationary measures $\{\mu
_{\delta }\}_{\delta \in \lbrack 0,\overline{\delta })}$. The linear response of the invariant measure $\mu _{0}$ of $S_{0}$ under the given perturbation is defined by the limit $$\dot{\mu}:=\lim_{\delta \rightarrow 0}\frac{\mu _{\delta }-\mu _{0}}{\delta }
\label{LRidea}$$where the meaning of this convergence can vary from system to system. In some systems and for a given perturbation, one may get $L^{1}$-convergence for this limit; in other systems or for other perturbations one may get convergence in weaker or stronger topologies. The linear response to the perturbation hence represents the first order term of the response of a system to the perturbation and in this case, a linear response formula can be written: $$\mu _{\delta }=\mu _{0}+\dot{\mu}\delta +o(\delta ) \label{lin}$$which holds in some weaker or stronger sense. We remark that given an observable function $c:X\rightarrow \mathbb{R}$, if the convergence in is strong enough with respect to the regularity of $c$, we get
$$\lim_{t\rightarrow 0}\frac{\int \ c\ d\mu _{t}-\int \ c\ d\mu _{0}}{t}=\int
\ c\ d\dot{\mu} \label{LRidea2}$$
showing how the linear response controls the behavior of observable averages. For instance the convergence in hold when $c\in
L^{\infty }$ and the convergence of the linear response is in $L^{1}$.
Once the first order (the linear part) of the response of a system to a perturbation is understood, it is natural to study further orders. The second order of the response may then be related to the second derivative and to other natural questions, as convexity aspects of the response of the system under perturbation, or the stability of the first order response. Hence, if the Linear Response $\dot{\mu}$ represents the first order term of the response (see ), the Quadratic Response $\ddot{\mu}$ will represent the second order term of this response, analogous to the second derivative in usual Taylor’s expansion:$$\mu _{\delta }=\mu _{0}+\dot{\mu}\delta +\frac{1}{2}\ddot{\mu}\delta
+o(\delta ^{2}). \label{Quad}$$
We refer to [@BB] for a recent survey on linear response for deterministic systems and perturbations and to the introduction of [@GS] for a very recent survey in the case of response for random systems and higher order terms in the response of a system to deterministic or random perturbations. Focusing on zero-noise limits, we point out the paper [GL]{}, where among other results, linear and high order response are proven for deterministic perturbations and zero-noise limits of uniformly hyperbolic systems (see also [@Li2] for some earlier examples of linear response in the zero-noise limit for expanding maps and [@BGNN] for rigorous numerical methods for its approximation including an example of zero-noise limit).
In the paper [@GS], a relatively simple and quite general approach to the first and second order terms in the response of a system to perturbations is proposed and applied to deterministic perturbations of deterministic systems and perturbations of random systems. In Subsection \[secgenres\] we recall the main general results of [@GS]. We then apply it to the zero-noise limit, providing precise quantitative information on the convergence of the zero-noise limit and proving some of the conjectures suggested by the numerical experiments and the heuristic exposed in [@Lin], in particular considering zero-noise limits of expanding and piecewise expanding maps.
In the literature, the general approach to this problem is often based on considering the family of transfer operators $\{L_{\delta }\}_{\delta \in
\lbrack 0,\overline{\delta })}$ associated to the dynamical system and its perturbations, remarking that invariant and stationary measures are fixed points of this family of operators. Quantitative perturbative statements about these operators and its spectral picture will hence give information on the perturbation of invariant measures. In this paper we use these tools to study the zero-noise limit from a quantitative point of view in two main cases: *smooth expanding maps* and *piecewise expanding maps* of the circle. In the following two subsections we enter in more details about our main results in these two cases.
**Smooth expanding maps, response and zero-noise limit.** We consider a smooth expanding map $T\in C^{8}(\mathbb{S}^{1}\rightarrow
\mathbb{S}^{1})$, with its associated transfer operator $L_{T}:SM(\mathbb{S}^{1})\rightarrow SM(\mathbb{S}^{1}),$ where $SM(\mathbb{S}^{1})$ is the space of finite Borel measures with sign on $\mathbb{S}^{1},$ defined by$$(L_{T}(\mu ))(A)=\mu (T^{-1}(A))$$for each signed measure $\mu \in SM(\mathbb{S}^{1}).$ $L_{T}$ is also called the transfer operator associated to $T$ or pushforward map associated to $T$. We consider an i.i.d. random perturbation distributed according to a kernel $\rho \in BV([-1,1])$. $\forall \delta \in \lbrack 0,\overline{\delta
})$ we denote by $\rho _{\delta }$ the rescaling of $\rho $ with amplitude $\delta $ by $$\rho _{\delta }(x)=\frac{1}{\delta }\rho \left( \frac{x}{\delta }\right) .$$The transfer operator associated to the randomly perturbed map is then defined as $$L_{\delta }=\rho _{\delta }\ast L_{T}$$where $\ast $ stands for the ordinary convolution operator on $\mathbb{S}^{1}.$ Note that we can extend this definition to $\delta =0$ with $\rho
_{0}=\delta _{0}$ the Dirac mass. It can be proved that (see Section [secQR]{}) each operator $L_{\delta }$ has a unique fixed point $h_{\delta }$ in the Sobolev space $W^{7,1}(\mathbb{S}^{1})$ and hence $h_{\delta }$ is the stationary measure of the perturbed system.
The idea is to prove that this family of operators admits a linear, and even quadratic response when $\delta$ tends to $0$. In particular, we prove and extend a result conjectured in [@Lin], in which the author predicted a convergence of order $\delta ^{2}$. We will precise the coefficients of the order two Taylor’s expansion of this zero-noise limit proving the following theorem:
\[thm:quadrespsmooth\] The map $\delta \mapsto h_{\delta }$ has an order two Taylor’s expansion at $\delta =0$, with$$\left\Vert \frac{h_{\delta }-h_{0}}{\delta ^{2}}-\frac{\sigma^2(\rho )}{2}(Id-L_{T})^{-1}h_{0}^{\prime \prime }\right\Vert _{W^{{1,1}}}\underset{\delta \rightarrow 0}{\longrightarrow }0.$$with $\sigma^2(\rho )=\int_{-1}^{1}x^{2}\rho (x)dx$.
Next Section \[secQR\] is essentially devoted to the proof of this result. We prove the theorem by the application of some general linear and quadratic response statements we recall in subsection \[secgenres\]. In subsection \[verif\] we verify the several assumptions needed to apply those theorems, completing the proof at the end of Section \[secQR\].
**Piecewise expanding maps, quantitative stability and zero-noise limit.** We have seen that for smooth expanding maps there is a quadratic speed of convergence in the zero-noise limit. This depend both on the smoothness and on the strong chaoticity of the system. When having less regularity, the speed of convergence changes. In the second part of the paper we consider indeed piecewise expanding maps, allowing discontinuities. In this case we have systems still having strong chaoticity, and exponential decay of correlations, but the speed of convergence in the zero limit is of order 1. We prove in fact the following
\[prop:linresppiecewise\] Let $T:\mathbb{S}^{1}\rightarrow \mathbb{S}^{1} $ be a piecewise expanding map having no periodic turning points (see Section \[sec:pwexp\] for the precise definitions). Let us suppose we perturb the associated dynamical system with noise of amplitude $\delta $ as above, let $L_{\delta }$ be the associated transfer operators and let $h_{\delta }$ be a family of invariant measures for $L_{\delta }$ then $h_{\delta }\in Lip[0,1]$ and there is $C\geq 0$ such that for each $\delta
\in \lbrack 0,\overline{\delta })$$$||h_{0}-h_{\delta }||_{L^{1}}\leq C\delta \log \delta .$$Furthermore, there are examples of piecewise expanding maps (with periodic turning points) for which there is a constant $C^{\prime }$ such that for each $\delta \in \lbrack 0,\overline{\delta })$, $$||h_{0}-h_{\delta }||_{L^{1}}\geq C^{\prime }\delta .$$
Quadratic response and the zero-noise limit of expanding maps[secQR]{}
======================================================================
In this section we consider the zero-noise limit of expanding maps on the circle. We get precise estimates on the speed of convergence of this limit, proving Theorem \[thm:quadrespsmooth\].
In this section we will consider maps $T:\mathbb{S}^{1}\rightarrow \mathbb{S}^{1}$ satisfying the following assumptions
1. $T\in C^{8}(\mathbb{S}^{1}\mathbb{)},$
2. $|T^{\prime }(x)|\geq \alpha ^{-1}>1$ $\forall x\in \mathbb{S}^{1}$.
To $T$ is associated a linear map $L_{T}:SM(\mathbb{S}^{1})\rightarrow SM(\mathbb{S}^{1}),$ where $SM(\mathbb{S}^{1})$ is the space of Borel measures with sign on $\mathbb{S}^{1},$ defined by$$(L_{T}(\mu ))(A)=\mu (T^{-1}(A))$$for each signed measure $\mu \in SM(\mathbb{S}^{1}).$ $L_{T}$ is also called the transfer operator associated to $T$ or pushforward map associated to $T$.
We consider a perturbation of this transfer operator by adding to the deterministic dynamics generated by $T$ a random independent and identically distributed perturbation: the noise. In other words we consider a random dynamical system, corresponding to the stochastic process $(X_{n})_{n\in
\mathbb{N}}$ defined by $$X_{n+1}=T(X_{n})+\Omega _{n}\mod 1 \label{eq:syswaddnoise}$$where $(\Omega _{n})_{n\in \mathbb{N}}$ are i.i.d random variables distributed according to a probability density $\rho _{\delta }$ (the noise kernel) where $\delta $ represent the “size” of the perturbation. We suppose that $\rho _{\delta }$ is obtained by rescaling a certain distribution $\rho
$ $\in BV([-1,1]),$ as follows $$\rho _{\delta }(x)=\frac{1}{\delta }\rho \left( \frac{x}{\delta }\right)
\label{eq:kerdef}$$for each $\ \delta \in (0,1].$ The (annealed) transfer operator associated to the perturbed random system is then defined by$$L_{\delta }=\rho _{\delta }\ast L_{T} \label{opdef}$$where $\ast $ is the convolution operator on $\mathbb{S}^{1}$ (see e.g.[@Viana], Section 5 or [@GMN], Section 8 for more details on the definition of the annealed transfer operator).
Remark that for each $f\in L^{1}(\mathbb{S}^{1})$, one has for almost every $x\in \mathbb{S}^{1}$: $$((\rho _{\delta }-\delta _{0})\ast f)(x)=\frac{1}{\delta }\int_{-\delta
}^{\delta }\rho \left( \frac{y}{\delta }\right)
f(x-y)dy-f(x)=\int_{-1}^{1}\rho (z)f(x-\delta z)dz-f(x). \label{eq:kernel}$$Also, one can remark that $\rho $ being a zero-average probability kernel, it verifies $$\int_{-1}^{1}\rho (z)dz=1\qquad \int_{-1}^{1}\rho (z)zdz=0\qquad
\int_{-1}^{1}\rho (z)z^{2}dz:=\sigma ^{2}(\rho ).$$
To keep the notation compact we will denote $L_{0}:=L_{T}$. We remark that the invariant measures of the map $T$ are fixed points of $L_{0}$ and the stationary measures of the random system constructed by the adding of the noise are fixed points of $L_{\delta }$. We are interested in the properties of these measures and how they vary as $\delta $ goes to $0$. Their characterization as fixed points of $L_{\delta }$ will be sufficient for our purposes. We recall that in the case we are considering (expanding maps) there can be a large set of invariant measures for the deterministic map $T$ but only one which is absolutely continuous with respect to the Lebesgue measure. On the other hand the stationary measures for $L_{\delta }$ will always be absolutely continuous. This is well known (see e.g. [@notes]) but it can also be easily derived from the regularization estimates we prove in the following.
General linear response and quadratic response results[secgenres]{}
-------------------------------------------------------------------
In this subsection we state some general results from [@GS] about linear and quadratic response of fixed points of Markov operators under suitable perturbations. These results will be applied to our zero-noise limit and to the operators $L_{\delta }$ to get precise estimates on the speed of convergence of $h_{\delta }$ to $h_{0}.$
In the following we consider four normed vectors spaces of signed Borel measures on $\mathbb{S}^{1}.$ The spaces $(B_{ss},\Vert ~\Vert
_{ss})\subseteq (B_{s},\Vert ~\Vert _{s})\subseteq (B_{w},\Vert ~\Vert
_{w})\subseteq (B_{w},\Vert ~\Vert _{ww})\subseteq BS(\mathbb{S}^{1})$ with norms satisfying$$\Vert ~\Vert _{ww}\leq \Vert ~\Vert _{w}\leq \Vert ~\Vert _{s}\leq \Vert
~\Vert _{ss}.$$
We will assume that the linear form $\mu \rightarrow \mu (\mathbb{S}^{1})$ is continuous on $B_{i}$, for $i\in \{ss,s,w,ww\}$. Since we will consider Markov operators[^1] acting on these spaces, the following (closed) spaces $V_{ss}\subseteq
V_{s}\subseteq V_{w}\subseteq V_{ww}$ of zero average measures will play an important role. We define $V_{i}$ as:$$V_{i}:=\{\mu \in B_{i}\mid \mu (\mathbb{S}^{1})=0\}$$where $i\in \{ss,s,w,ww\}$. Suppose hence we have a one parameter family of such Markov operators $L_{\delta }.$ The following is a general statement establishing linear response for suitable perturbations of such operators.
\[thm:linresp\] Suppose that the family of bounded Markov operators $L_{\delta }:B_{i}\rightarrow B_{i},$ where $i\in \{ss,s,w\}$ satisfy the following:
- (regularity bounds) for each $\delta \in \left[ 0,\overline{\delta }\right) $ there is $h_{\delta }\in B_{ss}$, a probability measure such that $L_{\delta }h_{\delta }=h_{\delta }$. Furthermore, there is $M\geq
0$ such that for each $\delta \in \left[ 0,\overline{\delta }\right) $ $$\Vert h_{\delta }\Vert _{ss}\leq M.$$
- (convergence to equilibrium for the unperturbed operator) There is a sequence $a_n\to 0$ such that for each $g\in V_{ss}$$$\Vert L_{0}^{n}g\Vert _{s}\leq a_n||g||_{ss};$$
- (resolvent of the unperturbed operator) $(Id-L_{0})^{-1}:=\sum_{i=0}^{\infty }L_{0}^{i}$ is a bounded operator $V_{w}\rightarrow V_{w}
$.
- (small perturbation and derivative operator) There is $K\geq 0$ such that $\left\vert |L_{0}-L_{\delta }|\right\vert _{B_{s}\rightarrow
B_{w}}\leq K\delta ,$ and $\left\vert |L_{0}-L_{\delta }|\right\vert
_{B_{ss}\rightarrow B_{s}}\leq K\delta $. There is ${\dot{L}h_{0}\in V_{w}}$ such that$$\underset{\delta \rightarrow 0}{\lim }\left\Vert \frac{(L_{\delta}-L_{0})}{\delta }h_{0}-\dot{L}h_{0}\right\Vert _{w}=0. \label{derivativeoperator}$$
Then we have the following Linear Response formula$$\lim_{\delta \rightarrow 0}\left\Vert \frac{h_{\delta }-h_{0}}{\delta }-(Id-L_{0})^{-1}\dot{L}h_{0}\right\Vert _{w}=0. \label{linresp}$$
The following is an abstract response result for the second derivative.
\[thm:quadresp\] Let $(L_{\delta })_{\delta \in \lbrack 0,\overline{\delta }]}:B_{i}\rightarrow B_{i}$, $i\in \{ss,...,ww\}$ be a family of Markov operators as in the previous theorem. Assume furthermore that:
1. The derivative operator $\dot{L}$ admits a bounded extension $\dot{L}:B_{w}\rightarrow V_{ww}$, such that $$\left\Vert \frac{1}{\delta }(L_{\delta}-L_{0})-\dot{L}\right\Vert
_{w\rightarrow ww}\underset{\delta \rightarrow 0}{\longrightarrow }0.
\label{eq:unifcvderivop}$$
2. There exists a “second derivative operator” at $h_{0}$, i.e. $\
\ddot{L}h_{0}\in V_{ww}$ such that $$\left\Vert \dfrac{(L_{\delta }-L_{0})h_{0}-\delta \dot{L}h_{0}}{\delta ^{2}}-\ddot{L}h_{0}\right\Vert _{ww}\underset{{\delta }\rightarrow 0}{\longrightarrow }0. \label{def:2ndderivop}$$
3. The resolvent operator $(Id-L_{0})^{-1}$ admits a bounded extension as an operator $V_{ww}\rightarrow V_{ww}$.
Then one has the following: the map $\delta \in \lbrack 0,\overline{\delta }]\mapsto h_{\delta }\in B_{ss}$ has an order two Taylor’s expansion at $\delta =0$, with $$\left\Vert \frac{h_{\delta }-h_{0}-\delta (Id-L_{0})^{-1}\dot{L}h_{0}}{\delta ^{2}}-(Id-L_{0})^{-1}\left[ \ddot{L}h_{0}+\dot{L}(Id-L_{0})^{-1}\dot{L}h_{0}\right] \right\Vert _{ww}\underset{{\delta }\rightarrow 0}{\longrightarrow }0. \label{eq:quadresp}$$
Given the family of transfer operators $L_{\delta }$ defined at \[opdef\], we will apply these response results using the sequence of stronger and weaker spaces$$W^{7,1}(\mathbb{S}^{1})\subset W^{5,1}(\mathbb{S}^{1})\subset W^{3,1}(\mathbb{S}^{1})\subset W^{1,1}(\mathbb{S}^{1})$$
where $W^{k,1}$ stands for the Sobolev space of functions having the $k^{th}$ derivative in $L^{1}$ (see [@ADA] for an introduction to these spaces).
In the following subsection we verify the assumptions needed to apply theorems \[thm:linresp\] and \[thm:quadresp\]. Theorem [thm:quadrespsmooth]{} will be proved at the end of the section.
Verifying the assumptions in the general response theorems[verif]{}.
--------------------------------------------------------------------
In this subsection we verify the assumptions needed to apply theorems [thm:linresp]{} and \[thm:quadresp\]. First we verify the spectral gap and existence of the resolvent assumptions for the unperturbed system. This is somewhat well known for circle expanding maps. However for completeness we recall the main steps of this construction in subsection \[LR2LR3QR3\]. In subsection \[LR1\] we prove a uniform Lasota Yorke inequality, to verify the assumption LR1. In subsection \[LR4QR1\] we compute the first derivative operator associated to the small-noise perturbation, verifying assumptions LR4 and QR1. In subsection \[QR2\] we compute the second derivative operator, verifying assumption QR2.
### Spectral gap and resolvent for the unperturbed operator (verifying LR2, LR3 and QR3)\[LR2LR3QR3\].
In this section we consider the tranfer operator $L_{0}$ of the unperturbed system acting on our Sobolev spaces and verify the convergence to equilibrium and the existence of the resolvent operator $(Id-L_{0})^{-1}$on the weak and weakest spaces $W^{3,1}(\mathbb{S}^{1}),$ $W^{1,1}(\mathbb{S}^{1})$ as required in assumptions LR2, LR3 and QR3. Since we are considering the transfer operator associated to an expanding map of the circle, these results are nowadays not surprising **(**see e.g. [@Li2]**).** The results follow from a standard construction, in which one can get information on the spectrum of $L_{0}$ acting on these spaces from the existence of a Lasota Yorke inequality on suitable functional spaces which are compactly embedded each other. As there are some variants of this contruction, for completeness we briefly recall some precise statements which we can apply to our case.
The following theorem (see Theorem 9 in [@GS] for a proof of the statement in this form) is a version of a classical tool to obtain spectral gap in systems satisfying a Lasota Yorke inequality. It allows one to estimate the contraction rate of zero average measures, and imply spectral gap when applied to Markov operators. Let us consider a Markov operator $L_{0}$ acting on two normed vector spaces of complex or signed measures $(B_{s},\Vert ~\Vert _{s}),~(B_{w},\Vert ~\Vert _{w}),$ $B_{s}\subseteq B_{w}$ with $\Vert ~\Vert _{s}\geq \Vert ~\Vert _{w}$. We furthermore assume that $\mu \mapsto \mu (X)$ is continuous in the $\Vert ~\Vert _{s}$ and $\Vert
~\Vert _{w}$ topologies, and let $V_{i}:=\{\mu \in B_{i},\mu (X)=0\}$, $i\in
\{w,s\}$.
\[gap\]Suppose:
1. (Lasota Yorke inequality). For each $g\in B_{s}$$$\|L_0^{n}g\|_{s}\leq A\lambda _{1}^{n}\|g\|_{s}+B\|g\|_{w};$$
2. (Mixing) for each $g\in V_{s}$, it holds $$\lim_{n\rightarrow \infty }\|L_0^{n}g\|_{w}=0;$$
3. (Compact inclusion) The image of the closed unit ball in $B_{s}$ under $L_0$ is relatively compact in $B_{w}$.
Under these assumptions, we have
a
: $L_{0}$ admits a unique fixed point in $h\in B_{s}$, satisfying $h(X)=1$.
b
: There are $C>0,\rho <1$ such that for all $f\in V_{s}$ and $m$ large enough, $$\Vert L_{0}^{m}f\Vert _{s}\leq C\rho ^{m}\Vert f\Vert _{s}. \label{gap2}$$
c
: The resolvent $(Id-L_{0})^{-1}:$ $V_{s}\rightarrow V_{s}$ is defined and continuous.
To apply this result to our case we recall the following (a proof can be found in [@GS]).
\[Lemsu\]A $C^{k+1}$ expanding map on $\mathbb{S}^{1}$ satisfies a Lasota-Yorke inequality on $W^{k,1}({\mathbb{S}}^{1})$: there is $\alpha <1$, $A_{k},~B_{k}\geq 0$ such that $$\left\{ \begin{aligned} &\|L^nf\|_{W^{k-1,1}}\leq A_k\|f\|_{W^{k-1,1}}\\
&\|L^{n}f\|_{W^{k,1}}\leq \alpha^{kn}\|f\|_{W^{k,1}}+B_k\|f\|_{W^{k-1,1}}
\end{aligned}\right. .$$
By the compact immersion of $W^{k,1}$ in $W^{s,1}$ when $k\geq s$ (see [ADA]{}) and the well-known fact that an expanding map satisfies the mixing assumption, we can apply Theorem \[gap\] to our transfer operator of a $C^{8}$ expanding map and deduce the following result.
\[propora\]For each $k\in \{1,2,\cdots ,7\}$ there are $C>0,\rho <1$ such that for each $g\in V_{k}$ it holds $$\Vert L^{n}g\Vert _{W^{k,1}}\leq C\rho ^{n}\Vert g\Vert _{W^{k,1}}.$$In particular, the resolvent $(Id-L)^{-1}=\sum_{i=0}^{\infty }L^{i}$ is a well-defined and bounded operator on $V_{k}$.
This is enough to verify assumptions LR2, LR3 and QR3 in our case.
### Uniform Lasota Yorke inequalities in the zero-noise limit (verifying LR1)\[LR1\]
In order to prove the assumption LR1, we use Theorem \[gap\] and show a uniform Lasota-Yorke inequality.
\[lem:contraction\] $\forall k \geq 0, f\in W^{k,1}(\mathbb{S}^1),$ $$\|L_\delta f\|_{W^{k,1}} \leq \|L_T f\|_{W^{k,1}}.$$
We first prove the general statement: $\forall f\in L^1(\mathbb{S}^1)$, $\|\rho_\delta * f\|_{L^1} \leq \|f\|_{L^1}$. $\forall x \in \mathbb{S}^1 $, we have $$\rho_\delta * f(x) = \int _{-\delta}^\delta \rho_\delta(y) f(x-y) dy.$$ Hence $$\|\rho_\delta* f\|_{L^1} \leq \int_{\mathbb{S}^1}\int _{-\delta}^\delta
\rho_\delta (y) |f(x-y)| dydx = \int _{-\delta}^\delta \rho_\delta(y)dy
\times \|f\|_{L^1} = \|f\|_{L^1}.$$ Using the fact that $\forall i\geq 0$, $$(L_\delta f)^{(i)} = (\rho_\delta * L_Tf)^{(i)} = \rho_\delta * \left(
L_Tf\right)^{(i)},$$ we then have $$\|(L_\delta f)^{(i)}\|_{L^1} \leq \|\left( L_Tf \right)^{(i)}\|_{L^1}.$$ Hence $$\|L_\delta f\|_{W^{k,1}} = \sum_{i=0}^k \|(L_\delta f)^{(i)}\|_{L^1} \leq
\|L_T f\|_{W^{k,1}}.$$
\[lem:LYNoise\] For $k\geq 1$, the family $(L_\delta)$ verifies a uniform Lasota-Yorke inequality on $W^{k,1}(\mathbb{S}^1)$, which is: there is $\alpha <1, C_k, D_k \geq 0$ such that: $$\left\{ \begin{aligned} &\|L_\delta^nf\|_{W^{k-1,1}}\leq
C_k\|f\|_{W^{k-1,1}}\\ &\|L_\delta^{n}f\|_{W^{k,1}}\leq
\alpha^{kn}\|f\|_{W^{k,1}}+D_k\|f\|_{W^{k-1,1}} \end{aligned}\right. .$$
We will prove this lemma by induction on $k\geq 1$. $L_T$ is a contraction on $L^1$: using Lemma \[lem:contraction\], it is also the case for $L_\delta$, proving the power-boundedness on $L^1$ (with $C_1 = 1$).
Then, using Lemma \[Lemsu\], we know that there is a $B_1 \geq 0$ such that $$\|L_\delta f\|_{W^{1,1}} \leq \|L_T f\|_{W^{1,1}} \leq
\alpha\|f\|_{W^{1,1}}+ B_1 \|f\|_1.$$ Applying this inequality to $L_\delta^2 f = L_\delta(L_\delta f)$ gives us $$\begin{aligned}
\|L_\delta^2 f\|_{W^{1,1}} &\leq \alpha\|L_\delta f\|_{W^{1,1}} + B_1
\|L_\delta f\|_{1} \\
&\leq \alpha^2 \|f\|_{W^{1,1}} + \alpha B_1 \|f\|_{1} + B_1 \|L_\delta
f\|_{1}.\end{aligned}$$ We can then iterate: $$\begin{aligned}
\|L_\delta^n f\|_{W^{1,1}} &\leq \alpha^n \|f\|_{W^{1,1}} + B_1
\sum_{i=0}^{n-1} \alpha^i \|L_\delta^{n-1-i} f\|_{1} \\
&\leq \alpha^n \|f\|_{W^{1,1}} + \frac{B_1 C_1}{1-\alpha} \|f\|_{1}\end{aligned}$$ giving us the property for $k=1$, with $C_1 = 1$ and $D_1 = \frac{B_1 C_1}{1-\alpha}$.
The induction is then analogous to the base case. Using the induction hypothesis on $k-1$, more precisely the Lasota-Yorke inequality, we have the power-boundedness of $L_\delta$ on $W^{k-1,1}$: $$\|L^n_\delta f\|_{W^{k-1,1}} \leq (1 + D_{k-1})\|f\|_{W^{k-1,1}}.$$ Hence $C_k = 1+D_{k-1}$. We can then use again Lemma \[Lemsu\] to have the first inequality $$\|L_\delta f\|_{W^{k,1}} \leq \alpha\|f\|_{W^{k,1}}+ B_k \|f\|_{W^{k-1}},$$ which we can iterate to $$\|L_\delta^n f\|_{W^{k,1}} \leq \alpha^n \|f\|_{W^{k,1}} + B_k
\sum_{i=0}^{n-1} \alpha^i \|L_\delta^{n-1-i} f\|_{W^{k-1,1}}.$$ We can finally use the power-boundedness result we just proved to conclude: $$\|L_\delta^n f\|_{W^{k,1}} \leq \alpha^n \|f\|_{W^{k,1}} + \frac{B_kC_k}{1-\alpha} \|f\|_{W^{k-1,1}}.$$ Hence the result for $k$, with $C_k = (1+D_{k-1})$ and $D_k =\frac{B_k C_k}{1-\alpha}$.
We can then extend this inequality to our spaces, $W^{2k+1,1}(\mathbb{S}^1)$.
For $k \geq 2$, the family $(L_\delta)$ verifies a uniform Lasota-Yorke inequality on $W^{k,1}(\mathbb{S}^1) \subset W^{k-2,1}(\mathbb{S}^1)$: there is $\alpha <1, E_k, F_k \geq 0$ such that $$\|L^n_\delta f\|_{W^{k,1}} \leq \alpha^{\frac{k-1}{2}n}E_k\|f\|_{W^{k,1}} +
F_k\|f\|_{W^{k-2,1}}.$$
By using Lemma \[lem:LYNoise\], we have that $\forall n,p \geq 0$, $$\begin{aligned}
\|L^{n+p}_\delta f\|_{W^{k,1}} &\leq \alpha^{kn}\|L^p_\delta f\|_{W^{k,1}} +
D_k \|L^p_\delta f\|_{W^{k-1,1}} \\
&\leq \alpha^{kn}C_{k+1}\|f\|_{W^{k,1}} + D_k\left(
\alpha^{(k-1)p}\|f\|_{W^{k-1,1}} + D_{k-1}\|f\|_{W^{k-2,1}}\right) \\
&\leq \left(\alpha^{kn}C_{k+1} + \alpha^{(k-1)p}D_k\right)\|f\|_{W^{k,1}} +
D_k D_{k-1}\|f\|_{W^{k-2,1}}.\end{aligned}$$ So in the case $p=n$ (an even exponent), we have $$\begin{aligned}
\|L^{2n}_\delta f\|_{W^{k,1}} &\leq \alpha^{(k-1)n}\left(\alpha^{n}C_{k+1} +
D_k\right)\|f\|_{W^{k,1}} + D_k B_{k-1}\|f\|_{W^{k-2,1}} \\
&\leq \alpha^{\frac{k-1}{2}2n}\left(C_{k+1} + D_k\right)\|f\|_{W^{k,1}} +
D_k D_{k-1}\|f\|_{W^{k-2,1}}.\end{aligned}$$ And in the case $p=n+1$ (an odd exponent), we have $$\begin{aligned}
\|L^{2n+1}_\delta f\|_{W^{k,1}} &\leq \alpha^{\frac{k-1}{2}(2n+1)}\left(\alpha^{n-\frac{k-1}{2}}C_{k+1} + \alpha^{\frac{k-1}{2}}D_k\right)\|f\|_{W^{k,1}} + D_k D_{k-1}\|f\|_{W^{k-2,1}} \\
&\leq \alpha^{\frac{k-1}{2}(2n+1)}\left(\alpha^{-\frac{k-1}{2}}C_{k+1} +
D_k\right)\|f\|_{W^{k,1}} + D_k D_{k-1}\|f\|_{W^{k-2,1}}.\end{aligned}$$ Taking the maximum of the two constants that differ finishes the proof: because $\alpha <1$, we can take $E_k = \left(\alpha^{-\frac{k-1}{2}}C_{k+1}+D_k\right)$ and $F_k = D_k D_{k-1}$.
Using the compact embedding of $W^{7,1}(\mathbb{S}^1)$ into $W^{5,1}(\mathbb{S}^1)$ (by Rellich-Kondrachov embedding theorem, see [@ADA]) and the Lasota-Yorke inequality we just proved, one can easily deduce assumption LR1 (an example of such reasoning can be found in [@notes]).
### First derivative operator (verifying LR4 and QR1)[LR4QR1]{}
In this subsection we prove LR4 and QR1. These assumptions concern the first derivative operator. We will first prove that this first derivative operator is zero.
Let $(\rho_{\delta })_{\delta }$ be the family of random kernels defined at $($\[eq:kerdef\]$)$. There exists $C\geq 0$ such that for all $f\in W^{2,1}(\mathbb{S}^1)$, the following inequality holds $$\left\|\frac{\rho_\delta - \delta_0}{\delta}*f \right\|_{L^1} \leq \delta
\|f\|_{W^{2,1}} C.$$
Let us use the following Taylor expansion for $f\in W^{2,1}(\mathbb{S}^1)$: $$\label{eq:taylor1}
f(x-\delta z) = f(x) - \delta z f^{\prime }(x) - \int_{x-\delta z}^x
(x-\delta z - t)f^{\prime \prime }(t)dt.$$ Using (\[eq:kernel\]), we have $$((\rho_\delta - \delta_0)*f)(x) = -\int_{-1}^1\int_{x-\delta z}^x
\rho(z)(x-\delta z-t)f^{\prime \prime }(t)dtdz$$ By using the substitution $y = \frac{x-t}{\delta}$ in the last integral, we can re-write the result as $$((\rho_\delta - \delta_0)*f)(x) = \delta^2 \int_{-1}^1\int_z^0
\rho(z)(y-z)f^{\prime \prime }(x-\delta y)dydz$$ In particular, $$\begin{aligned}
\left\|\frac{\rho_\delta - \delta_0}{\delta}*f \right\|_{L^1} &\leq \delta
\int_{\mathbb{S}^1} \int_{-1}^1\int_{[0,z]} \rho(z)|z-y||f^{\prime
\prime}(x-\delta y)| dydzdx \\
&= \delta \int_{-1}^1\int_{[0,z]} \rho(z)|z-y| \left( \int_{\mathbb{S}^1}
|f^{\prime \prime }(x-\delta y)|dx \right) dydz \\
\left\|\frac{\rho_\delta - \delta_0}{\delta}*f \right\|_{L^1} &\leq \delta
\|f\|_{W^{2,1}} \underbrace{\int_{-1}^1\int_{[0,z]} \rho(z)|z-y| dydz}_{= C}\end{aligned}$$
We use this lemma to prove QR1 with a zero first derivative operator.
\[1der\] Let $(L_{\delta })_{\delta }$ be the family of operators defined at $($\[opdef\]$)$. The following limit, defining the first derivative operator holds $$\lim_{\delta \rightarrow 0}\left\Vert \frac{L_{\delta }-L_{0}}{\delta }\right\Vert _{W^{3,1}\rightarrow W^{1,1}}=0.$$
Let us consider $f\in W^{3,1}(\mathbb{S}^{1}),$ we get $$\begin{aligned}
\left\Vert \frac{L_{\delta }-L_{0}}{\delta }f\right\Vert _{W^{1,1}}&
=\left\Vert \frac{\rho _{\delta }-\delta _{0}}{\delta }\ast
(L_{T}f)\right\Vert _{W^{1,1}} \\
& =\left\Vert \frac{\rho _{\delta }-\delta _{0}}{\delta }\ast
(L_{T}f)\right\Vert _{L^{1}}+\left\Vert \frac{\rho _{\delta }-\delta _{0}}{\delta }\ast (L_{T}f)^{\prime }\right\Vert _{L^{1}} \\
& \leq \delta C \left( \Vert L_{T}f \Vert_{W^{2,1}} + \Vert (L_{T}f)^{\prime
} \Vert_{W^{2,1}} \right) \\
& \leq 2\delta C\Vert L_{T}f\Vert _{W^{3,1}} \\
& \leq 2\delta C\Vert L_{T}\Vert _{W^{3,1}\rightarrow W^{3,1}}\Vert
f\Vert_{W^{3,1}}.\end{aligned}$$The operator norm is then bounded by $2\delta C \Vert L_T\Vert$, which tends to 0 when $\delta$ does.
To finish verifying assumption LR4: we can remark that $\forall k\geq 0$, $$\begin{aligned}
\Vert (\rho _{\delta }-\delta _{0})\ast f\Vert _{W^{k,1}}
&=\sum_{i=0}^{k}\Vert \left( (\rho _{\delta }-\delta _{0})\ast
f\right)^{(i)}\Vert _{L^{1}} \\
& \leq \delta ^{2}C\sum_{i=0}^{k}\Vert f^{(i)}\Vert _{W^{2,1}} \\
& \leq \delta ^{2}C(k+1)\Vert f\Vert _{W^{k+2,1}} \\
\Vert (\rho _{\delta }-\delta _{0})\ast f\Vert _{W^{k,1}} & \leq \delta C
(k+1) \Vert f\Vert _{W^{k+2,1}}.\end{aligned}$$So $\forall k\geq 0$, $$\Vert (L_{\delta }-L_{0})f\Vert _{W^{k,1}}\leq \delta C (k+1) A_{k+2}\Vert
f\Vert _{W^{k+2,1}}$$with $A_{k+2} $ the constant from Lemma \[Lemsu\]. We then have LR4, with the result for $k\in \{3,5\}$: $$\Vert L_{\delta }-L_{0}\Vert _{W^{k+2,1}\rightarrow W^{k,1}}\leq \delta C
(k+1) A_{k+2}.$$
### Second derivative operator (verifying QR2)\[QR2\]
In this subsection we prove assumption QR2, computing the second derivative operator and showing its relation with the variance of $\rho $.
Let $(\rho_\delta)_\delta$ be the family of random kernels described in [eq:kerdef]{}. Then there exists $C >0$ such that for all $f \in W^{3,1}(\mathbb{S}^1)$, the following inequality holds $$\left\|\frac{\rho_\delta - \delta_0}{\delta^2}*f - \frac{\sigma^2(\rho)}{2}
f^{\prime \prime }\right\|_{L^1} \leq \delta C \|f\|_{W^{3,1}}.$$
We can extend for $f\in W^{3,1}(\mathbb{S}^1)$ the Taylor expansion ([eq:taylor1]{}): $$\label{eq:taylor2}
f(x-\delta z)=f(x)-\delta zf^{\prime }(x)+\frac{\delta ^{2}z^{2}}{2}f^{\prime \prime }(x)-\int_{x-\delta z}^{x}\frac{(x-\delta z-t)^{2}}{2}f^{(3)}(t)dt.$$We then have $$((\rho _{\delta }-\delta _{0})\ast f)(x)=\delta ^{2}\frac{f^{\prime \prime
}(x)}{2}\sigma^2(\rho )-\int_{-1}^{1}\int_{x-\delta z}^{x}\rho (z)\frac{(x-\delta z-t)^{2}}{2}f^{(3)}(t)dtdz.$$
By using the substitution $y = \frac{x-t}{\delta}$ in the last integral, we can re-write the result as $$((\rho_\delta - \delta_0)*f)(x) = \delta^2 \frac{f^{\prime \prime }(x)}{2}\sigma^2(\rho) + \frac{\delta^3}{2}\int_{-1}^1\int_z^0
\rho(z)(z-y)^2f^{(3)}(x-\delta y)dydz$$ i.e. $$\left(\frac{\rho_\delta - \delta_0}{\delta^2}*f\right)(x) = \frac{f^{\prime
\prime }(x)}{2}\sigma^2(\rho) + \frac{\delta}{2}\int_{-1}^1\int_z^0
\rho(z)(z-y)^2f^{(3)}(x-\delta y)dydz$$
Once again we can use Fubini theorem to exchange the last integrals: $$\int_{-1}^1\int_z^0\rho(z)(z-y)^2f^{(3)}(x-\delta y)dydz = R_1(x,\delta) +
R_2(x,\delta)$$ with $$\begin{aligned}
R_1(x,\delta) &= \int_{-1}^0\int_z^0\rho(z)(z-y)^2f^{(3)}(x-\delta y)dydz \\
&=\int_{-1}^0 f^{(3)}(x-\delta y) \left( \int_{-1}^y\rho(z)(z-y)^2 dz
\right)dy\end{aligned}$$ and $$\begin{aligned}
R_2(x,\delta) &= \int_0^1\int_z^0\rho(z)(z-y)^2f^{(3)}(x-\delta y)dydz \\
&=\int_0^1 f^{(3)}(x-\delta y) \left(\int_y^1\rho(z)(z-y)^2 dz \right)dy.\end{aligned}$$ So $$\left(\frac{\rho_\delta - \delta_0}{\delta^2}*f\right)(x) = \frac{f^{\prime
\prime }(x)}{2}\sigma^2(\rho) + \frac{\delta}{2}\int_{-1}^1 f^{(3)}(x-\delta
y) \Omega(y)dy$$ with $$\Omega(y) = \left\{
\begin{array}{ll}
\int_y^1 -\rho(z)(z-y)^2dz & \text{if } y\geq 0 \\
\int_{-1}^y \rho(z)(z-y)^2dz & \text{if } y < 0\end{array}
\right. .$$
We can then conclude that $$\begin{aligned}
\left\|\frac{\rho_\delta - \delta_0}{\delta^2}*f - \frac{\sigma^2(\rho)}{2}
f^{\prime \prime }\right\|_{L^1} &\leq \frac{\delta}{2} \int_{\mathbb{S}^1}\int_{-1}^1 \left|f^{(3)}(x-\delta y) \Omega(y) \right|dy dx \\
&\leq \frac{\delta}{2} \int_{-1}^1 |\Omega(y)| dy \times \|f\|_{W^{3,1}}.\end{aligned}$$
As in Subsection \[LR4QR1\], we can apply this lemma to our problem, obtaining the following.
Suppose $T$ is a $C^{5}$ expanding map on the circle $\mathbb{S}^{1}$. Let $h_{0}\in \mathbb{S}^{1}$, its invariant probability density and let $L_{\delta }$ be the family of operators defined in $($\[opdef\]$)$ then the following holds $$\left\Vert \frac{(L_{\delta }-L_{0})h_{0}}{\delta ^{2}}-\frac{\sigma^2(\rho )}{2}h_{0}^{\prime \prime }\right\Vert _{W^{1,1}}\underset{\delta \rightarrow
0}{\longrightarrow }0.$$
Remark that because $h_0$ is the invariant probability measure of $T$ and the property of derivation of a convolution product, $$\frac{(L_{\delta }-L_{0})h_{0}}{\delta ^{2}} = \frac{\rho_{\delta }-\delta
_{0}}{\delta ^{2}}\ast h_{0} \qquad \text{and} \qquad \left(\frac{(L_{\delta
}-L_{0})h_{0}}{\delta ^{2}}\right)^\prime = \frac{\rho_{\delta }-\delta _{0}}{\delta ^{2}}\ast h^\prime_{0}.$$
$T$ being a $C^{5}$ expanding map imply that $h_0$ is $C^4$ (see [@notes]): we can then apply our lemma to both $h_0$ and $h_0^\prime$, giving us $$\left\Vert \frac{(L_{\delta }-L_{0})h_{0}}{\delta ^{2}}-\frac{\sigma^2(\rho )}{2}h_{0}^{\prime \prime }\right\Vert _{L^{1}} \leq \delta C
\|h_0\|_{W^{3,1}}$$ and $$\left\Vert \left(\frac{(L_{\delta }-L_{0})h_{0}}{\delta ^{2}}-\frac{\sigma^2(\rho )}{2}h_{0}^{\prime \prime }\right)^\prime\right\Vert _{L^{1}}
\leq \delta C \|h^\prime_0\|_{W^{3,1}}.$$ The result then follows.
We have then verified the assumption QR2. Since all the assumptions are verified, we can hence apply Theorem \[thm:quadresp\] to the family of perturbed operators $L_{\delta }$, proving Theorem \[thm:quadrespsmooth\].
We apply Theorem \[thm:quadresp\], with the spaces $B_i = W^{i,1}(\mathbb{S}^1)$, with $i \in \{ss,s,w,ww\} = \{7,5,3,1\}$. We showed that our family of operator verifies the assumptions of both Theorem \[thm:linresp\] and \[thm:quadresp\] in the previous subsections: assumptions LR2, LR3 and QR3 in subsection \[LR2LR3QR3\], LR1 in subsection \[LR1\], LR4 and QR1 in subsection \[LR4QR1\], and QR2 in subsection \[QR2\].
Quantitative zero-noise limit of piecewise expanding maps[sec:pwexp]{}
======================================================================
In this section we prove that for a certain family of piecewise expanding maps, the invariant densities in the zero-noise limit have a speed of convergence “of order at least about 1”, as stated more precisely in Proposition \[prop:linresppiecewise\], confirming the numerical findings of [@Lin]. In this paper, the author shows numerically one example of piecewise expanding map having a discontinuous invariant density, where the speed of convergence is of order 1. This is due to the presence of discontinuities in the map and in the corresponding invariant densities. We remark that, as shown in the previous section, the exponent can be larger than 1 for smoother maps. The proof of Proposition [prop:linresppiecewise]{} is composed of three parts: in section \[UFO\] we introduce the concept of Uniform Family of Operators and state their link with the speed of convergence to equilibrium . We then show that the family of perturbations we consider in the small noise limit is uniform in this sense. Finally, in section \[LB\] we show a lower bound on the speed of convergence, based on the the approximation of a discontinuity by Lipschitz functions.
Upper bounds: Convergence to equilibrium and stability\[UB\]
------------------------------------------------------------
In this section we provide the upper bounds sufficient to prove Proposition \[prop:linresppiecewise\]. We start by defining the class of maps we mean to consider.
A map $T:\mathbb{S}^{1}\rightarrow \mathbb{S}^{1}$ is said to be piecewise $C^{2}$ if there exists a finite set of points $d_{1}=0<d_{2}<...<d_{n}=1$ such that for each $0\leq i<n$, $T_{i}:=T_{(d_{i},d_{i+1})}$ extends to a $C^{2}$ function on the closure. Its expanding constant is defined as $\lambda _{T}=\inf_{i,x\in \lbrack d_{i},d_{i+1}]}\left\vert T^{\prime
}(x)\right\vert $.
A piecewise $C^2$ map is called piecewise expanding if there is a integer $k
>0$ such that $\lambda_{T^k} >1$, where $T^k$ is the $k^{th}$ iterate of $T$.
A turning point of a map $T$ is a point where the derivative of the map is not well defined.
### Uniform Family of Operators, exponential convergence to equilibrium and quantitative statistical stability\[UFO\]
In this subsection we present a general quantitative result relating the stability of the invariant measure of an uniform family of operator and the speed of convergence to equilibrium.
Let $L$ be a Markov operator acting on two vector subspaces of signed measures on $\mathbb{S}^{1}$, $L:(B_{s},||~||_{s})\longrightarrow
(B_{s},||~||_{s})$ and $L:(B_{w},||~||_{w})\longrightarrow (B_{w},||~||_{w})$, endowed with two norms, $||~||_{s}$ on $B_{s}, $ and $||~||_{w}$ on $B_{w}$, such that $||~||_{s}\geq ||~||_{w}$. Suppose that, $$B_{s}\mathcal{\subseteq }B_{w}\mathcal{\subseteq }BS(\mathbb{S}^{1}),$$where again $BS(\mathbb{S}^{1})$ denotes the space of Borel finite signed measures on $\mathbb{S}^{1}.$ Let us consider again the space of zero average measures$$V_{s}=\{f\in B_{s},f(\mathbb{S}^{1})=0\}. \label{vs}$$This space is preserved by any Markov opertator.
We say that $L$ has convergence to equilibrium with at least speed $\Phi $ and with respect to the norms $||~||_{s}$ and $||~||_{w}$, if for each $f\in
V_{s}$ it holds $$||Lf||_{w}\leq \Phi (n)||f||_{s}, \label{wwe}$$where $\Phi (n)\longrightarrow 0$ as $n\longrightarrow \infty $.
\[UF\] A one parameter family of transfer operators $\{L_{\delta
}\}_{\delta \in \left[ 0,1\right) }$ is said to be an **uniform family of operators** with respect to the weak space $(B_{w},||~||_{w})$ and the strong space $(B_{s},||~||_{s})$ if $||~||_{s}\geq ||~||_{w}$ and it satisfies
1. Let $h_{\delta }\in B_{s}$ be a probability measure fixed under the operator $L_{\delta }$. Suppose there is $M>0$ such that for all $\delta \in \lbrack 0,1)$, it holds $$||h_{\delta }||_{s}\leq M;$$
2. $L_{\delta }$ approximates $L_{0}$ when $\delta $ is small in the following sense: there is $C\in \mathbb{R}^{+}$ such that: $$||(L_{0}-L_{\delta })h_{\delta }||_{w}\leq \delta C;$$
3. $L_{0}$ has exponential convergence to equilibrium with respect to the norms $||~||_{s}$ and $||~||_{w}$: there exists $0<\rho
_{2}<1 $ and $C_{2}>0$ such that $$\forall \ f\in V_{s}:=\{f\in B_{s}:f(X)=0\}$$it holds $$||L_{0}^{n}f||_{w}\leq \rho _{2}^{n}C_{2}||f||_{s};$$
4. The iterates of the operators are uniformly bounded for the weak norm: there exists $M_{2}>0$ such that $$\forall \delta ,n,g\in B_{s}\ \mathnormal{it~holds}\ ||L_{\delta
}^{n}g||_{w}\leq M_{2}||g||_{w}.$$
Under these assumptions we can ensure that the invariant measure of the system varies continuously (in the weak norm) when $L_{0}$ is perturbed to $L_{\delta }$, for small values of $\delta $. Moreover, the modulus of continuity can be estimated. The following result was indeed proved in [GLu]{}.
\[dlogd\]Suppose $\{\func{L}_{\delta }\}_{\delta \in \left[0, 1 \right)}$ is an uniform family of operators as in Definition \[UF\], where $h_{0}$ is the unique fixed point of $\func{L}_{0}$ in $B_{w}$ and $h_{\delta }$ is a fixed point of $\func{L} _{\delta }$. Then, there exists $\delta _0 \in
(0,1)$ such that for all $\delta \in [0,\delta _0)$, it holds
$$||h_{\delta }-h_{0}||_{w}=O(\delta \log \delta ).$$
It is worth to remark that such a statement can be generalized to other speed of convergence to equilibrium, obtaining for example Holder bounds to the statistical stability of systems having a power law speed of convergence to equilibrium (see [@Gcsf],[@Gpre]).
In the next section, we will prove that our small noise perturbation gives us a uniform family of operator. We can then apply Proposition \[dlogd\] to our family to prove an upper bound on the speed of convergence of invariant densities. Note that it does not give us a purely linear upper bound $O(\delta)$; however a convergence in $\delta \log \delta$ (up to a multiplicative constant) would still give an exponent 1 if extracted as a power law behavior: $$\lim_{\delta \to 0} \frac{\log \|h_\delta - h_0\|_1}{-\log(\delta)} = 1.$$
Proof that the small noise perturbation gives a uniform family of operators
---------------------------------------------------------------------------
### UF3 and UF4
Assumption **UF4** is immediate, as transfer operators are contractions on $L^1$. As showed earlier, we have that for all $f \in L^1$, $\|\rho_\delta * f\|_{L^1} \leq \|f\|_{L^1}$. $L_T$ being a contraction on $L^1$, we then have $$\|L_\delta f\|_{L^1} \leq \|L_T f\|_{L^1} \leq \|f\|_{L^1}.$$ $L_\delta$ is then also a contraction on $L^1$, hence the result for all $n$: $\|L^n_\delta f\|_{L^1} \leq \|f\|_{L^1}$.
Assumption **UF3** is verified for our spaces $BV$ and $L^1$ for piecewise expanding maps of the circle (see [@notes]).
### UF2
We first prove a similar result, but only for smooth functions. The calculations are basically the same as the ones we had for the derivative operator in the smooth expanding maps case.
\[lem:Cinftyresult\] There exists a $C >0$ such that for all $f \in
C^\infty$, $$\|\rho_\delta*f - f\|_{L^1} \leq C\|f^{\prime }\|_{L^1} \delta .$$
$$\begin{aligned}
\|\rho_\delta \ast f - f\|_{L^1} &= \int_{\mathbb{S}^1} \frac{1}{\delta}\left| \int_{-\delta}^\delta \rho\left(\frac{y}{\delta}\right)[f(x-y) -
f(x)]dy \right| dx \\
&\leq \int_{-\delta}^\delta \frac{1}{\delta}\rho\left(\frac{y}{\delta}\right) \left( \int_{\mathbb{S}^1} \left| \int_y^0 f^{\prime }(x-t)dt
\right| dx \right) dy \\
&\leq \|f^{\prime }\|_1 \times \int_{-\delta}^\delta \frac{1}{\delta}\rho\left(\frac{y}{\delta}\right)|y| dy \\
\|\rho_\delta \ast f - f\|_{L^1} &\leq \|f^{\prime }\|_{L^1} \times \delta
\times \underbrace{\int_{-1}^1\rho(z)|z| dz}_{=C}\end{aligned}$$
To extend the result for all $BV$ functions, we will use the following lemma, which proof can be found in [@EG]. The set of smooth functions are not dense in $BV$ for their norm; however we can still approximate $BV$ functions by smooth ones in a weaker sense.
\[lem:smoothApprox\] For all $f \in BV$, there exists a sequence $(f_n)\in (C^\infty \cap BV)^{\mathbb{N}}$ such that $$\left\{\begin{array}{ll}
\|f-f_n\|_{L^1} & \underset{n\rightarrow \infty}{\longrightarrow} 0 \\
\operatorname{Var}(f_n) & \underset{n\rightarrow \infty}{\longrightarrow} \operatorname{Var}(f)\end{array}
\right.$$
We can then extend our result from Lemma \[lem:Cinftyresult\].
\[prop:BVresult\] There exists a $C >0$ such that for all $f \in BV$, $$\|\rho_\delta*f - f\|_{L^1} \leq C\operatorname{Var}(f) \delta .$$
Lemma \[lem:Cinftyresult\] gives us the result for all $f \in C^\infty$. Indeed, for them $\operatorname{Var}(f) = \|f^{\prime }\|_{L^1}$. Now let $g$ be a $BV$ function, and $\epsilon >0$ be arbitrarily small. Let us prove $$\|\rho_\delta*g - g\|_{L^1} \leq C \operatorname{Var}(g)\delta +\epsilon.$$ with $C$ the same constant from Lemma \[lem:Cinftyresult\].
Using Lemma \[lem:smoothApprox\], there is a $f\in C^\infty$ such that $$\left\{
\begin{array}{l}
\|f-g \|_{L^1} \leq \frac{\epsilon}{3} \\
\operatorname{Var}(f) \leq \operatorname{Var}(g) + \frac{\epsilon}{3C\delta}\end{array}
\right.$$ We then have $$\begin{aligned}
\|\rho_\delta*g - g\|_{L^1} &\leq \|\rho_\delta*(g - f)\|_{L^1} +
\|g-f\|_{L^1} + \|\rho_\delta*f -f\|_{L^1} \\
&\leq \frac{\epsilon}{3} + \frac{\epsilon}{3} + C\|f^{\prime }\|_{L^1}\delta
\\
&= \frac{2\epsilon}{3} + C\operatorname{Var}(f)\delta \\
\|\rho_\delta*g - g\|_{L^1} &\leq \epsilon + C\operatorname{Var}(g)\delta\end{aligned}$$
We then have **UF2** assuming **UF1**: indeed, because $L_T
h_\delta \in BV$, we can use the result from Proposition \[prop:BVresult\] as: $$\begin{aligned}
\|(L_\delta - L_0)h_\delta\|_{L^1} &= \|\rho_\delta* L_T h_\delta - L_T
h_\delta\|_{L^1} \\
&\leq C\operatorname{Var}(L_T h_\delta) \delta \\
&\leq C |||L_T|||_{BV \to BV} \|h_\delta\|_{BV} \delta \\
\|(L_\delta - L_0)h_\delta\|_1 &\leq C |||L_T|||_{BV \to BV} M \delta\end{aligned}$$ with $M$ the constant of **UF1**.
### UF1
To prove the strong boundedness of the family of BV functions $\{h_\delta\}$, we can use a uniform L-Y inequality on $L_\delta$.
In the first section, the unperturbed operator $L_0$ verified a Lasota-Yorke inequality of type $$\|L_0^n f\|_s \leq \alpha^n \|f\|_s + C \|f\|_w,$$ where the proof is based on iterating the case $n=1$: we proved the uniform Lasota-Yorke inequality using $\|L_\delta f\| \leq \|L_T f\|$. In the general case of a piecewise expanding map, we have (see [@notes]) $$\|L_0^n f\|_s \leq \alpha^n A \|f\|_s + B \|f\|_w.$$ Proving a uniform Lasota-Yorke inequality is then more complex.
We use the result stated in [@BK2], which was proved in [@BK1].
A transition probability is a linear positive (sub)-Markovian operator $Q:
L^1 \to L^1$ such that $\|Q\|_1 \leq 1$. $Q^*$ denote its dual operator on $L^\infty$. A transition probability can be represented via a (sub)-Markov transition kernel on $[0,1]$ into itself: $$Q^*h(x) = \int h(y)Q(x,dy) \qquad \text{and} \qquad Qh(y) = \left( \frac{d}{dm}\int h(x) Q(x,.)m(dx)\right)(y).$$
If $Q(x,.) \ll m$ for each $x$, we note $q(x, y) = \frac{d}{dm}Q(x,.)(y)$.
\[prop:BlankKeller\] Let $T$ be a piecewise expanding map with no periodic turning point. Suppose it is perturbed by a family of transition probabilities $\{Q_\delta\}_\delta$ (i.e. $L_\delta = Q L_T$) verifying the following assumptions: $$\text{(Small perturbation) } \mathbf{d}(Q_\delta) := \sup\{\|Q_\delta f -
f\|_{L^1} \mid \|f\|_{BV}\leq 1\} \underset{\delta\rightarrow 0}{\longrightarrow} 0.$$ $$\text{(Locality) }\forall x,A \text{ s.t. dist}(x,A) > \delta, \quad
Q_\delta(x,A) = 0$$ $$\text{(Regularity) } \forall f \in BV, \quad \operatorname{Var}(Q_\delta f) \leq \operatorname{Var}(f) +
C\|f\|_{L^1}$$ where $Q_\delta f$ represent the density of $A \mapsto \int
Q_\delta(x,A)f(x)dx$ with respect to the Lebesgue measure.
Then there exists constants $C, \delta_0, \alpha <1$ and $N\in {\mathbb{N}}$ such that $$\operatorname{Var}(L_\delta^N f) \leq \alpha \operatorname{Var}(f) + C \|f\|_1$$ $\forall \delta \leq \delta_0$ and $f\in BV$.
In the case of an additive noise $\rho _{\delta }$, the Markov kernels have densities $q_{\delta }$ with respect to the Lebesgue measure, defined as (with the subtraction on $\mathbf{S}^1$) $$Q_{\delta }(x,A)=\int_{A}q_{\delta }(x,y)dy\qquad \text{with}\qquad
q_{\delta }(x,y):=\rho _{\delta }(y-x).$$Then $Q_{\delta }f=\rho _{\delta }\ast f$. The *small perturbation* assumption is a simple application of Proposition \[prop:BVresult\], as it gives us that $\mathbf{d}(Q_{\delta })\leq C\delta \underset{\delta
\rightarrow 0}{\longrightarrow }0$. The *locality* assumption is verified as the support of $\rho _{\delta }$ is included in the interval $[-\delta ,+\delta ]$. The *regularity* assumption is easily verified by our noise kernel via the following lemma.
$\forall \delta$ and $f\in BV$, $$\operatorname{Var}(\rho_\delta * f) \leq \operatorname{Var}(f)$$
One equivalent definition of Var is the following: $$\operatorname{Var}(f) = \sup \left\{ \int_{\mathbb{S}^1} \phi^{\prime}(x) f(x) dx ~\mid
\phi \in C^1_c \text{ s.t. } \|\phi\|_\infty \leq 1\right\}$$ Let $\phi \in C^1_c$. We then have $$\begin{aligned}
\int \phi^{\prime }\times (\rho_\delta*f) dx &= \int \int_{-\delta}^\delta
\frac{1}{\delta}\rho\left(\frac{y}{\delta}\right)\phi^{\prime }(x)f(x-y) dydx
\\
&= \int_{-\delta}^\delta \frac{1}{\delta}\rho\left(\frac{y}{\delta}\right)\left( \int\phi^{\prime }(x)f(x-y) dx\right) dy \\
&= \int_{-\delta}^\delta \frac{1}{\delta}\rho\left(\frac{y}{\delta}\right)
\left( \int\phi^{\prime }(\tilde{x}+y)f(\tilde{x}) d\tilde{x}\right) dy \\
&\leq \int_{-\delta}^\delta \frac{1}{\delta}\rho\left(\frac{y}{\delta}\right) \operatorname{Var}(f) dy \\
\int \phi^{\prime }\times (\rho_\delta*f) dx &\leq \operatorname{Var}(f).\end{aligned}$$ Hence $\operatorname{Var}(\rho_\delta*f) \leq \operatorname{Var}(f)$.
Because our noise verifies all the assumptions, we can apply Proposition [prop:BlankKeller]{}. Using the contracting property of $L_\delta$ on $L^1$, we easily deduce the following.
\[prop:LYInequalityTurning\] Let $T$ be a piecewise expanding map with no periodic turning point. Then there exists constants $C, \overline{\delta}, \alpha <1$ and $N\in {\mathbb{N}}$ such that $$\|L_\delta^N f\|_{BV} \leq \alpha \|f\|_{BV} + C^{\prime }\|f\|_{L^1}$$ $\forall \delta \leq \overline{\delta}$ and $f\in BV$.
This can give us an uniform L-Y inequality.
Under the same assumptions as before, we have that $\forall p\in {\mathbb{N}}, \newline
\delta \leq \overline{\delta}, \ 0 \leq k < N, \ f\in BV$, $$\|L_\delta^{pN+k}f\|_{BV} \leq \alpha^p |\|L_T|\|^k_{BV \to BV} \|f\|_{BV} +
\frac{C}{1-\alpha}\|f\|_{L^1}$$ which then leads to $$\|L_\delta^n f\|_{BV} \leq \alpha^n A \|f\|_{BV} + B\|f\|_{L^1} \qquad
\forall n\in {\mathbb{N}}.$$
The previous proposition gives us $$\|L_\delta^N f\|_{BV} \leq \alpha \|f\|_{BV} + C^{\prime }\|f\|_{L^1}.$$ Using the same type of induction as in the proof of Lemma \[lem:LYNoise\], we have the following result $\forall p\in {\mathbb{N}}$: $$\|L_\delta^{pN}f\|_{BV} \leq \alpha^p \|f\|_{BV} + C\sum_{i=0}^{p-1}
\alpha^i\|L_\delta^{p-1-i}f\|_{L^1} \leq \alpha^p \|f\|_{BV} + \frac{C}{1-\alpha}\|f\|_{L^1}.$$
Note that using the regularity assumption on our noise, we have that for all $f\in BV$, $\|L_\delta f \|_{BV} \leq \|L_T f\|_{BV}$. Then, $\forall k < N$, $\|L_\delta^k f \|_{BV} \leq |\|L_T|\|_{BV\to BV}^k \|f\|_{BV}$. We can then conclude that $$\begin{aligned}
\|L_\delta^{pN+k}f\|_{BV} &\leq \alpha^p \|L_\delta^k f\|_{BV} + \frac{C}{1-\alpha}\|L_\delta^k f\|_{L^1} \\
&\leq \alpha^p |\|L_\delta|\|_{BV\to BV}^k \|f\|_{BV} + \frac{C}{1-\alpha}\|f\|_{L^1}.\end{aligned}$$ Note that the final inequality is assuming that $|\|L_T|\| \geq 1$. If it is $\leq 1$, the constant is just $1$, but the L-Y inequality is then trivial: using the previous lemma, we would also have the norm of $L_\delta$ being 1, and have the inequality as in the smooth expanding case.
Having proven a uniform Lasota-Yorke inequality, we can conclude that our family of operators verifies also assumption $\mathbf{UF1}$. We then have proved that the dynamics resulting from a piecewise expanding maps of the circle with no periodic turning point perturbed by an additive noise have an upper bound on their modulus of continuity. More explicitly, $$\|h_\delta - h_0\|_{L^1} = O(\delta \log\delta).$$
The Lasota-Yorke inequality used in the verification of $\mathbf{UF1}$ might be extended to piecewise expanding maps having periodic turning points by the results in [@BK1], therefore extending our conclusion to all piecewise expanding maps of the circle.
Lower bounds: approximation of a discontinuity by Lipschitz functions\[LB\]
---------------------------------------------------------------------------
Until now, we only proved an upper bound on the modulus of continuity. Here, we show examples of piecewise expanding map of the circle for which the speed of approximation in the zero-noise limit is in fact of order $1$, providing the lower bound sufficient to prove Proposition [prop:linresppiecewise]{}. Let us consider the following map$$T:x\mapsto \left\{
\begin{array}{ll}
x+\frac{1}{2} & 0\leq x\leq \frac{1}{2} \\
2(1-x) & \frac{1}{2}\leq x\leq 1\end{array}\right. . \label{mapT}$$One gets easily that $T$ has the following invariant density, which is discontinuous: $$h_{0}:x\mapsto \left\{
\begin{array}{ll}
\frac{2}{3} & ~0\leq x\leq \frac{1}{2} \\
\frac{4}{3} & ~\frac{1}{2}< x\leq 1\end{array}\right. . \label{densH}$$This example has already been studied in [@Lin], where the author numerically found linear speed of convergence in the zero-noise limit. Note that $T$ admits $\{0, \frac{1}{2}, 1\}$ as periodic turning points, we cannot apply the upper bound result proven in the previous section.
In this section we prove the following proposition
\[propapprox\]Let $T$ be the map defined in $($\[mapT\]$)$ and $h_{0}$ be its invariant density, as in $(\ref{densH}).$ Let $L_{\delta }$ be the annealed transfer operator of the system with noise as defined at $(\ref{opdef})$ with $\rho $ $\in BV[-1,1]$ and let $h_{\delta }\in L^{1}$ be an invariant density for $L_{\delta }$. Then there exists a constant $C\in
\mathbb{R}$ such that $$\Vert h_{\delta }-h_{0}\Vert _{L^{1}}\geq C\delta .$$
We prove the proposition by showing in section \[lipcon\] that $h_{\delta
} $ is Lipschitz and providing an estimate for its Lipschitz constant, showing that $h_{\delta }$ is $\frac{C^{\prime }}{\delta }$-Lipschitz for some constant $C^{\prime }$. Then in section \[approx\] we prove that there is a $C^{\prime \prime }$ such that $\Vert f-h_{0}\Vert _{L^{1}}\geq
\frac{C^{\prime \prime }}{a}$ for any function $f$ which is $a$-Lipschitz, completing the proof.
### Estimating the Lipschitz constant of $h_{\protect\delta }$\[lipcon\]
In this section we prove that under the assumptions proposition [propapprox]{} for any $\delta >0$, the invariant density $h_{\delta }$ of the perturbed system is $\frac{C^{\prime }}{\delta }$-Lipschitz. This will be proved in Proposition \[proplip2\].
Before the main proposition we need two technical lemmas.
For $f \in BV$, $h\geq 0$, we have $$\int |f(x+h)-f(x)|dx \leq \operatorname{Var}(f)|h|.$$
We first prove it for $f \in C^\infty \cap BV$: $$\begin{aligned}
\int |f(x+h)-f(x)|dx &= \int_\mathbb{R} \left| \int f^{\prime }(y) \chi_{x
\leq y \leq x+h} dy \right| dx \\
&\leq \int_\mathbb{R} |f^{\prime }(y)| dy \times h \\
&= \operatorname{Var}(f) \times h\end{aligned}$$
Then for all $f \in BV$: let us set $\epsilon >0$. Using Lemma [lem:smoothApprox]{}, we can have $g \in C^\infty \cap BV$ such that $\|g-f\|_{L^1} \leq \epsilon$ and $\operatorname{Var}(g) \leq \operatorname{Var}(f) + \epsilon$. We then have $$\begin{aligned}
\int |f(x+h) - f(x)|dx &\leq \int |f(x+h) - g(x+h)| + |g(x) - f(x)| +
|g(x+h) - g(x)| dx \\
&\leq 2\epsilon + \operatorname{Var}(g)h \\
&\leq (2+h)\epsilon + \operatorname{Var}(f).\end{aligned}$$ We have the inequality for all $\epsilon >0$, so we have our result.
$$\operatorname{Var}(\rho_\delta) = \frac{\operatorname{Var}(\rho)}{\delta}.$$
These two lemmas easily give us a Lipschitz constant for the convolution product.
\[sop1\]For all $f\in L^{\infty }$, the function $\rho _{\delta }\ast f$ is $\frac{\operatorname{Var}(\rho )\Vert f\Vert _{\infty }}{\delta }$-Lipschitz.
We use the first lemma to write that, for all $x\in \mathbb{S}^{1},h\geq 0$,$$\begin{aligned}
|\rho _{\delta }\ast f(x+h)-\rho _{\delta }\ast f(x)|& \leq \int
|f(y)|\times |\rho _{\delta }(x-y+h)-\rho _{\delta }(x-y)|dy \\
& \leq \Vert f\Vert _{\infty }\operatorname{Var}(\rho _{\delta })h.\end{aligned}$$The second lemma then allows us to conclude.
We now want to use this result to bound the Lipschitz constant of $h_{\delta
}$, the invariant density of the perturbed system.
\[proplip2\]There is a $C^{\prime }>0$ such that for all $\delta >0$, the invariant density of the perturbed system $h_{\delta }$ is $\frac{C^{\prime }}{\delta }$-Lipschitz.
By definition, $h_{\delta }=L_{\delta }h_{\delta }=\rho _{\delta }\ast
L_{T}h_{\delta }$. Proposition \[sop1\] gives us that $h_{\delta }$ is $\frac{\operatorname{Var}(\rho )}{\delta }\Vert L_{T}h_{\delta }\Vert _{\infty }$-Lipschitz. Another well known result is the existence of a constant $A>0$ such that for all $f\in BV(\mathbb{S}^1), \|f\|_\infty \leq A \|f\|_{BV}$. Hence: $$\begin{aligned}
\Vert L_{T}h_{\delta }\Vert _{\infty }& \leq A\Vert L_{T}h_{\delta
}\Vert_{BV} & \\
& \leq AB\Vert h_{\delta }\Vert _{BV} & & \text{because }L_{T}\text{ is
bounded on }BV \\
& \leq ABM & & \text{by property UF1 proven earlier}.\end{aligned}$$We then have our result, as all the constants are independent from $\delta $.
### Approximation of a discontinuity \[approx\]
We prove here the lower bound on the approximation of $h_0$ by $a$-Lipschitz functions, with $a>0$ fixed.
Recall that $h_0$ is defined as (Figure \[fig:zeroslope\]) $$h_{0}:x\mapsto \left\{
\begin{array}{ll}
\frac{2}{3} & ~0\leq x\leq 0.5 \\
\frac{4}{3} & ~0.5< x\leq 1\end{array}\right. .$$
The intuitive “best approximation” function that is $a$-Lip would then be the linear path, $$f_{a}:x\mapsto \left\{
\begin{array}{cl}
\frac{2}{3} & \text{if }~x\leq 0.5-\frac{1}{3a} \\
1 + ax - \frac{a}{2} & \text{if }~0.5-\frac{1}{3a}\leq x\leq 0.5+\frac{1}{3a}
\\
\frac{4}{3} & \text{if }~x\geq 0.5+\frac{1}{3a}\end{array}\right. .$$We now prove that this is the best approximation in $L^{1}$, in the sense of the following proposition.
; ; ; ;
Let $f$ be a real-valued $a$-Lipschitz function of $[0,1]$. The following inequality holds: $$\|f-h_0\|_{L^1} \geq \|f_a-h_0\|_{L^1} = \frac{1}{9a}.$$
The first step is to only consider the difference in the neighborhood of the discontinuity where $f_a \neq h_0$: for all real-valued $f$, $$\|f-h_0\|_{L^1[0,1]} \geq \|f-h_0\|_{L^1([0.5-\frac{1}{3a}, ~0.5+\frac{1}{3a}])}.$$
We can then simplify our problem by only considering functions with values on the interval $[\frac{2}{3}, \frac{4}{3}]$. Indeed, for every real-valued $a$-Lip function $f$, if we denote by $\tilde{f}$ the function defined by $\tilde{f}:x \mapsto \min( \max(f(x),\frac{2}{3}), ~\frac{4}{3})$, the latter is a better approximation of the discontinuity (in the sense $\|f-h_0\|_{L^1} \geq \|\tilde{f}-h_0\|_{L^1}$) while also being $a$-Lipschitz.
By a linear change of coordinates, one can see that proving the result on the window $[0.5 - \frac{1}{3a},0.5+\frac{1}{3a}] \times [\frac{2}{3},\frac{4}{3}]$ for $a$-Lip functions is equivalent to proving it on $[0,1]\times[0,1]
$ for $1$-Lip functions, with $h_0$ and $f_a$ now being (Figure [fig:zeroslope2]{}) $$h_{0}:x\mapsto \left\{
\begin{array}{ll}
0 & ~0\leq x\leq 0.5 \\
1 & ~0.5< x\leq 1\end{array}\right. \qquad \text{and} \qquad f_a:x\mapsto x.$$
; ; ; ;
Let $f$ be a $1$-Lip function of $[0,1]$, with values in $[0,1]$. $$\|f-h_0\|_{L^1} = \int_{0}^{0.5} f(x) ~dx + \int_{0.5}^{1} 1-f(x) ~dx.$$ Using the $1$-Lip property, we have that for all $x > 0.5$, $$f(x) - f(0.5) \leq \left|f(x) - f(0.5)\right| \leq x-0.5 \qquad \text{i.e.}
\qquad -f(x) \geq 0.5-x - f(0.5).$$ Hence $$\begin{aligned}
\|f-h_0\|_{L^1} &\geq \int_{0}^{0.5} f(x) ~dx + \int_{0.5}^{1} 1-x ~dx +
\frac{1}{2}(0.5 - f(0.5)) \\
&= \int_{0}^{0.5} f(x) + 0.5 - f(0.5) ~dx + \int_{0.5}^{1} 1-x ~dx.\end{aligned}$$ We can re-use the $1$-Lip property on $x < 0.5$ to obtain $$f(x) - f(0.5) + 0.5 \geq x$$ and conclude $$\|f-h_0\|_{L^1} \geq \int_{0}^{0.5} x ~dx + \int_{0.5}^{1} 1-x ~dx =
\|f_a-h_0\|_{L^1}.$$
We are now ready to prove Proposition \[propapprox\].
We showed that there is a $C^{\prime }>0$ such that the invariant density of $L_{\delta }$ is $\frac{C^{\prime }}{\delta }$-Lipschitz. We can apply the last proposition to state the following lower bound on the modulus of continuity: $$\Vert h_{\delta }-h_{0}\Vert _{L^{1}}\geq \frac{\delta }{9C^{\prime }}=C\delta .$$Note that this lower bound result could easily be applied to all piecewise expanding maps with a discontinuity in their unperturbed invariant density, with a different constant for each map.
**Acknowledgments.** S.G. is partially supported by the research project PRIN 2017S35EHN\_004 “Regular and stochastic behavior in dynamical systems” of the Italian Ministry of Education and Research. The authors whish to thank Ecole Normale Paris Saclay and Università di Pisa for the organization of the international master stage “Stage d’initiation à la recherche M1” in which framework the work has been done. The authors also whish to thank W. Bahsoun and J. Sedro for useful discussions about zero noise limits and response.
[99]{} R.A. Adams, *Sobolev Spaces*, Academic Press, New York (1975)
Alves, J F and Araújo, V. *Random perturbations of nonuniformly expanding maps*. Astérisque, 286: 25–62 (2003)
Alves, J F. *Strong statistical stability of non-uniformly expanding maps*. Nonlinearity, 17: 1193–1215. (2004)
J. Alves H. Vilarinho *Strong stochastic stability for non-uniformly expanding maps* Erg. Th. Dyn. Sys. 33, 3 pp. 647-692 (2013)
JF Alves, MA Khan *Statistical instability for contracting Lorenz flows* Nonlinearity, V. 32, 11, (2019)
Araujo, V and Tahzibi, A. 2005. Stochastic stability at the boundary of expanding maps. Nonlinearity, 18: 939–958. doi:10.1088/0951-7715/18/3/001
Baladi V and Young L-S 1993 *On the spectra of randomly perturbed expanding maps* Commun. Math. Phys. 156 355–85
Baladi, V and Viana, M. 1996. *Strong stochastic stability and rate of mixing for unimodal.* Ann Scient Ec Norm Sup, 29(4): 483–517.
W. Bahsoun, S. Galatolo, I. Nisoli, X. Niu *A Rigorous Computational Approach to Linear Response* Nonlinearity, Volume 31, Number 3 pp. 1073–1109 (2018)
V. Baladi *Linear response, or else* ICM Seoul 2014 talk (arXiv:1408.2937)
M. Blank, G. Keller *Stochastic stability versus localization in chaotic dynamical systems* Nonlinearity, Volume 10, pp. 91–107 (1997)
M. Blank, G. Keller *Random perturbations of chaotic dynamical systems: stability of the spectrum* Nonlinearity, Volume 11, pp. 1351–1364 (1998)
L. Evans, R. Gariepy, *Measure theory and fine properties of functions.* CRC Press, 1992
J.-P. Eckmann and D. Ruelle *Ergodic theory of chaos and strange attractors *Rev. Mod. Phys. 57, 617 (1985)
S. Galatolo, *Statistical properties of dynamics. Introduction to the functional analytic approach*, arXiv:1510.02615
S. Galatolo *Quantitative statistical stability and speed of convergence to equilibrium for partially hyperbolic skew products* J. Éc. Pol. Math., 5, 377–405 (2018)
S. Galatolo, M. Monge, I. Nisoli *Existence of Noise Induced Order, a Computer Aided Proof *arXiv:1702.07024
S. Gouëzel, Liverani, C. *Banach spaces adapted to Anosov systems.* Ergodic Theory and Dynamical Systems, 26(1), 189-217 (2006)
S. Galatolo, and J. Sedro, *Quadratic response of random and deterministic dynamical systems*, Chaos 30, 023113 (2020).
S. Galatolo, R. Lucena *Spectral gap and quantitative statistical stability for systems with contracting fibers and Lorenz-like maps* DCDS 40(3): 1309-1360 (2020) doi: 10.3934/dcds.2020079
S. Galatolo *Quantitative statistical stability and convergence to equilibrium. An application to maps with indifferent fixed points* Chaos, Solitons & Fractals 103, 596-601 (2017) https://doi.org/10.1016/j.chaos.2017.07.005
Yu. Kifer, On small random perturbations of some smooth dynamical systems, Math. USSR Ivestija 8:1083–1107 (1974).
K. K. Lin *Convergence of invariant densities in the small-noise limit* Nonlinearity 18, 2 pp. 659-683 (2005).
R. J. Metzger *Stochastic Stability for Contracting Lorenz Maps and Flows *Comm. Math. Phys. 212, pp. 277–296 (2000)
C.Liverani, *Invariant measures and their properties. A functional analytic point of view*. Dynamical systems. Part II, pp. 185-237, Pubbl. Cent. Ric. Mat. Ennio Giorgi, Scuola Norm. Sup., Pisa, (2003)
C. Liverani, B. Saussol, S. Vaienti. *A probabilistic approach to intermittency.* Ergodic Theory and Dynamical Systems, 19(3), 671-685. (1999)
W. Shen, S. van Strien *On stochastic stability of expanding circle maps with neutral fixed points* Dynamical systems 28, 3 pp. 423-452 (2013)
W. Shen, *On stochastic stability of non-uniformly expanding interval maps.* Proc London Math Soc, doi:10.1112/plms/pdt013
M.Viana *Lectures on Lyapunov Exponents*, Cambridge Studies in Advanced Mathematics 145, Cambridge University Press (2014)
L.S. Young *What Are SRB Measures, and Which Dynamical Systems Have Them?* J. of Stat. Phys., Vol. 108, No. 5/6 (2002)
L.-S. Young, Stochastic stability of hyperbolic attractors, Erg. Th. & Dynam. Sys.6 311–319 (1986).
[^1]: A Markov operator is a linear operator preserving positive measures and such that for each positive measure $\mu $, it holds $[L(\mu )](X)=\mu (X)$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Maintaining peer-to-peer connectivity with low energy overhead is a key requirement for several emerging Internet of Things (IoT) applications. It is also desirable to develop such connectivity solutions for non-static network topologies, so that resilience to device failures can be fully realized. Decentralized clustering has emerged as a promising technique to address this critical challenge. Clustering of nodes around cluster heads (CHs) provides an energy-efficient two-tier framework for peer-to-peer communication. At the same time, decentralization ensures that the framework can quickly adapt to a dynamically changing network topology. Although some decentralized clustering solutions have been proposed in the literature, they either lack guarantees on connectivity or incur significant energy overhead to maintain the clusters. In this paper, we present Decentralized Connected Resilient IoT Clustering (DeCoRIC), an energy-efficient clustering scheme that is self-organizing and resilient to network changes while guaranteeing connectivity. Using experiments implemented on the Contiki simulator, we show that our clustering scheme adapts itself to node faults in a time-bound manner. Our experiments show that DeCoRIC achieves 100% connectivity among all nodes while improving the power efficiency of nodes in the system compared to the state-of-the-art techniques BEEM and LEACH by up to 110% and 70%, respectively. The improved power efficiency also translates to longer lifetime before first node death with a best-case of 109% longer than BEEM and 42% longer than LEACH.'
author:
- '\'
bibliography:
- 'all.bib'
title: |
DeCoRIC: Decentralized Connected Resilient IoT Clustering\
[^1]
---
IoT, Clustering, Resiliency, Decentralization.
[^1]: This work was financially supported in part by the Singapore National Research Foundation under its Campus for Research Excellence And Technological Enterprise (CREATE) programme. With the support of the Technische Universit[ä]{}t [M]{}unchen - Institute for Advanced Study, funded by the German Excellence Initiative and the European Union Seventh Framework Programme under grant agreement n 291763
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper we introduce the concept of completeness of sets. We study this property on the set of integers. We examine how this property is preserved as we carry out various operations compatible with sets. We also introduce the problem of counting the number of complete subsets of any given set. That is, given any interval of integers $\mathcal{H}:=[1,N]$ and letting $\mathcal{C}(N)$ denotes the complete set counting function, we establish the lower bound $\mathcal{C}(N)\gg N\log N$.'
address: 'Department of Mathematics, African Institute for Mathematical science, Ghana '
author:
- Theophilus Agama
title: Complete sets
---
[^1].
INTRODUCTION
============
The development of set theory dates back to the days of the German mathematician George Cantor. Infact he was one of the major pioneers of set theory and it’s development, and so, he is thought today as the major force behind it [@ferreiros2008labyrinth]. Today it is widely studied in many areas of mathematics, including number theory, combinatorics, computer science, algebra etc. Intuititively, a set can be thought of as a collection of well-defined objects. The objects in the set can be seen as it’s members or elements. These elements do characterize and tell us more about the nature of the set in question. The elements of a set can either be finite or infinite. For example the set $\mathcal{A}:=\{2,5,9,1,-54\}$ denotes a finite set of integers, since all the elements are integers. The set of $\mathbb{R}$ of real numbers and the set $\mathbb{Z}$ of integers are examples of infinite sets.\
\
In what follows we set $\mathcal{A}\pm\mathcal{B}:=\{a_i\pm b_i:a_i\in \mathcal{A}~\text{and}~b_i\in \mathcal{B}\}$, $\mathcal{A\cdot B}:=\{a_ib_j:a_i\in \mathcal{A}, b_i\in \mathcal{B}\}$ and $c\cdot \mathcal{A}:=\{ca:a\in \mathcal{A}\}$, $\mathcal{A}\setminus\mathcal{B}:=\mathcal{A}-\mathcal{B}$ for finite sets of integers $\mathcal{A}$ and $\mathcal{B}$. We recall an arithmetic progression of length $n$ to be the set $\mathcal{A}$ of the form $\mathcal{A}=\{a_0, a_0+q, a_0+2q, \ldots, a_0+(n-1)q\}$. In a more special case we have the $\mathcal{A}:=\{q, 2q,\ldots, nq\}$, a homogenous arithmetic progression. For the set $\mathcal{A}:=\{a_0,a_1, \ldots, a_n\}$, we call $\mathcal{A}^{(N)}:=\{a_0', a_1', \ldots, a_n'\}$, where $a_i'=\frac{a_i-a_0}{d(\mathcal{A})}$ and where $d(\mathcal{A})=(a_0-a_0, a_1-a_0, \ldots, a_n-a_0)$ with $0=a_0'<a_1'<\ldots <a_n'$, the normal form of $\mathcal{A}$.
There are various classifications concerning set of integers. For example, the theory of multiple sets and primitive sets is very vast and rich (See [@nathanson2000elementary]). A set $\mathcal{F}$ can also be classed as sumfree if the relation $a+b=c$ is not satisfied in $\mathcal{F}$, for $a,b,c\in \mathcal{F}$. In this paper, however, we study a particular class of sets of integers.
COMPLETE SETS
=============
In this section we introduce the concept of completeness of a set. Using various operations compatible with sets, we investigate how this property is preserved.
\[def1\] Let $\mathcal{A}:=\{a_{1},a_2,\cdots, a_n\}$ be a finite set of elements in $\mathcal{B}$, where addition and multiplication is well defined in $\mathcal{B}$. Then $\mathcal{A}$ is said to be complete in $\mathcal{B}$ if there exists some $b\in \mathcal{B}$ such that $\prod \limits_{i=1}^{n}a_i=b\sum \limits_{i=1}^{n}a_i$.
It follows from the above definition the nature of a complete set will depend on the set $\mathcal{B}$. If we take $\mathcal{B}:=\mathbb{R}$, then the complete set in question will be a complete set of real numbers. Again if we take $\mathcal{B}:=\mathbb{N}$, then the complete set in question will be a complete set of natural numbers. If $\mathcal{B}:=\mathbb{F}[x]$, then any complete set under $\mathcal{B}$ will be a complete set of polynomials. Let us consider the finite set natural numbers $\mathcal{P}:=\{3,5,7\}$. It is easily seen that this set is a complete set of natural numbers. Again we notice that the set $\mathcal{F}:=\{x^2, -x^2, 2x^2\}$ is a complete set of polynomials in $\mathbb{Z}[x]$. However, the set $\{4x^3,7x^3, 10x^3 \}$ is not complete in $\mathbb{Z}[x]$. So therefore there are, if not infinitely many, complete sets under any given type of set. Every finite set of real numbers is easily seen to be complete in $\mathbb{R}$, hence the concept of completeness is not very interesting in this setting. Thus we examine this concept on the set of integers $\mathbb{Z}$, where it is very strong.
COMPLETENESS IN $\mathbb{Z}$
============================
In this section we study the concept of completeness of finite sets of integers. In this case the set $\mathcal{B}$ in definition \[def1\] reduces to the set of all integers. Hence we can rewrite the definition in this particular setting as follows:
Let $\mathcal{A}:=\{a_1, a_2, \ldots, a_n\}$ be a finite set of integers. Then $\mathcal{A}$ is said to be complete in $\mathbb{Z}$ if there exists some $b\in \mathbb{Z}$ such that $\prod \limits_{i=1}^{n}a_i=b\sum \limits_{i=1}^{n}a_i$.
EXAMPLES OF COMPLETE SETS IN $\mathbb{Z}$
-----------------------------------------
1. The sets $\{3,5,7\}$, $\{-2, 5,3,-1\}$, $\{1,3,2\}$, $\{3,7,11\}$ are examples of sets of integers complete in $\mathbb{Z}$.
2. The sets $\{3,7,9,4, 2\}$, $\{7,11, 13, 15\}$ , $\{1,18, 17,3\}$ are not complete in $\mathbb{Z}$.
3. The sets $\{2,4,6\}$, $\{7,14,21,28, 35\}$, $\{3,5,12\}$ are all complete in $\mathbb{Z}$.
PROPERTIES OF COMPLETENESS OF SETS IN $\mathbb{Z}$
--------------------------------------------------
In this section we examine some properties of completeness of finite set of integers in $\mathbb{Z}$. We examine how this property is preserved as we perform various algebras compatible with sets.
Let $\mathcal{A}_1:=\{a_1, a_2, \ldots, a_n\}$ and $\mathcal{A}_2:=\{b_1, b_2, \ldots, b_n\}$ be complete sets in $\mathbb{Z}$. Then the following remain valid:
1. The prodset $\mathcal{A}_1\cdot \mathcal{A}_2$ is also complete in $\mathbb{Z}$.
2. The union $\mathcal{A}_1\cup \mathcal{A}_2$ is complete in $\mathbb{Z}$ provided there exist some $t\in \mathbb{Z}$ such that $a_ib_j=t(a_i+b_j)$ for each $1\leq i,j\leq n$.
3. Let $\mathcal{H}:=\{c_1,c_2, \ldots, c_n\}$. Then the set $\mathcal{A}_1\cup \mathcal{H}$ is complete in $\mathbb{Z}$ provided $\mathcal{A}_1\cap \mathcal{H}=\emptyset$ and $c_1+c_2+\cdots c_n=0$\[agama\].
4. Let $\mathcal{H}:=\{a_1, a_2,\ldots, a_n, b_1, b_2, \ldots, b_n\}$. Then the set $\mathcal{H}\setminus \mathcal{A}_1:=\{b_1, b_2, \ldots b_n\}$ is also complete in $\mathbb{Z}$ provided for each $i=1,\ldots n$, $b_i=ta_i$ for some fixed $t\in \mathbb{Z}$.
5. Let $\mathcal{A}:=\{d_1, d_2, \ldots, d_n\}$ be complete in $\mathbb{Z}$ and suppose $|2A|=\frac{n(n+1)}{2}$. Then the two fold sumset $2A$ is complete in $\mathbb{Z}$ provided $(n+1)|(d_i+d_j)$ with $i\neq j$ for some $1\leq i,j \leq n$.
6. The set $q\cdot \mathcal{A}_1:=\{qa_1, qa_2,\ldots, qa_n\}$ is also complete in $\mathbb{Z}$.
$(i)$ Suppose $\mathcal{A}_1:=\{a_1,a_2,\ldots, a_n\}$ and $\mathcal{A}_2:=\{b_1, b_2, \ldots, b_n\}$. Then the prodset $\mathcal{A}_1\cdot \mathcal{A}_2:=\{a_1b_1, \ldots, a_1b_n,$ $ a_2b_1, a_2b_2, \ldots, a_2b_n, \ldots, a_nb_1, a_nb_2, \ldots, a_nb_n\}$. Now, it follows that$$\begin{aligned}
\prod \limits_{i,j=1}^{n}a_ib_j&=(b_1b_2\cdots b_n)(a_1a_2\cdots a_n)\bigg((b_1b_2\cdots b_n)^{n-1}\prod \limits_{i=1}^{n}a_i^{n-1}\bigg).\nonumber \end{aligned}$$Since $\mathcal{A}_1$ and $\mathcal{A}_2$ are complete in $\mathbb{Z}$, it follows that $\prod \limits_{i,j=1}^{n}a_ib_j=K\bigg(\sum \limits_{i=1}^{n}a_i\bigg)\bigg(\sum \limits_{j=1}^{n}b_j\bigg)$ $\bigg((b_1b_2\cdots b_n)^{n-1}\prod \limits_{i=1}^{n}a_i^{n-1}\bigg)$, for some $K\in \mathbb{Z}$. Thus we can write $$\begin{aligned}
\prod \limits_{i,j=1}^{n}a_ib_j=R\sum \limits_{i,j=1}^{n}a_ib_j,\nonumber\end{aligned}$$where it is easily seen that $R\in \mathbb{Z}$. Hence the conclusion follows immediately.
$(ii)$ Suppose $\mathcal{A}_1:=\{a_1,a_2, \ldots, a_n\}$ and $\mathcal{A}_2:=\{b_1,b_2,\ldots, b_n\}$ be complete in $\mathbb{Z}$. Now the union $\mathcal{A}_1\cup \mathcal{A}_2=\{a_1, a_2, \ldots, a_n,b_1, b_2, \ldots, b_n\}$. Let us take the product of all the elements of the set, given by $(a_1a_2\cdots a_n)(b_1b_2\cdots b_n)$. Bearing in mind each of the sets is complete in $\mathbb{Z}$, it follows that $(a_1a_2\cdots a_n)(b_1b_2\cdots b_n)=R\bigg(\sum\limits_{i=1}^{n}a_i\bigg)\bigg(\sum \limits_{j=1}^{n}b_j\bigg)=R\sum \limits_{i,j=1}^{n}a_ib_j$. It follows from the hypothesis that $\sum \limits_{i,j=1}^{n}a_ib_j=S\sum \limits_{i=1}^{n}a_i+S\sum \limits_{j=1}^{n}b_j$, where $S\in \mathbb{Z}$. Hence it is easy to see that the conclusion follows immediately.
$(iii)$ Suppose $\mathcal{A}_1:=\{a_1, a_2, \ldots, a_n\}$ be a complete set in $\mathbb{Z}$ and let $\mathcal{H}:=\{c_1, c_2, \ldots, c_n\}$ such that $c_1+c_2+\cdots +c_n=0$. Assume the set $\mathcal{A}_1\cup \mathcal{H}:=\{a_1,a_2, \ldots, a_n, c_1, c_2, \ldots, c_n\}$ so that $\mathcal{A}_1\cap \mathcal{H}=\emptyset$. Since $\mathcal{A}_1$ is complete in $\mathbb{Z}$, we see that $a_1\cdot a_2 \cdots a_nc_1\cdot c_2\cdots c_n=R(c_1c_2\cdots c_n)(a_1+a_2\cdots a_n+c_1+c_2+\cdots +c_n)=R_1(a_1+a_2\cdots a_n+c_1+c_2+\cdots +c_n)$ and it follows that $\mathcal{A}_1\cup \mathcal{H}$ is also complete in $\mathbb{Z}$.
$(iv)$ Consider the set $\mathcal{H}\setminus \mathcal{A}_1:=\{b_1, b_2, \ldots, b_n\}$. Using the hypothesis and the fact that $\mathcal{A}_1$ is complete in $\mathbb{Z}$, it follows that $b_1\cdot b_2\cdots b_n=t^n(a_1\cdot a_2\cdots a_n)=t^nk(a_1+a_2+\cdots +a_n)$. Hence it follows that $\mathcal{H}\setminus \mathcal{A}_1$ is also complete in $\mathbb{Z}$.
$(v)$ Suppose $\mathcal{A}:=\{d_1, d_2, \ldots, d_n\}$ is complete in $\mathbb{Z}$ and let $|2\mathcal{A}|=\frac{n(n+1)}{2}$. Then the two fold sumset $2\mathcal{A}:=\{d_1+d_1, d_1+d_2, \ldots, d_1+d_n, d_2+d_2, d_2+d_3, \ldots, d_2+d_n, \ldots, d_{n-1}+d_{n-1}, d_{n-1}+d_{n}, d_{n}+d_{n}\}$. The product of the elements in $2\mathcal{A}$ is given by $2^{n}(d_{1}\cdot d_2\cdots d_{n})\prod \limits_{\substack{i,j=1\\i\neq j}}^{n}(d_i+d_j)$. Since $\mathcal{A}$ is complete in $\mathbb{Z}$ and $n+1|(d_i+d_j)$ for some $1\leq i,j \leq n$ with $i\neq j$, we have that $2^{n}(d_{1}\cdot d_2\cdots d_{n})\prod \limits_{\substack{i,j=1\\i\neq j}}^{n}(d_i+d_j)=R(n+1)\sum \limits_{i=1}^{n}d_i$ with $R\in \mathbb{Z}$, and we see that the result follows immediately.
$(vi)$ The result follows immediately, since $\mathcal{A}_1$ is complete in $\mathbb{Z}$.
Property $(iii)$ in Theorem \[agama\] is very important and useful for construction purposes. It tells us we only need to find a complete set of small size as we seek for a large complete set, since a larger complete set can be obtained by adding well-balanced elements of any size we wish into the set. Again property $(v)$ in Theorem \[agama\] informs us that if a set is complete, then the two fold sumset has a very high chance of being complete provided the size is not too small. Finally property $(ii)$ tells us that if the product of any two elements from any two complete sets, not necessarily distinct, can be controlled additively then their union will certainly be complete in $\mathbb{Z}$.
Let $\mathcal{G}$ be a finite set of integers. Then the normal form of $\mathcal{G}$, denoted $\mathcal{G}^{(N)}$ is always complete in $\mathbb{Z}$.
Consider the set $\mathcal{G}=\{a_0, a_1, \ldots, a_n\}$ of integers, where $a_0<a_1<\cdots <a_n$. The normal form of $\mathcal{G}$ is given by $\mathcal{G}^{(N)}:=\{a_0', a_1', \ldots, a_n'\}$, where $a_i'=\frac{a_i-a_0}{d(\mathcal{G})}$ and where $d(\mathcal{G})=(a_0-a_0, a_1-a_0, \ldots, a_n-a_0)$ with $0=a_0'<a_1'<\ldots <a_n'$. It follows immediately that $\mathcal{G}^{N}$ is complete, thereby ending the proof.
We have seen in Theorem \[agama\] in order to construct a large complete set we only need to first find a small complete set and then add terms of a well-balanced finite sequence into the set, thereby obtaining a complete set. This process requires adding negative integers. We can avoid the negative integers by examining the following result encapsulated in the following theorem.
\[Agama1\] Let $\mathcal{F}$ be a finite set of integers. If $\mathcal{F}$ is a homogenous arithmetic progression of odd length, then $\mathcal{F}$ is complete in $\mathbb{Z}$.
Let us consider the homogenous arithmetic progression $\mathcal{F}=\{d, 2d, \ldots ,(n-1)d, nd\}$. Clearly we see that $(d\cdot 2d\cdots nd)=d^{n}n!$. We observe that $$\begin{aligned}
d^{n}n!=\frac{n(n+1)}{2}d^{n}\bigg(2(n-2)!-4\frac{(n-2)!}{n+1}\bigg).\nonumber\end{aligned}$$Suppose $\mathcal{F}$ is of odd length, then it is easy to see that $(n+1)|4(n-2)!$. Thus, $\bigg(2(n-2)!-4\frac{(n-2)!}{n+1}\bigg)\in \mathbb{Z}$, and it follows that $\mathcal{F}$ is complete in $\mathbb{Z}$, as required.
This result , albeit easy to state, is very useful for the theory. It helps us to construct complete sets of integers of any length we wish. More than this, it relates two important concepts of sets of integers, one of which is widely studied in the whole of mathematics and has led to massive developments, arithmetic progression. It is also worth pointing out that the converse of Theorem \[Agama1\] is not true, since there are complete sets that are not homogenous arithmetic progressions.
Theorem \[Agama1\] informs us that the sets $\{3,6,9\}$, $\{23,46, 69, 92, 115\}$, $\{4,8, 12, 16, 20\}$, $\{7,14, 21,28,35\}$, $\{101,202,303, 404, 505\}$, $\{11, 22, 33,\}$, $\{9,18,27,$ $36,45,54,63,72,81\}$ are all complete in $\mathbb{Z}$.
Every interval $[1,N]$ of positive integers of odd length $N$ is complete in $\mathbb{Z}$.
The result follows immediately from Theorem \[Agama1\], since all integers in the interval $[1,N]$ form a homogenous arithmetic progression.
Let $\mathcal{H}:=\{p_1,p_2, \ldots, p_n\}$ be a set of odd length of the first $n$ odd primes, with $3=p_1<p_2\cdots <p_n$. Then either $\mathcal{H}$ is a complete set or $$\begin{aligned}
\sum \limits_{i=1}^{n}p_i=L,\nonumber\end{aligned}$$is prime or $\omega(L)=2$, where $\omega(L)=\sum \limits_{p|L}1$.
Every finite set $\mathcal{T}\subset \mathbb{N}$ can be completed in $\mathbb{N}$.
THE NUMBER OF COMPLETE SETS IN $\mathbb{Z}$
===========================================
In this section we turn our attention to counting the number complete subsets that can be formed from any finite set of integers. We begin addressing the problem from a narrower perspective, which is to say we seek the maximum number of complete subsets of the set $\{1,2, \ldots, N\}$ of integers. We obtain a lower bound in the following results.
\[harmonic\] The estimate $$\begin{aligned}
\sum \limits_{n\leq x}\frac{1}{n}=\log x+O(1),\nonumber\end{aligned}$$is valid.
See Theorem 6.9 in the book of Nathanson [@nathanson2000elementary].
\[counting1\] Let $\mathcal{C}(N)$ denotes the total number of complete subsets of the set $\mathcal{H}=\{1,2,\ldots, N\}$ of integers, then $$\begin{aligned}
\mathcal{C}(N)\gg N\log N.\nonumber\end{aligned}$$
Using Theorem \[Agama1\], we only need to count the number of homogenous arithmetic progression of length $k$ that can be formed from the interval $[1,N]$, where $k$ runs through the odd numbers no bigger than $N$. Let us consider all the homogenous arithmetic progressions of length $3$ that can be formed from the interval $[1,N]$; clearly there are $\left \lfloor \frac{N}{3}\right \rfloor$ such number of sets. We have the total count for those of length $5$ to be $\left \lfloor \frac{N}{5}\right \rfloor$. The total count for those of length $j$ is given by $\left \lfloor \frac{N}{j}\right \rfloor$. This culminates into the assertion that the total number of such complete sets is given by $$\begin{aligned}
\sum \limits_{j\geq 3}\left \lfloor \frac{N}{j}\right \rfloor, \nonumber\end{aligned}$$ where $j$ runs over the odd numbers no bigger that $N$. Hence the assertion follows immediately by applying Lemma \[harmonic\].
Let us consider the interval $[1,10]$. The complete sets $\{1,2,3\}$, $\{2,4,6\}$ and $\{3,6,9\}$, represents complete sets of size $3$ that can be formed from the interval $[1,10]$. Clearly there are $3$ of them. Again the set $\{1,2,3,4,5\}$ and $\{2,4,6,8,10\}$ represents complete sets of size $5$ that can be formed from the interval. Similarly, there is only one complete set of size $7$ that can be formed from $[1,10]$ and $\{1,2,3,4,5,6,7\}$ is an example. The set $\mathcal{L}=\{1,2,3,4,5,6,7,8,9\}$ represents a complete set of size $9$ that can be formed from $[1,10]$. Thus in total there are at least $7$ constructible complete sets that can be formed from the interval $[1,10]$.
\[counting 2\] Let $\mathcal{C}(N)$ be the number of complete subsets of the set $\mathcal{H}:=\{1,2,\ldots, N\}$. Then $$\begin{aligned}
\mathcal{C}(N)=O(N(\log N)(\log \log N)).\nonumber\end{aligned}$$
Conjecture \[counting 2\] tells us a great deal about the distribution of complete subsets of the set $\{1,2\ldots, N\}$. In probabilistic language, it tells us that the chance of any subset of the set $\{1,2,\ldots, N\}$ to be complete is very minimal, since $$\begin{aligned}
\lim \limits_{N\longrightarrow \infty}\frac{N(\log N)(\log \log N)}{2^{N}}=0. \nonumber\end{aligned}$$
END REMARKS
===========
As mentioned earlier, there are some complete sets that are not homogeneous arithmetic progressions. Given the interval $[1, 10]$ it turns out that the sets $\{3,5,7\}$, $\{2,5,7\}$ and $\{2,3,5\}$, that were not taken into account in Theorem \[counting1\], are also complete in $\mathbb{Z}$. Such a loss becomes very significant an $N$ is taken sufficiently large. This significant loss indicates something wierd unfolding as $N$ increases without bound, and does suggest the lower bound $\mathcal{C}(N)\gg N\log N$ is not the best possible and can be improved. To this end we raise some questions whose answer may be attributed to such a loss.
Does there exist complete sets of the form $\{r,r^2,r^3,\ldots, r^n\}$?
If the set $\mathcal{F}$ is complete, does there exist some integer $s<M$ such that $\mathcal{F}+\{s\}$ is complete?
[10]{}
, *Labyrinth of thought: A history of set theory and its role in modern mathematics,* Springer Science & Business Media, 2008.
Nathanson, M.B, *Graduate Texts in Mathematics*, New York, NY: Springer New York, 2000.
[^1]:
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper we define Bessel potentials in Ahlfors regular spaces using a Coifman type approximation of the identity, and show they improve regularity for Lipschitz, Besov and Sobolev-type functions. We prove density and embedding results for the Sobolev potential spaces defined by them. Finally, via fractional derivatives, we find that for small orders, these Bessel potentials are inversible, and show a way to characterize potential spaces, using singular integrals techniques, such as the $T1$ theorem. Moreover, this characterization allows us to prove these spaces in fact coincide with the classical potential Sobolev spaces in the Euclidean case.'
author:
- 'Miguel Andrés Marcos [^1]'
title: Bessel potentials in Ahlfors regular metric spaces
---
0.3 truecm
Introduction
============
Riesz and Bessel potentials of order $\alpha>0$ in ${\ensuremath{\mathbb{R}}}^n$ are defined as the operators $\mathcal{I}_\alpha=(-\Delta)^{-\alpha/2}$ and $\mathcal{J}_\alpha=(I-\Delta)^{-\alpha/2}$ respectively, where $\Delta$ is the Laplacian and $I$ the identity. By means of the Fourier transform, it can be shown they are given by multipliers $$\begin{aligned}
\left(\mathcal{I}_\alpha f\right)^{\ensuremath{\wedge}}(\xi)= (2\pi|\xi|)^{-\alpha}\hat{f}(\xi), \hspace{1cm} \left(\mathcal{J}_\alpha f\right)^{\ensuremath{\wedge}}(\xi)= (1+4\pi^2|\xi|^2)^{-\alpha/2}\hat{f}(\xi).\end{aligned}$$
These frequency representations of Riesz and Bessel potentials, as well as of their associated fractional differential operators, depend on the existence of Fourier Transforms on the underlying space. In more general settings alternative tools are needed. Spaces such as self similar fractals are more general, but are still Ahlfors regular. In spaces with this type of regularity, scales are a good substitute of frequencies.
Both the Riesz potential and its inverse the fractional derivative $\mathscr{D}_\alpha=(-\Delta)^{\alpha/2}$, which on the frequency side is given by $$\begin{aligned}
\left(\mathscr{D}_\alpha f\right)^{\ensuremath{\wedge}}(\xi)= (2\pi|\xi|)^\alpha\hat{f}(\xi),\end{aligned}$$ have an immediate generalization to metric measure spaces, as they take the form $$\begin{aligned}
\mathcal{I}_\alpha f(x)=c_{\alpha,n}\int \frac{f(y)}{|x-y|^{n-\alpha}}dy, \hspace{1cm} \mathscr{D}_\alpha f(x)= \tilde{c}_{\alpha,n}\int\frac{f(y)-f(x)}{|x-y|^{n+\alpha}}dy,\end{aligned}$$ at least for functions of certain integrability or regularity and $\alpha<2$. One can just replace $|x-y|^\alpha$ by a distance or quasi-distance $d(x,y)^\alpha$, Lebesgue measure by a general measure and $|x-y|^n$ by the measure of the ball of center $x$ and radius $d(x,y)$.
For spaces of homogeneous type, fractional integrals (i.e. Riesz potentials) and derivatives, as well as their composition, have been widely studied. In the absence of Fourier transform, other techniques have been developed, such as the use of a Coifman type approximation of the identity (see for instance [@HS], [@DJS]). It has been proven that even though the composition of a fractional integral and a fractional derivative (of the same order) is not necessarily the identity, at least for small orders of regularity it is an inversible singular integral. See [@GSV], [@Ga] for the study of this composition in $L^2$ and [@HV] for Besov and Triebel-Lizorkin spaces.
Bessel potentials have essentially the same local behavior than Riesz potentials, but behave much better globally. For instance, they are bounded in every $L^p$ space, whereas $\mathcal{I}_\alpha$ is bounded from $L^p$ only to $L^q$ with $\frac{1}{p}-\frac{1}{q}=\frac{\alpha}{n}$. This leads to define potential spaces $\mathcal{L}^{\alpha,p}=\mathcal{J}_\alpha(L^p)$, and these coincide with Sobolev spaces when $\alpha$ is an integer.
For $\alpha>0$, as $$\begin{aligned}
2^{-\alpha/2}\leq\frac{1+(2\pi|\xi|)^\alpha}{(1+4\pi^2|\xi|^2)^{\alpha/2}}\leq 2,\end{aligned}$$ the composition $(I+\mathscr{D}_\alpha)\mathcal{J}_\alpha$ is inversible in $L^2$. In fact, as shown in [@S], for $1<p<\infty$ and $0<\alpha<2$, $$\begin{aligned}
\label{deriv}f\in \mathcal{L}^{\alpha,p} \text{ if and only if } f,\mathscr{D}_\alpha f\in L^p,\end{aligned}$$ and in terms of Riesz potentials, $$\begin{aligned}
\label{integ}f\in \mathcal{L}^{\alpha,p} \text{ if and only if } f\in L^p \text{ and there exists $\gamma\in L^p$ with }f=\mathcal{I}_\alpha\gamma.\end{aligned}$$
Bessel operators have been rarely studied in the metric setting, although in ${\ensuremath{\mathbb{R}}}^n$ they can be represented as $$\begin{aligned}
\mathcal{J}_\alpha f(x)=f*G_\alpha (x)=\int f(y)G_\alpha(x-y)dy,\end{aligned}$$ where $G_\alpha$ is a radial function, so their definition does not present a limitation. In this paper we define Bessel-type potentials using the same construction found in [@GSV].
All the known tools and definitions used in this paper are described in section 2, such as approximations of the identity and singular integrals. In section 3 we define a Bessel-type potential operator and prove it increases the regularity of Lipschitz, Besov and Sobolev functions. In section 4 we describe the potential space obtained with this operator, and find relationships with Lipschitz, Besov and Sobolev functions, as well as a Sobolev embedding theorem. In section 5 we prove an inversion result for the Bessel operator using the techniques from [@GSV] and [@HV]. We finish this paper characterizing the potential space with the fractional derivative analogous to the Euclidean version in \[deriv\] and with the fractional integral, analogous to \[integ\], and analyze the case of ${\ensuremath{\mathbb{R}}}^n$.
Preliminaries
=============
In this section we describe the geometric setting and basic results from harmonic analysis on spaces of homogeneous type needed to prove our results.
The geometric setting
---------------------
We say $(X,\rho,m)$ is a space of homogeneous type if $\rho$ is a quasi-metric on $X$ and $m$ a measure such that balls and open sets are measurable and there exists a constant $C>0$ such that $$\begin{aligned}
m_\rho(B(x,2r))\leq Cm(B_\rho(x,r))\end{aligned}$$ for each $x\in X$ and $r>0$.
If $m(\{x\})=0$ for each $x\in X$, by [@MS] there exists a metric $d$ giving the same topology as $\rho$ and a number $N>0$ such that $(X,d,m)$ satisfies $$\begin{aligned}
\label{ahlf}m(B_d(x,2r))\sim r^N\end{aligned}$$ for each $x\in X$ and $0<r<m(X)$.
Spaces that satisfy condition \[ahlf\] are called Ahlfors $N$-regular. Besides ${\ensuremath{\mathbb{R}}}^n$ (with $N=n$), examples include self-similar fractals such as the Cantor ternary set or the Sierpiński gasket.
Throughout this paper we will assume $(X,d,m)$ is Ahlfors $N$-regular. One useful property these spaces have is regarding the integrability of the distance function:
- $\int_{B(x,r)}d(x,y)^sdm(y)<\infty$ if and only if $-N<s<\infty$, and here $$\begin{aligned}
\int_{B(x,r)}d(x,y)^sdm(y)\sim r^{s+N};\end{aligned}$$
- $\int_{X{\symbol{92}}B(x,r)}d(x,y)^sdm(y)<\infty$ if and only if $-\infty<s<-N$, and here $$\begin{aligned}
\int_{X{\symbol{92}}B(x,r)}d(x,y)^sdm(y)\sim r^{s+N}.\end{aligned}$$
If we add (locally integrable) functions we get
- if $-N<s<\infty$, $$\begin{aligned}
\int_{B(x,r)}f(y)d(x,y)^sdm(y)\leq C r^{s+N}Mf(x);\end{aligned}$$
- if $-\infty<s<-N$, $$\begin{aligned}
\int_{X{\symbol{92}}B(x,r)}f(y)d(x,y)^sdm(y)\leq C r^{s+N}Mf(x),\end{aligned}$$
where $Mf$ is the Hardy-Littlewood maximal function of $f$.
Aproximations of the identity
-----------------------------
In Ahlfors spaces of infinite measure (and thus unbounded), Coifman-type aproximations of the identity can be constructed. In this paper we will work with a continuous version, as presented in [@GSV]. See [@HS] for the discrete version. The construction is as follows.
Let $(X,d,m)$ be an Ahlfors $N$-regular space with $m(X)=\infty$. Let $h:[0,\infty){\ensuremath{\rightarrow}}{\ensuremath{\mathbb{R}}}$ be a non-negative decreasing $C^\infty$ function with $h\equiv 1$ in $[0,1/2]$ and $h\equiv 0$ in $[2,\infty)$. For $t>0$ and $f\in L^1_{loc}$, define
- $T_tf(x)=\frac{1}{t^N}\int_X h\left(\frac{d(x,y)}{t}\right)f(y)dm(y)$;
- $M_tf(x)=\varphi(x,t)f(x)$, with $\varphi(x,t)=\frac{1}{T_t1(x)}$;
- $V_tf(x)=\psi(x,t)f(x)$, with $\psi(x,t)=\frac{1}{T_t\left(\frac{1}{T_t1}\right)(x)}$;
- $S_t f(x)=M_tT_tV_tT_tM_t f(x)=\int_X s(x,y,t)f(y)dm(y)$, where $$\begin{aligned}
s(x,y,t)=\frac{\varphi(x,t)\varphi(y,t)}{t^{2N}}\int_X h\left(\frac{d(x,z)}{t}\right)h\left(\frac{d(y,z)}{t}\right)\psi(z,t)dm(z). \end{aligned}$$
$(S_t)_{t>0}$ will be our aproximation of the identity, with kernel $s$. We now list some of the properties they possess, they can be found in [@GSV] for the case $N=1$.
1. $S_t1\equiv1$ for all $t>0$;
2. $s(x,y,t)=s(y,x,t)$ for $x,y\in X$, $t>0$;
3. $s(x,y,t)\leq C/t^N$ for $x,y\in X$, $t>0$;
4. $s(x,y,t)=0$ if $d(x,y)>4t$;
5. $s(x,y,t)\geq C'/t^N$ if $d(x,y)<t/4$;
6. $|s(x,y,t)-s(x',y,t)|\leq C''\frac{1}{t^{N+1}}d(x,x')$;
7. $S_t$ is linear and continuous from $L^p$ to $L^p$;
8. $S_tf{\ensuremath{\rightarrow}}f$ pointwise when $t{\ensuremath{\rightarrow}}0$ if $f$ is continuous;
9. $|S_tf(x)-f(x)|\leq Ct^\gamma$ for each $x$ if $f$ is Lipschitz-$\gamma$;
10. $S_tf(x){\ensuremath{\rightarrow}}0$ uniformly in $x$ when $t{\ensuremath{\rightarrow}}\infty$ if $f\in L^1$;
11. $s$ is continuously differentiable with respect to $t$.
Continuity of a linear operator $T$ from $A$ to $B$ will be denoted throughout this paper as $$\begin{aligned}
T:A{\ensuremath{\rightarrow}}B.\end{aligned}$$
To include an interesting example of an Ahlfors space satisfying $m(X)=\infty$ (and thus having a Coifman-type approximation of the identity), we can modify the Sierpiński gasket $T$ by taking dilations (powers of 2): $\tilde{T}=\cup_{k\geq 1}2^kT$. This $\tilde{T}$ preserves some properties of the original triangle, including the Ahlfors character.
(-0.5,0)–(1.5,0)–(0.5,1.74)–cycle; (0.5,0)–(0,0.87)–(1,0.87)–cycle; (2,0)–(1.5,0); (0.75,2.175)–(0.5,1.74); (1.5,1.16)–(1,0.87); (-0.5,0)–(0.5,0)–(0,0.87)–cycle; (0,0)–(0.25,0.435)–(-0.25,0.435)–cycle; (-0.25,0)–(-0.125,0.2175)–(-0.375,0.2175)–cycle; (0.25,0)–(0.375,0.2175)–(0.125,0.2175)–cycle; (0,0.435)–(0.125,0.6525)–(-0.125,0.6525)–cycle; (-0.375,0)–(-0.4375,0.10875)–(-0.3125,0.10875)–cycle; (0.375,0)–(0.4375,0.10875)–(0.3125,0.10875)–cycle; (-0.125,0)–(-0.1875,0.10875)–(-0.0675,0.10875)–cycle; (0.125,0)–(0.1875,0.10875)–(0.0675,0.10875)–cycle; (-0.25,0.2175)–(-0.3125,0.32625)–(-0.1875,0.32625)–cycle; (0.25,0.2175)–(0.3125,0.32625)–(0.1875,0.32625)–cycle; (-0.125,0.4350)–(-0.1875,0.54375)–(-0.0675,0.54375)–cycle; (0.125,0.4350)–(0.1875,0.54375)–(0.0675,0.54375)–cycle; (0,0.6525)–(-0.0675,0.76125)–(0.0675,0.76125)–cycle;
Calderón reproducing formulas
-----------------------------
With this approximation of the identity $(S_t)_{t>0}$ we will construct our Bessel potential $J_\alpha$. For the proof relating $J_\alpha$ with the fractional derivative $D_\alpha$, we will follow the proof for the fractional integral as presented in [@GSV] and [@HV], which requires the derivative of $S_t$ (that exists because $s$ is continuously differentiable with respect to $t$): let $$\begin{aligned}
\frac{d}{dt}S_tf(x)=-\frac{1}{t}Q_tf(x),\end{aligned}$$ so $$\begin{aligned}
Q_tf(x)=\int_X q(x,y,t)f(y)dm(y), \hspace{0.75cm}\text{with}\hspace{0.75cm}q(x,y,t)=-t\frac{d}{dt}s(x,y,t).\end{aligned}$$
Some of their properties mirror those from $S_t$ and $s$:
1. $Q_t1\equiv0$ for all $t>0$;
2. $q(x,y,t)=q(y,x,t)$ for $x,y\in X$, $t>0$;
3. $|q(x,y,t)|\leq C/t^N$ for $x,y\in X$, $t>0$;
4. $q(x,y,t)=0$ if $d(x,y)>4t$;
5. $|q(x,y,t)-q(x',y,t)|\leq C'\frac{1}{t^{N+1}}d(x,x')$;
6. $Q_t: L^p{\ensuremath{\rightarrow}}L^p$;
7. *Calderón-type reproducing formulas.* (see [@C]) $$\begin{aligned}
f=\int_0^\infty Q_tf\frac{dt}{t}, \hspace{1cm}f=\int_0^\infty\int_0^\infty Q_tQ_sf\frac{dt}{t}\frac{ds}{s}.\end{aligned}$$
Singular Integrals
------------------
In Ahlfors $N$-regular spaces, the following version of the $T1$ theorem hold (see for instance [@Ga]). Once again we require $m(X)=\infty$.
A continuous function $K:X\times X{\symbol{92}}\Delta{\ensuremath{\rightarrow}}{\ensuremath{\mathbb{R}}}$ (where $\Delta=\{(x,x):x\in X\}$) is a standard kernel if there exist constants $0<\eta\leq 1$, $C>0$ such that
- $|K(x,y)|\leq Cd(x,y)^{-N}$;
- for $x\neq y$, $d(x,x')\leq cd(x,y)$ (with $c<1$) we have $$\begin{aligned}
|K(x,y)-K(x',y)|\leq Cd(x,x')^\eta d(x,y)^{-(N+\eta)};\end{aligned}$$
- for $x\neq y$, $d(y,y')\leq cd(x,y)$ (with $c<1$) we have $$\begin{aligned}
|K(x,y)-K(x,y')|\leq Cd(y,y')^\eta d(x,y)^{-(N+\eta)}.\end{aligned}$$
Let $C_c^\gamma$ denote the space of Lipschitz-$\gamma$ functions with compact support. A linear continuous operator $T:C_c^\gamma{\ensuremath{\rightarrow}}(C_c^\gamma)'$ for $0<\gamma\leq 1$ is a singular integral operator with associated standard kernel $K$ if it satisfies $$\begin{aligned}
\langle Tf,g\rangle = \iint K(x,y)f(y)g(x)dm(y)dm(x),\end{aligned}$$ for $f,g\in C_c^\gamma$ with disjoint supports. If a singular integral operator can be extended to a bounded operator on $L^2$ it is called a Calderón-Zygmund operator or CZO.
Every CZO is bounded in $L^p$ for $1<p<\infty$, of weak type $(1,1)$, and bounded from $L^\infty$ to $BMO$.
The $T1$ theorem characterizes CZO’s. We say that an operator is weakly bounded if $$\begin{aligned}
|\langle Tf,g\rangle|\leq Cm(B)^{1+2\gamma/N}[f]_\gamma[g]_\gamma,\end{aligned}$$ for $f,g\in C_c^\gamma(B)$, for each ball $B$.
\[tedeuno\] ($\mathbf{T1}$) Let $T$ be a singular integral operator. Then $T$ is a CZO if and only if $T1,T^*1\in BMO$ and $T$ is weakly bounded.
Besov spaces
------------
In metric measure spaces $(X,d,m)$, Besov spaces can be defined through a modulus of continuity, as seen in [@GKS]. For $1\leq p<\infty$ and $t>0$, the $p$-modulus of continuity of a locally integrable function $f$ is defined as $$\begin{aligned}
E_pf(t)=\left(\int_X\fint_{B(x,t)}|f(x)-f(y)|^pdm(y)dm(x)\right)^{1/p},\end{aligned}$$ where $\fint_A fdm$ denotes the average $\frac{1}{m(A)}\int_A fdm$, and the Besov space $B^\alpha_{p,q}$ for $\alpha>0$ and $1\leq q\leq \infty$ is the space of functions $f$ with the following finite norm $$\begin{aligned}
\|f\|_{B^\alpha_{p,q}}=\|f\|_p+\left(\int_0^\infty t^{-\alpha q}E_pf(t)^q\frac{dt}{t}\right)^{1/q}\end{aligned}$$ (with the usual modification for $q=\infty$).
For the case $p=q$, if the measure is doubling, an equivalent definition of the norm is $$\begin{aligned}
\|f\|_{B^\alpha_{p,q}}=\|f\|_p+\left(\iint\frac{|f(x)-f(y)|^p}{d(x,y)^{\alpha p}m(B(x,d(x,y))}dm(y)dm(x)\right)^{1/q}.\end{aligned}$$
Sobolev spaces
--------------
A way of defining Sobolev spaces in arbitrary metric measure spaces is Hajasz approach (see [@H1] for the case $\beta=1$): a nonnegative function $g$ is a $\beta$-Hajasz gradient of a function $f$ it the following inequality holds for almost every pair $x,y\in X$ $$\begin{aligned}
|f(x)-f(y)|\leq d(x,y)^\beta (g(x)+g(y)).\end{aligned}$$
For $1\leq p\leq \infty$, the Hajasz-Sobolev (fractional) space $M^{\beta,p}$ is defined as the space of functions $f\in L^p$ that have a gradient in $L^p$. Its norm is defined as $$\begin{aligned}
\|f\|_{M^{\beta,p}}=\|f\|_p+\inf_g \|g\|_p\end{aligned}$$ where the infimum is taken over all $\beta$-Hajasz gradients of $f$.
For the case $p=\infty$, the space $M^{\beta,\infty}$ coincides with the space $C^\beta$ of bounded Lipschitz-$\beta$ functions.
Functions with $\beta$-Hajasz gradients satisfy the following Poincaré inequality $$\begin{aligned}
\fint_B |f-f_B|dm\leq C\text{diam}(B)^\beta\fint_B gdm,\end{aligned}$$ for all balls $B$ (again, see [@H1] for the case $\beta=1$).
If the measure is doubling and $1\leq p<\infty$, then the following relationships hold between Besov and Sobolev spaces, for $\beta>0$ and $0<\epsilon<\beta$ $$\begin{aligned}
B^\beta_{p,p}{\ensuremath{\hookrightarrow}}M^{\beta,p}{\ensuremath{\hookrightarrow}}B^{\beta-\epsilon}_{p,p}\end{aligned}$$ (see [@GKS]). Here the expression $A{\ensuremath{\hookrightarrow}}B$ means $A\subset B$ with continuous inclusion.
Bessel potentials
=================
In this section we define the kernel $k_\alpha(x,y)$, to replace the convolution kernel $G_\alpha$ in the definition of $\mathcal{J}_\alpha$, and prove some properties this new Bessel-type potential operator $J_\alpha$ possesses, emulating those from $\mathcal{J}_\alpha$.
The convolution kernel $G_\alpha$ takes the form $$\begin{aligned}
G_\alpha(x-y)=c_{n,\alpha}\int_0^\infty \left(t^\alpha e^{-t^2}\right)\left(t^{-n}e^{-\frac{1}{4}\left(\frac{|x-y|}{t}\right)^2}\right)\frac{dt}{t},\end{aligned}$$ where $\varphi_t(x)=t^{-n}e^{-\frac{1}{4}\left(\frac{|x-y|}{t}\right)^2}$ is the Gaussian approximation of the identity. This provides us with a way to define the kernel in our context.
Let $(X,d,m)$ be our fixed Ahlfors $N$-regular space with $m(X)=\infty$, and $(S_t)_{t>0}$ an approximation of the identity as constructed in the previous section.
For $\alpha>0$, we define $$\begin{aligned}
k_\alpha(x,y)=\alpha\int_0^\infty \frac{t^\alpha}{(1+t^\alpha)^2}s(x,y,t)\frac{dt}{t}.\end{aligned}$$
Observe that the factor multiplying the approximation of the identity is $\frac{t^\alpha}{(1+t^\alpha)^2}$, as opposed to $t^\alpha e^{-t^2}$ in $G_\alpha$. It presents the same local behaviour, but near infinity it has only integrable decay. However, the properties obtained for $k_\alpha$ will be sufficient for our purposes.
The following properties follow immediately from definition and the properties of the kernel $s$, listed in section 2.
\[propk\]$k_\alpha$ satisfies:
1. $k_\alpha\geq 0$;
2. $k_\alpha(x,y)=k_\alpha(y,x)$
3. $k_\alpha(x,y)\leq C d(x,y)^{-(N-\alpha)}$;
4. $k_\alpha(x,y)\leq C d(x,y)^{-(N+\alpha)}$ if $d(x,y)\geq4$;
5. $|k_\alpha(x,z)-k_\alpha(y,z)|\leq C d(x,y)(d(x,z){\ensuremath{\wedge}}d(y,z))^{-(N+1-\alpha)}$;
6. $|k_\alpha(x,z)-k_\alpha(y,z)|\leq C d(x,y)(d(x,z){\ensuremath{\wedge}}d(y,z))^{-(N+1+\alpha)}$ if $d(x,z)\geq 4$ and $d(y,z)\geq 4$;
7. $\int_X k_\alpha(x,z)dm(z)=\int_X k_\alpha(z,y)dm(z)=1$ $\forall x,y$.
All results that will be presented in sections 3 and 4 involving the kernel $k_\alpha$ can be derived from just these properties. The actual need for the definition will become clear in section 5. We are now able to define our Bessel potential $$\begin{aligned}
J_\alpha g(x)=\int_X g(z)k_\alpha(x,z)dm(z).\end{aligned}$$
Observe that from property $7$ of the last lemma, we get $$\begin{aligned}
\|J_\alpha g\|_p\leq \|g\|_p\end{aligned}$$ for $1\leq p\leq\infty$.
As expected, we can compare this operator with the Riesz potential, which is be defined from the kernel $$\begin{aligned}
k'(x,y)=\int_0^\infty \alpha t^\alpha s(x,y,t)\frac{dt}{t}\sim \frac{1}{d(x,y)^{N-\alpha}}\end{aligned}$$ as $$\begin{aligned}
I_\alpha f(x)=\int_X f(y)k'(x,y)dm(y),\end{aligned}$$ (see [@GSV]) and we obtain $|J_\alpha g(x)|\leq C I_\alpha |g|(x)$.
We now proceed to prove $J_\alpha$ improves regularity on Lipschitz, Besov and Hajasz-Sobolev functions. We start with the Lipschitz case
If $f=J_\alpha g$ and $\alpha+\beta<1$ for $\alpha,\beta>0$, $$\begin{aligned}
|f(x)-f(y)|\leq C[g]_\beta d(x,y)^{\alpha+\beta}.\end{aligned}$$ In particular, as $J_\alpha$ is bounded in $L^\infty$, $$\begin{aligned}
J_\alpha: C^\beta{\ensuremath{\rightarrow}}C^{\alpha+\beta}.\end{aligned}$$
We will prove only the first part, the second follows immediately. What we will show also holds true for $I_\alpha$, as shown in [@GSV]. As $\int k_\alpha=1$, we have $$\begin{aligned}
f(x)-f(y) &=\int_X g(z)\left(k_\alpha(x,z)-k_\alpha(y,z)\right)dm(z)\\
&=\int_X (g(z)-g(x))\left(k_\alpha(x,z)-k_\alpha(y,z)\right)dm(z),\end{aligned}$$ and if we call $d=d(x,y)$ $$\begin{aligned}
|f(x)-f(y)|&\leq C\int_{B(x,2d)}\frac{|g(x)-g(z)|}{d(x,z)^{N-\alpha}}dm(z)\\
&\hspace{1cm}+C\int_{B(y,3d)}\frac{|g(x)-g(z)|}{d(y,z)^{N-\alpha}}dm(z)\\
&\hspace{1cm}+C\int_{X{\symbol{92}}B(x,2d)}|g(z)-g(x)|\left|k_\alpha(x,z)-k_\alpha(y,z)\right|dm(z)\\
&= I + II + III.\end{aligned}$$ Then for $I$ and $II$, as $\alpha,\beta>0$, $$\begin{aligned}
I\leq C[g]_\beta\int_{B(x,2d)}\frac{d(x,z)^\beta}{d(x,z)^{N-\alpha}}dm(z)\leq C[g]_\beta d^{\alpha+\beta},\end{aligned}$$ $$\begin{aligned}
II\leq C[g]_\beta d^\beta\int_{B(y,3d)}\frac{1}{d(y,z)^{N-\alpha}}dm(z)\leq C[g]_\beta d^{\alpha+\beta}.\end{aligned}$$ Finally, as $d(x,z)\sim d(y,z)$ for $z\in X{\symbol{92}}B(x,2d)$, and as $\alpha+\beta<1$, $$\begin{aligned}
III\leq C[g]_\beta d\int_{X{\symbol{92}}B(x,2d)}d(x,z)^\beta d(x,z)^{-(N+1-\alpha)}dm(z)\leq C[g]_\beta d^{\alpha+\beta}.\end{aligned}$$
Before proving the increase in Besov regularity, we need the following lemma, that follows from properties 3 and 5 of \[propk\]:
\[ineqk\]For $q>0$ and $x,y\in X$,
- if $q(N-\alpha)<N$, $$\begin{aligned}
\int_{d(x,z)<2d(x,y)}|k_\alpha(x,z)-k_\alpha(y,z)|^qdm(z)\leq C d(x,y)^{N-q(N-\alpha)};\end{aligned}$$
- if $N<q(N-\alpha+1)$, $$\begin{aligned}
\int_{d(x,z)\geq2d(x,y)}|k_\alpha(x,z)-k_\alpha(y,z)|^qdm(z)\leq C d(x,y)^{N-q(N-\alpha)}.\end{aligned}$$
If $f=J_\alpha g$ and $\alpha+\beta<1$ for $\alpha,\beta>0$, $$\begin{aligned}
\iint_{X\times X}\frac{|f(x)-f(y)|^p}{d(x,y)^{N+(\alpha+\beta)p}}dm(y)dm(x)\leq C\iint_{X\times X}\frac{|g(x)-g(z)|^p}{d(x,z)^{N+\beta p}}dm(z)dm(x).\end{aligned}$$ In particular, as $J_\alpha$ is bounded in $L^p$, $$\begin{aligned}
J_\alpha: B^\beta_{p,p}{\ensuremath{\rightarrow}}B^{\alpha+\beta}_{p,p}.\end{aligned}$$
Using $\int k_\alpha=1$, by Hölder’s inequality we have $$\begin{aligned}
|f(x)-f(y)|^p &\leq\\
&\hspace{-1cm}\leq C\left(\int_{B(x,2d(x,y))} |g(x)-g(z)|^p|k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\right)\\
&\hspace{0cm}\times\left(\int_{B(x,2d(x,y))} |k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\right)^{p/p'}\\
&\hspace{-0.5cm}+C\left(\int_{B(x,2d(x,y))^c} |g(x)-g(z)|^p|k_\alpha(x,z)-k_\alpha(y,z)|^{\theta p}dm(z)\right)\\
&\hspace{0cm}\times\left(\int_{B(x,2d(x,y))^c} |k_\alpha(x,z)-k_\alpha(y,z)|^{(1-\theta)p'}dm(z)\right)^{p/p'}.\end{aligned}$$ By the previous lemma, if we find $0\leq \theta\leq 1$ such that $N<(1-\theta)p'(N-\alpha+1)$, we get $$\begin{aligned}
|f(x)-f(y)|^p &\leq\\
&\hspace{-1cm}\leq C d(x,y)^{p\alpha-\alpha}\int_{B(x,2d(x,y))} |g(x)-g(z)|^p|k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\\
&\hspace{-0.5cm}+C d(x,y)^{-N+p\alpha+\theta p(N-\alpha)}\\
&\hspace{0cm}\times\int_{B(x,2d(x,y))^c} |g(x)-g(z)|^p|k_\alpha(x,z)-k_\alpha(y,z)|^{\theta p}dm(z).\end{aligned}$$ With this, to conclude the theorem it will be enough to prove $$\begin{aligned}
\int_{d(x,z)<2d(x,y)} \frac{|k_\alpha(x,z)-k_\alpha(y,z)|}{d(x,y)^{N+\beta p+\alpha}}dm(y)\leq C \frac{1}{d(x,z)^{N+\beta p}}.\end{aligned}$$ and for the other part $$\begin{aligned}
\int_{d(x,z)\geq 2d(x,y)} \frac{|k_\alpha(x,z)-k_\alpha(y,z)|^{\theta p}}{d(x,y)^{2N+\beta p-\theta p(N-\alpha)}}dm(y)\leq C \frac{1}{d(x,z)^{N+\beta p}}.\end{aligned}$$
- For the first one, if $d(x,z)<2d(x,y)$ then $d(y,z)<3d(x,y)$ and by using the bound for $k_\alpha$, $$\begin{aligned}
\hspace{-1cm}\int_{d(x,z)<2d(x,y)} \frac{|k_\alpha(x,z)-k_\alpha(y,z)|}{d(x,y)^{N+\beta p+\alpha}}dm(y) &\leq\\
&\hspace{-6cm}\leq C\int_{d(x,z)<2d(x,y)} \frac{1}{d(x,y)^{N+\beta p+\alpha}}\left(\frac{1}{d(x,z)^{N-\alpha}}+\frac{1}{d(y,z)^{N-\alpha}}\right)dm(y),\end{aligned}$$ then we consider two cases,
- if $d(y,z)<\frac{3}{2}d(x,z)<3d(x,y)$, then $$\begin{aligned}
\hspace{-2cm}\int_{d(y,z)<\frac{3}{2}d(x,z)<3d(x,y)}\frac{1}{d(x,y)^{N+\beta p+\alpha}}\left(\frac{1}{d(x,z)^{N-\alpha}}+\frac{1}{d(y,z)^{N-\alpha}}\right)dm(y) &\leq\\
&\hspace{-8cm}\leq C\frac{1}{d(x,z)^{N+\beta p+\alpha}} \int_{d(y,z)<\frac{3}{2}d(x,z)}\frac{1}{d(y,z)^{N-\alpha}}dm(y)\\
&\hspace{-8cm}\leq C\frac{1}{d(x,z)^{N+\beta p}};\end{aligned}$$
- if $\frac{3}{2}d(x,z)\leq d(y,z)<3d(x,y)$, $$\begin{aligned}
\hspace{-2cm}\int_{\frac{3}{2}d(x,z)\leq d(y,z)<3d(x,y)}\frac{1}{d(x,y)^{N+\beta p+\alpha}}\left(\frac{1}{d(x,z)^{N-\alpha}}+\frac{1}{d(y,z)^{N-\alpha}}\right)dm(y) &\leq\\
&\hspace{-8cm}\leq C\frac{1}{d(x,z)^{N-\alpha}}\int_{d(x,y)>d(x,z)/2}\frac{1}{d(x,y)^{N+\beta p+\alpha}}dm(y)\\
&\hspace{-8cm}\leq C\frac{1}{d(x,z)^{N+\beta p}}.\end{aligned}$$
- For the second one, if $d(x,z)\geq 2d(x,y)$, then $d(x,z)\sim d(y,z)$ and by property 5 in \[propk\], $$\begin{aligned}
\int_{d(x,z)\geq2d(x,y)} \frac{|k_\alpha(x,z)-k_\alpha(y,z)|^{\theta p}}{d(x,y)^{2N+\beta p-\theta p(N-\alpha)}}dm(y) &\leq\\
&\hspace{-6cm}\leq C\frac{1}{d(x,z)^{\theta p(N-\alpha+1)}}\int_{d(x,z)\geq2d(x,y)} \frac{d(x,y)^{\theta p}}{d(x,y)^{2N+\beta p-\theta p(N-\alpha)}}dm(y)\\
&\hspace{-6cm}\leq C\frac{1}{d(x,z)^{N+\beta p}}\end{aligned}$$ as long as $N+\beta p<\theta p(N-\alpha+1)$.
Finally, both conditions over $\theta$ can be rewritten as $$\begin{aligned}
N+\beta p<\theta p(N-\alpha+1)<N+(1-\alpha)p,\end{aligned}$$ and there is always a value for $\theta$ satisfying them, for $\beta<1-\alpha$.
We have now the following result regarding Sobolev regularity.
Let $f,g$ satisfy, for a.e. $x,y$, $$\begin{aligned}
|f(x)-f(y)|\leq d(x,y)^\beta(g(x)+g(y)),\end{aligned}$$ with $g\geq0$, $\beta>0$. Then for $\alpha>0$ and $\alpha+\beta<1$, $$\begin{aligned}
|J_\alpha f(x)-J_\alpha f(y)|\leq Cd(x,y)^{\alpha+\beta}(Mg(x)+Mg(y)).\end{aligned}$$ In particular, if $p>1$, $$\begin{aligned}
J_\alpha: M^{\beta,p}{\ensuremath{\rightarrow}}M^{\alpha+\beta,p}.\end{aligned}$$
Once again, using $\int k_\alpha=1$, and proceeding as in the Lipschitz case, $$\begin{aligned}
|J_\alpha f(x)-J_\alpha f(y)| &\leq \int_X |f(x)-f(z)||k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\\
&\hspace{-2.75cm}\leq C\int_{B(x,2d(x,y))}d(x,z)^\beta(g(x)+g(z))\left(\frac{1}{d(x,z)^{N-\alpha}}+\frac{1}{d(y,z)^{N-\alpha}}\right)dm(z)\\
&\hspace{-2.25cm}+C\int_{B(x,2d(x,y))^c}d(x,z)^\beta(g(x)+g(z))\frac{d(x,y)}{d(x,z)^{N-\alpha+1}}dm(z)\\
&\hspace{-2.75cm}\leq Cg(x)d(x,y)^{\alpha+\beta}+Cd(x,y)^{\alpha+\beta}Mg(x)\\
&\hspace{-2.25cm} +Cd(x,y)^\beta g(x)d(x,y)^\alpha+Cd(x,y)^\beta Mg(y)d(x,y)^\alpha\\
&\hspace{-2.25cm} +Cd(x,y)g(x)\frac{1}{d(x,y)^{1-(\alpha+\beta)}}+Cd(x,y)\frac{1}{d(x,y)^{1-(\alpha+\beta)}}Mg(x)\\
&\hspace{-2.75cm}\leq Cd(x,y)^{\alpha+\beta}(Mg(x)+Mg(y)).\end{aligned}$$
Potential spaces $L^{\alpha,p}$
===============================
In this section we define potential spaces $L^{\alpha,p}$ and see they are Banach spaces. We prove they are embedded in certain Sobolev and Besov spaces, and that Lipschitz functions are dense. We finish the section with Sobolev embedding theorems for $L^{\alpha,p}$.
For $\alpha>0$, we define the **potential space** $$\begin{aligned}
L^{\alpha,p}(X)=\{f\in L^p:\exists g\in L^p, f=J_\alpha g\}=J_\alpha(L^p)\end{aligned}$$ and equip it with the following norm $$\begin{aligned}
\|f\|_{\alpha,p}=\|f\|_p+\inf_{g\in J_\alpha^{-1}(\{f\})}\|g\|_p.\end{aligned}$$
$L^{\alpha,p}$ is Banach.
To prove completeness, we will show the convergence of every absolutely convergent series. Let $(f_n)$ be a sequence in $L^{\alpha,p}$ such that $$\begin{aligned}
\sum_n\|f_n\|_{\alpha,p}<\infty.\end{aligned}$$ In particular, $\sum_n\|f_n\|_p<\infty$, so the series $\sum_n f_n$ converges in $L^p$ to some function $f$. For each $n$, take $g_n$ in $L^p$ with $f_n=J_\alpha g_n$ and $$\begin{aligned}
\|g_n\|_p\leq \|f_n\|_{\alpha,p}+2^{-n},\end{aligned}$$ then clearly $\sum_n \|g_n\|_p<\infty$ and $\sum_n g_n$ converges to some $g\in L^p$. Finally, as $J_\alpha$ is continuous in $L^p$, $$\begin{aligned}
f=\sum_n f_n=\sum_n J_\alpha g_n=J_\alpha\left(\sum_n g_n\right)=J_\alpha g\end{aligned}$$ so $f\in L^{\alpha,p}$, and $$\begin{aligned}
\left\|f-\sum_{k=1}^nf_k\right\|_{\alpha,p}\leq \left\|f-\sum_{k=1}^nf_k\right\|_p+\left\|g-\sum_{k=1}^ng_k\right\|_p{\ensuremath{\rightarrow}}0.\end{aligned}$$
\[linf\]$\|J_\alpha g\|_{\alpha,p}\leq 2\|g\|_p$, so it is continuous from $L^p$ onto $L^{\alpha,p}$. In particular, as $L^\infty\cap L^p$ is dense in $L^p$ for $1\leq p\leq\infty$, we get that $J_\alpha(L^\infty\cap L^p)$ is dense in $L^{\alpha,p}$.
The following theorem shows that ‘potential functions’ have Hajasz gradients, and this leads to some interesting results, such as Lipschitz density and embeddings in Sobolev spaces.
\[eleeme\]Let $f=J_\alpha g$ for some $g$ such that $f$ is finite $a.e.$. Then if $0<\alpha<1$, $$\begin{aligned}
|f(x)-f(y)|\leq C_\alpha d(x,y)^\alpha (Mg(x)+Mg(y))\end{aligned}$$ for every $x,y$ outside a set of measure zero. If $\alpha\geq1$, then for each $\beta<1$ we get $$\begin{aligned}
|f(x)-f(y)|\leq C_{\alpha,\beta}d(x,y)^\alpha (Mg(x)+Mg(y))\end{aligned}$$ for every $x,y$ outside a set of measure zero.
Assume first $\alpha<1$. Let $d=d(x,y)$, $$\begin{aligned}
|f(x)-f(y)| &\leq \int_X |g(z)||k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\\
&\leq \int_{B(x,2d)}+\int_{X{\symbol{92}}B(x,2d)}=I+II.\end{aligned}$$ In $I$ we have $$\begin{aligned}
I &\leq C\int_{B(x,2d)}|g(z)|\frac{1}{d(x,z)^{N-\alpha}}dm(z)+C\int_{B(y,3d)}|g(z)|\frac{1}{d(y,z)^{N-\alpha}}dm(z)\\
&\leq Cd^\alpha (Mg(x)+Mg(y)),\end{aligned}$$ and for $II$, as $d(x,z)\sim d(y,z)$ we get $$\begin{aligned}
II &\leq Cd\int_{B(x,2d)^c}|g(z)|d(x,z)^{-(N+1-\alpha)}dm(z)\\
&\leq Cdd^{-(1-\alpha)}Mg(x)=Cd^\alpha Mg(x).\end{aligned}$$ Let now $\alpha\geq 1$ and fix $0<\beta<1$. Observe that the bound for $I$ also holds in this case, and for $d(x,y)<1$ we get $$\begin{aligned}
I\leq Cd^\alpha (Mg(x)+Mg(y))\leq Cd^\beta (Mg(x)+Mg(y)).\end{aligned}$$ We now divide $X{\symbol{92}}B(x,2d)$ in two regions (and use in both cases the fact that $d(x,z)\sim d(y,z)$) $$\begin{aligned}
\int_{2d\leq d(x,z)<5}|g(z)|\frac{d}{d(x,z)^{N-\alpha+1}}dm(z) &\leq\\
&\hspace{-1cm}\leq \int_{2d\leq d(x,z)<5}|g(z)|\frac{d^\beta}{d(x,z)^{N-(\alpha-\beta)}}dm(z)\\
&\hspace{-1cm}\leq Cd^\beta Mg(x);\end{aligned}$$ and if $d(x,z)\geq 5$, as $d(y,z)\geq 4$ we can use the other bound for differences of $k_\alpha$ (property 6 in \[propk\]) $$\begin{aligned}
\int_{d(x,z)\geq 5}|g(z)|\frac{d}{d(x,z)^{N+\alpha+1}}dm(z)\leq CdMg(x)\leq Cd^\beta Mg(x).\end{aligned}$$ Finally, if $d(x,y)\geq 1$, as $|f|\leq Mg$, $$\begin{aligned}
|f(x)-f(y)|\leq C(Mg(x)+Mg(y))\leq Cd(x,y)^\beta (Mg(x)+Mg(y)).\end{aligned}$$
Let $1<p<\infty$. If $0<\alpha<1$, then $L^{\alpha,p}{\ensuremath{\hookrightarrow}}M^{\alpha,p}$. For $\alpha\geq 1$, $L^{\alpha,p}{\ensuremath{\hookrightarrow}}M^{\beta,p}$ for all $0<\beta<1$.
\[eleeme2\]Let $p=\infty$. If $0<\alpha<1$, then $L^{\alpha,\infty}{\ensuremath{\hookrightarrow}}C^\alpha$. For $\alpha\geq 1$, $L^{\alpha,\infty}{\ensuremath{\hookrightarrow}}C^\beta$ for all $0<\beta<1$. In particular, functions in $L^{\alpha,\infty}$ are continuous for all $\alpha>0$ (after eventual modification on a null set).
From this last result and remark \[linf\], we get the following density property.
Let $1\leq p\leq\infty$ and $\alpha>0$. Then $C^\beta\cap L^{\alpha,p}$ is dense in $L^{\alpha,p}$ for all $0<\beta\leq \alpha$ if $\alpha<1$, and for all $0<\beta<1$ if $\alpha\geq1$.
As a last corollary of theorem \[eleeme\], since $Mg$ is a Hajasz gradient for potential functions, we get the following Poincaré inequality.
Let $0<\alpha<1$ and $f=J_\alpha g$ for some $g$ such that $f\in L^1_{loc}$, then for each ball $B$ we get $$\begin{aligned}
\fint_B |f-f_B|\leq C\text{diam}(B)^\alpha \fint_B Mg.\end{aligned}$$
Now, regarding Besov spaces, as $M^{\alpha,p}{\ensuremath{\hookrightarrow}}B^{\alpha-\epsilon}_{p,p}$ for $1\leq p<\infty$ and $0<\epsilon<\alpha$, from \[eleeme2\] we obtain for $\alpha<1$ $L^{\alpha,p}{\ensuremath{\hookrightarrow}}B^{\alpha-\epsilon}_{p,p}$. This also holds true for $B^{\alpha-\epsilon}_{p,q}$. First, a lemma.
\[besovk\]Let $0<\alpha<1$ and $q>0$ satifying $q(N-\alpha)<N<q(N+q-\alpha)$. Then there exists $C>0$ such that, for every $z\in X$ and $t>0$ $$\begin{aligned}
\int_X\fint_{B(x,t)}|k_\alpha(x,z)-k_\alpha(y,z)|^q dm(y)dm(x)\leq C t^{N-q(N-\alpha)}.\end{aligned}$$
Consider $$\begin{aligned}
A_1=\{(x,y):d(x,y)<t,d(x,z)<2t\};\end{aligned}$$ $$\begin{aligned}
A_2=\{(x,y):d(x,y)<t,2t\leq d(x,z)\}.\end{aligned}$$ Integrating over $A_1$, we get $$\begin{aligned}
\iint_{A_1}\frac{1}{t^N}|k_\alpha(x,z)-k_\alpha(y,z)|^q dm(y)dm(x) &\leq C\int_{B(z,3t)}|k_\alpha(x,z)|^q dm(x)\\
&\leq C\int_{B(z,3t)}\frac{1}{d(x,z)^{q(N-\alpha)}}dm(x)\\
&\leq Ct^{N-q(N-\alpha)},\end{aligned}$$ and the last inequality holds because $N>q(N-\alpha)$.
In $A_2$ we have $d(x,z)\sim d(y,z)$, and then, as $d(x,y)<t$, $$\begin{aligned}
\iint_{A_2}\frac{1}{t^N}|k_\alpha(x,z)-k_\alpha(y,z)|^qdm(y)dm(x) &\leq\\
&\hspace{-2cm}\leq Ct^q\iint_{A_2}\frac{1}{t^N}\frac{1}{d(x,z)^{q(N+1-\alpha)}}dm(y)dm(x)\\
&\hspace{-2cm}\leq Ct^q\int_{X{\symbol{92}}B(z,2t)}\frac{1}{d(x,z)^{q(N+1-\alpha)}}dm(x)\\
&\hspace{-2cm}\leq C t^qt^{N-q(N+1-\alpha)}\leq Ct^{N-q(N-\alpha)},\end{aligned}$$ given $N<q(N+1-\alpha)$.
Let $f=J_\alpha g$, $0<\alpha<1$ and $1\leq p\leq\infty$, then for $t>0$ we get $$\begin{aligned}
E_pf(t)\leq Ct^\alpha\|g\|_p\end{aligned}$$
If $p<\infty$, $$\begin{aligned}
|f(x)-f(y)|^p &\leq\left(\int_X|k_\alpha(x,z)-k_\alpha(y,z)|^{\frac{1}{p}+\frac{1}{p'}}|g(z)|dm(z)\right)^p\\
&\leq\left(\int_X|k_\alpha(x,z)-k_\alpha(y,z)||g(z)|^pdm(z)\right)\\
&\hspace{1cm}\times\left(\int_X|k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\right)^{p/p'}.\end{aligned}$$ By lemma \[ineqk\] for $q=1$, as $d(x,y)<t$ and $\alpha<1$, $$\begin{aligned}
\int_X|k_\alpha(x,z)-k_\alpha(y,z)|dm(z)\leq Ct^\alpha\end{aligned}$$ so $$\begin{aligned}
\int_X\fint_{B(x,t)}|f(x)-f(y)|^pdm(y)dm(x) &\leq\\
&\hspace{-6cm}\leq C t^{\alpha p/p'}\int_X\left(\int_X\fint_{B(x,t)}|k_\alpha(x,z)-k_\alpha(y,z)|dm(y)dm(x)\right)|g(z)|^pdm(z)\end{aligned}$$ and by lemma \[besovk\] (also taking $q=1$) $$\begin{aligned}
\int_X\fint_{B(x,t)}|f(x)-f(y)|^pdm(y)dm(x)\leq Ct^{\alpha p/p'} t^\alpha \|g\|_p^p=C t^{\alpha p}\|g\|_p^p.\end{aligned}$$ For $p=\infty$, as $\alpha<1$, $$\begin{aligned}
E_\infty f(t) &=\sup_{d(x,y)<t}|f(x)-f(y)|\\
&\leq C\sup_{d(x,y)<t} d(x,y)^\alpha(Mg(x)+Mg(y))\\
&\leq Ct^\alpha\|g\|_\infty.\end{aligned}$$
We can now conclude the following embedding in Besov spaces.
Let $1\leq p\leq\infty$ and $0<\alpha<1$. Then for $1\leq q<\infty$ and $0<\epsilon<\alpha$ we have $L^{\alpha,p}{\ensuremath{\hookrightarrow}}B^{\alpha-\epsilon}_{p,q}$. For $q=\infty$ we obtain $L^{\alpha,p}{\ensuremath{\hookrightarrow}}B^\alpha_{p,\infty}$.
Let $f=J_\alpha g$. By the previous proposition, if $q=\infty$, $$\begin{aligned}
\|f\|_{B^\alpha_{p,\infty}}=\|f\|_p+\sup_{t>0}t^{-\alpha}E_pf(t)\leq C\|f\|_{\alpha,p}.\end{aligned}$$ And for $1\leq q<\infty$, as we also have $E_p f\leq C\|f\|_p$, $$\begin{aligned}
\|f\|_{B^{\alpha-\epsilon}_{p,q}} &\leq C\|f\|_p+C\left(\int_0^1 t^{-(\alpha-\epsilon)q}E_pf(t)^q\frac{dt}{t}\right)^{1/q}\\
&\leq C\|f\|_p+C\|g\|_p\left(\int_0^1 t^{\epsilon q}\frac{dt}{t}\right)\leq \frac{C}{\epsilon^{1/q}}\|f\|_{\alpha,p}.\end{aligned}$$
We finish this section with Sobolev-type embedding theorems for potential spaces. First we need a lemma.
\[boundk\]For $\alpha>0$ and $q>0$ satisfying $q(N-\alpha)<N<q(N+\alpha)$, there exists $C>0$ such that for every $x\in X$, $$\begin{aligned}
\int_X k_\alpha(x,y)^qdm(y)\leq C<\infty.\end{aligned}$$
By lemma \[propk\], $$\begin{aligned}
k_\alpha(x,y)^q\leq C\frac{\chi_{B(x,4)}(y)}{d(x,y)^{q(N-\alpha)}}+C\frac{\chi_{X{\symbol{92}}B(x,4)}(y)}{d(x,y)^{q(N+\alpha)}},\end{aligned}$$ and restrictions over $q$ guarantee integrability.
Let $1<p<\infty$ and $\alpha>0$. The following embeddings hold for $L^{\alpha,p}$
a.
: If $p<\frac{N}{\alpha}$, $$\begin{aligned}
L^{\alpha,p}{\ensuremath{\hookrightarrow}}L^q\end{aligned}$$ for $p\leq q\leq p^*$ where $\frac{1}{p^*}=\frac{1}{p}-\frac{\alpha}{N}$.
b.
: If $p=\frac{N}{\alpha}$, then for $p\leq q<\infty$, $$\begin{aligned}
L^{\alpha,p}{\ensuremath{\hookrightarrow}}L^q.\end{aligned}$$ If in addition $\alpha<1$, $$\begin{aligned}
L^{\alpha,p}{\ensuremath{\hookrightarrow}}BMO.\end{aligned}$$
c.
: If $p>\frac{N}{\alpha}$,then for $p\leq q\leq \infty$ $$\begin{aligned}
L^{\alpha,p}{\ensuremath{\hookrightarrow}}L^q.\end{aligned}$$ If in addition $\alpha<1+N/p$, $$\begin{aligned}
L^{\alpha,p}{\ensuremath{\hookrightarrow}}C^{\alpha-N/p}.\end{aligned}$$
<!-- -->
a.
: We know $L^{\alpha,p}{\ensuremath{\hookrightarrow}}L^p$ (for $\|f\|_p\leq \|f\|_{\alpha,p}$), then if we prove $L^{\alpha,p}{\ensuremath{\hookrightarrow}}L^{p*}$, by an interpolation argument we are done. This follows from $|J_\alpha f|\leq CI_\alpha|f|$, as $(N-\alpha)p'>N$ and for any $t>0$ we get $$\begin{aligned}
|J_\alpha f(x)| &\leq C\int_{B(x,t)}\frac{|f(y)|}{d(x,y)^{N-\alpha}}dm(y)+C\int_{X{\symbol{92}}B(x,t)}\frac{|f(y)|}{d(x,y)^{N-\alpha}}dm(y)\\
&\leq Ct^\alpha Mf(x)+Ct^{(N-(N-\alpha)p')/p'}\|f\|_p\\
&= Ct^\alpha Mf(x)+Ct^{-N/p^*}\|f\|_p.\end{aligned}$$ This last expression attains its minimum for $t=CMf(x)^{-p/N}\|f\|_p^{p/N}$, and for this value of $t$ we obtain $$\begin{aligned}
|J_\alpha f(x)|\leq C Mf(x)^{p/p^*}\|f\|_p^{1-p/p^*},\end{aligned}$$ and as $p>1$, boundedness of the maximal function implies $$\begin{aligned}
\int_X |J_\alpha f|^{p^*}dm\leq C\|f\|_p^{p^*-p}\int_X (Mf)^pdm\leq C \|f\|_p^{p^*}.\end{aligned}$$
b.
: Let $N/\alpha=p<q<\infty$, so there exists $a>1$ such that $$\begin{aligned}
1+\frac{1}{q}=\frac{\alpha}{N}+\frac{1}{a}.\end{aligned}$$ In particular $a(N-\alpha)<N$ (and also $N<a(N+\alpha)$, as $a>1$), so by the previous lemma $$\begin{aligned}
\int_X k_\alpha(x,y)^a dm(y)\leq C<\infty.\end{aligned}$$ Let now $f=J_\alpha g$ with $g\in L^p$, as $\frac{1}{q'}=\frac{1}{p'}+\frac{1}{a'}$ by Hölder’s inequality we obtain the following Young-type inequality $$\begin{aligned}
|f(x)| &\leq \int_X k_\alpha(x,y)^{a/q+a/p'}|g(y)|^{p/q+p/a'}dm(y)\\
&\leq\left(\int_X k_\alpha(x,y)^a|g(y)|^pdm(y)\right)^{1/q}\left(\int_X|g(y)|^pdm(y)\right)^{1/a'}\\
&\hspace{1cm}\times\left(\int_Xk_\alpha(x,y)^adm(y)\right)^{1/p'}\\
&\leq C\|g\|_p^{p/a'}\left(\int_X k_\alpha(x,y)^a|g(y)|^pdm(y)\right)^{1/q}\end{aligned}$$ (here we use $a/q+a/p'=1$ and $p/q+p/a'=1$) and $$\begin{aligned}
\int_X|f(x)|^qdm(x) &\leq C\|g\|_p^{qp/a'}\int_X\int_X k_\alpha(x,y)^a|g(y)|^pdm(y)dm(x)\\
&\leq C\|g\|_p^{p(q/a'+1)}=C\|g\|_p^q.\end{aligned}$$
Moreover, if $\alpha<1$, by Poincaré inequality for any ball $B$, $$\begin{aligned}
\fint_B |f-f_B| &\leq C\text{diam}(B)^\alpha \fint_B Mg\leq Cm(B)^{\alpha/N}\left(\fint_B (Mg)^{N/\alpha}\right)^{\alpha/N}\\
&\leq C\left(\int_B (Mg)^{N/\alpha}\right)^{\alpha/N}\leq C\|g\|_{N/\alpha}\end{aligned}$$ and we conclude $$\begin{aligned}
\|f\|_{BMO}\leq C\|f\|_{\alpha,N/\alpha}.\end{aligned}$$
c.
: For the first part, again by interpolation it is enough to prove $L^{\alpha,p}{\ensuremath{\hookrightarrow}}L^\infty$. If $f=J_\alpha g$ with $g\in L^p$, $$\begin{aligned}
|f(x)| &=|J_\alpha g(x)|\leq \int_X k_\alpha(x,y)|g(y)|dm(y)\\
&\leq \|g\|_p\left(\int_X k_\alpha(x,y)^{p'}dm(y)\right)^{1/p'}\\
&\leq C\|g\|_p\leq C\|f\|_{\alpha,p}\end{aligned}$$ as long as $p'(N-\alpha)<N<p'(N+\alpha)$. The second inequality is trivial for $p'\geq 1$ and the first one is equivalent to $p\alpha>N$.
Assume now $\alpha<1+N/p$. Then $$\begin{aligned}
|f(x)-f(y)| &\leq \int_X |k_\alpha(x,z)-k_\alpha(y,z)||g(z)|dm(z)\\
&\leq \|g\|_p\left(\int_X |k_\alpha(x,z)-k_\alpha(y,z)|^{p'}dm(z)\right)^{1/p'}\\
&\leq C\|g\|_p d(x,y)^{\frac{N-p'(N-\alpha)}{p'}}=C\|g\|_p d(x,y)^{\alpha-N/p}\end{aligned}$$ if $p'(N-\alpha)<N<p'(N-\alpha+1)$. The first inequality is once again equivalent to $p>N/\alpha$, and the second to $\alpha<1+N/p$.
The inverse of $J_\alpha$
=========================
In this section, with the fractional derivative $D_\alpha$ as defined in [@GSV], we prove conditions for the composition $(I+D_\alpha)J_\alpha$ to be inversible in $L^p$ for $1<p<\infty$, which in turn will lead to inversibility of $J_\alpha$. We follow the techniques used in [@Hz], proving $$\begin{aligned}
\|I-(I+D_\alpha)J_\alpha\|_{L^p{\ensuremath{\rightarrow}}L^p}<1\end{aligned}$$ by rewriting the operators in terms of $(Q_t)_{t>0}$ instead of $(S_t)_{t>0}$, and applying the $T1$ theorem for Ahlfors spaces (see [@Ga]).
Let $\alpha>0$. Define $$\begin{aligned}
n_\alpha(x,y)=\int_0^\infty \alpha t^{-\alpha}s(x,y,t)\frac{dt}{t}.\end{aligned}$$ This kernel satisfies $$\begin{aligned}
n_\alpha(x,y)\sim \frac{1}{d(x,y)^{N+\alpha}}\end{aligned}$$ and $$\begin{aligned}
|n_\alpha(x,y)-n_\alpha(x',y)|\leq C d(x,x')(d(x,y){\ensuremath{\wedge}}d(x',y))^{-(N+1+\alpha)}.\end{aligned}$$
The fractional derivative can be then defined as $$\begin{aligned}
D_\alpha f(x)=\int_X n_\alpha(x,y)(f(x)-f(y))dm(y),\end{aligned}$$ see for instance [@GSV], whenever this integral makes sense (for instance if $f$ has sufficient regularity of Lipschitz or Besov type).
Let us now rewrite the operators with $Q_t=-t\frac{d}{dt}S_t$. Assume $f\in C_c^\gamma$ for some $\alpha<\gamma\leq 1$, then $$\begin{aligned}
J_\alpha f(x) &=\int_X k_\alpha(x,y)f(y)dm(y)=\int_X\int_0^\infty \frac{\alpha t^{\alpha-1}}{(1+t^\alpha)^2}s(x,y,t)f(y)dtdm(y)\\
&=\int_0^\infty \frac{\alpha t^{\alpha-1}}{(1+t^\alpha)^2}S_tf(x)dt=\int_0^\infty \frac{d}{dt}\left(\frac{1}{1+t^{-\alpha}}\right)S_tf(x)dt\\
&=\left.\frac{S_tf(x)}{1+t^{-\alpha}}\right|_0^\infty+\int_0^\infty \frac{1}{1+t^{-\alpha}}\left(-t\frac{d}{dt}S_tf(x)\right)\frac{dt}{t}\\
&=\int_0^\infty \frac{1}{1+t^{-\alpha}}Q_tf(x)\frac{dt}{t}\end{aligned}$$ where we have used $S_tf{\ensuremath{\rightarrow}}f$ when $t{\ensuremath{\rightarrow}}0$ and $S_tf{\ensuremath{\rightarrow}}0$ when $t{\ensuremath{\rightarrow}}\infty$.
On the other hand, we obtain $$\begin{aligned}
D_\alpha f(x) &=\int_X n_\alpha(x,y)(f(x)-f(y))dm(y)\\
&=\int_X\int_0^\infty \alpha t^{-\alpha-1}s(x,y,t)(f(x)-f(y))dtdm(y)\\
&=\int_0^\infty \alpha t^{-\alpha-1}(f(x)-S_tf(x))dt=\int_0^\infty \frac{d}{dt}\left(t^{-\alpha}\right)(S_tf(x)-f(x))dt\\
&=\left.\frac{(S_tf(x)-f(x))}{t^\alpha}\right|_0^\infty+\int_0^\infty t^{-\alpha}\left(-t\frac{d}{dt}S_tf(x)\right)\frac{dt}{t}\\
&=\int_0^\infty t^{-\alpha}Q_tf(x)\frac{dt}{t},\end{aligned}$$ where we have used that $S_tf{\ensuremath{\rightarrow}}0$ when $t{\ensuremath{\rightarrow}}\infty$ and that $|S_tf(x)-f(x)|\leq Ct^\gamma$. Since we also have $$\begin{aligned}
f(x)=-\int_0^\infty\frac{d}{dt}S_tf(x)dt=\int_0^\infty Q_tf(x)\frac{dt}{t},\end{aligned}$$ we get $$\begin{aligned}
(I+ D_\alpha)f(x)=\int_0^\infty (1+t^{-\alpha})Q_tf(x)\frac{dt}{t}.\end{aligned}$$
This way, $$\begin{aligned}
(I+D_\alpha) J_\alpha f=\int_0^\infty\int_0^\infty\frac{1+s^{-\alpha}}{1+t^{-\alpha}}Q_sQ_tf\frac{dt}{t}\frac{ds}{s},\end{aligned}$$ and as we also have $$\begin{aligned}
f=\int_0^\infty\int_0^\infty Q_sQ_tf\frac{dt}{t}\frac{ds}{s},\end{aligned}$$ we conclude $$\begin{aligned}
(I-(I+D_\alpha)J_\alpha)f &=\int_0^\infty\int_0^\infty \left(1-\frac{1+s^{-\alpha}}{1+t^{-\alpha}}\right) Q_sQ_tf\frac{dt}{t}\frac{ds}{s}\\
&=\int_0^\infty\int_0^\infty \frac{t^{-\alpha}-s^{-\alpha}}{1+t^{-\alpha}} Q_sQ_tf\frac{dt}{t}\frac{ds}{s}\\
&=\int_0^\infty (1-v^\alpha)\left(\int_0^\infty \frac{1}{1+(uv)^\alpha}Q_uQ_{uv}f\frac{du}{u}\right)\frac{dv}{v}.\end{aligned}$$
For each $v>0$ we define $$\begin{aligned}
T_{\alpha,v}f=\int_0^\infty \frac{1}{1+(uv)^\alpha}Q_uQ_{uv}f\frac{du}{u},\end{aligned}$$ and, following [@Hz], if we can prove $$\begin{aligned}
\|T_{\alpha,v}f\|_p\leq C_{\alpha,p}(v)\|f\|_p,\end{aligned}$$ with $$\begin{aligned}
\int_0^\infty |1-v^\alpha| C_{\alpha,p}(v)\frac{dv}{v}<1\end{aligned}$$ for $\alpha$ small enough, we will obtain $$\begin{aligned}
\|(I-(I+D_\alpha) J_\alpha)f\|_p\leq \int_0^\infty |1-v^\alpha|\|T_{v,\alpha}f\|_p\frac{dv}{v}< \|f\|_p\end{aligned}$$ and therefore $(I+D_\alpha) J_\alpha$ will be inversible for those values of $\alpha$.
To prove the boundedness of $T_{\alpha,v}$, we will use the $T1$ theorem as presented in \[tedeuno\]. As a first step, we need to show $T_{\alpha,v}$ is a singular integral operator, for which we need to find its kernel.
For $u,v>0$, $x,z\in X$, $$\begin{aligned}
\left|\int_X q(x,y,u)q(y,z,uv)dm(y)\right|\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v^{N+1}}\right)\frac{1}{u^N}\chi_{\left(\frac{d(x,z)}{4(v+1)},\infty\right)}(u).\end{aligned}$$ As a consequence, $$\begin{aligned}
\left|\int_0^\infty \frac{1}{1+(uv)^\alpha}\int_X q(x,y,u)q(y,z,uv)dm(y)\frac{du}{u}\right|\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right) \frac{1}{d(x,z)^N}.\end{aligned}$$
The second inequality follows immediately from the first one. For this one, as $$\begin{aligned}
q(x,y,u)=0 \text{ when } d(x,y)\geq 4u; \quad q(y,z,uv)=0 \text{ when } d(y,z)\geq 4uv,\end{aligned}$$ for the product to be non zero $d(x,z)<4u(v+1)$ must hold. If $v\geq 1$, as $\int_X q(x,y,u)q(x,z,uv)dm(y)=0$, $$\begin{aligned}
\left|\int_X q(x,y,u)q(y,z,uv)dm(y)\right| &=\\
&\hspace{-1cm}=\left|\int_X q(x,y,u)(q(y,z,uv)-q(x,z,uv))dm(y)\right|\\
&\hspace{-1cm}\leq C\int_{B(x,4u)}\frac{1}{u^N}\frac{d(x,y)}{(uv)^{N+1}}dm(y)\leq C\frac{1}{u^N}\frac{1}{v^{N+1}};\end{aligned}$$ and if $v<1$, as $\int_X q(x,z,u)q(y,z,uv)dm(y)=0$, $$\begin{aligned}
\left|\int_X q(x,y,u)q(y,z,uv)dm(y)\right| &=\\
&\hspace{-1cm}=\left|\int_X (q(x,y,u)-q(x,z,u))q(y,z,uv)dm(y)\right|\\
&\hspace{-1cm}\leq C\int_{B(z,4uv)}\frac{d(y,z)}{u^{N+1}}\frac{1}{(uv)^N}dm(y)\leq C\frac{1}{u^N}v.\end{aligned}$$
Let now $f,g\in C^\beta_c$ with disjoint supports, and let $x\in supp(g)$. Then $$\begin{aligned}
T_{\alpha,v}f(x) &=\int_0^\infty \frac{1}{1+(uv)^\alpha}Q_uQ_{uv}f\frac{du}{u}\\
&\hspace{-0.75cm}=\int_0^\infty \frac{1}{1+(uv)^\alpha}\left(\int_X q(x,y,u)\left(\int_X q(y,z,uv)f(z)dm(z)\right)dm(y)\right)\frac{du}{u}\end{aligned}$$ and from the previous lemma we have this integral converges absolutely, so we can change the order of integration and obtain $$\begin{aligned}
\left\langle T_{\alpha,v}f,g\right\rangle=\int_X\int_X N_{\alpha,v}(x,z)f(z)g(x)dm(z)dm(x),\end{aligned}$$ where $$\begin{aligned}
N_{\alpha,v}(x,z)=\int_0^\infty \frac{1}{1+(uv)^\alpha}\int_X q(x,y,u)q(y,z,uv)dm(y)\frac{du}{u}.\end{aligned}$$
From the previous lemma, $N_{\alpha,v}(x,z)\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)\frac{1}{d(x,z)^N}$. To see that $T_{\alpha,v}$ is a singular integral operator we need to check the smoothness conditions for the kernel $N_{\alpha,v}$.
For $u,v>0$, $x,x',z\in X$ and $0<\delta<1$, it holds $$\begin{aligned}
\left|\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\right| &\leq\\
&\hspace{-5cm}\leq C\left(\frac{d(x,x')}{u}\right)^{1-\delta}\left(v^\delta{\ensuremath{\wedge}}\frac{1}{v^{N+1}}\right)\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u).\end{aligned}$$ From this we obtain $$\begin{aligned}
\left|\int_0^\infty \frac{1}{1+(uv)^\alpha}\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\frac{du}{u}\right| &\leq\\
&\hspace{-7cm}\leq C\frac{d(x,x')^{1-\delta}}{(d(x,z){\ensuremath{\wedge}}d(x',z))^{N+1-\delta}}\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta.\end{aligned}$$
As in the other lemma, the second inequality follows from the first one. We consider two cases: If $v\geq1$ y $d(x,x')\geq u$, by that same lemma, $$\begin{aligned}
\left|\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\right| &\leq\\
&\hspace{-3cm}\leq C\frac{1}{v^{N+1}}\frac{1}{u^N}\left(\chi_{\left(\frac{d(x,z)}{4(v+1)},\infty\right)}(u)+\chi_{\left(\frac{d(x',z)}{4(v+1)},\infty\right)}(u)\right)\\
&\hspace{-3cm}\leq C\frac{1}{v^{N+1}}\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u)\\
&\hspace{-3cm}\leq C\frac{1}{v^{N+1}}\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u) \left(\frac{d(x,x')}{u}\right)^{1-\delta}.\end{aligned}$$ And for $d(x,x')<u$, the integrand will be nonzero only if $d(x,z)<4u(v+1)$ or $d(x',z)<4u(v+1)$, so $$\begin{aligned}
\left|\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\right|=\\
&\hspace{-6cm}=\left|\int_X (q(x,y,u)-q(x',y,u))(q(y,z,uv)-q(x,z,uv))dm(y)\right|\\
&\hspace{-6cm}\leq Cd(x,x')\frac{1}{u^{N+1}}\frac{1}{(uv)^{N+1}}\int_{B(x,4u)\cup B(x',4u)}d(x,y)dm(y)\\
&\hspace{-6cm}\leq C\frac{1}{v^{N+1}}\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u) \left(\frac{d(x,x')}{u}\right)\\
&\hspace{-6cm}\leq C\frac{1}{v^{N+1}}\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u) \left(\frac{d(x,x')}{u}\right)^{1-\delta}.\end{aligned}$$ For the case $v<1$, on one hand by the previous lemma we obtain $$\begin{aligned}
\left|\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\right|\leq\\
&\hspace{-1cm}\leq Cv\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u),\end{aligned}$$ on the other hand $$\begin{aligned}
\left|\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\right|\leq\\
&\hspace{-2cm}\leq C \frac{d(x,x')}{u}\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u),\end{aligned}$$ and by combining both inequalities we get $$\begin{aligned}
\left|\int_X (q(x,y,u)-q(x',y,u))q(y,z,uv)dm(y)\right| &\leq\\
&\hspace{-2cm}\leq C v^\delta \left(\frac{d(x,x')}{u}\right)^{1-\delta}\frac{1}{u^N}\chi_{\left(\frac{d(x,z){\ensuremath{\wedge}}d(x',z)}{4(v+1)},\infty\right)}(u).\end{aligned}$$
For the rest of the section, we fix $0<\delta<1$. Joining both lemmas we conclude
\[singint\]$T_{\alpha,v}$ is a singular integral operator. Its kernel $N_{\alpha,v}$ satisfies $$\begin{aligned}
|N_{\alpha,v}(x,z)|\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta \frac{1}{d(x,z)^N};\end{aligned}$$ and for $3d(x,x')<d(x,z)$, $$\begin{aligned}
|N_{\alpha,v}(x,z)-N_{\alpha,v}(x',z)|\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta\frac{d(x,x')^{1-\delta}}{d(x,z)^{N+1-\delta}}\end{aligned}$$ and $$\begin{aligned}
|N_{\alpha,v}(z,x)-N_{\alpha,v}(z,x')|\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta\frac{d(x,x')^{1-\delta}}{d(x,z)^{N+1-\delta}}.\end{aligned}$$
To prove each $T_{\alpha,v}$ is a Calderón-Zygmund operator, and thus bounded in $L^p$, we will use the $T1$ theorem. The next lemma proves the other conditions needed.
\[weak\]$T_{\alpha,v}$ satisfies $$\begin{aligned}
T_{\alpha,v}1=0,\end{aligned}$$ $$\begin{aligned}
T_{\alpha,v}^*1=0,\end{aligned}$$ and for $f,g\in C^\beta_c(B)$, for some ball $B$, $$\begin{aligned}
|\langle T_{\alpha,v}f,g\rangle|\leq C \left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta m(B)^{1+\frac{2\beta}{N}}[f]_\beta[g]_\beta.\end{aligned}$$
The first equality is immediate, the second uses the fact that $q$ is symmetrical. $$\begin{aligned}
\langle T_{\alpha,v}f,g\rangle &= \int_X\left(\int_X N_{\alpha,v}(x,z)f(z)dm(z)\right)g(x)dm(x)\\
&=\int_X \int_X\int_0^\infty \int_X \frac{1}{1+(uv)^\alpha}\\
&\hspace{1cm}\times q(x,y,u)q(y,z,uv)dm(y)\frac{du}{u}f(z)dm(z)g(x)dm(x)\\
&=\int_Xf(z)\left(\int_X N^*_{\alpha,v}(z,x)g(x)dm(x)\right)dm(z)=\langle f,T_{\alpha,v}^*g\rangle\end{aligned}$$ so clearly $T_{\alpha,v}^*1=0$.
For the third one, as $$\begin{aligned}
\left\langle T_{\alpha,v}f,g \right\rangle &=\\
&\hspace{-1.5cm}=\int\limits_0^\infty\frac{1}{1+(uv)^\alpha}\int_X\int_X\int_X q(x,y,u)q(y,z,uv)f(z)g(x)dm(y)dm(z)dm(x)\frac{du}{u}\end{aligned}$$ we observe that the triple integral inside may be estimated in three different ways
- Firstly, $$\begin{aligned}
A &=\left|\int_X\int_X\int_X q(x,y,u)q(y,z,uv)f(z)g(x)dm(y)dm(z)dm(x)\right|\\
&\leq C\|f\|_\infty\|g\|_\infty\left(v{\ensuremath{\wedge}}\frac{1}{v^{N+1}}\right)\frac{1}{u^N}\int_B\int_B \chi_{B(x,4u(v+1))}(z)dm(z)dm(x)\\
&\leq C[f]_\beta[g]_\beta m(B)^{2\beta/N}\left(v{\ensuremath{\wedge}}\frac{1}{v^{N+1}}\right)m(B)(v+1)^N\\
&\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)[f]_\beta[g]_\beta m(B)^{1+2\beta/N}.\end{aligned}$$
- Secondly, using the fact that $\int_X q(x,y,u)q(y,z,uv)f(y)g(x)dm(z)=0$, $$\begin{aligned}
A&=\left|\int_X\int_X\int_X q(x,y,u)q(y,z,uv)(f(z)-f(y))g(x)dm(z)dm(y)dm(x)\right|\\
&\leq C[f]_\beta\|g\|_\infty \int_B\fint_{B(x,4u)}\fint_{B(y,4uv)}d(z,y)^\beta dm(z)dm(y)dm(x)\\
&\leq C[f]_\beta[g]_\beta m(B)^{1+\beta/N}(uv)^\beta\\
&\leq C\left(\frac{uv}{m(B)^{1/N}}\right)^\beta [f]_\beta[g]_\beta m(B)^{1+2\beta/N}.\end{aligned}$$
- And lastly, it also holds $$\begin{aligned}
A &\leq C\|f\|_\infty\|g\|_\infty \frac{m(B)^2}{(uv)^N}\\
&\leq C\left(\frac{uv}{m(B)^{1/N}}\right)^{-N} [f]_\beta[g]_\beta m(B)^{1+2\beta/N}.\end{aligned}$$
By taking an appropriate combination of the previous three inequalities, we have $$\begin{aligned}
A &=\left|\int_X\int_X\int_X q(x,y,u)q(y,z,uv)f(z)g(x)dm(y)dm(z)dm(x)\right|\\
&\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta \left(\left(\frac{uv}{m(B)^{1/N}}\right)^\beta{\ensuremath{\wedge}}\left(\frac{uv}{m(B)^{1/N}}\right)^{-N}\right)^{1-\delta}[f]_\beta[g]_\beta m(B)^{1+2\beta/N},\end{aligned}$$ and conclude $$\begin{aligned}
\left|\left\langle T_{\alpha,v}f,g \right\rangle\right|\leq C\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta [f]_\beta[g]_\beta m(B)^{1+2\beta/N}.\end{aligned}$$
Thus the $T1$ theorem holds for each $T_{\alpha,v}$, and we get
For $1<p<\infty$ and $0<\delta<1$ the following holds $$\begin{aligned}
\|T_{\alpha,v}f\|_p\leq C_p\left(v{\ensuremath{\wedge}}\frac{1}{v}\right)^\delta\|f\|_p.\end{aligned}$$
The fact that the $L^p$-constant of $T_{\alpha,v}$ is bounded by the constants appearing in Theorem \[singint\] and Lemma \[weak\] follows the same ideas that the Euclidean case (see for instance [@Gr]).
From this result, as for $\alpha<\delta$ we have $$\begin{aligned}
\|I-(I+D_\alpha)J_\alpha\|_{L^p{\ensuremath{\rightarrow}}L^p}\leq \int_0^\infty |1-v^\alpha|\|T_{\alpha,v}\|_{L^p{\ensuremath{\rightarrow}}L^p}\frac{dv}{v}\leq C_p\frac{\alpha}{\delta^2-\alpha^2}\end{aligned}$$ so we obtain the estimate we were looking for and we can conclude
- For any $0<\alpha<1$, $I-(I+D_\alpha)J_\alpha$, and thus $(I+D_\alpha)J_\alpha$, is bounded in $L^p$
- There exists $\alpha_0<1$ such that, for $\alpha<\alpha_0$, $$\begin{aligned}
\|I-(I+D_\alpha)J_\alpha\|_{L^p{\ensuremath{\rightarrow}}L^p}<1,\end{aligned}$$ and thus $(I+D_\alpha)J_\alpha$ is inversible (with bounded inverse) in $L^p$. As $J_\alpha$ maps $L^p$ *onto* $L^{\alpha,p}$, $$\begin{aligned}
\left[(I+D_\alpha)J_\alpha\right]^{-1}(I+D_\alpha)J_\alpha = Id_{L^p}\end{aligned}$$ so $J_\alpha$ is inversible with inverse $J_\alpha^{-1}:L^{\alpha,p}{\ensuremath{\rightarrow}}L^p$ given by $$\begin{aligned}
J_\alpha^{-1}=\left[(I+D_\alpha)J_\alpha\right]^{-1}(I+D_\alpha).\end{aligned}$$
A characterization of $L^{\alpha,p}$ in terms of $D_\alpha$
===========================================================
For $0<\alpha<1$ and $1<p<\infty$, we proved that, if $f\in L^{\alpha,p}$, then $f\in L^p$ (this holds for any $\alpha>0$ and $1\leq p\leq\infty$) and $(I+D_\alpha)f\in L^p$, so $$\begin{aligned}
\text{If } f\in L^{\alpha,p}\text{, then } f,D_\alpha f\in L^p,\end{aligned}$$ moreover, $$\begin{aligned}
\|D_\alpha f\|_p\leq C\|f\|_{\alpha,p}.\end{aligned}$$
For the case $\alpha<\alpha_0$, we obtain the reciprocal.
Let $1<p<\infty$ and $0<\alpha<\alpha_0$. Then $$\begin{aligned}
f\in L^{\alpha,p}\text{ if and only if } f,D_\alpha f\in L^p,\end{aligned}$$ Furthermore, $$\begin{aligned}
\|f\|_{\alpha,p}\sim \|(I+D_\alpha)f\|_p.\end{aligned}$$
We have already seen in this case $J_\alpha:L^p{\ensuremath{\rightarrow}}L^{\alpha,p}$ is bijective, and therefore $I+D_\alpha$ is also bijective. If $f,D_\alpha f\in L^p$, define $$\begin{aligned}
g=\left[(I+D_\alpha)J_\alpha\right]^{-1}(I+D_\alpha)f,\end{aligned}$$ we get $g\in L^p$ and $$\begin{aligned}
J_\alpha g= J_\alpha\left[(I+D_\alpha)J_\alpha\right]^{-1}(I+D_\alpha)f=J_\alpha J_\alpha^{-1}(I+D_\alpha)^{-1}(I+D_\alpha)f=f.\end{aligned}$$ We also get $$\begin{aligned}
\|f\|_{\alpha,p} &=\|f\|_p+\|J_\alpha^{-1}f\|_p\\
&\leq C\|J_\alpha^{-1}f\|_p=C\|\left[(I+D_\alpha)J_\alpha\right]^{-1}(I+D_\alpha) f\|_p\\
&\leq C\|(I+D_\alpha)f\|_p.\end{aligned}$$
We can also characterize functions in $L^{\alpha,p}$ in terms of the Riesz potential $I_\alpha$ as follows. In [@GSV] and [@Ga], it is proven there exists $0<\tilde{\alpha}_0$ such that, for $\alpha<\tilde{\alpha}_0$, the operator $D_\alpha I_\alpha$ is inversible in $L^p$, $1<p<\infty$. Thus we obtain
For $\alpha>0$ satisfying $\alpha<\alpha_0{\ensuremath{\wedge}}\tilde{\alpha}_0$ and $1<p<\infty$, we get $$\begin{aligned}
f\in L^{\alpha,p}\text{ if and only if } f\in L^p \text{ and there exists $\gamma\in L^p$ with }f=I_\alpha \gamma.\end{aligned}$$
As another corolary, the following embeddings hold, which follow from the fact that $D_\alpha f\in L^p$ for $f$ smooth enough.
- If $0<\alpha<\alpha_0$ and $\epsilon>0$ satisfies $0<\alpha+\epsilon<1$, for $1<p<\infty$ we have $$\begin{aligned}
M^{\alpha+\epsilon,p}{\ensuremath{\hookrightarrow}}L^{\alpha,p}{\ensuremath{\hookrightarrow}}M^{\alpha,p}.\end{aligned}$$
- If $0<\alpha<\alpha_0$ and $0<\epsilon<\alpha$ satisfies $0<\alpha+\epsilon<1$, for $1<p<\infty$ we have $$\begin{aligned}
B^{\alpha+\epsilon}_{p,p}{\ensuremath{\hookrightarrow}}L^{\alpha,p}{\ensuremath{\hookrightarrow}}B^{\alpha-\epsilon}_{p,p}.\end{aligned}$$
- If $0<\alpha<\alpha_0$ and $\beta>0$ satisfies $\alpha<\beta<1$, for $1<p<\infty$ we have $$\begin{aligned}
L^{\beta,p}{\ensuremath{\hookrightarrow}}L^{\alpha,p}.\end{aligned}$$
As a final result, we show that in ${\ensuremath{\mathbb{R}}}^n$, for $\alpha<\alpha_0$, the space $L^{\alpha,p}$ coincides with the classical $\mathcal{L}^{\alpha,p}$.
Let $(S_t)_{t>0}$ be an approximation of the identity as constructed in the introduction, from a function $h$. Let $H(x)=h(|x|)$ and $H_t(x)=t^{-n}H(x/t)$. Then
- $T_tf(x)=\frac{1}{t^n}\int h\left(\frac{|x-y|}{t}\right)f(y)dy=\int H_t(x-y)f(y)dy=H_t*f(x)$;
- $T_t1\equiv\int H_t=\int H=c_H$ for every $t>0$ and $x\in {\ensuremath{\mathbb{R}}}^n$, then $\varphi\equiv\frac{1}{c_H}$ and $\psi\equiv 1$.
- $S_t f=\frac{1}{c_H^2}H_t*H_t*f=\int \left(\frac{1}{c_H^2}H_t*H_t\right)(x-y)f(y)dy$.
- $s(x,y,t)=\left(\frac{1}{c_H^2}H_t*H_t\right)(x-y)$.
We will see that $$\begin{aligned}
s(x,y,t)=\varphi_t(x-y)\end{aligned}$$ where $\varphi$ is radial. Observe $$\begin{aligned}
H_t*H_t(x) &=\frac{1}{t^{2n}}\int H\left(\frac{x-y}{t}\right)H\left(\frac{y}{t}\right)dy=\frac{1}{t^n}\int H\left(\frac{x}{t}-z\right)H(z)dz\\
&=\frac{1}{t^n}(H*H)(x/t)=\left(H*H\right)_t(x).\end{aligned}$$ Besides, if $\rho$ is a rotation, as $H$ is radial, we get $$\begin{aligned}
H*H(\rho x) &=\int H(\rho x-y)H(y)dy=\int H(\rho(x-\rho^{-1}y))H(\rho\rho^{-1}y)dy\\
&=\int H(x-\rho^{-1}y)H(\rho^{-1}y)dy=H*H(x).\end{aligned}$$
This way, if $\phi=\frac{1}{c_H^2}H*H$, we will have $$\begin{aligned}
\frac{1}{c_H^2}H_t*H_t=\phi_t.\end{aligned}$$
With this expression for $s$, we obtain $$\begin{aligned}
n_\alpha(x,y) &=\int_0^\infty \alpha t^{-\alpha}s(x,y,t)\frac{dt}{t}=\int_0^\infty \alpha t^{-\alpha}\frac{1}{t^n}\phi\left(\frac{x-y}{t}\right)\frac{dt}{t}\\
&=\frac{1}{|x-y|^{n+\alpha}}\int_0^\infty \alpha u^{n+\alpha}\phi(ue_1)\frac{du}{u}=\frac{c_{n,\alpha,\phi}}{|x-y|^{n+\alpha}}\end{aligned}$$ and the last integral converges because $\phi$ is bounded and compactly supported.
Now, recall that for $0<\alpha<2$, $$\begin{aligned}
\mathscr{D}_\alpha f(x)=\text{p.v. } c_{\alpha,n}\int\frac{f(y)-f(x)}{|x-y|^{n+\alpha}}dy\end{aligned}$$ and that for those values of $\alpha$, $$\begin{aligned}
f\in \mathcal{L}^{\alpha,p}\text{ if and only if } f,\mathscr{D}_\alpha f\in L^p.\end{aligned}$$
From the previous result, we get $$\begin{aligned}
D_\alpha f=C_{n,\alpha,h} \mathscr{D}_\alpha f,\end{aligned}$$ and thus $$\begin{aligned}
f\in \mathcal{L}^{\alpha,p}\text{ if and only if } f,D_\alpha f\in L^p.\end{aligned}$$
In conclusion, for $0<\alpha<\alpha_0$, by the characterization theorem the spaces $L^{\alpha,p}({\ensuremath{\mathbb{R}}}^n)$ are independent from the choice of $h$ in the aproximation of the identity $(S_t)$, and they coincide with the classical space $$\begin{aligned}
L^{\alpha,p}=\mathcal{L}^{\alpha,p}.\end{aligned}$$
Acknowledgements {#acknowledgements .unnumbered}
================
The author is infinitely indebted to his advisors Eleonor ‘Pola’ Harboure and Hugo Aimar for their guidance and support throughout the development of his doctoral thesis and its resulting papers.
[AAA]{}
R.R. Coifman, *Multiresolution analysis in non homogeneous media*. In *Wavelets, Time Frequency Methods and Phase Space*, Proceedings of the International Conference, Marseille, December 14-18, 1987 (1990 2nd ed.), J.M. Combes, A. Grossman, Ph. Tchamitchian, eds., Springer Verlag, 1990. G. David, J.L. Journé, S. Semmes, *Opérateurs de Calderón-Zygmund, conctions para-accrétives et interpolation*, Rev. Mat. Iberoamericana 1 (1985), 1-56. A. Gatto, *On fractional calculus associated to doubling and non-doubling measures*, dedicated to S. Vági, Mathematics Subject Classification, 2000. A. Gatto , C. Segovia, S. Vági, *On fractional differentiation and integration on spaces of homogeneous type*, Revista Matemática Iberoamericana 12, (1996), 111-145. Gogatishvili, Amiran; Koskela, Pekka; Shanmugalingam, Nageswari. *Interpolation properties of Besov spaces defined on metric spaces*. (English summary) Math. Nachr. 283 (2010), no. 2, 215-231. Grafakos, Loukas. *Modern Fourier analysis*. Third edition. Graduate Texts in Mathematics, 250. Springer, New York, 2014. xvi+624 pp. ISBN: 978-1-4939-1229-2; 978-1-4939-1230-8 P. Hajasz, *Sobolev spaces on an arbitrary metric space*, Potential Anal. 5 (1995), 403-415. Y.S. Han, E.T. Sawyer, *Littlewood Paley theory on spaces of homogeneous type and the classical function spaces*, Mem. Amer. Math. Soc. 110 (530) (1994). S. Hartzstein, *Acotación de operadores de Calderón-Zygmund en espacios de Triebel-Lizorkin y de Besov generalizados sobre espacios de tipo homogéneo*. Tesis para la obtención del Grado Académico de Doctor en Matemática. Advisor: B. Viviani. Universidad Nacional del Litoral, Facultad de Ingeniería Química, 2000. Argentina. S. Hartzstein, B. Viviani, *Homeomorphisms acting on Besov and Triebel-Lizorkin spaces of Local Regularity $\psi(t)$*, Collectanea Mathematica; (2005) vol. 56 p. 27 - 45. R.A. Macías, C. Segovia, *Lipschitz functions on spaces of homogeneous type*, Advances in Math 33 (1979), 257-270. Stein, E. *Singular integrals and differentiability properties of functions*, Princeton University Press (1971). 1984.
*E-mail address:* `[email protected]`
[^1]: The author was supported by Consejo Nacional de Investigaciones Científicas y Técnicas, Agencia Nacional de Promoción Científica y Tecnológica and Universidad Nacional del Litoral.Keywords and phrases: Bessel potential, Ahlfors spaces, fractional derivative, Sobolev spaces 2010 Mathematics Subject Classification: Primary 43A85.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Non-perturbative phenomena in four-wave mixing spectra of semiconductors are studied using the exact solution of a widely used phenomenological non-linear equation of motion of the exciton polarization. It is shown that Coulomb interaction, included in the nonlinearity, leads to two characteristic effects, which are essentially of dynamical origin, — a split of the exciton peak and a non-monotonous dependence of the response at the exciton frequency on the magnitude of the external field. Relations between the spectral features and the parameters of the system is obtained. It is found that the transition from perturbative to non-perturbative regimes is controlled by parameters inversely proportional to the decay rate. It implies that the condition of low excitation density does not necessarily warrant applicability of the perturbational approach.'
author:
- Mikhail Erementchouk
- 'Michael N. Leuenberger'
- 'L. J. Sham'
title: 'Non-perturbative phenomena in semiconductor four-wave mixing spectra'
---
Introduction
============
One of the clearest manifestations of the many-body effects in semiconductors is the the phenomenon of the non-linear optical response. The Pauli blocking and the Coulomb interaction between the quasiparticles lead to the dependence of the polarization dynamics on its spatial distribution.[@BINDER:1995; @OSTREICH:1994; @OSTREICH:1998; @CHERNYAK:1998; @CIUTI:2000; @SAVASTA:2001; @TAKAYAMA:2002] When a semiconductor quantum well is excited by two successive pulses, they produce the signal in directions which are prohibited in the linear regime while still conserve momentum in four-wave and multi-wave mixing. The dynamical origin[@WEGENER:1990] of the formation of the four-wave mixing signal is naturally incorporated into the description in terms of the exciton modes characterized by the frequency $\omega_0$ and the in-plane wave vector $\mathbf{k}$. From this perspective the effects of the Pauli blocking and of the Coulomb interaction are clearly different. The Pauli blocking reduces *locally* the intensity of the field-matter interaction according to the magnitude of the local polarization. As a result, the excitation field with particular value of the wave-vector $\mathbf{q}$ becomes coupled to the exciton modes with different $\mathbf{k}\ne \mathbf{q}$. The Coulomb interaction between the excitons leads to the direct coupling between the exciton modes. In particular, when two modes with $\mathbf{k}_1$ and $\mathbf{k}_2$ are excited, the relevant coupled modes are characterized by $\mathbf{k}^{(n)} = \mathbf{k}_2 + n\Delta\mathbf{k}$, where $-\infty <
n < \infty$ is an integer number and $\Delta\mathbf{k} = \mathbf{k}_1 -
\mathbf{k}_2$. Although initially all energy is concentrated in the modes $\mathbf{k}^{(0)}$ and $\mathbf{k}^{(1)}$ during the evolution the energy is redistributed among the coupled modes. In particular it leads to the formation of the four-wave mixing (FWM) signals, which correspond to $n=2$ and $n = -1$.
It follows from this picture that the redistribution of the energy between the modes characterized by different $\mathbf{k}$ should not be the only manifestation of the mode coupling. This coupling should also lead to the modification of the frequencies of the exciton modes. Indeed, without the interaction and neglecting the dispersion of the exciton modes one has a many-fold degeneracy at the exciton resonance frequency $\omega_0$. The interaction between the modes should lift the degeneracy and, if the excitation is sufficiently strong, make the split of the exciton peak visible in the FWM spectrum.
The nonlinear dynamics of the exciton polarization
==================================================
The effect of the modification of the frequencies of the exciton modes caused by the coupling is a nonperturbative effect.[@SHACKLETTE:2002; @WUHR:2004] In order to study the problem we present the dynamics of the exciton polarization as the exact solution to the phenomenological nonlinear equation[@SCHAFER:1996] $$\label{eq:eq_of_motion}
\begin{split}
\dot{P}(t; {\mathbf{r}}) = -(i\omega_0 + \gamma)& P(t; {\mathbf{r}}) - i \beta |P(t; {\mathbf{r}})|^2P(t; {\mathbf{r}}) \\
-i & E(t; {\mathbf{r}})\left[1 - |P(t; {\mathbf{r}})|^2/P_{sat}^2\right],
\end{split}$$ where $\omega_0$ is the detuning, i.e. the difference between the frequency of the external field and the exciton frequency in the stationary frame, $\gamma$ is the phenomenological decay rate, $E(t; {\mathbf{r}})$ is the envelope magnitude of the external field, $P_{sat}^2$ is the exciton saturation density and, finally, $\beta = \beta' - i \beta''$, with $\beta', \beta''
\geq 0$, is a phenomenological parameter quantifying the interaction between the excitons. The positive real and negative imaginary parts of this parameter constitute the excitation induced shift (EIS) and the excitation induced decay (EID), respectively. This phenomenological equation has provided valuable interpretations for nonlinear measurements.[@SCHAFER:1996; @KNER:1997; @FU:1997]. The terms may be viewed as the short-time limit of the memory function,[@OSTREICH:1998] which is exact to the third order in the exciting electric field.[@AXT:1994] The nonpertubative solutions have been considered by Refs. and many papers referenced in Ref. .
![The semiconductor is excited by two short pulses separated by the delay time $\tau = t_-''-t_+'$.[]{data-label="fig:excitation_setup"}](Excitation_setup.eps "fig:"){width="3"}\
We consider the excitation of the semiconductor by two short pulses (see Fig. \[fig:excitation\_setup\]) acting on the system during the intervals $t_-' < t < t_+'$ and $t_-'' < t < t_+''$, respectively. Between the pulses, $t_+' < t < t_-''$ and after the second pulse $t > t_+''$ the dynamics of the exciton polarization is free and satisfies Eq. (\[eq:eq\_of\_motion\]) with $E=0$. We study the free dynamics polarization $P_0(t)$ before calculating the polarization $P_{1,2}(t)$ of the driven dynamics. Then we are going to solve for the polarization in the following order: $P_1(t)$ between $t_-' < t < t_+'$, $P_{10}(t)$ between $t_+' < t < t_-''$, $P_2(t)$ between $t_-'' < t < t_+''$, and at the end the final solution $P_{20}(t)$ for $t
> t_+''$.
The free dynamics \[$E(t)=0$\] can be solved exactly by noting that Eq. in this regime is reduced to the simple form $$\label{eq:effective_linear}
\dot{P_0}(t;{\mathbf{r}}) = -\left[i\omega_0 + \gamma + i\Omega(t;{\mathbf{r}})\right]P_0(t;{\mathbf{r}}),$$ where $\Omega(t;{\mathbf{r}}) = \beta|P_0(t; {\mathbf{r}})|^2$, and by observing that Eq. yields an equation for the magnitude of the polarization $$\label{eq:p_amp_equation}
\frac{d}{dt}|P_0(t; {\mathbf{r}})|^2 = -2[\gamma + \beta'' |P_0(t; {\mathbf{r}})|^2]|P_0(t;
{\mathbf{r}})|^2$$ with the solution $$\label{eq:p_amp_solution}
|P_0(t; {\mathbf{r}})| = |P_0(0; {\mathbf{r}})| e^{-\gamma t} A(t; |P_0(0,\mathbf{r})|),$$ where $$\label{eq:amplitude_factor}
A^2(t; |P_0(0,\mathbf{r})|) = \left[1+ \frac{\beta''|P_0(0; {\mathbf{r}})|^{2}}{\gamma}
\left(1 - e^{-2\gamma t}\right)\right]^{-1}.$$ The free propagator may be expressed in terms of the amplitude modulation $A(t;\mathbf{r})$ and the phase modulation $\varphi(t;\mathbf{r})$, $$P_0(t;\mathbf{r})/ P_0(0;\mathbf{r}) =: \theta\left(t; |P_0(0,\mathbf{r})|\right)
= A(t;|P_0(0,\mathbf{r})|) e^{-i\omega_0 t -\gamma t + i \varphi(t;\mathbf{r})},$$ where $$\label{eq:phase_shift}
\varphi(t;\mathbf{r}) = \frac{\beta'}{\beta''} \ln A\left(t;|P_0(0,\mathbf{r})|\right).$$ The amplitude term shows the nonpertubative effect of EID. It is interesting to note that EID does not lead to a mere modification of the decay rate, $\gamma$. Instead the amplitude modulation $A(t;\mathbf{r})$ decreases to a fraction of the initial amplitude at twice the linear rate, i.e., $2\gamma$. The phase modulation shows an oscillation dependent on the nonlinear quality factor $\beta'/\beta''$. This is related to the Goldstone mode in the excitons studied in Ref. .
The initial conditions for Eq. (\[eq:effective\_linear\]) are determined by the polarization distribution right after the external field is switched off. We find the immediate response of the system assuming that the duration of the excitation pulses is much shorter than the typical dynamical time scales determined by detuning and the decay rate.
We turn now to the driven ($E(t)\ne 0$) time evolution of the polarization, i.e. we consider the time interval $t_- < t < t_+$ where the particular excitation pulse does not vanish and factor out the term $\exp(i\mathbf{k}\cdot{\mathbf{r}})$ so that we can consider the excitation pulse to be spatially homogeneous. Neglecting the contribution to the phase $\sim
\int_{t_-}^{t_+}\left[i\omega_0 + \gamma +
i\Omega(t;{\mathbf{r}})\right]P_{1,2}(t;{\mathbf{r}}) dt$ we solve the dynamical equation and find the relation between the polarization at the instances $t_-$ and $t_+$
$$\label{eq:immediate response}
P_{1,2}(t_+) = {P'}_{1,2}(t_-) - i
\sqrt{P_{sat}^2-{P'}_{1,2}^2(t_-)}\tanh\left\{\sqrt{P_{sat}^2-{P'}_{1,2}^2(t_-)}\frac{{\epsilon}}{P_{sat}^2}
-\mathrm{atanh}\left[\frac{{P''}_{1,2}(t_-)}{\sqrt{P_{sat}^2-{P'}_{1,2}^2(t_-)}}\right]\right\},$$
where ${P'}_{1,2}(t_-) = \mathrm{Re}[P_{1,2}(t_-)]$ and ${P''}_{1,2}(t_-) =
\mathrm{Im}[P_{1,2}(t_-)]$, and we have introduced the area of the exciting pulse $\epsilon = \int dt E(t)$.
{width="6in"}
Using the solutions for the free and driven polarization in Eqs. (\[eq:effective\_linear\]) and (\[eq:immediate response\]), we solve the time evolution of the polarization in all the four regions $t_-'
< t < t_+'$, $t_+' < t < t_-''$, $t_-'' < t < t_+''$, and $t > t_+''$. Assuming that the system initially is in the ground state $P(t_-')\equiv0$ we find the exciton polarization created by the first pulse $$\label{eq:first_pulse_response}
P_1(t_+';{\mathbf{r}}) = -i e^{i \mathbf{k}_1\cdot {\mathbf{r}}}P_1(t_+'),$$ where $$\label{eq:amplitude_first_response}
P_1(t_+') = P_{sat} \tanh\left(\frac{{\epsilon}}{P_{sat}}\right).$$ It follows from this equation that the saturation effect for the first pulse reduces to a simple (although nonlinear) renormalization of the signal area. Using Eq. (\[eq:first\_pulse\_response\]) in Eq. (\[eq:p\_amp\_solution\]) we can see that the effective frequency $\Omega(t; {\mathbf{r}})$ determining the dynamics of the polarization is constant across the sample. As a result there is no coupling between the exciton modes characterized by different $\mathbf{k}$.
Denoting the delay time, the time separation between the pulses, by $\tau=t_-''-t_+'$ we obtain the polarization right before the arrival of the second pulse $$\label{eq:second_initial_condition}
P_{10}(t_-''; {\mathbf{r}}) = P_1(t_+';{\mathbf{r}}) \theta_{10}(\tau),$$ where $\theta_{10}(\tau) = \theta(\tau; |P_1(t_+')|)$. The spatial distribution of the polarization $P_{10}(t_-''; {\mathbf{r}})$ plays the role of the initial condition for the immediate response with respect to the second pulse according to Eq. (\[eq:immediate response\]). For the analysis of the time evolution of the polarization during the second pulse it is convenient to factor $\exp(i\mathbf{k}_2\cdot {\mathbf{r}})$ out of $P_2(t;{\mathbf{r}})$ introducing $$\label{eq:factorized_P}
\widetilde{P}_2(t; \Delta \mathbf{k}\cdot {\mathbf{r}}) = e^{-i\mathbf{k}_2\cdot
{\mathbf{r}}} P_2(t; {\mathbf{r}})$$ with $\Delta\mathbf{k} = \mathbf{k}_1 - \mathbf{k}_2$. The reduced distribution $\widetilde{P}_2(t; \Delta \mathbf{k}\cdot {\mathbf{r}})$ satisfies Eq. (\[eq:eq\_of\_motion\]) with the modified spatial profile of the external field $\widetilde{E}(t) = e^{-i\mathbf{k}_2\cdot {\mathbf{r}}}E(t;{\mathbf{r}})$. The form of the initial conditions in Eq. (\[eq:immediate response\]) changes according to $$\label{eq:initial_conditions_new}
\widetilde{P}_{10}(t_-''; \Delta \mathbf{k}\cdot {\mathbf{r}}) = -i\theta_{10}(\tau)P_1(t_+')
e^{i\Delta\mathbf{k}\cdot {\mathbf{r}}}.$$
It follows from Eqs. (\[eq:immediate response\]) and (\[eq:initial\_conditions\_new\]) that for the second pulse the role of the saturation effect is two-fold. It modifies the pulse area and excites all modes $\mathbf{k}^{(n)}$ rather than just a single mode as we had for the first pulse. We present the polarization as a superposition of the multi-wave mixing modes $$\label{eq:superposition_modes}
\widetilde{P}_2(t; \kappa) = \sum_n {P}_2^{(n)}(t)
e^{in \kappa},$$ where $\kappa = \Delta\mathbf{k}\cdot {\mathbf{r}}$ and $P_2^{(n)}(t)$ are the amplitudes of the multi-wave mixing polarizations. Substituting this representation into Eq. (\[eq:effective\_linear\]) we can see that in terms of the multi-wave mixing modes the free dynamics of the polarization can be presented as the dynamics of a system with an infinite number of degrees of freedom coupled to each other due to the nonlinearity $$\label{eq:binding_model}
\frac{d}{dt}{P_2}^{(n)}(t) = -(i\omega_0 + \gamma)P_2^{(n)}(t) -
i\sum_{m}\Omega_{n-m}(t) P_2^{(m)}(t),$$ where $ \Omega_{n}(t) = \beta(2\pi)^{-1}\int_{-\pi}^\pi
|\widetilde{P}_2(t; \kappa)|^2 e^{-i\kappa n}d\kappa$. The initial conditions for Eqs. (\[eq:binding\_model\]) are constituted by the immediate response to the second pulse thanks to the saturation effect. This picture clearly illustrates the difference between the effect of the Pauli blocking and the Coulomb interaction on formation of the multi-wave mixing response.
Using the solution of the equation of motion for free polarization dynamics we find $$\label{eq:final_solution}
P_{2}^{(n)}(t) = \frac{1}{2\pi} \int_{-\pi}^\pi \widetilde{P}_2(t_+''; \kappa)
\theta\left(t; |\widetilde{P}_2(t_+''; \kappa)|\right)e^{-i\kappa n} d\kappa.$$ This equation and Eqs. (\[eq:immediate response\]) and (\[eq:initial\_conditions\_new\]) give the exact evolution of the exciton polarization in the limit of short excitation pulses in the two-pulses scheme. Formally, one can obtain from Eq. (\[eq:final\_solution\]) the spectrum $P_{2}^{(n)}(\omega)$ using the Fourier transform of $\theta(\tau; |\widetilde{P}_2(t_+''; \kappa)|)$ with respect to time $$\label{eq:pi_transform}
\theta(\omega; |\widetilde{P}_2(t_+''; \kappa)|) = \frac{i}{\gamma(w + i/2)}\,
{}_2 F_1\left(1; \frac{1}{2}+iX; \frac{3}{2} - iw;
-\frac{\beta''|\widetilde{P}_2(t_+''; \kappa)|^2}{\gamma}\right),$$ where $w=(\omega - \omega_0)/2\gamma$, $X=\beta'/2\beta''$ and ${}_2 F_1$ is the hypergeometric function. Technically, however, because the second argument of this function is a complex number it may be more efficient to calculate the spectrum using the time series $P_{2}^{(n)}(t)$.
The spectrum corresponding to the four-wave mixing directions $2\mathbf{k}_2 - \mathbf{k}_1$ and $2\mathbf{k}_1 - \mathbf{k}_2$ is obtained from Eq. (\[eq:final\_solution\]) taking $n=-1$ and $n=2$, respectively, and for $n=2$ is shown in Fig. \[fig:typical\_spectrum\] as a function of the pulse area. In order to estimate the relation between the decay rate and the non-linear parameter we have used the expressions following from the microscopic consideration[@TAKAYAMA:2002] $\beta'
\approx 1.52 a_B^2 E_b $ and $P_{sat}^2 = 7/4\pi a_B^2$, where $E_b$ is the exciton Rydberg and $a_B$ is the exciton Bohr radius.
The spectrum has two typical features. The first one consists of branches detached from the exciton frequency with increasing pulse area. These branches for sufficiently high amplitudes of the excitation field may manifest themselves on the spectrum in the form of multiple resonances. The second interesting feature is the oscillatory character of the field dependence of $|P_{2}^{(2)}(\omega_0;\epsilon)|^2$, that is the FWM response at the exciton frequency.
The FWM spectrum in the limit of negligible EID
-----------------------------------------------
We start the discussion of these features from the simplest case $\beta''=0$ (vanishing EID) and high $P_{sat}$ (low saturation regime). The first assumption simplifies the effect of the initial conditions on the polarization dynamics while the second simplifies the relation between the excitation field and the polarization of the immediate response, so that $$\label{eq:immediate_response_low_saturation}
\widetilde{P}_2(t_+''; \kappa) = -i\left(|{\epsilon}_1\theta_{10}(\tau)|e^{-i\kappa+i\phi_1}
+ |{\epsilon}_2|e^{i\phi_2}\right),$$ where $\phi_1 = \arg({\epsilon}_1\theta_{10}(\tau))$ and $\phi_2 =
\arg({\epsilon}_2)$.
It is interesting to note that according to Eqs. (\[eq:binding\_model\]) and (\[eq:p\_amp\_solution\]) in this approximation since only $\Omega_{\pm1}(t)$ differ from zero the dynamics of the polarization is described by the equation of motion for a 1D tight-binding model with the time-dependent coupling between neighboring sites $\propto
\beta|\epsilon_2\epsilon_1 \theta_{10}(\tau)| e^{-2\gamma t}$. The relation between the amplitudes of the excitation at different sites gives the relation between the amplitudes of the signals corresponding to multi-wave mixing. Initially the excitation is localized on sites $n=0,1$ and with time it propagates along the chain giving
$$\label{eq:solution_simplest}
\begin{split}
P_{2}^{(n)}(t) = -e^{in(\phi_2 - \phi_1 - \pi/2)+i\phi_2}
\exp\left\{-i\omega_0 t - \gamma t - \frac{i\beta I}{2\gamma}\left(1 - e^{-2\gamma t}\right)\right\}
\times
\\
\left\{|\epsilon_1 \theta_{10}(\tau)| J_{n-1}\left[\eta(1 - e^{-2\gamma t})\right]
+ i|\epsilon_2| J_{n}\left[\eta(1 - e^{-2\gamma t})\right]
\right\},
\end{split}$$
where $J_n$ are the Bessel functions of the first kind, $I = |\epsilon_1
\theta_{10}(\tau)|^2 + |\epsilon_2|^2$, and $\eta = \beta
|\epsilon_2\epsilon_1 \theta_{10}(\tau)|/\gamma$. Deriving Eq. (\[eq:solution\_simplest\]) we have used the Jacobi-Anger expansion[@Arfken_Weber] $\exp(i z \cos \kappa) =
\sum_{n=-\infty}^\infty i^n J_n(z) e^{in\kappa}$.
The spectrum of the four-wave mixing signal corresponding to $n=2$ is shown in Fig. \[fig:spectrum\_simple\]. Shortly after the excitation, for $t \ll 1/2\gamma$, the amplitude of the multi-wave mixing signal drops exponentially with the order of mixing, $\sim (\eta \gamma t/2)^n$. The exponential drop holds asymptotically in time if $\eta < 1$. This result agrees with the perturbational approach. The situation, however, drastically changes if the parameters of the system are such that $\eta
\gg1$. In this case, the intensity of the multi-wave mixing signal becomes independent of its order starting time $t\gtrsim-\ln(1-\eta^{-1})/2\gamma
\approx 1/2\beta |\epsilon_2\epsilon_1 \theta_{10}(\tau)|$. This consideration suggests naturally to identify $\eta = 1$ as a critical value that separates perturbative and non-perturbative regimes.
![The form of the four-wave mixing spectrum in low saturation regime in the case of negligible EID ($\beta''=0$) and $\beta'/\gamma = 2.5\cdot 10^2$. The solid line shows the position of the resonant frequency determined by $\omega = \omega_0 + \beta I + \eta\gamma$. []{data-label="fig:spectrum_simple"}](simple_model.eps){width="3in"}
The detailed form of the spectrum is essentially determined by the fact that the effective coupling between the modes vanishes with time. Right after arrival of the second pulse the polarization oscillates with the frequency detuned from the exciton frequency by the value $\sim \beta I +
\eta\gamma$, where the second term estimates the contribution of the Bessel functions. This detuned frequency qualitatively describes the dependence of the frequencies of the detached resonances on the signal area (see the bold solid line in Fig. \[fig:spectrum\_simple\]). In particular it shows that the resonances in the spectrum widen with increasing nonlinear parameter $\beta$. In the opposite limit, $t \gg
1/\gamma$, the coupling between the modes vanishes and, as a result, the exciton polarization oscillates with the non-modified exciton frequency $\omega_0$. These oscillations give rise to the resonant behavior at the exciton frequency, for which dependence on the pulse area is determined by the asymptotic value of the Bessel functions $J(\eta)$. Therefore, the response at the exciton frequency essentially depends on whether the system is in perturbative or non-perturbative regime. We illustrate the difference between these regimes considering the effect of the positive and negative delay time. The expression for the negative delay time can be obtained from Eq. (\[eq:solution\_simplest\]) by considering the signal in the “conjugate" direction $\bar{n} = 1-n$ and exchanging $\epsilon_1
\leftrightarrow \epsilon_2$.
In the perturbative regime, $\eta \ll1$, we obtain $$\label{eq:perturb_positive_negative}
\begin{split}
\left|P_{2}^{(n)}(\omega_0; \tau>0)\right|^2 \sim 4\eta_0^{2n-2}
\left|\epsilon_1 n \theta_{10}^n(\tau)\right|^2, \\
\left|P_{2}^{(n)}(\omega_0; \tau<0)\right|^2 \sim 4\eta_0^{2n-2}
\left|\epsilon_1 n \theta_{10}^{n-1}(-\tau)\right|^2,
\end{split}$$ where $\eta_0 = \beta |\epsilon_1\epsilon_2|/\gamma$. In this regime $|\theta_{10}(\tau)| \approx \exp(-\gamma \tau)$, thus, the positive delay signal decays with time constant $2n\gamma$ while for the negative delay it decays more slowly with the constant $2(n-1)\gamma$. For the case of FWM signal this result agrees with the perturbational calculations.[@WEGENER:1990]
In the non-perturbative regime the positive and negative delay signals are determined by the oscillating asymptotics of the Bessel functions[@Gradshteyn] $$\label{eq:nonpert_positive_negative}
\begin{split}
\left|P_{2}^{(n)}(\omega_0)\right|^2 \sim & \frac{1}{\pi \eta}
\left[\left|\epsilon_1 \theta_{10}(\tau)\right|^2 + \left|\epsilon_2\right|^2 \right. \\
&\left.\mp (-1)^n
\left(\left|\epsilon_1 \theta_{10}(\tau)\right|^2 - \left|\epsilon_2\right|^2\right)
\sin(2\eta)
\right],
\end{split}$$ where “$-$" (“$+$") sign corresponds to the positive (negative) delay. Writing Eq. (\[eq:nonpert\_positive\_negative\]) we have neglected the oscillating term $\propto \cos(2\eta)$ vanishing as $(\left|\epsilon_1
\theta_{10}(\tau)\right|^2 + \left|\epsilon_2\right|^2)/\eta^2$ with increasing signal area. Thus, for both positive and negative delays the response at the exciton frequency saturates at the oscillations with the period $\eta_T = 2\pi$. The strong asymmetry between these cases specific for the perturbative regime does not hold any longer and the only difference is the phase of the oscillations. It should be noted that in Ref. the saturation of the FWM response was attributed to the renormalization of the pulse area by the EID and EIS. The present consideration, however, suggests that the origin of the non-trivial dependence of the response on the pulse area is the redistribution of the excitation over multi-wave mixing modes. Considering the identity $\sum_n J_n^2(z) = 1$ for the limits $z\ll1$ and $z\gg1$ it can be seen that such redistribution is especially effective in the nonperturbative regime resulting in essential suppression of the FWM response.
It should be emphasized that the mechanism of the oscillatory dependence of the response at the exciton frequency is different from the Rabi oscillations,[@BINDER:1990; @OSTREICH:1993]which would correspond to the non-monotonous dependence of the immediate response on the excitation field. In the case under consideration the oscillations are the result of the free dynamics of the exciton polarization when the external field is turned off. The physics of the Rabi oscillations and of the oscillations of $|P^{(n)}_{2}(\omega;
\eta)|^2$, of course, are essentially the same. As noted above the dynamics of the polarization in the case under consideration appears analogous to a 1D tight-binding model with vanishing coupling between the neighboring sites. In quantum mechanical terms it can be described as a multiple level system, where the levels correspond to the exciton modes, with time dependent field $V_{ij}(t)$, which couples different levels. Depending on the “area" of the off-diagonal elements, $\int dt
V_{ij}(t)$, one has the oscillations of the final populations of the different levels. Translated to the language of the multi-mixing signals $P_{2}^{(n)}$ this result implies the oscillations of $|P_{2}^{(n)}(\omega_0)|^2$ since asymptotically, as has been noted, one has the dynamics determined by the non-perturbed exciton frequency.
The effect of the excitation induced decay
------------------------------------------
In order to study the effect of the EID on the spectrum (compare Figs. \[fig:typical\_spectrum\] and \[fig:spectrum\_simple\]) we use Eq. (\[eq:pi\_transform\]) assuming that $|{\epsilon}_1
\theta_{10}(\tau)| = |{\epsilon}_2| = {\epsilon}$. Considering the asymptotic values of the hypergeometric function[@Abramowitz] in the limit $\eta''=2\beta''{\epsilon}^2/\gamma \gg 1$ we can approximately present the spectrum in the form $$\label{eq:factorized_spectrum}
P_{2}^{(n)}(\omega) \approx \frac{\epsilon}{\sqrt{\eta''}}\left[
C(\omega; X)A_2 e^{-iX \ln \eta''}+\frac{1}{\eta''}\,
\frac{A_1}{iX-1/2}\right],$$ where $$\label{eq:C_def}
C(\omega; X) = \frac{\Gamma(1/2-iw)\Gamma(1/2-iX)}{\Gamma(1-iw-iX)}$$ and $A_1=A_n(1)$, $A_2=A_n(1/2)$ with $A_n(p)$ depending on $X$ only $$\label{eq:def_Anp}
\begin{split}
A_n(p) = \frac 1{2\pi}\int_{-\pi}^\pi & d\kappa (1+e^{-i\kappa})e^{-in\kappa} \\
& \times \exp\left\lbrace -(iX+p)\ln\left[2\cos^2\left(\frac{\kappa}{2}\right)\right]\right\rbrace.
\end{split}$$
Similarly to the case of negligible EID, the response oscillates and reaches the saturation in the high excitation limit $\eta''\gg1$. At the exciton frequency the magnitude of the signal is $$\label{eq:EID_limit}
\left|P_{2}^{(n)}(\omega_0)\right|^2 \sim \frac{\pi\epsilon^2}{X\eta''} \left|A_2 \right|^2
\tanh(\pi X).$$ This saturation also is of the dynamical origin since we work in the regime of weak saturation due to the Pauli blocking ($\epsilon/P_{sat}^2
\ll 1$). The significant difference with the previously considered situation is that now the minima of the spectrum as a function of the pulse area are not equally spaced as the oscillating part has the form $\propto \cos(X \ln \eta'')$. Thus, with the decreasing ratio between the real and imaginary parts of the non-linear parameter the crossover from linear to logarithmic scale occurs.
Finally, we discuss the effect of the saturation parameter. Qualitatively, this effect can be understood as follows. The non-monotonous behavior of the semiconductor response on the pulse area studied above is supported by unrestricted increase of the polarization of the immediate response \[see Eq. (\[eq:immediate\_response\_low\_saturation\])\]. However, the saturation effect renormalizes the pulse area so that the magnitude of the polarization can not exceed $P_{sat}$. For example, from the perspective of the discussion of the effect of EID this means that Eq. (\[eq:factorized\_spectrum\]) remains valid only if two restrictions are met $2\beta''{\epsilon}^2/\gamma \gg 1$ and ${\epsilon} < P_{sat}$. This imposes the restriction of the decay rate to be sufficiently small $\gamma \ll \beta'' P_{sat}^2$.
Conclusion
==========
We have studied non-perturbative effects in four-wave mixing spectra of semiconductors. These effects are analyzed using the exact solution of the non-linear equation of motion of the exciton polarization taking into account excitation induced shift (EIS), excitation induced decay (EID) and the saturation effect phenomenologically. We found that the interaction between the excitons accounted by EIS leads to two specific spectral features — a split of the exciton peak and a non-monotonous dependence of the response at the exciton frequency $\omega_0$ on the magnitude of the external field. The important characteristic of the splitting is that new spectral features should appear at frequencies higher than $\omega_0$. This allows one to make a distinction between the effect of interaction of the exciton modes and the manifestation of bound biexciton states, which should modify the spectrum at frequencies lower than $\omega_0$.
We would like to emphasize that these effects do not appear in any order of the perturbational ($\chi^{(n)}$) approach. It can be shown that the appearance of additional spectral features can be traced as a divergence of the perturbational series. It should be stressed out that the crossover from the perturbative to non-perturbative regimes is governed by parameters that essentially depend on the decay rate (e.g. $\beta
|{\epsilon}_1\theta_{10}(\tau){\epsilon}_2|/\gamma \sim 1$ in the case $\beta''=0$). This means that the condition of low excitation itself does not necessarily warrant the validity of the perturbation theory. As an ultimate example one can consider the model with $\gamma = 0$ when the spectrum (for $\tau =0$) has the form $P(\omega)\propto 1/\sqrt{(\omega -
\omega_0 - \beta I)^2 - \beta^2 I^2}$ with the exciton peak being splitted for arbitrary low excitations.
We would like to acknowledge the support for this work from the National Science Foundation under grant number ECCS-0725514 and through the DARPA/MTO Young Faculty Award under Grant No. HR0011-08-1-0059.
[23]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, ****, ().
, , , , , , ****, ().
, , , , , ****, ().
, ****, ().
, ** (, , ), ed.
, ** (, , ).
, ****, ().
, , , , , ****, ().
, ****, ().
, eds., ** (, , ).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We consider a single server queueing system with two classes of jobs: [*eager*]{} jobs with small sizes that require service to begin almost immediately upon arrival, and [*tolerant*]{} jobs with larger sizes that can wait for service. While blocking probability is the relevant performance metric for the eager class, the tolerant class seeks to minimize its mean sojourn time. In this paper, we discuss the performance of each class under dynamic scheduling policies, where the scheduling of both classes depends on the instantaneous state of the system. This analysis is carried out under a certain fluid limit, where the arrival rate and service rate of the eager class are scaled to infinity, holding the offered load constant. Our performance characterizations reveal a [*(dynamic) pseudo-conservation law*]{} that ties the performance of both the classes to the standalone blocking probabilities of the eager class. Further, the performance is robust to other specifics of the scheduling policies. We also characterize the Pareto frontier of the achievable region of performance vectors under the same fluid limit, and identify a (two-parameter) class of [*Pareto-complete*]{} scheduling policies.'
author:
- 'Kiran Chaudhary[^1]'
- Veeraruna Kavitha
- 'Jayakrishnan Nair[^2]'
bibliography:
- 'refs.bib'
title: 'Dynamic scheduling in a partially fluid, partially lossy queueing system'
---
[^1]: Kiran Chaudhary and Veeraruna Kavitha are with the Industrial Engineering and Operations Research (IEOR) Department at IIT Bombay.
[^2]: Jayakrishnan Nair is with the Electrical Engineering Department at IIT Bombay.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We propose a Bayesian regression method that accounts for multi-way interactions of arbitrary orders among the predictor variables. Our model makes use of a factorization mechanism for representing the regression coefficients of interactions among the predictors, while the interaction selection is guided by a prior distribution on random hypergraphs, a construction which generalizes the Finite Feature Model. We present a posterior inference algorithm based on Gibbs sampling, and establish posterior consistency of our regression model. Our method is evaluated with extensive experiments on simulated data and demonstrated to be able to identify meaningful interactions in applications in genetics and retail demand forecasting.[^1]'
author:
- |
Mikhail Yurochkin\
Department of Statistics\
University of Michigan\
`[email protected]` XuanLong Nguyen\
Department of Statistics\
University of Michigan\
`[email protected]` Nikolaos Vasiloglou\
LogicBlox\
`[email protected]`\
bibliography:
- 'MY\_ref.bib'
title: 'Multi-way Interacting Regression via Factorization Machines'
---
Introduction
============
A fundamental challenge in supervised learning, particularly in regression, is the need for learning functions which produce accurate prediction of the response, while retaining the explanatory power for the role of the predictor variables in the model. The standard linear regression method is favored for the latter requirement, but it fails the former when there are complex interactions among the predictor variables in determining the response. The challenge becomes even more pronounced in a high-dimensional setting – there are exponentially many potential interactions among the predictors, for which it is simply not computationally feasible to resort to standard variable selection techniques (cf. @fan2010selective).
There are numerous examples where accounting for the predictors’ interactions is of interest, including problems of identifying epistasis (gene-gene) and gene-environment interactions in genetics [@cordell2009detecting], modeling problems in political science [@brambor2006understanding] and economics [@ai2003interaction]. In the business analytics of retail demand forecasting, a strong prediction model that also accurately accounts for the interactions of relevant predictors such as seasons, product types, geography, promotions, etc. plays a critical role in the decision making of marketing design.
A simple way to address the aforementioned issue in the regression problem is to simply restrict our attention to lower order interactions (i.e. 2- or 3-way) among predictor variables. This can be achieved, for instance, via a support vector machine (SVM) using polynomial kernels [@cristianini2000introduction], which pre-determine the maximum order of predictor interactions. In practice, for computational reasons the degree of the polynomial kernel tends to be small. Factorization machines [@rendle2010factorization] can be viewed as an extension of SVM to sparse settings where most interactions are observed only infrequently, subject to a constraint that the interaction order (a.k.a. interaction depth) is given. Neither SVM nor FM can perform any selection of predictor interactions, but several authors have extended the SVM by combining it with $\ell_1$ penalty for the purpose of feature selection [@zhu2004one] and gradient boosting for FM [@cheng2014gradient] to select interacting features. It is also an option to perform linear regression on as many interactions as we can and combine it with regularization procedures for selection (e.g. LASSO [@tibshirani1996regression] or Elastic net [@zou2005regularization]). It is noted that such methods are still not computationally feasible for accounting for interactions that involve a large number of predictor variables.
In this work we propose a regression method capable of adaptive selection of multi-way interactions of arbitrary order (MiFM for short), while avoiding the combinatorial complexity growth encountered by the methods described above. MiFM extends the basic factorization mechanism for representing the regression coefficients of interactions among the predictors, while the interaction selection is guided by a prior distribution on random hypergraphs. The prior, which does not insist on the upper bound on the order of interactions among the predictor variables, is motivated from but also generalizes Finite Feature Model, a parametric form of the well-known Indian Buffet process (IBP) [@ghahramani2005infinite]. We introduce a notion of the hypergraph of interactions and show how a parametric distribution over binary matrices can be utilized to express interactions of unbounded order. In addition, our generalized construction allows us to exert extra control on the tail behavior of the interaction order. IBP was initially used for infinite latent feature modeling and later utilized in the modeling of a variety of domains (see a review paper by @griffiths2011indian). In developing MiFM, our contributions are the following: (i) we introduce a Bayesian multi-linear regression model, which aims to account for the multi-way interactions among predictor variables; part of our model construction includes a prior specification on the hypergraph of interactions — in particular we show how our prior can be used to model the incidence matrix of interactions in several ways; (ii) we propose a procedure to estimate coefficients of arbitrary interactions structure; (iii) we establish posterior consistency of the resulting MiFM model, i.e., the property that the posterior distribution on the true regression function represented by the MiFM model contracts toward the truth under some conditions, without requiring an upper bound on the order of the predictor interactions; and (iv) we present a comprehensive simulation study of our model and analyze its performance for retail demand forecasting and case-control genetics datasets with epistasis. The unique strength of the MiFM method is the ability to recover meaningful interactions among the predictors while maintaining a competitive prediction quality compared to existing methods that target prediction only.
The paper proceeds as follows. Section \[bg\] introduces the problem of modeling interactions in regression, and gives a brief background on the Factorization Machines. Sections \[ifm\_model\] and \[hyps\] carry out the contributions outlined above. Section \[experiment\] presents results of the experiments. We conclude with a discussion in Section \[dis\].
Background and related work {#bg}
===========================
Our starting point is a model which regresses a response variable $y\in\mathbb{R}$ to observed covariates (predictor variables) $x\in\mathbb{R}^D$ by a non-linear functional relationship. In particular, we consider a multi-linear structure to account for the interactions among the covariates in the model: $$\label{yhat}
\E(Y|x) = w_0 + \sum_{i=1}^Dw_ix_i + \sum_{j=1}^J\beta_j\prod_{i\in Z_j}x_i.$$ Here, $w_i$ for $i=0,\ldots,D$ are bias and linear weights as in the standard linear regression model, $J$ is the number of multi-way interactions where $Z_j, \beta_j$ for $j=1,\ldots,J$ represent the interactions, i.e., sets of indices of interacting covariates and the corresponding interaction weights, respectively. Fitting such a model is very challenging even if dimension $D$ is of magnitude of a dozen, since there are $2^D - 1$ possible interactions to choose from in addition to other parameters. The goal of our work is to perform interaction selection and estimate corresponding weights. Before doing so, let us first discuss a model that puts a priori assumptions on the number and the structure of interactions.
Factorization Machines
----------------------
Factorization Machines (FM) [@rendle2010factorization] is a special case of the general interactions model defined in Eq. . Let $J=\sum_{l=2}^d\binom{D}{l}$ and $Z:=\bigcup_{j=1}^J Z_j=\bigcup_{l=2}^d \{(i_1,\ldots,i_l) | i_1<\ldots<i_l;\,i_1,\ldots,i_l\in\{1,\ldots,D\}\}$. I.e., restricting the set of interactions to $2,\ldots,d$-way, so becomes: $$\label{fm-svm}
\E(Y|x) = w_0 + \sum_{i=1}^Dw_ix_i + \sum_{l=2}^d \sum _{i_1=1}^D \ldots \sum _{i_l=i_{l-1} + 1}^D \beta_{i_1,\ldots,i_l} \prod_{t=1}^l x_{i_t},$$ where coefficients $\beta_j := \beta_{i_1,\ldots,i_l}$ quantify the interactions. In order to reduce model complexity and handle sparse data more effectively, [@rendle2010factorization] suggested to factorize interaction weights using PARAFAC [@harshman1970foundations]: $\beta_{i_1,\ldots,i_l}:= \sum_{f=1}^{k_l}\prod_{t=1}^l v_{i_t, f}^{(l)}$, where $V^{(l)} \in \mathbb{R}^{D\times k_l}$, $k_l \in \mathbb{N}$ and $k_l \ll D$ for $l=2,\ldots,d$. Advantages of the FM over SVM are discussed in details by @rendle2010factorization. FMs turn out to be successful in the recommendation systems setups, since they utilize various context information [@rendle2011fast; @nguyen2014gaussian]. Parameter estimation is typically achieved via stochastic gradient descent technique, or in the case of Bayesian FM [@freudenthaler2011bayesian] via MCMC. In practice only $d=2$ or $d=3$ are typically used, since the number of interactions and hence the computational complexity grow exponentially. We are interested in methods that can adapt to fewer interactions but of arbitrarily varying orders.
MiFM: Multi-way Factorization Machine {#ifm_model}
=====================================
We start by defining a mathematical object that can encode sets of interacting variables $Z_1,\ldots,Z_J$ of Eq. and selecting an appropriate prior to model it.
Modeling hypergraph of interactions {#hyper}
-----------------------------------
Multi-way interactions are naturally represented by hypergraphs, which are defined as follows.
\[hyp\] Given $D$ vertices indexed by $S=\{1,\ldots,D\}$, let $Z=\{Z_1,\ldots,Z_J\}$ be the set of $J$ subsets of $S$. Then we say that $G=(S,Z)$ is a hypergraph with $D$ vertices and $J$ hyperedges.
A hypergraph can be equivalently represented as an incidence binary matrix. Therefore, with a bit abuse of notation, we recast $Z$ as the matrix of interactions, i.e., $Z\in\{0,1\}^{D\times J}$, where $Z_{i_1 j} = Z_{i_2 j} = 1$ iff $i_1$ and $i_2$ are part of a hyperedge indexed by column/interaction $j$.
Placing a prior on multi-way interactions is the same as specifying the prior distribution on the space of binary matrices. We will at first adopt the Finite Feature Model (FFM) prior [@ghahramani2005infinite], which is based on the Beta-Bernoulli construction: $\pi_j|\gamma_1, \gamma_2 \stackrel{iid} {\thicksim} \text{Beta}(\gamma_1,\gamma_2)$ and $Z_{ij}|\pi_j \stackrel{iid}{\thicksim} \text{Bernoulli}(\pi_j)$. This simple prior has the attractive feature of treating the variables involved in each interaction (hyperedge) in an symmetric fashion and admits exchangeabilility among the variables inside interactions. In Section \[hyps\] we will present an extension of FFM which allows to incorporate extra information about the distribution of the interaction degrees and explain the choice of the parametric construction.
Modeling regression with multi-way interactions
-----------------------------------------------
Now that we know how to model unknown interactions of arbitrary order, we combine it with the Bayesian FM to arrive at a complete specification of MiFM, the Multi-way interacting Factorization Machine. Starting with the specification for hyperparameters: $$\begin{aligned}
& \sigma\thicksim\Gamma(\alpha_1/2,\beta_1/2),
\qquad
\lambda\thicksim\Gamma(\alpha_0/2,\beta_0/2),
\qquad
\mu\thicksim\mathcal{N}(\mu_0, 1/\gamma_0),\\
&\lambda_k\thicksim\Gamma(\alpha_0/2,\beta_0/2),
\qquad
\mu_k\thicksim\mathcal{N}(\mu_0, 1/\gamma_0)\text{ for }k=1,\ldots,K.\end{aligned}$$ Interactions and their weights: $$\begin{aligned}
& w_i|\mu, \lambda\thicksim\mathcal{N}(\mu, 1/\lambda)\text{ for }i=0,\ldots,D,\,\,\,\,\,\,\,\,
Z\thicksim\text{FFM}(\gamma_1,\gamma_2),\\
& v_{ik}|\mu_k, \lambda_k \thicksim\mathcal{N}(\mu_k, 1/\lambda_k)\text{ for }i=1,\ldots,D;\,k=1,\ldots,K.\end{aligned}$$ Likelihood specification given data pairs $(y_n,x_n=(x_{n1},\ldots,x_{nD}))_{n=1}^{N}$: $$\begin{aligned}
\label{ifm_y}
& y_n|\Theta\thicksim \mathcal{N}(y(x_n, \Theta),\sigma)\text{, where }y(x, \Theta):= w_0 + \sum_{i=1}^Dw_ix_i + \sum_{j=1}^J\sum_{k=1}^K\prod_{i\in Z_j}x_i v_{ik},\end{aligned}$$ for $n=1,\ldots,N,$ and $\Theta=\{Z,V,\sigma,w_{0,\ldots,D}\}$. Note that while the specification above utilizes Gaussian distributions, the main innovation of MiFM is the idea to utilize incidence matrix of the hypergraph of interactions $Z$ with a low rank matrix $V$ to model the mean response as in Eq. \[yhat\]. Therefore, within the MiFM framework, different distributional choices can be made according to the problem at hand — e.g. Poisson likelihood and Gamma priors for count data or logistic regression for classification. Additionally, if selection of linear terms is desired, $\sum_{i=1}^Dw_ix_i$ can be removed from the model since FFM can select linear interactions besides higher order ones.
MiFM for Categorical Variables {#ifm_cat}
------------------------------
In numerous real world scenarios such as retail demand forecasting, recommender systems, genotype structures, most predictor variables may be categorical (e.g. color, season). Categorical variables with multiple attributes are often handled by so-called “one-hot encoding”, via vectors of binary variables (e.g., IS\_blue; IS\_red), which must be mutually exclusive. The FFM cannot immediately be applied to such structures since it assigns positive probability to interactions between attributes of the same category. To this end, we model interactions between categories in $Z$, while with $V$ we model coefficients of interactions between attributes. For example, for an interaction between “product type” and “season” in $Z$, $V$ will have individual coefficients for “jacket-summer” and “jacket-winter” leading to a more refined predictive model of jackets sales (see examples in Section \[real\]).
We proceed to describe MiFM for the case of categorical variables as follows. Let $U$ be the number of categories and $d_u$ be the set of attributes for the category $u$, for $u=1,\ldots, U$. Then $D=\sum_{u=1}^U \operatorname*{card}(d_u)$ is the number of binary variables in the one-hot encoding and $\bigsqcup_{u=1}^U d_u = \{1,\ldots,D\}$. In this representation the input data of predictors is $X$, a $N\times U$ matrix, where $x_{nu}$ is an active attribute of category $u$ of observation $n$. Coefficients matrix $V\in\mathbb{R}^{D\times K}$ and interactions $Z\in \{0,1\}^{U\times J}$. All priors and hyperpriors are as before, while the mean response is replaced by: $$\label{ifm_cat_y}
y(x, \Theta):= w_0 + \sum_{u=1}^Uw_{x_u} + \sum_{k=1}^{K} \sum_{j=1}^J \prod_{u \in Z_j} v_{x_uk}.$$ Note that this model specification is easy to combine with continuous variables, allowing MiFM to handle data with different variable types.
Posterior Consistency of the MiFM
---------------------------------
In this section we shall establish posterior consistency of MiFM model, namely: the posterior distribution $\Pi$ of the conditional distribution $P(Y|X)$, given the training $N$-data pairs, contracts in a weak sense toward the truth as sample size $N$ increases.
Suppose that the data pairs $(x_n,y_n)_{n=1}^{N} \in \mathbb{R}^D\times \mathbb{R}$ are i.i.d. samples from the joint distribution $P^*(X,Y)$, according to which the marginal distribution for $X$ and the conditional distribution of $Y$ given $X$ admit density functions $f^*(x)$ and $f^*(y|x)$, respectively, with respect to Lebesgue measure. In particular, $f^*(y|x)$ is defined by $$\begin{aligned}
\label{true}
\left.\begin{aligned}
&Y=y_n|X=x_n,\Theta^*\thicksim \mathcal{N}(y(x_n, \Theta^*),\sigma),\text{ where }\Theta^*=\{\beta^*_1,\ldots,\beta^*_{J},Z^*_1,\ldots,Z^*_{J}\},\\
&y(x, \Theta^*):= \sum_{j=1}^{J}\beta^*_j\prod_{i\in Z^*_j}x_i,\text{ and }x_n\in \mathbb{R}^D, y_n\in \mathbb{R}, \beta^*_j\in \mathbb{R}, Z^*_j\subset \{1,\ldots,D\}
\end{aligned}\right.\end{aligned}$$ for $n=1,\ldots,N,j=1,\ldots,J$. In the above $\Theta^*$ represents the *true* parameter for the conditional density $f^*(y|x)$ that generates data sample $y_n$ given $x_n$, for $n=1,\ldots,N$. A key step in establishing posterior consistency for the MiFM (here we omit linear terms since, as mentioned earlier, they can be absorbed into the interaction structure) is to show that our PARAFAC type structure can approximate arbitrarily well the true coefficients $\beta^*_1,\ldots,\beta^*_{J}$ for the model given by .
\[elim2\] Given natural number $J\geq 1$, $\beta_j \in \mathbb{R}\setminus\{0\}$ and $Z_j\subset \{1,\ldots,D\}$ for $j=1,\ldots J$, exists $K_0 < J$ such that for all $K\geq K_0$ system of polynomial equations $\beta_j = \sum_{k=1}^K \prod_{i \in Z_j} v_{ik} \text{, }j=1,\ldots,m$ has at least one solution in terms of $v_{11},\ldots,v_{DK}$.
The upper bound $K_0=J-1$ is only required when *all* interactions are of the depth $D-1$. This is typically not expected to be the case in practice, therefore smaller values of $K$ are often sufficient. By conditioning on the training data pairs $(x_n,y_n)$ to account for the likelihood induced by the PARAFAC representation, the statistician obtains the posterior distribution on the parameters of interest, namely, $\Theta := (Z,V)$, which in turn induces the posterior distribution on the conditional density, to be denoted by $f(y|x)$, according to the MiFM model without linear terms. The main result of this section is to show that under some conditions this posterior distribution $\Pi$ will place most of its mass on the true conditional density $f^*(y|x)$ as $N\rightarrow \infty$. To state the theorem precisely, we need to adopt a suitable notion of weak topology on the space of conditional densities, namely the set of $f(y|x)$, which is induced by the weak topology on the space of joint densities on $X,Y$, that is the set of $f(x,y) = f^*(x) f(y|x)$, where $f^*(x)$ is the true (but unknown) marginal density on $X$ (see @ghosal1999posterior, Sec. 2 for a formal definition).
\[th1\] Given any true conditional density $f^*(y|x)$ given by , and assuming that the support of $f^*(x)$ is bounded, there is a constant $K_0< J$ such that by setting $K \geq K_0$, the following statement holds: for any weak neighborhood $U$ of $f^*(y|x)$, under the MiFM model, the posterior probability $\Pi(U|(X_n,Y_n)_{n=1}^{N}) \rightarrow 1$ with $P^*$-probability one, as $N\rightarrow \infty$.
The proof’s sketch for this theorem is given in the .
Prior constructions for interactions: FFM revisited and extended {#hyps}
================================================================
The adoption of the FFM prior on the hypergraph of interactions carries a distinct behavior in contrast to the typical Latent Feature modeling setting. In a standard Latent Feature modeling setting [@griffiths2011indian], each row of $Z$ describes one of the data points in terms of its feature representation; controlling row sums is desired to induce sparsity of the features. By contrast, for us a column of $Z$ is identified with an interaction; its sum represents the interaction depth, which we want to control a priori.
#### Interaction selection using MCMC sampler
One interesting issue of practical consequence arises in the aggregation of the MCMC samples (details of the sampler are in the ). When aggregating MCMC samples in the context of *latent feature modeling* one would always obtain exactly $J$ latent features. However, in *interaction modeling*, different samples might have no interactions in common (i.e. no exactly matching columns), meaning that support of the resulting posterior estimate can have up to $\min\{2^D-1, IJ\}$ unique interactions, where $I$ is the number of MCMC samples. In practice, we can obtain marginal distributions of all interactions across MCMC samples and use those marginals for selection. One approach is to pick $J$ interactions with highest marginals and another is to consider interactions with marginal above some threshold (e.g. 0.5). We will resort to the second approach in our experiments in Section \[experiment\] as it seems to be in more agreement with the concept of “selection”. Lastly, we note that while a data instance may a priori possess unbounded number of features, the number of possible interactions in the data is bounded by $2^D-1$, therefore taking $J\rightarrow \infty$ might not be appropriate. In any case, we do not want to encourage the number of interactions to be too high for regression modeling, which would lead to overfitting. The above considerations led us to opt for a parametric prior such as the FFM for interactions structure $Z$, as opposed to going fully nonparametric. $J$ can then be chosen using model selection procedures (e.g. cross validation), or simply taken as the model input parameter.
#### Generalized construction and induced distribution of interactions depths
We now proceed to introduce a richer family of prior distributions on hypergraphs of which the FFM is one instance. Our construction is motivated by the induced distribution on the column sums and the conditional probability updates that arise in the original FFM. Recall that under the FFM prior, interactions are a priori independent. Fix an interaction $j$, for the remainder of this section let $Z_i$ denote the indicator of whether variable $i$ is present in interaction $j$ or not (subscript $j$ is dropped from $Z_{ij}$ to simplify notation). Let $M_i = Z_1 + \ldots + Z_i$ denote the number of variables among the first $i$ present in the corresponding interaction. By the Beta-Bernoulli conjugacy, one obtains $\P(Z_i=1|Z_1,\ldots,Z_{i-1}) = \frac{M_{i-1} + \gamma_1}{i-1+\gamma_1+\gamma_2}$. This highlights the “rich-gets-richer” effect of the FFM prior, which encourages the existence of very deep interactions while most other interactions have very small depths. In some situations we may prefer a relatively larger number of interactions of depths in the medium range. An intuitive but somewhat naive alternative sampling process is to allow a variable to be included into an interaction according to its present “shallowness” quantified by $(i-1-M_{i-1})$ (instead of $M_{i-1}$ in the FFM). It can be verified that this construction will lead to a distribution of interactions which concentrates most its mass around $D/2$; moreover, exchangeability among $Z_i$ would be lost. To maintain exchangeability, we define the sampling process for the sequence $Z=(Z_1,\ldots,Z_D) \in
\{0,1\}^D$ as follows: let $\sigma(\cdot)$ be a random uniform permutation of $\{1,\ldots,D\}$ and let $\sigma_1 = \sigma^{-1}(1),\ldots,\sigma_D = \sigma^{-1}(D)$. Note that $\sigma_1,\ldots,\sigma_D$ are discrete random variables and $\P(\sigma_k = i)=1/D$ for any $i,k = 1,\ldots,D$. For $i=1,\ldots,D$, set $$\begin{aligned}
\label{nex}
&\P(Z_{\sigma_i}=1|Z_{\sigma_1},\ldots,Z_{\sigma_{i-1}}) = \frac{\alpha M_{i-1} + (1-\alpha)(i-1-M_{i-1})+\gamma_1}
{i-1+\gamma_1+\gamma_2},\nonumber\\
&\P(Z_{\sigma_i}=0|Z_{\sigma_1},\ldots,Z_{\sigma_{i-1}}) = \frac{(1-\alpha)M_{i-1} + \alpha(i-1-M_{i-1})+\gamma_2}
{i-1+\gamma_1 + \gamma_2},\end{aligned}$$ where $\gamma_1>0,\gamma_2>0,\alpha \in [0,1]$ are given parameters and $M_i = Z_{\sigma_1} + \ldots + Z_{\sigma_i}$. The collection of $Z$ generated by this process shall be called to follow FFM$_\alpha$. When $\alpha=1$ we recover the original FFM prior. When $\alpha=0$, we get the other extremal behavior mentioned at the beginning of the paragraph. Allowing $\alpha\in [0,1]$ yields a richer spectrum spanning the two distinct extremal behaviors. Details of the process and some of its properties are given in the . Here we briefly describe how FFM$_\alpha$ *a priori* ensures “poor gets richer” behavior and offers extra flexibility in modeling interaction depths compared to the original FFM. The depth of an interaction of $D$ variables is described by the distribution of $M_D$. Consider the conditionals obtained for a Gibbs sampler where index of a variable to be updated is random and based on $\P(\sigma_D = i|Z)$ (it is simply $1/D$ for FFM$_1$). Suppose we want to assess how likely it is to *add* a variable into an existing interaction via the expression $\sum_{i:Z^{(k)}_i = 0}\P(Z^{(k+1)}_i = 1, \sigma_D = i| Z^{(k)})$, where $k+1$ is the next iteration of the Gibbs sampler’s conditional update. This probability is a function of $M_D^{(k)}$; for small values of $M_D^{(k)}$ it quantifies the tendency for the “poor gets richer” behavior. For the FFM$_1$ it is given by $\frac{D-M_D^{(k)}}{D}\frac{M_D^{(k)} + \gamma_1}{D-1+\gamma_1+\gamma_2}$. In Fig. \[fig:hip\_ibp\](a) we show that FFM$_1$’s behavior is opposite of “poor gets richer”, while $\alpha\leq 0.7$ appears to ensure the desired property. Next, in Fig.\[fig:hip\_ibp\] (b-f) we show the distribution of $M_D$ for various $\alpha$, which exhibits a broader spectrum of behavior.
{width="\textwidth"}
-0.2in
Experimental Results {#experiment}
====================
Simulation Studies
------------------
We shall compare MiFM methods against a variety of other regression techniques in the literature, including Bayesian Factorization Machines (FM), lasso-type regression, Support Vector Regression (SVR), multilayer perceptron neural network (MLP).[^2] The comparisons are done on the basis of prediction accuracy of responses (Root Mean Squared Error on the held out data), quality of regression coefficient estimates and the interactions recovered.
{width="\textwidth"}
-0.2in
### Predictive Performance {#pred}
In this set of experiments we demonstrate that MiFMs with either $\alpha=0.7$ or $\alpha=1$ have dominant predictive performance when high order interactions are in play. In Fig. \[fig:all\](a) we analyzed 70 random interactions of varying orders. We see that MiFM can handle arbitrary complexity of the interactions, while other methods are comparative only when interaction structure is simple (i.e. linear or 2-way on the right of the Fig. \[fig:all\](a)).
Next, to assess the effectiveness of MiFM in handling categorical variables (cf. Section \[ifm\_cat\]) we vary the number of continuous variables from 1 (and 29 attributes across categories) to 30 (no categorical variables). Results in Fig. \[fig:all\](b) demonstrate that our models can handle both variable types in the data (including continuous-categorical interactions), and still exhibit competitive RMSE performance.
### Interactions Quality
#### Coefficients of the interactions
This experiment verifies the posterior consistency result of Theorem \[th1\] and validates our factorization model for coefficients approximation. In Fig. \[fig:all\](c) we compare MiFMs versus OLS fitted with the corresponding sets of chosen interactions. Additionally we benchmark against Elastic net [@zou2005regularization] based on the expanded data matrix with interactions of all depths included, that is $2^D - 1$ columns, and a corresponding OLS with only selected interactions.
#### Selection of the interactions
In this experiments we assess how well MiFM can recover true interactions. We consider three interaction structures: a realistic one with five linear, five 2-way, three 3-way and one of each $4,\ldots,8$-way interactions, and two artificial ones with 15 either only 4- or only 6-way interactions to challenge our model. Both binary and continuous variables are explored. Fig. \[fig:all\](d) shows that MiFM can *exactly* recover up to 83% of the interactions and with $\alpha=0.8$ it recovers 75% of the interaction in 4 out of 6 scenarios. Situation with 6-way interactions is more challenging, where 36% for binary data is recovered and almost half for continuous. It is interesting to note that lower values of $\alpha$ handle binary data better, while higher values are more appropriate for continuous, which is especially noticeable on the “only 6-way” case. We think it might be related to the fact that high order interactions between binary variables are very rare in the data (i.e. product of 6 binary variables is equal to 0 most of the times) and we need a prior eager to explore ($\alpha=0$) to find them.
Real world applications {#real}
-----------------------
### Finding epistasis
Identifying epistasis (i.e. interactions between genes) is one of the major questions in the field of human genetics. Interactions between multiple genes and environmental factors can often tell a lot more about the presence of a certain disease than any of the genes individually [@templeton2000epistasis]. Our analysis of the epistasis is based on the data from @himmelstein2011evolving. These authors show that interactions between single nucleotide polymorphisms (SNPs) are often powerful predictors of various diseases, while individually SNPs might not contain important information at all. They developed a model free approach to simulate data mimicking relationships between complex gene interactions and the presence of a disease. We used datasets with five SNPs and either 3-,4- and 5-way interactions or only 5-way interactions. For this experiment we compared MiFM$_1$, MiFM$_0$; refitted logistic regression for each of our models based on the selected interactions (LMiFM$_1$ and LMiFM$_0$), Multilayer Perceptron with 3 layers and Random Forest.[^3] Results in Table \[genes\] demonstrate that MiFM produces competitive performance compared to the very best black-box techniques on this data set, while it also selects interacting genes (i.e. finds epistasis). We don’t know which of the 3- and 4-way interactions are present in the data, but since there is only one possible 5-way interaction we can check if it was identified or not — both MiFM$_1$ and MiFM$_0$ had a 5-way interaction in at least *95%* of the posterior samples.
-0.1in
MiFM$_1$ MiFM$_0$ LMiFM$_1$ LMiFM$_0$ MLP RF
--------------- ---------- ---------- ----------- ----------- ------- -------
3-, 4-, 5-way 0.775 0.771 0.883 0.860 0.870 0.887
only 5-way 0.649 0.645 0.628 0.623 0.625 0.628
-0.05in
### Understanding retail demand
We finally report the analysis of data obtained from a major retailer with stores in multiple locations all over the world. This dataset has 430k observations and 26 variables spanning over 1100 binary variables after the one-hot encoding. Sales of a variety of products on different days and in different stores are provided as response. We will compare MiFM$_1$ and MiFM$_0$, both fitted with $K=12$ and $J=150$, versus Factorization Machines in terms of adjusted mean absolute percent error $\text{AMAPE}=100\frac{\sum_n|\hat{y}_n-y_n|}{\sum_n y_n}$, a common metric for evaluating sales forecasts. FM is currently a method of choice by the company for this data set, partly because the data is sparse and is similar in nature to the recommender systems. AMAPE for MiFM$_1$ is 92.4; for MiFM$_0$ - 92.45; for FM - 92.0.
-0.05in
{width="\columnwidth"}
-0.2in
#### Posterior analysis of predictor interactions
The unique strength of MiFM is the ability to provide valuable insights about the data through its posterior analysis. MiFM$_1$ recovered 62 non-linear interactions among which there are five 3-way and three 4-way. MiFM$_0$ selected 63 non-linear interactions including nine 3-way and four 4-way. We note that choice $\alpha=0$ was made to explore deeper interactions and as we see MiFM$_0$ has more deeper interactions than MiFM$_1$. Coefficients for a 3-way interaction of MiFM$_1$ for two stores in France across years and months are shown in Fig. \[fig:sales\](a,b). We observe different behavior, which would not be captured by a low order interaction. In Fig. \[fig:sales\](c,d) we plot coefficients of a 4-way MiFM$_0$ interaction for the same two stores in France. It is interesting to note negative correlation between Saturday and Sunday coefficients for the store in Merignac, while the store in Perols is not affected by this interaction - this is an example of how MiFM can select interactions between attributes across categories.
Discussion {#dis}
==========
We have proposed a novel regression method which is capable of learning interactions of arbitrary orders among the regression predictors. Our model extends Finite Feature Model and utilizes the extension to specify a hypergraph of interactions, while adopting a factorization mechanism for representing the corresponding coefficients. We found that MiFM performs very well when there are some important interactions among a relatively high number (higher than two) of predictor variables. This is the situation where existing modeling techniques may be ill-equipped at describing and recovering. There are several future directions that we would like to pursue. A thorough understanding of the fully nonparametric version of the FFM$_\alpha$ is of interest, that is, when the number of columns is taken to infinity. Such understanding may lead to an extension of the IBP and new modeling approaches in various domains.
### Acknowledgments {#acknowledgments .unnumbered}
This research is supported in part by grants NSF CAREER DMS-1351362, NSF CNS-1409303, a research gift from Adobe Research and a Margaret and Herman Sokol Faculty Award.
Supplementary material
======================
In the Supplementary material we will start by proving consistency of the MiFM theorem, then we will show several important results related to FFM$_\alpha$: how exchangeability is achieved using uniform permutation prior on the order in which variables enter the process, how it leads to a Gibbs sampler using distribution of the index of the variable entering FFM$_\alpha$ last and how to obtain distribution of the interaction depths $M_D$ and compute its expectation. Lastly we will present a Gibbs sampling algorithm for the MiFM under the FFM$_\alpha$ prior on interactions structure $Z$. \[supp\]
Proof of the Consistency Theorem \[th1\] {#supp_th}
----------------------------------------
First let us remind the reader of the problem setup. Suppose that the data pairs $(x_n,y_n)_{n=1}^{N} \in \mathbb{R}^D\times \mathbb{R}$ are i.i.d. samples from the joint distribution $P^*(X,Y)$, according to which marginal distribution for $X$ and the conditional distribution of $Y$ given $X$ admit density functions $f^*(x)$ and $f^*(y|x)$, respectively, with respect to Lebesgue measure. In particular, $f^*(y|x)$ is defined as in Eq. : $$\begin{aligned}
\left.\begin{aligned}
&Y=y_n|X=x_n,\Theta^*\thicksim \mathcal{N}(y(x_n, \Theta^*),\sigma),\text{ where }\Theta^*=\{\beta^*_1,\ldots,\beta^*_{J},Z^*_1,\ldots,Z^*_{J}\},\\
&y(x, \Theta^*):= \sum_{j=1}^{J}\beta^*_j\prod_{i\in Z^*_j}x_i,\text{ and }x_n\in \mathbb{R}^D, y_n\in \mathbb{R}, \beta^*_j\in \mathbb{R}, Z^*_j\subset \{1,\ldots,D\},\\
&\text{for }n=1,\ldots,N,j=1,\ldots,J.
\end{aligned}\right.\end{aligned}$$ In the above $\Theta^*$ represents the *true* parameter for the conditional density $f^*(y|x)$ that generates data sample $y_n$ given $x_n$, for $n=1,\ldots,N$. On the other hand, the statistical modeler has access only to the MiFM: $$\begin{aligned}
\label{ifm_proof}
\left.\begin{aligned}
& Z\thicksim\text{FFM$_\alpha$}(\gamma_1, \gamma_2), \ v_{ik}|\mu_k, \lambda_k \thicksim\mathcal{N}(\mu_k, \frac{1}{\lambda_k})\text{ for }i=1,\ldots,D;\,k=1,\ldots,K,\\
& y_n|\Theta\thicksim \mathcal{N}(y(x_n, \Theta),\sigma)\text{, where }y(x, \Theta):= \sum_{j=1}^J\sum_{k=1}^K\prod_{i\in Z_j}x_i v_{ik},\\
& \text{for }n=1,\ldots,N,\text{ and }\Theta=(Z,V).
\end{aligned}\right.\end{aligned}$$ We omitted linear terms in the MiFM since they can naturally be parts of the interaction structure $Z$ and discarded hyperpriors for the ease of representation. Now we show that under some conditions posterior distribution $\Pi$ will place most of its mass on the true conditional density $f^*(y|x)$ as $N\rightarrow \infty$.
Given any true conditional density $f^*(y|x)$ given by , and assuming that the support of $f^*(x)$ is bounded, there is a constant $K_0< J$ such that by setting $K \geq K_0$, the following statement holds: for any weak neighborhood $U$ of $f^*(y|x)$, under the MiFM model , the posterior probability $\Pi(U|(X_n,Y_n)_{n=1}^{N}) \rightarrow 1$ with $P^*$-probability one, as $N\rightarrow \infty$.
A key part in the proof of this theorem is to clarify the role of parameter $K$, and the fact that under model , the regression coefficient $\beta_j$ associated with interaction $j$ is parameterized by $\beta_j := \sum_{k=1}^K \prod_{i \in Z_j} v_{ik}$, for $j=1,\ldots,J$, which for some suitable choice of $\Theta = (Z,V)$ can represent exactly the true parameters $\beta_1^*,\ldots,\beta^*_{J}$, provided that $K$ is sufficiently large. The following basic lemma is informative.
\[elim\] Let $m \in [1,J]$ be a natural number, $\beta_j \in \mathbb{R}\setminus\{0\}$ for $j=1,\ldots, m$. Suppose that the $m$ subsets $Z_j\subset \{1,\ldots,D\}$ for $j=1,\ldots m$ have non-empty intersection, then as long as $K\geq m$, the system of polynomial equations $$\label{syst}
\sum_{k=1}^K \prod_{i \in Z_j} v_{ik} = \beta_j\text{, }j=1,\ldots,m$$ has at least one solution in terms of $v_{11},\ldots,v_{DK}$ such that the following collection of $K$ vectors in $\mathbb{R}^m$, namely, $\{(\prod_{i \in Z_1}v_{ik},\ldots,\prod_{i \in Z_m}v_{ik}), k=1,\ldots,K\}$ contains $m$ linearly independent vectors.
Let $i_0$ be an element of the intersection of all $Z_j$, for $j=1,\ldots, m$. We consider system as linear with respect to $\{v_{i_01},\ldots,v_{i_0K}\}$, where corresponding coefficients are given by $\prod_{i \in Z_j\setminus\{i_0\}} v_{i,k}$, which we can pick to form a matrix of nonzero determinant. Hence by Rouché–Capelli theorem the system has at least one solution if $K\geq m$ and, since $\beta_j\neq 0$ for $\forall j$, the resulting $\{(\prod_{i \in Z_1}v_{ik},\ldots,\prod_{i \in Z_m}v_{ik}), k=1,\ldots,K\}$ contains at least $m$ linearly independent vectors.
Given natural number $J\geq 1$, $\beta_j \in \mathbb{R}\setminus\{0\}$ and $Z_j\subset \{1,\ldots,D\}$ for $j=1,\ldots J$, exists $K_0 < J: \forall K\geq K_0$ system of polynomial equations has at least one solution in terms of $v_{11},\ldots,v_{DK}$.
The proof proceeds by performing an elimination process on the collection of variables $v_{ik}$ according to an ordering that we now define. Let $J_i = \operatorname*{card}(\{Z_j|i\in Z_j\})$ for $i=1,\ldots,D$. Define $J^0=\min\limits_i J_i$ and $i_0=\operatorname*{argmin}\limits_i J_i$. If $K \geq J^0$ by Lemma \[elim\] we can find a solution of the reduced system of equations $$\sum_{k=1}^K \prod_{i \in Z_j} v_{i,k} = \beta_j\text{, }j\in\{j|i_0\in Z_j\},$$ while maintaining the linear independence needed to apply Lemma \[elim\] again further. Now we know that we can find a solution for equations indexed by $\{j|i_0\in Z_j\}$. We remove them from system and recompute $J^1=\min\limits_{i\neq i_0} J_i$ and $i_1=\operatorname*{argmin}\limits_{i\neq i_0} J_i$ to apply Lemma \[elim\] again. Iteratively we will remove all the equations, meaning that there is at least one solution. Note that $J_i$ are decreasing since whenever we remove equations, number of $Z_j$s containing certain $i$ can only decrease. Therefore, we will need $K\geq K_0 :=
\max(J^0, J^1, \ldots, 0)$ in order to apply Lemma \[elim\] on every elimination step.
From the proof of Lemma \[elim2\], it can be observed that $K_0 = \max(J^0,J^1,\ldots)
\ll J$ when we anticipate only few interactions per variable, whereas the upper bound $K_0=J-1$ is attained when there are only $(D-1)$-way interactions. Now we are ready to present a proof of the main theorem.
(of main theorem). By Lemma \[elim2\] and the fact that the probability of a finite number of independent continuous random vectors being linearly dependent is 0 it follows that under the MiFM prior on $V$ as in and $\forall \beta_1,\ldots,\beta_{J} \in \mathbb{R}\setminus\{0\}$, distinct $Z_1,\ldots,Z_{J}$ and $\epsilon>0$ $$\label{step2}
\Pi\left(\sum_{j=1}^{J}(\beta_j - \sum_k \prod_{i \in Z_j} v_{ik})^2 < \epsilon\,| Z_1,\ldots,Z_{J}\right) > 0.$$ From Eq. it follows that for any $Z_1,\ldots,Z_{J}$, the prior probability of the corresponding incidence matrix is bounded away from 0. Combining this with , we now establish that the probability of the true model parameters to be arbitrary close to the MiFM parameters under the MiFM prior as in : $$\label{step3}
\Pi\left((\sum_{j=1}^{J}\beta_j - \sum_{j=1}^J\sum_k \prod_{i \in Z_j} v_{ik})^2 < \epsilon\right) > 0,\,\forall \epsilon>0.$$ We shall appeal to Schwartz’s theorem (cf. [@ghosal1999posterior]), which asserts that the desired posterior consistency holds as soon as we can establish that the true joint distribution $P^*(X,Y)$ lies in the Kullback-Leibler support of the prior $\Pi$ on the joint distribution $P(X,Y)$. That is, $$\label{kl}
\Pi\left(\text{KL}(P^*||P)<\epsilon\right)>0,\text{ for }\forall \epsilon>0.$$ Since the KL divergence of the two Gaussian distributions is proportional to the mean difference, we have ($\mathbb{E}_X^*$ denotes expectation with respect to the true marginal distribution of $X$) $$\begin{aligned}
\label{gaus_kl}
\left.\begin{aligned}
& \text{KL}(P^*||P) \propto \mathbb{E}_X^* \frac{1}{2}(y(X,\Theta)-y(X,\Theta^*))^2 \propto \\
& \mathbb{E}_X^* (\sum_{j=1}^{J} \beta_j\prod_{i\in Z_j}x_{i} - \sum_{j=1}^{J}\sum_k \prod_{i \in Z_j} v_{ik}x_{i})^2 \lesssim (\sum_{j=1}^{J}\beta_j - \sum_{j=1}^J\sum_k \prod_{i \in Z_j} v_{ik})^2.
\end{aligned}\right.\end{aligned}$$ Due to this quantity can be made arbitrarily close to 0 with positive probability. Therefore and then Schwartz theorem hold, which concludes the proof.
Analyzing FFM$_\alpha$ {#ffm_all}
----------------------
### Model definition and exchangeability
Here we remind the reader the construction of FFM$_\alpha$ — the distribution over finite collection of binary random variables that we used to model interactions. Let $D$ be the number of variables in the data and $Z\in \{0,1\}^D$ is $j$-th interaction (subscript $j$ is dropped to simplify notation). Let $\sigma(\cdot)$ be a random uniform permutation of $\{1,\ldots,D\}$ and let $\sigma_1 = \sigma^{-1}(1),\ldots,\sigma_D = \sigma^{-1}(D)$. Note that $\sigma_1,\ldots,\sigma_D$ are discrete random variables and $\P(\sigma_k = i)=1/D$ for any $i,k = 1,\ldots,D$. Next recall FFM$_\alpha$ from Eq. : $$\begin{aligned}
&\P(Z_{\sigma_i}=1|Z_{\sigma_1},\ldots,Z_{\sigma_{i-1}}) = \frac{\alpha M_{i-1} + (1-\alpha)(i-1-M_{i-1})+\gamma_1}
{i-1+\gamma_1+\gamma_2},\nonumber\\
&\P(Z_{\sigma_i}=0|Z_{\sigma_1},\ldots,Z_{\sigma_{i-1}}) = \frac{(1-\alpha)M_{i-1} + \alpha(i-1-M_{i-1})+\gamma_2}
{i-1+\gamma_1 + \gamma_2},\end{aligned}$$ where $\gamma_1>0,\gamma_2>0,\alpha \in [0,1]$ are given parameters and $M_i = Z_{\sigma_1} + \ldots + Z_{\sigma_i}$. Due to the random permutation of indices, distribution of $Z_1,\ldots,Z_D$ is exchangeable because any ordering of variables entering the process has same probability. Next, we need to integrate the permutation part out to obtain a tractable full conditional representation.
### Gibbs sampling for FFM$_\alpha$ and distribution of interaction depths $M_D$ {#sample_fmm}
To construct a Gibbs sampler for the the FFM$_\alpha$ we will use an additional latent variable - index of the variable entering the process last, $\sigma_D$. Additionally observe that when permutation is integrated out $\P(Z_1,\ldots, Z_D)=\P(M_D=Z_1+\ldots + Z_D)$ since $\P(M_D=m)$ is precisely the summation over all possible orderings of $Z_1,\ldots,Z_D$ such that $Z_1 + \ldots + Z_D = m$. $$\begin{aligned}
\label{g_last}
\left.\begin{aligned}
&\P(\sigma_D = i|Z_1,\ldots, Z_D)\propto \\
& Z_i\P(\sigma_D = i|Z_{\sigma_D}=1,Z)\P(Z_{\sigma_D}=1|M_{D-1}=\sum_{k=1}^D Z_k - 1)\P(M_{D-1}=\sum_{k=1}^D Z_k - 1) + \\
& + (1-Z_i)\P(\sigma_D = i|Z_{\sigma_D}=0,Z)\P(Z_{\sigma_D}=0|M_{D-1}=\sum_{k=1}^D Z_k)\P(M_{D-1}=\sum_{k=1}^D Z_k),
\end{aligned}\right.\end{aligned}$$ then if $Z_i=1$ and $\sum_{k=1}^D Z_k = m$ we obtain $$\begin{aligned}
\label{last_1}
\left.\begin{aligned}
\P(\sigma_D = i|Z_{-i},Z_i=1) & = \P(\sigma_D = i| M_D=m, Z_i=1) = \\
& = \frac{\P(M_{D-1} = m - 1)\P(Z_{\sigma_D}=1|M_{D-1}=m-1)}{m\P(M_{D} = m)},
\end{aligned}\right.\end{aligned}$$ where $\P(Z_{\sigma_D}=1|M_{D-1}=m-1)$ and $\P(Z_{\sigma_D}=0|M_{D-1}=m)$ can be computed as in Eq. \[nex\]. Our next step is to analyze probability $\P(M_D = m)$. Indeed it is easy to obtain this distribution recursively: $$\begin{aligned}
\label{mn_rec}
\left.\begin{aligned}
\P(M_D = m) & = \P(M_{D-1}=m)\P(Z_{\sigma_D}=0|M_{D-1}=m) + \\ &+ \P(M_{D-1}=m-1)\P(Z_{\sigma_D}=1|M_{D-1}=m-1).
\end{aligned}\right.\end{aligned}$$ The base of recursion is given by the following identities: $$\begin{aligned}
\label{mn_rec_base}
\left.\begin{aligned}
& \P(M_0 = 0) = 1, \\
& \P(M_i = 0) = \prod_{k=0}^{i-1}\frac{\alpha(i-1-k) + \gamma_2}{k+\gamma_1+\gamma_2} = \prod_{k=0}^{i-1}\frac{\alpha k + \gamma_2}{k+\gamma_1+\gamma_2},\\
& \P(M_i = i) = \prod_{k=0}^{i-1}\frac{\alpha k + \gamma_1}{k+\gamma_1+\gamma_2}.
\end{aligned}\right.\end{aligned}$$
The above formulation allows us compute $\P(M_i=k), D\geq i\geq k$ dynamically (computations are very fast since we only need to perform $\frac{(D+1)(D+2)}{2}-1$ calculations) *before* running MiFM inference and utilize the table of probabilities during it. The last step of the Gibbs sampler is clearly the update of the $Z_i|\sigma_D=i,Z_{-i}$ which is done simply using the FFM$_\alpha$ definition \[nex\]. Recall Figure 1 (a) of the main text which illustrates the behavior of $$\begin{aligned}
\left.\begin{aligned}
\sum_{i:Z^{(k)}_i = 0}\P(Z^{(k+1)}_i = 1, \sigma_D = i| Z^{(k)}) = \P(Z_{\sigma_D}=0|Z)\P(Z_i=1|\sigma_D=i,Z_{-i}),
\end{aligned}\right.\end{aligned}$$ and since we choose index of a variable to update based on the probability of it being last, the expression above reads as the probability that we choose to update a variable not present in the interaction and then add it to the interaction, therefore increasing the depth of the interaction.
### Mean Behavior of the FFM$_\alpha$
From Eq. it follows that $$\begin{aligned}
\label{mean}
\left.\begin{aligned}
& \E M_D = \sum_{m=0}^{D}m\P(M_D = m) = \\
& = \frac{1}{D-1+\gamma_1+\gamma_2} \biggr \{
(1-2\alpha)\E M_{D-1}^2
+ (\alpha(D-1)+ \gamma_2) \E M_{D-1} + \\
& + (2\alpha-1) \E (M_{D-1}+1)^2
+ ((1-\alpha)D - \alpha + \gamma_1)\E(M_{D-1}+1) \biggr \} \\
& = \frac{1}{D-1+\gamma_1+\gamma_2}
\biggr \{\E M_{D-1}(D+2\alpha+\gamma_1+\gamma_2-2) +D(1-\alpha)+\alpha + \gamma_1-1 \biggr \}.
\end{aligned}\right.\end{aligned}$$ For $\alpha = 0$, this relation is simplified to be $$\begin{aligned}
\label{mean0}
\left.\begin{aligned}
(D-1+\gamma_1+\gamma_2) \E M_D & = \E M_{D-1}(D+\gamma_1+\gamma_2-2) + (D+\gamma_1 -1) = \\
& = (D+\gamma_1-1) + \ldots + \gamma_1 = \frac{1}{2}D(D+2\gamma_1-1).
\end{aligned}\right.\end{aligned}$$
Gibbs Sampler for the MiFM {#supp_gibbs}
--------------------------
Our Gibbs sampling algorithm consists of two parts — updating factorization coefficients $V$ (based on the results from @freudenthaler2011bayesian) and then updating interactions $Z$ based on the analysis of Section \[sample\_fmm\]. Recall the MiFM model construction. First we have a layer of hyperpriors: $$\begin{aligned}
& \sigma\thicksim\Gamma(\frac{\alpha_1}{2},\frac{\beta_1}{2}),
\qquad
\lambda\thicksim\Gamma(\frac{\alpha_0}{2},\frac{\beta_0}{2}),
\qquad
\mu\thicksim\mathcal{N}(\mu_0, \frac{1}{\gamma_0}),
\\
&\lambda_k\thicksim\Gamma(\frac{\alpha_0}{2},\frac{\beta_0}{2}),\
\mu_k\thicksim\mathcal{N}(\mu_0, \frac{1}{\gamma_0})\text{ for }k=1,\ldots,K,\end{aligned}$$ Then interactions and their weights: $$\begin{aligned}
& w_i|\mu, \lambda\thicksim\mathcal{N}(\mu, \frac{1}{\lambda})\text{ for }i=0,\ldots,D,
\qquad
Z\thicksim\text{FFM}_\alpha(\gamma_1,\gamma_2),\\
& v_{ik}|\mu_k, \lambda_k \thicksim\mathcal{N}(\mu_k, \frac{1}{\lambda_k})\text{ for }i=1,\ldots,D;\,k=1,\ldots,K,\end{aligned}$$ And finally the model’s likelihood from Eq. $$\begin{aligned}
\left.\begin{aligned}
& y_n|\Theta\thicksim \mathcal{N}(y(x_n, \Theta),\frac{1}{\sigma})\text{, where}\\
& y(x, \Theta):= w_0 + \sum_{i=1}^Dw_ix_{i} + \sum_{j=1}^J\sum_{k=1}^K\prod_{i\in Z_j}x_i v_{ik},\\
& \text{for }n=1,\ldots,N,\text{ and }\Theta=\{Z,V,\sigma,w_{0,\ldots,D}\}.
\end{aligned}\right.\end{aligned}$$ Inference in the context of Bayesian modeling is often related to learning the posterior distribution $\P(\Theta|X,Y)$. Then, if one wants point estimates, certain statistics of the posterior can be used, i.e. mean or median. In most situations (including MiFM) analytical form of the posterior is intractable, but with the help of Bayes rule it is often possible to compute it up to a proportionality constant: $$\begin{aligned}
\label{prop_ll}
\left.\begin{aligned}
&\P(\Theta,\mu,\gamma,\mu_1,\ldots\mu_K,\lambda_1,\ldots,\lambda_K|Y) \propto \prod_{n=1}^N \P(y_n|Z,V,\sigma,w_{0,\ldots,D})\cdot\\
&\cdot\P(Z)\P(V|\mu_1,\ldots,\mu_K, \lambda_1,\ldots,\lambda_K)\P(\sigma,\mu,\gamma,\mu_1,\ldots\mu_K,\lambda_1,\ldots,\lambda_K).
\end{aligned}\right.\end{aligned}$$ One can maximize this quantity to obtain MAP estimate, but this is very complicated due to the combinatorial complexity of interactions in $Z$ and, additionally, often leads to overfitting. We use Gibbs sampling procedure for learning the posterior of our model. Due to normal-normal conjugacy and a priori independence of $Z$ and other latent variables, we can derive closed form full conditional (i.e. variable given all the rest and the data) distributions for each of the latent variables in the model.\
#### Updating hyperprior parameters
$$\begin{aligned}
&\sigma \thicksim \Gamma\left(\frac{\alpha_1 + N}{2};\, \frac{\sum_{n=1}^N(y_n - y(x_n,\Theta))^2 + \beta_1)}{2}\right),\\
& \lambda \thicksim \Gamma\left(\frac{\alpha_0 + D + 1}{2};\, \frac{\sum_{i=0}^D(w_i - \mu)^2 + \beta_0}{2}\right),\\
& \mu \thicksim \mathcal{N}\left(\frac{\sum_{i=0}^D w_i + \gamma_0\mu_0}{D+1+\gamma_0};\, \frac{1}{\lambda(D+1+\gamma_0)}\right),\\
& \lambda_k \thicksim \Gamma\left(\frac{\alpha_0 + D}{2};\, \frac{\sum_{i=1}^D(v_{ik} - \mu_k)^2 + \beta_0}{2}\right),\\
& \mu_k \thicksim \mathcal{N}\left(\frac{\sum_{i=1}^D v_{ik} + \gamma_0\mu_0}{D+\gamma_0};\, \frac{1}{\lambda_k(D+\gamma_0)}\right),\\
& \text{for }k=1,\ldots ,K.\nonumber\end{aligned}$$
#### Updating factorization coefficients $V$
For updating coefficients of the model we can utilize the multi-linear property also used for the Factorization Machines MCMC updates [@freudenthaler2011bayesian]. Note that for any $\theta \in \{w_0,\ldots, w_D, v_{11},\ldots, v_{DK}\}$ we can write $y(x,\Theta) = l_\theta(x) + \theta m_\theta(x)$, where $l_\theta(\cdot)$ are all the terms independent of $\theta$ and $m_\theta(\cdot)$ are the terms multiplied by $\theta$. For example, if $\theta=w_0$, then $m_\theta(x)=1$ and $l_\theta(x)=\sum_{i=1}^Dw_ix_{i} + \sum_{j=1}^J\sum_{k=1}^K\prod_{i\in Z_j}x_i v_{ik}$. Next we give updating distribution that can be used for any $\theta \in \{w_0,\ldots, w_D, v_{11},\ldots, v_{DK}\}$. $$\begin{aligned}
\label{mult_theta}
\left.\begin{aligned}
&\theta \thicksim \mathcal{N}(\mu^*_\theta, \sigma_\theta^2),\text{where }\sigma_\theta^2 = \left(\sigma\sum_{n=1}^N m_\theta(x_n)^2 + \lambda_\theta\right)^{-1},\\
& \mu^*_\theta = \sigma_\theta^2\left(\sigma\sum_{n=1}^N (y_n - l_\theta(x_n))m_\theta(x_n) + \mu_\theta\lambda_\theta\right),
\end{aligned}\right.\end{aligned}$$ and $\mu_\theta,\lambda_\theta$ are the corresponding hyperprior parameters.
#### Updating interactions $Z$
Posterior updates of $Z$ can be decomposed into prior times the likelihood: $$\P(Z_i|Z_{-i},V,Y) \propto \P(Z_i|Z_{-i})\P(Y|V,Z),$$ where second part is the Gaussian likelihood as in Eq. . To sample $Z_i|Z_{-i}$ we use the construction from Section \[ffm\_all\], where we first sample the value of $Z_{\sigma_D}$ for fixed $j$: $$\begin{aligned}
\left.\begin{aligned}
\P(Z_{\sigma_D}=1|Z) & = \P(\sigma_D = i| M_D=m, Z_i=1) = \\
& = \frac{\P(M_{D-1} = m - 1)\P(Z_{\sigma_D}=1|M_{D-1}=m-1)}{\P(M_{D} = m)},
\end{aligned}\right.\end{aligned}$$ and then uniformly choose and index $i$ to update among $\{i:Z_i=Z_{\sigma_D}\}$. Next $Z_i$ can simply be updated using the process construction \[nex\] assuming it to be last. Recall that $\P(M_{D} = m)$ should be computed beforehand using Eq. .
[^1]: Code is available at <https://github.com/moonfolk/MiFM>.
[^2]: Random Forest Regression and optimization based FM showed worse results than other methods.
[^3]: FM, SVM and logistic regression had low accuracy of around 50% and are not reported.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Developing efficient and guaranteed nonconvex algorithms has been an important challenge in modern machine learning. Algorithms with good empirical performance such as stochastic gradient descent often lack theoretical guarantees. In this paper, we analyze the class of homotopy or continuation methods for global optimization of nonconvex functions. These methods start from an objective function that is efficient to optimize (e.g. convex), and progressively modify it to obtain the required objective, and the solutions are passed along the homotopy path. For the challenging problem of tensor PCA, we prove global convergence of the homotopy method in the “high noise” regime. The signal-to-noise requirement for our algorithm is tight in the sense that it matches the recovery guarantee for the [*best*]{} degree-$4$ sum-of-squares algorithm. In addition, we prove a phase transition along the homotopy path for tensor PCA. This allows us to simplify the homotopy method to a local search algorithm, viz., tensor power iterations, with a specific initialization and a noise injection procedure, while retaining the theoretical guarantees.'
author:
- 'Anima Anandkumar[^1] Yuan Deng[^2] Rong Ge[^3] Hossein Mobahi[^4]'
bibliography:
- 'bib.bib'
title: Homotopy Analysis for Tensor PCA
---
Introduction
============
Non-convex optimization is a critical component in modern machine learning. Unfortunately, theoretical guarantees for nonconvex optimization have been mostly negative, and the problems are computationally hard in the worst case. Nevertheless, simple local-search algorithms such as stochastic gradient descent have enjoyed great empirical success in areas such as deep learning. As such, recent research efforts have attempted to bridge this gap between theory and practice.
For example, one property that can guarantee the success of local search methods over nonconvex functions is when all local minima are also the global minima. Interestingly, it has been recently proven that many well known nonconvex problems do have this property, under mild conditions. Consequently, local-search methods, which are designed to find a local optimum, automatically achieve global optimality. Examples of such problems include matrix completion [@ge2016matrix], orthogonal tensor decomposition [@anandkumar2014tensor; @ge2015escaping], phase retrieval [@DBLP:journals/corr/SunQW16], complete dictionary learning [@sun2015complete], and so on. However, such a class of nonconvex problems is limited, and there are many practical problems with poor local optima, where local search methods can fail.
The above property, while very helpful, imposes a strong assumption on the nonconvex problem. A less restrictive requirement for the success of local search methods is the ability to initialize local search in the basin of attraction of the global optimum using another polynomial-time algorithm. This approach does not require all the local optima to be of good quality, and thus can cover a broader set of problems. Efficient initialization strategies have recently been developed for many nonconvex problems such as overcomplete dictionary learning [@arora2014new; @agarwal2014learning], tensor decomposition [@anandkumar2015learning], robust PCA [@netrapalli2014non], mixed linear regression [@DBLP:journals/corr/YiCS16] and so on.
Although the list of such tractable nonconvex problems is growing, currently, the initialization algorithms are problem-specific and as such, cannot be directly extended to new problems. An interesting question is whether there exist common principles that can be used in designing efficient initialization schemes for local search methods. In this paper, we demonstrate how a class of homotopy continuation methods may provide such a framework for efficient initialization of local search schemes.
Homotopy Method
---------------
The homotopy method is a general and a problem independent technique for tackling nonconvex problems. It starts from an objective function that is efficient to optimize (e.g. convex function), and progressively transforms it to the required objective [@MobahiLink]. Throughout this progression, the solution of each intermediate objective is used to initialize a local search on the next one. A particular approach for constructing this progression is to smooth the objective function. Precisely, the objective function is convolved with the Gaussian kernel and the amount of smoothing is varied to obtain the set of transformations. Intuitively, smoothing “erases wiggles” on the objective surface (which can lead to poor local optima), thereby resulting in a function that is easier to optimize. Global optimality guarantees for the homotopy method have been recently established [@HosseinAAAI; @HazanICML16]. However, the assumptions in these results are either too restrictive [@HosseinAAAI] or extremely difficult to check [@HazanICML16]. In addition, homotopy algorithms are generally slow since local search is repeated within each instantiation of the smoothed objective.
In this paper, we address all the above issues for the nonconvex tensor PCA problem. We analyze the homotopy method and guarantee convergence to global optimum under a set of transparent conditions. Additionally, we demonstrate how the homotopy method can be drastically simplified without sacrificing the theoretical guarantees. Specifically, by taking advantage of the phase transitions in the homotopy path, we can avoid the intermediate transformations of the objective function. In fact, we can start from the extreme case of “easy” (convex) function of the homotopy, and use its solution to initialize local search on the original objective. Thus, we show that the homotopy method can serve as a problem independent principle for obtaining a smart initialization which is then employed in local search methods. Although we limit ourselves to the problem of tensor PCA in this paper, we expect the developed techniques to be applicable for a broader set of nonconvex problems.
Tensor PCA
----------
Tensor PCA problem is an extension of the matrix PCA. The statistical model for tensor PCA was first introduced by [@richard2014statistical]. This is a single spike model where the input tensor $ \bm T \in {\mathbb{R}}^{n\times n\times n}$ is a combination of an unknown rank-$1$ tensor and a Gaussian noise tensor $\bm A$ with $\bm A_{ijk}\sim {\mathcal{N}}(0,1)$ for $i,j,k\in [n].$ $$\bm T = \tau \bm v\otimes \bm v \otimes \bm v + \bm A,$$ where $\bm v\in {\mathbb{R}}^n$ is the signal that we would like to recover.
Tensor PCA belongs to the class of “needle in a haystack” or high dimensional denoising problems, where the goal is to separate the unknown signal from a large amount of random noise. Recovery in the high noise regime has intimate connections to computational hardness, and has been extensively studied in a number of settings. For instance, in the spiked random matrix model, the input is an additive combination of an unknown rank-$1$ matrix and a random noise matrix. The requirement on the signal-to-noise ratio for simple algorithms, such as principal component analysis (PCA), to recover the unknown signal has been studied under various noise models [@perry2016optimality; @bloemendal2013limits] and sparsity assumptions on the signal vector [@berthet2013optimal].
Previous algorithms for tensor PCA belong to two classes: local search methods such as tensor power iterations [@richard2014statistical], and global methods such as sum of squares [@hopkins2015tensor]. Currently, the best signal-to-noise guarantee is achieved by the sum-of-squares algorithm and the flattening algorithm, which are more expensive compared to power iterations (see Table \[table:comparison\]). In this paper, we analyze the Gaussian homotopy method for tensor PCA, and prove that it matches the best known signal-to-noise performance. [@hopkins2015tensor] also showed a lowerbound that no degree-$4$ (or lower) sum-of-squares algorithm can achieve better signal-to-noise ratio, implying that our analysis is likely to be tight.
Method Bound on $\tau$ Time Space
---------------------------------------------------------------- ---------------------------- ------------------ ------------------
[**Power method + initialization + noise injection (ours)**]{} $\tilde{\Omega}(n^{3/4}) $ $\tilde{O}(n^3)$ $\tilde{O}(n)$
Power method, random initialization $\tilde{\Omega}(n)$ $\tilde{O}(n^3)$ $\tilde{O}(n)$
Sum-of-Squares $\tilde{\Omega}(n^{3/4})$ $>\Omega(n^6)$ $>\Omega(n^6)$
Recover and Certify $\tilde{\Omega}(n^{3/4})$ $\tilde{O}(n^5)$ $O(n^4)$
Eigendecomposition of flattened matrix $\tilde{\Omega}(n^{3/4})$ $\tilde{O}(n^3)$ $\tilde{O}(n^2)$
Information-theoretic $\tilde{\Omega}(\sqrt{n})$ Exp $O(n)$
Contributions
-------------
We analyze a simple variant of the popular tensor power method, which is a local search method for finding the best rank-$1$ approximation of the input tensor. We modify it by introducing a specific initialization and injecting appropriate random noise in each iteration. This runs almost in linear time; see Table \[table:comparison\] for more details.
There is an almost linear time algorithm for tensor PCA that finds the signal $\bm v$ as long as the signal strength $\tau = \tilde{\Omega}(n^{3/4})$.
Our algorithm achieves the [*best possible trade-offs*]{} among all known algorithms (see Table \[table:comparison\]).
Our algorithm is inspired by the homotopy framework. In particular, we establish a phase transition along the homotopy path.
Under a plausible independence conjecture, there is a threshold $\theta$ such that if the radius of smoothing is significantly larger than $\theta$, the smoothed function will have a unique local and global maximum. If the radius of smoothing is smaller, then the smoothed function can have multiple local maxima, but one of them is close to the signal vector $\bm v$.
The above result allows us to skip the intermediate steps in the homotopy path. We only need two end points of the homotopy path: the original objective function with no smoothing and with an infinite amount of smoothing. The optimal solution for the latter can be obtained through any local search method; in fact, in our case, it has a closed form. This serves as initialization for the original objective function. In the proof we also design a new noise injection procedure that breaks the dependency between the steps. This allows for simpler analysis and our algorithm does not rely on the independence conjecture. We discuss this in more detail in Section \[sec:resampling\].
The comparison of all the current algorithms for tensor PCA is given in Table \[table:comparison\]. Note that the space in the table does not include the space for storing the tensor, this is because the more practical algorithms only access the tensor for a very small number of passes, which allows the algorithms to be implemented online and do not need to keep the whole tensor in the memory. We see that our algorithm has the best performance across all the measures. In our synthetic experiments (see Section \[sec:exp\], we find that our method significantly outperforms the other methods: it converges to a better solution faster and with a lower variance.
Preliminaries {#sec:prelim}
=============
In this section, we formally define the tensor PCA problem and its associated objective function. Then we show how to compute the smoothed versions of these objective functions.
Tensors and Polynomials
-----------------------
Tensors are higher dimensional generalization of matrices. In this paper we focus on 3rd order tensors, which correspond to a 3 dimensional arrays. Given a vector $\bm v\in {\mathbb{R}}^n$, similar to rank one matrices $\bm v \bm v^\top$, we consider rank 1 tensors $\bm v^{\otimes 3}$ to be a $n\times n\times n$ array whose $i,j,k$-th entry is equal to $\bm v_i \bm v_j \bm v_k$.
For a matrix $\bm M$, we often consider the quadratic form it defines: $\bm x^\top \bm M \bm x$. Similarly, for a tensor $\bm T \in {\mathbb{R}}^{n\times n\times n}$, we define a degree 3 polynomial $\bm T(\bm x, \bm x, \bm x) = \sum_{i,j,k=1}^n \bm T_{i,j,k} \bm x_i \bm x_j \bm x_k$. This polynomial is just a special trilinear form defined by the tensor. Given three vectors $\bm x,\bm y, \bm z$, the trilinear form $\bm T(\bm x, \bm y, \bm z) = \sum_{i,j,k=1}^n \bm T_{i,j,k} \bm x_i \bm y_j \bm z_k$. Using this trilinear form, we can also consider the tensor as an operator that maps vectors to matrices, or two vectors into a single vector. In particular, $\bm T(\bm x, :, :)$ is a matrix whose $i,j$-th entry is equal to $\bm T(\bm x, \bm e_i, \bm e_j)$ where $\bm e_i$ is the $i$-th basis vector. Similarly, $\bm T(\bm x, \bm y, :)$ is a vector whose $i$-th coordinate is equal to $\bm T(\bm x, \bm y, \bm e_i)$.
Since the tensor $\bm T$ we consider is not symmetric ($\bm A_{ijk}$ is not necessarily equal to $\bm A_{jik}$ or other permutations), we also define the symmetric operator
$$\delta(\bm x) = \bm A(\bm x,\bm x,:)+\bm A(\bm x,:,\bm x)+\bm A(:,\bm x,\bm x).$$
Objective Functions for Tensor PCA
----------------------------------
We first define the tensor PCA problem formally.
Given input tensor $\bm T = \tau \cdot \bm v^{\otimes 3} + \bm A$, where $\bm v \in \mathbb R^n$ is an arbitrary unit vector, $\tau \geq 0$ is the signal-to-noise ratio, and $\bm A$ is a random noise tensor with iid standard Gaussian entries, recover the signal $\bm v$ approximately (find a vector $\|\bm x\| = 1$ such that ${\langle \bm x, \bm v \rangle} \ge 0.8$).
Similar to the Matrix PCA where we maximize the quadratic form, for tensor PCA we can focus on optimizing the degree 3 polynomial $f(\bm x) = \bm T(\bm x, \bm x, \bm x)$ over the unit sphere. $$\begin{aligned}
\max_{\bm x} \quad &f(\bm x) = \tau \langle \bm v, \bm x \rangle^3 + \bm A(\bm x,\bm x,\bm x) \label{eq:obj}\\
& \|\bm x\| = 1 \nonumber \end{aligned}$$
The optimal value of this program is known as the spectral norm of the tensor. It is often solved in practice by tensor power method. [@richard2014statistical] noticed that:
When $\tau \ge C\sqrt{n}$ for large constant $C$, the global optimum of (\[eq:obj\]) is close to the signal $\bm v$.
Unfortunately, solving this optimization problem is NP-hard in the worst-case [@Hillar]. Currently, the best known algorithm uses sum-of-squares hierarchy and works when $\tau \ge Cn^{3/4}$. There is a huge gap between what’s achievable information theoretically ($O(\sqrt{n})$) and what can be achieved algorithmically ($\Omega(n^{3/4})$).
Gaussian Smoothing for the Objective Function
---------------------------------------------
Guaranteed homotopy methods rely on smoothing the objective function by the Gaussian kernel [@MobahiLink; @HosseinAAAI]. More precisely, smoothing the objective (\[eq:obj\]) requires convolving it with the Gaussian kernel. Let $g : \mathcal X \times \mathcal {\mathbb{R}}^+ \to \mathbb R$ be a mapping such that $$g(\bm x, t) = [f \star k_{t}](\bm x)$$ Here, $k_{t}$ is the Gaussian density function for $\mathcal{N}(\boldsymbol{0}, t^2 \boldsymbol{I}_n)$, satisfying $$k_{t}(\bm x) = \dfrac{1}{(\sqrt{2 \pi}t)^n} \cdot e^{-\frac{\|\bm x\|_2^2}{2 t^2}}.$$
It is known that convolution of polynomials with the Gaussian kernel has a closed form expression [@ClosedGauss16]. In particular, the objective function of the Tensor PCA has the following smoothed form.
\[Smoothed Tensor PCA Objective\] \[lem:smooth\] The smoothed objective has the form $$g(\bm x, t) = \tau \langle \bm v, \bm x \rangle^3 + t^2 \langle 3 \tau \bm v + \bm u, \bm x \rangle + \bm A(\bm x,\bm x,\bm x),$$ where the vector $\bm u$ is defined by $\bm u_j = \sum_{i=1}^n (\bm A_{iij}+\bm A_{iji} + \bm A_{jii})$. Moreover, it is easy to compute vector $\bm z = 3\tau \bm v+\bm u$ given just the tensor $\bm T$, as $\forall j, \bm z_j=\sum_{i=1}^n (\bm T_{iij}+\bm T_{iji}+\bm T_{jii})$.
The proof of this Lemma is based on interpreting the convolution as an expectation ${\mathbb{E}}_{y\sim N(\bm 0, \mathbf I_n)}[f(x+y)]$. We defer the detailed calculation to Appendix \[app:lemma1\]
Tensor PCA by Homotopy Initialization {#sec:simplealg}
=====================================
In this section we give a simple smart initialization algorithm for tensor PCA. Our algorithm only uses two points in homotopy path – the infinite smoothing $t\rightarrow \infty$ and the no smoothing $t \rightarrow 0$. This is inspired by our full analysis of the homotopy path (see Section \[sec:path\]), where we show there is a [*phase transition*]{} in the homotopy path. When the smoothing parameter is larger than a threshold, the function behaves like the infinite smoothing case; when the smoothing parameter is smaller than the threshold, the function behaves like the no smoothing case.
Recall that the smoothed function $g(\bm x, t)$ is: $$g(\bm x, t) = \tau \langle \bm v, \bm x \rangle^3 + t^2 \langle 3 \tau \bm v + \bm u, \bm x \rangle + \bm A(\bm x,\bm x,\bm x)$$ with $\bm u$ as a vector such that $\bm u_j = \sum_i \bm A_{iij} + \bm A_{iji} + \bm A_{jii}$. When $t \rightarrow \infty$, the solution of the smoothed problem has the special form $\bm x^{\dag} = \frac{3 \tau \bm v + \bm u}{\|3 \tau \bm v + \bm u\|}$. That is because the term $t^2 \langle 3 \tau \bm v + \bm u, \bm x \rangle$ dominates $g$ and thus its maximizer under $\|\bm x \|_2 = 1$ yields $\bm x^{\dag}$.
Note that by Lemma \[lem:smooth\], we can compute vector $\bm z$ $\bm z_j = \sum_{i=1}^n \bm T_{iij}+\bm T_{iji} + \bm T_{jii}$, and we know $\bm z = 3\tau \bm v+\bm u$. Therefore we know $\bm x^\dag = \frac{\bm z}{\|\bm z\|_2}$ can be computed just from the tensor. We use this point as an initialization, and then run power method on the original function. The resulting algorithm is described in Algorithm \[algo:simple\].
In order to analyze the algorithm, we use the following [*independence*]{} condition, which states that the “random”-looking vectors $\bm u$ and $\delta(\bm x^p) = \bm A(\bm x^p, \bm x^p, :)+\bm A(\bm x^p, :, \bm x^p)+\bm A(:, \bm x^p, \bm x^p)$ indeed have some properties satisfied by random vectors:
\[assump\]\[Independence Condition\] The norm and correlation with $\bm v$ for the vectors $\bm u$ and $\delta(\bm x^p)$ are not far from expectation. More precisely: (1) $\|\bm u\|_2 = O(n \sqrt{m})$ and $|\langle \bm u, \bm v \rangle| = O(\sqrt {n m \log n})$; (2) for the sequence computed by Algorithm \[algo:simple\], $\bm x^0, \bm x^1, \cdots, \bm x^m$, $\forall 0 \leq p \leq m$, $\|\delta(\bm x^p)\|_2 = O(\sqrt{n m}) \|\bm x^p\|_2^2$ and $|\langle \delta(\bm x^p), \bm v \rangle| = O(\sqrt{m \log n}) \|\bm x^p\|_2^2$.
Note that if in every step of the algorithm, the noise tensor $\bm A$ is [*resampled*]{} to be a fresh random tensor, independent of the previous step $\bm x^p$, then $\delta(\bm x^p)$ is just a random Gaussian vector. In this case the condition is trivially satisfied. Of course, in reality $\bm x^i$’s are dependent on $\bm A$. However, we are able to modify the algorithm by a [*noise injection*]{} procedure, that adds more noise to the tensor $\bm T$, and make the noise tensor “look” as if they were independent. The extra dependency on $m$ in Condition \[assump\] comes from noise injection procedure and will be more clear in Section \[sec:resampling\]. We will first show the correctness of the algorithm assuming independence condition here, and in Section \[sec:resampling\] we discuss the noise injection procedure and prove the independence condition.
\[thm:converge\] When $\tau \ge Cn^{3/4}\log n$ for a large enough constant $C$, under the Independence Condition (Condition \[assump\]), Algorithm \[algo:simple\] finds a vector $\bm x^m$ such that ${\langle \bm x^m, \bm v \rangle} \geq 0.8$ in $O(\log \log n)$ iterations.
$m = O(\log \log n)$;\
$\forall~j, \bm x^0_j = \sum_i \bm T_{iij} + \bm T_{iji} + \bm T_{jii}$;\
$\bm x^0 = \bm x^0 / \|\bm x^0\|$; // Now $\bm x^0 = \bm x^\dag$\
;
The main idea is to show the correlation of $\bm x^k$ and $\bm v$ increases in every step. In order to do this, first notice that the initial point $\bm x^\dag$ itself is equal to a normalization of $3\tau \bm v+\bm u$, where the norm of $\bm u$ and its correlation with $\bm v$ are all bounded by the Independence Condition. It is easy to check that ${\langle \bm x^0,\bm v \rangle} \gg n^{-1/4}$, which is already non-trivial because a random vector would only have correlation around $n^{-1/2}$. For the later iterations, let $\hat{\bm x}^k$ be the vector $\bm x^k$ before normalization and we have $\hat{\bm x}^{k+1} = 3 \tau \langle \bm v, \bm x^k \rangle^2 \bm v + \delta(\bm x^k)$. Notice that the first term is in the direction $\bm v$, and the Independence Condition bounds the norm and correlation with $\bm v$ for the second term. We can show that the correlation with $\bm v$ increases in every iteration, because the initial point already has a large inner product with $\bm v$. The detailed proof is deferred to Appendix \[app:theorem3\].
Noise Injection Procedure {#sec:resampling}
-------------------------
$m = O(\log \log n)$;\
Sample $\bm B^0, \bm B^1, ..., \bm B^{m - 1} \in {\mathbb{R}}^{n\times n\times n}$ whose entries are $\mathcal N(0,m)$.\
Let $\overline{\bm B} = \frac{1}{m} \sum_{p=0}^{m-1} \bm B^p$.\
Let $\bm T^p = \bm T-\bar{\bm B}+\bm B^p$\
$\forall~j, \bm x^0_j = \sum_i \bm T^0_{iij} + \bm T^0_{iji} + \bm T^0_{jii}$;\
$\bm x^0 = \bm x^0 / \|\bm x^0\|$;\
;
In order to prove the Independence Condition, we slightly modify the algorithm (see Algorithm \[algo:resample\]). In particular, we add more noise in every step as follows
- Get the input tensor $\bm T = \tau \cdot \bm v^{\otimes 3} + \bm A$;
- Draw a sequence of $\bm B^p \in \mathbb R^{n \otimes 3}$ such that $\bm B_{ijk}^p \sim \mathcal N(0, m)$;
- Let $\bm T^p = \bm T - \overline{\bm B} + \bm B^p$ with $\overline{\bm B} = \frac{1}{m} \sum_{p=0}^{m-1} \bm B^p$, run Algorithm \[algo:resample\] by using $\bm T^p$ in the $p$-th iteration;
Intuitively, by adding more noise the new noise will overwhelm the original noise $\bm A$, and every time it looks like a fresh random noise. We prove this formally by the following lemma:
\[lem:resample\] Let the sequence $\bm T^0, \cdots, \bm T^{m-1}$ be generated according to Algorithm \[algo:resample\]. Let $\bm Q^i = \tau \bm v^{\otimes 3}+\bm C^i$, where $\bm C^i$’s are tensors with independent Gaussian entries. Each entry in $\bm C^i$ is distributed as $N(0,m)$. The two sets of variables $\{\bm T^i\}$ and $\{\bm Q^i\}$ has the same distribution.
This Lemma states that after our noise injection procedure, the tensors $\bm T^0,...,\bm T^{m-1}$ look [*exactly the same*]{} as tensors where the noise $\bm A$ is sampled independently. The basic idea for this lemma is that for two multivariate Gaussians to have the same distribution, we only need to show that they have the same first and second moments. We defer the details to Appendix \[app:theorem3\].
Using Lemma \[lem:resample\] we can create a sequence of $T^p$ such that its noise tensor $\bm A^p = \bm A - \overline{\bm B} + \bm B^p$ is redrawn independently and each element is according to $\mathcal N(0, m)$. Now, because each $\bm T^i$ behave as if it is drawn independently, we can prove the Independence Condition:
\[lem:resamplingassump\] Let $\bm T^p$ be generated according to Algorithm \[algo:resample\] and $\bm A^p = \bm T^p - \tau \bm v^{\otimes 3}$. Let $\bm u^0$ be a vector such that $\bm u^0_j = \sum_i \bm A_{iij}^0 + \bm A_{iji}^0 + \bm A_{jii}^0$, and $\delta^p(\bm x^p) = \bm A^p(\bm x^p, \bm x^p, :)+\bm A^p(\bm x^p, :, \bm x^p)+\bm A^p(:, \bm x^p, \bm x^p)$. With high probability[^5], (1) $\|\bm u^0\|_2 = \Theta(n \sqrt{m})$ and $|\langle \bm u^0, \bm v \rangle| = O(\sqrt {n m \log n})$; (2) for the sequence computed by Algorithm \[algo:resample\], $\bm x^0, \bm x^1, \cdots, \bm x^{m-1}$, $\forall~ 0 \leq p \leq m-1$, $\|\delta^p(\bm x^p)\|_2 = \Theta(\sqrt{n m}) \|\bm x^p\|_2^2$ and $|\langle \delta^p(\bm x^p), \bm v \rangle| = O(\sqrt{m \log n}) \|\bm x^p\|_2^2$. As a result Condition \[assump\] is satisfied.
This Lemma is now true because by Lemma \[lem:resample\], we know the noise tensors $\bm A^p$ is independent of $\bm A^0,...,\bm A^{p-1}$. As a result $\bm A^p$ is [*independent*]{} of $\bm x^p$! This lemma then follows immediately from standard concentration inequalities. We defer the full proof to Appendix \[app:theorem3\].
The noise injection technique is mostly a technicality that we need in order to make sure different steps are independent. This is standard in analyzing nonconvex optimization algorithms. As an example, previous works on alternating minimization for matrix completion [@jain2013low] relied on the availability of different subsamples in different iterations to obtain the theoretical guarantees. Our noise injection procedure is very similar, however this is the first application of this idea for the case of Gaussian noise. The main usage of the noise injection is to get rid of the dependence of the noise matrix between different iterations. Moreover, this technique is designed to simplify the proof and rarely used in the real applications. In practice, an algorithm without noise injection, like Algorithm \[algo:simple\], usually performs well enough.
Combining Lemma \[lem:resamplingassump\] and Theorem \[thm:converge\], we know Algorithm \[algo:resample\] solves the tensor PCA problem when $\tau \ge Cn^{3/4}\log n$.
In the above analysis, we assume the variance of entries of $\bm A$ is $1$. In practice, we can estimate the variance $\sigma^2$ of entries of $\bm A$ from $\bm T$ by computing its Frobenius norm. Note that when $\tau$ is large, the simple power method already performs well. The interesting case is when $\tau$ is small, say $\tau < n$. In this case, the square of the Frobenius norm of $\tau \bm v^{\otimes 3} = \tau^2$ while the square of the Frobenius norm of the noise matrix $\bm A$ in expectation is $\sigma^2 n^3$ with variance $\sigma^2 O(n^3)$. Therefore, we can get a good estimation of $\sigma^2$ by computing the square of the Frobenius norm of $\bm A$ divided by $n^3$.
Characterizing the Homotopy Path
================================
{width="0.9\linewidth"}\
[(a) $t \gg n^{-1}$]{}
{width="0.9\linewidth"}\
[(b) $t \approx n^{-1}$]{}
{width="0.9\linewidth"}\
[(c) $t \ll n^{-1}$]{}
This section analyzes the behavior of the smoothed objective function $g$ as $t$ varies. Under a plausible conjecture, we prove that a phase transition occurs: when $t$ is large $g(\bm x ,t)$ behaves very similarly to $g(\bm x ,\infty)$ and when $t$ is small $g(\bm x ,t)$ behaves very similarly to $g(\bm x ,0)$. This motivates the algorithms in the previous section, as the phase transition suggests the most important regimes are very large $t$ and $t=0$.
In this section we first describe how the homotopy method works in more details. Then we present an alternative objective function of Tensor PCA and derive its smoothed version. Finally, we prove that when $t \gg n^{-1}$, the smoothed function retains its maximizer around $\bm x^\dag$. However, when $t \ll n^{-1}$, the configuration of critical points change, with only one of the critical points being close to the solution $\bm v$. Importantly, we can find our way from the vicinity of $\bm x^\dag$ toward this critical point via the dominant curvature direction of the function.
\[sec:path\]
Homotopy
--------
In the homotopy method, we start from the maximizer of the function with large amount of smoothing $t \rightarrow \infty$. We earlier denoted this maximizer as $\bm x^\dag$. Then we continuously decrease the amount of smoothing $t$, while following the maximizer throughout this process, until reaching $t = 0$. We call the path taken by the maximizer the homotopy path. It is formally defined as follows.
A homotopy path $\bm x(t)$ is a continuous function $\bm x : \mathcal T \to \mathcal X$ satisfying $\lim_{t \to \infty} \bm x(t) = \bm x^{\dag}$ and $\forall ~ t \geq 0$, $\nabla g(\bm x(t), t) = 0$, where the gradient $\nabla$ is w.r.t. to the first argument of $g$.
In practice, to search a homotopy path, one computes the initial point $\bm x^{\dag}$ by analytical derivation or numerical approximation as $\arg \max_{\bm x} g(\bm x, t)$ and then successively minimizes the smoothed functions over a finite sequence of decreasing numbers $t_0$ to $t_m$, where $t_0$ is sufficiently large, and $t_m=0$. The resulted procedure is listed in Algorithm \[algo:homotopy\].
$\bm x^{t_0} = \mbox{global maximizer of }g(\bm x, t_0)$;\
$\bm x^{t_m}$.
Alternative Objective Function and Its Smoothing
------------------------------------------------
Turning a constrained problem into an unconstrained problem can facilitate the computation of the effective gradient and Hessian of $g(\bm x, t)$. In this section, we consider the alternative objective function: we modify $f(\bm x)$ by adding the penalty term $-\frac {3 \tau} 4 \|\bm x\|_2^4$: $$f_r(\bm x) = \tau \langle \bm v, \bm x \rangle^3 + \bm A(\bm x,\bm x,\bm x) - \frac {3 \tau} 4 \|\bm x\|_2^4$$
Thus we consider the following unconstrained optimization problem, $$\max \quad f_r(\bm x) = \bm T(\bm x, \bm x, \bm x) - \frac{3\tau}{4}\|\bm x\|_2^4. \label{eq:obj2}$$
If we fix the magnitude $\|\bm x\| = 1$, the function $f_r(\lambda \bm x)$ is $\lambda^3 \bm T(\bm x, \bm x, \bm x) - \frac{3\tau}{4} \lambda^4$. The optimizer of this is an increasing function of $\bm T(\bm x, \bm x, \bm x)$. Therefore the maximizer of (\[eq:obj2\]) is exactly in the same direction as the constrained problem (\[eq:obj\]). The $3\tau/4$ factor here is just to make sure the optimal solution has roughly unit norm; in practice we can choose any coefficient in front of $\|\bm x\|^4$ and the solution will only differ by scaling.
Moreover, note that if in the absence of noise tensor $\bm A$, then $$\nabla f_r(\bm x) = 3 \tau \langle \bm v, \bm x \rangle^2 \bm v - \frac {3 \tau} 4 \cdot 4 \|\bm x\|_2^2 \bm x \\$$ To get the stationary point, we have $$\bm x = \frac {3 \tau} 4 \cdot \frac {\langle \bm v, \bm x \rangle^2}{\frac {3 \tau} 4 \cdot \|\bm x\|_2^2} \bm v = \bm v$$ Therefore, the new function $f_r(\bm x)$ is defined on $\mathbb R^n$ and the maximizer of $\mathbb R^n$ is close to $\bm v$. We also compute the smoothed version of this problem:
\[lem:smoothalternative\] The smoothed version of the alternative objective is $$\begin{aligned}
g_r(\bm x, t) = \tau \langle \bm v, \bm x \rangle^3 + t^2 \langle 3 \tau \bm v + \bm u, \bm x \rangle + \bm A(\bm x,\bm x,\bm x) - \frac {3 \tau} 4 \left(\|x\|_2^4 + 2 t^2 (n+2) \|x\|_2^2 + t^4 (n^2+2n)\right)
\end{aligned}$$ Its gradient and Hessian are equal to $$\label{eq:gradient}
\nabla g_r(\bm x, t) = 3 \tau \langle \bm v, \bm x \rangle^2 \bm v + t^2 (3 \tau \bm v + \bm u)
+ \delta(\bm x) - 3 \tau (\|\bm x\|_2^2 \bm x + t^2 (n+2) \bm x).$$ and $$\begin{aligned}
\nabla^2 g_r(\bm x, t) &~= -3 \tau ((\|\bm x\|_2^2 + t^2 (n+2)) \bm I - 2 \langle \bm v, \bm x \rangle \bm v \bm v^T + 2 \bm x \bm x^T) \nonumber \\
&~+ P_{sym}[\bm A(\bm x, :, :) + \bm A(:, \bm x, :) + \bm A(:, :, \bm x)]. \label{eq:hessian}
\end{aligned}$$ Here $P_{sym} \bm M = \frac{\bm M+\bm M^\top}{2}$ is the projection to symmetric matrices.
The proof of this Lemma is very similar to Lemma \[lem:smooth\] and is deferred to Appendix \[app:path\].
Phase Transition on the Homotopy Path
-------------------------------------
Notice that when $t\to \infty$, the dominating terms in $g_r(\bm x,t)$ are $t^2$ terms (the only $t^4$ term is a constant). Therefore, $g_r(\bm x,t)$ forms a quadratic function, so it has a unique global maximizer equal to $\frac{3\tau \bm v+\bm u}{3 \tau (n+2)}$, denoted as $\bm x^\dag$. Notice that this vector has different norm compared to the $\bm x^\dag$ in previous section.
Before we state the Theorem, we need a counterpart of the Independence Condition. We call this the Strong Independence Conjecture:
\[assump2\]\[Strong Independence Conjecture\] Suppose $\bm T = \tau \bm v^{\otimes 3} + \bm A$ and $\bm u_j = \bm A_{iij}+\bm A_{iji}+\bm A_{jii}$, $\delta(\bm x) = \bm A(\bm x,\bm x,:)+\bm A(\bm x,:,\bm x)+\bm A(:, \bm x,\bm x)$ be defined as before. With high probability, (1) $\|\bm u\|_2 = \Theta(n)$ and $|\langle \bm u, \bm v \rangle| = O(\sqrt {n\log n})$; (2) for all $\bm x^{t_k}$ on the homotopy path, $\|\delta(\bm x^{t_k})\|_2 = \Theta(\sqrt{n}) \|\bm x^{t_k}\|_2^2$ and $|\langle \delta(\bm x^{t_k}), \bm v \rangle| = O(\sqrt{\log n}) \|\bm x^{t_k}\|_2^2$,
Intuitively, this assumes that the noise is not adversarially correlated with the signal $\bm v$ on the entire homotopy path. The main difference between the strong independence conjecture and the weak independence conjecture is that they apply to different algorithms with different number of iterations. The strong independence conjecture applies to the general Homotopy method, which may have a large number of iterations, and thus a conjecture that depends on the number of iterations does not provide us any useful properties. We use the strong independence conjecture to analyze the general Homotopy method to gain intuitions in order to design our algorithm. The weak conjecture is for our Algorithm \[algo:simple\], which only has $O(\log \log n)$ rounds, and can be satisfied using the noise injection technique. Although we cannot use noise injection to prove the strong independence conjecture, similar conjectures are often used to get intuitions about optimization problems [@donoho2009message; @javanmard2013state; @choromanska2015loss].
\[thm:homotopypath\] Assuming the Strong Independence Conjecture (Conjecture \[assump2\]), when $\tau = n^{3/4}\log n$,
1. When $t \ge Cn^{-1}$ for a large enough constant $C$, there exists a local maximizer $\bm x^t$ of $g_r(\bm x, t)$ such that $\|\bm x^t - \bm x^\dag\|_2 = o(1) \|\bm x^\dag\|_2$;
2. When $t < n^{-1}\log^{-2} n$, we know there are two types of local maximizers $\bm x^t$:
- $\|\bm x^t\|_2 = \Theta(1)$ and ${\langle \bm v, \bm x^t \rangle} = \Theta(1)$. This corresponds to a local maximizer near the true signal $\bm v$.
- $\|\bm x^t\|_2 = \Theta(n^{-\frac 1 4} \log^{-1} n)$ and ${\langle \bm v, \bm x^t \rangle} = O(n^{-\frac 1 2} \log^{-1} n)$. These local maximizers have poor correlation with the true signal.
3. When $t < n^{-1}\log^{-2} n$, let $\bm b$ be the top eigenvector of $\nabla^2 (g_r(\bm x^\dag, t))$, we know $\sin\theta(\bm b, \bm v) \le 1/\log^2 n$.
Intuitively, this theorem shows that in the process of homotopy method, if we consider a continuous path in the sense that $t_{k+1} - t_k$ is close to $0$ for all $k$, then (1) at the beginning, $\bm x^k$ is close to $\bm x^\dag$; (2) at some point $k^*$, $\bm x^{k^*}$ is a saddle point in the function $g(\bm x, t_{k^*+1})$ and from the saddle point we are very likely to follow the Hessian direction to actually converge to the good local maximizer near the signal. This phenomenon is illustrated in Figure \[fig:phasetransition\]:
Figure \[fig:phasetransition\](a) has large smoothing parameter, and the function has a unique local/global maximizer. Figure \[fig:phasetransition\](b) has medium smoothing parameter, the original global maximizer now behaves like a local minimizer in one dimension, but it in general could be a saddle point in high dimensions. The Hessian at this point leads the direction of the homotopy path. In Figure \[fig:phasetransition\](c) the smoothing is small and the algorithm should go to a different maximizer.
Next, we are going to show that, when $t = \Theta(n^{-1} \log^{-2c} n)$, starting from $\bm x^\dag$, with high probability, a gradient descent algorithm will converge to the local maximizers with $\|\bm x^t\|_2^2 = \Theta(1)$ and $\langle \bm v, \bm x^t \rangle = \Theta(1)$.
Let $\bm x^0 = \bm x^\dag$;\
the final $\bm x^{k}$.
With high probability, Algorithm \[algo:phase\] returns a vector $\bm x^*$ such that $\langle \bm x^*, \bm v \rangle = \Theta(1)$ and $\|\bm x^*\|_2 = \Theta(1)$.
In the Algorithm \[algo:phase\], the first step is to compute $\alpha$ such that $g_r(\alpha \hat{\bm x}^k, t)$ is maximized. That is, the gradient at point $\alpha \hat{\bm x}^k$ should be perpendicular to $\bm x$, i.e. $\langle \nabla g_r(\alpha \hat{\bm x}^k, t), \hat{\bm x}^k \rangle = 0$. Henceforth, we have $$3 \tau \alpha^2 \langle \bm v, \hat {\bm x}^k \rangle^3 + t^2 (3 \tau \langle \bm v, \hat {\bm x}^k \rangle + \langle \bm u, \hat {\bm x}^k \rangle) + 3 \alpha^2 \bm A(\hat {\bm x}^k, \hat {\bm x}^k, \hat {\bm x}^k) = 3 \tau (\alpha^3 \|\hat {\bm x}^k\|_2^4 + t^2 n \alpha \|\hat {\bm x}^k\|_2^2)$$ Since $\hat {\bm x}^k$ is normalized, we have $$\label{eq:find_alpha}
3 \tau \alpha^2 \langle \bm v, \hat {\bm x}^k \rangle^3 + t^2 (3 \tau \langle \bm v, \hat {\bm x}^k \rangle + \langle \bm u, \hat {\bm x}^k \rangle) + O(1) \alpha^2 = 3 \tau (\alpha^3 + t^2 n \alpha)$$ Recall that $\bm x^\dag = \frac{\bm v} n + \frac{\bm u}{3 \tau n}$ with norm $\Theta(\frac 1 \tau)$ and correlation $\langle \bm x^\dag, \bm v \rangle = \Theta(\frac 1 n)$. For the very first iteration with $k = 0$, we have $$\alpha^2 \Theta(\log^{4c} n) + \Theta(n^{-1} \varepsilon^2(n)) = \alpha^3 \Theta(n^{\frac 3 4} \log^c n) + \alpha \Theta(n^{-\frac 1 4} \log^c n \cdot \varepsilon^2(n))$$ By scale analysis, we can get $\alpha = \Theta(n^{-\frac 3 4} \log^{-c} n)$.
Partition the gradient into three parts: $g_r(\hat{\bm x}^0, t) = 3 \tau \alpha_0^2 \langle \bm v, \hat{\bm x}^0 \rangle^2 \bm v + \alpha_0^2 \delta(\hat{\bm x}^0)$, $t^2 (3 \tau \bm v + \bm u)$ and $- 3 \tau(\alpha_0^3 + t^2 n \alpha_0) \hat{\bm x}^k$. Notice that, $\|g_r(\hat{\bm x}^0, t)\|_2 = \Theta(n^{-1} \log^{-2c} n)$ dominates $\|t^2 (3 \tau \bm v + \bm u)\|_2 = \Theta(n^{-1} \varepsilon^2(n))$, $\|- 3 \tau(\alpha_0^3 + t^2 n \alpha_0) \hat{\bm x}^k\|_2 = \Theta(n^{-1} \varepsilon^2(n))$. Therefore, in $\nabla g_r(\alpha_0 \hat{\bm x}^k, t)$, $g_r(\hat{\bm x}^0, t)$ is the dominant term. Moreover, we have $$\langle \frac{g_r(\hat{\bm x}^0, t)}{\|g_r(\hat{\bm x}^0, t)\|_2}, \bm v \rangle = \Theta(n^{-\frac 1 4} \log^{3c} n)$$ Therefore, after the first step, the algorithm makes improvement, compared to the initial correlation $\langle \hat{\bm x}^0, \bm v \rangle = \Theta(n^{-\frac 1 4} \log^{c} n)$.
In order to show this for the future steps, we claim that for each step $k$, $\langle \hat{\bm x}^k, \bm v \rangle \leq \langle \hat{\bm x}^{k+1}, \bm v \rangle$. If the claim is true, according to Lemma \[lem:middle\_t\], the process can only converge to a good maximizer.
The claim is clearly true for $k = 0$. By induction, assume this is true for $k \geq 1$, in the next iteration, we partition the gradient into three parts, $g_r(\hat{\bm x}^k, t) = 3 \tau \alpha_k^2 \langle \bm v, \hat{\bm x}^k \rangle^2 \bm v + \alpha_k^2 \delta(\hat{\bm x}^k)$, $t^2 (3 \tau \bm v + \bm u)$ and $- 3 \tau(\alpha_k^3 + t^2 n \alpha_k) \hat{\bm x}^k$, as before.
For $g_r(\hat{\bm x}^k, t)$, if $3 \tau \alpha_k^2 \langle \bm v, \hat{\bm x}^k \rangle^2 \bm v$ dominates $\alpha_k^2 \delta(\hat{\bm x}^k)$, then $\langle g_r(\hat{\bm x}^k, t), \bm v \rangle = \Theta(1)$. Otherwise, due to induction hypothesis, $\langle \hat{\bm x}^k, \bm v \rangle = \Omega(n^{-\frac 1 4} \log^{c} n)$ and we have $$\langle \frac{g_r(\hat{\bm x}^k, t)}{\|g_r(\hat{\bm x}^k, t)\|_2}, \bm v \rangle = \Theta(n^{\frac 1 4} \log^c n) \langle \hat{\bm x}^k, \bm v \rangle^2 \geq \langle \hat{\bm x}^k, \bm v \rangle$$ Therefore, the only part that could decrease the correlation is $t^2 (3 \tau \bm v + \bm u)$. However, notice that $\|t^2 (3 \tau \bm v + \bm u)\|_2 = \Theta(t^2 n)$ and $\|- 3 \tau(\alpha_k^3 + t^2 n \alpha_k) \hat{\bm x}^k\|_2 \geq \Theta(\tau \alpha_k t^2 n)$. It suffices to show that $\tau \alpha_k$ dominates $\Theta(1)$. To figure out $\alpha_k$, recall (\[eq:find\_alpha\]) $$\label{eq:alpha_full}
3 \tau \alpha^2 \langle \bm v, \hat {\bm x}^k \rangle^3 + t^2 (3 \tau \langle \bm v, \hat {\bm x}^k \rangle + \langle \bm u, \hat {\bm x}^k \rangle) = 3 \tau (\alpha^3 + t^2 n \alpha)$$ where $\hat{\bm x}^k = g_r(\hat{\bm x}^{k-1}, t) / \|g_r(\hat{\bm x}^{k-1}, t)\|_2$. Recall that $g_r(\hat{\bm x}^{k-1}, t) = 3 \tau \alpha_{k-1}^2 \langle \bm v, \hat{\bm x}^{k-1} \rangle^2 \bm v + \alpha_{k-1}^2 \delta(\hat{\bm x}^{k-1})$, $\|g_r(\hat{\bm x}^{k-1}, t)\|_2 = \alpha_{k-1}^2 \Theta(\sqrt n)$. Henceforth, we have $$\langle \bm u, \hat {\bm x}^k \rangle = \langle \bm v, \hat {\bm x}^k \rangle \langle \bm u, \bm v \rangle + O(n^{-\frac 1 2}) = O(\sqrt n) \langle \bm v, \hat {\bm x}^k \rangle + O(n^{-\frac 1 2})$$ Therefore, (\[eq:alpha\_full\]) becomes $$\label{eq:alpha_final}
\alpha^2 \langle \bm v, \hat {\bm x}^k \rangle^3 + t^2 \langle \bm v, \hat {\bm x}^k \rangle = \alpha^3 + t^2 n \alpha$$ If $\alpha = o(\frac 1 {\tau}) = o(n^{-\frac 3 4} \log^{-c} n)$, then the right hand side of (\[eq:find\_alpha\]) is dominated by $t^2 n \alpha$. But, notice that if $t^2 \langle \bm v, \hat {\bm x}^k \rangle$ dominates the left hand side of (\[eq:find\_alpha\]), $\alpha = \Omega()$ [[\[Yuan: proof breaks down here, steep descent does not work? but can we find a way to guarantee the length of current solution does not decrease?\]]{}]{}
Experiments {#sec:exp}
===========
{width="0.9\linewidth"} [Homotopy PCA]{} \[fig:homotopy\]
{width="0.9\linewidth"} [Power Method]{} \[fig:power\]
{width="0.9\linewidth"} [Flatten Algorithm]{} \[fig:flatten\]
{width="0.9\linewidth"} [$\alpha = 1.1$]{}
{width="0.9\linewidth"} [$\alpha = 1.5$]{}
{width="0.9\linewidth"} [$\alpha = 2$]{}
For brevity we refer to our Tensor PCA with homotopy initialization method (Algorithm \[algo:simple\]) as HomotopyPCA. We compare that with two other algorithms: the Flatten algorithm and the Power method. The Flatten algorithm was originally proposed by [@richard2014statistical], where they show it works when $\tau = \Omega(n)$. [@hopkins2015tensor] accelerated the Flatten algorithm to near-linear time, and improved the analysis to show it works when $\tau = \tilde{\Omega}(n^{3/4})$. The Power method is similar to our algorithm, except it does not use intuitions from homotopy, and initialize at a random vector. Note that there are other algorithms proposed in [@hopkins2015tensor], however they are based on the Sum-of-Squares SDP hierarchy, and even the fastest version runs in time $O(n^5)$ (much worse than the $O(n^3)$ algorithms compared here).
We first compare how often these algorithms successfully find the signal vector $v$, given different values of $\tau$ and $n$. The results are in Figure \[fig:success\_rate\], in which $y$-axis represents $n$ and $x$-axis represents $\tau$. We run 50 experiments for each values of $(n,\tau)$, and the grayness in each grid shows how frequent each algorithm succeeds: black stands for “always fail” and white stands “always succeed”. For every algorithm, we say it fails if (1) when it converges, i.e., the result at two consecutive iterations are very close, the correlation with the signal $\bm v$ is less than $80\%$; (2) the number of iterations exceeds 100. In the experiments for Power Method, we observe there are many cases where situation (1) is true, although our new algorithms can always find the correct solution. In these cases the function indeed have a local maximizer. From Figure \[fig:success\_rate\], our algorithm outperforms both Power Method and the Flatten algorithm in practice. This suggests the constant hiding in our algorithm is possibly smaller.
Next we compare the number of iterations to converge with $n = 500$ and $\tau = \alpha n^{\frac 3 4}$, where $\alpha$ varies in $[1.1, 1.5, 2]$. In Figure \[fig:iteration\_comparison\], the x-axis is the number of iterations, and the $y$ axis is the correlation with the signal $\bm v$ (error bars shows the distribution from 50 independent runs). For all $\alpha$, Homotopy PCA performs well — converges in less than $5$ iterations and finds the signal $\bm v$. The Power Method converges to a result with good correlations with the signal $\bm v$, but has large variance because it sometimes gets trapped in local optima. As for the Flatten algorithm, the algorithm always converges. However, it takes more iterations compared to our algorithm. Also when $\alpha$ is small, the converged result has bad correlation with $\bm v$.
Omitted Proofs
==============
Omitted Proof in Section \[sec:prelim\] {#app:lemma1}
---------------------------------------
\[Lemma \[lem:smooth\] restated\] $$g(\bm x, t) = \tau \langle \bm v, \bm x \rangle^3 + t^2 \langle 3 \tau \bm v + \bm u, \bm x \rangle + \bm A(\bm x,\bm x,\bm x),$$ where the vector $\bm u$ is defined by $\bm u_j = \sum_{i=1}^n (\bm A_{iij}+\bm A_{iji} + \bm A_{jii})$. Moreover, let $\bm z$ be a vector where $\bm z_j = \sum_{i=1}^d(\bm T_{iij}+\bm T_{iji}+\bm T_{jii})$, then we have $\bm z = 3\tau \bm v+\bm u$.
We can write $g(x,t)$ as an expectation $$\begin{aligned}
g(\bm x, t) = \int_{{\mathbb{R}}^n} f(\bm x + \bm y) k_{t}(\bm y) dy & = \mathbb E_{y \sim N(\bm 0, t^2 \mathbf I_n)} [f(\bm x + \bm y)] = \mathbb E_{y \sim N(\bm 0, \mathbf I_n)} [ f(\bm x + t\bm y)]
\end{aligned}$$ Since $f$ is just a degree 3 polynomial, we can expand it and use the lower moments of Gaussian distributions: $$\begin{aligned}
g(\bm x, t) &~= \mathbb E [f(\bm x + t \bm y)] \\
&~= \mathbb E [\tau \langle \bm v, (\bm x + \bm t y) \rangle^3 + \bm A(\bm x + t \bm y, \bm x + t \bm y, \bm x + t \bm y)] \\
&~= \tau \langle \bm v, \bm x \rangle^3 + 3 \tau t^2 \langle \bm v, \bm x \rangle \cdot \mathbb E[\langle \bm v, \bm y \rangle^2] + \mathbb E[\bm A(\bm x + t \bm y, \bm x + t \bm y, \bm x + t \bm y)] \\
&~= \tau \langle \bm v, \bm x \rangle^3 + 3 \tau t^2 \langle \bm v, \bm x \rangle + t^2 \sum_{i,j} (\bm A_{iij} + \bm A_{iji} + \bm A_{jii}) \bm x_j + \bm A(\bm x,\bm x,\bm x)
\end{aligned}$$ Therefore the first part of the lemma holds if we define $\bm u$ to be the vector $\bm u_j = \sum_i \bm A_{iij} + \bm A_{iji} + \bm A_{jii}$. In order to compute the vector $3\tau \bm v+\bm u$, notice that the term ${\langle 3\tau \bm v+\bm u, \bm x \rangle}$ is the linear term on $\bm x$, and it is equal to $${\langle 3\tau \bm v+\bm u, \bm x \rangle} = {\mathbb{E}}_{y\sim N(\bm 0,\mathbf I_n)}[\bm T(\bm x,\bm y,\bm y)+\bm T(\bm y,\bm x,\bm y) + \bm T(\bm y,\bm y, \bm x)].$$ This means $$(3\tau \bm v+\bm u)_j = {\mathbb{E}}_{y\sim N(\bm 0,\mathbf I_n)}[\bm T(\bm e_j,\bm y,\bm y)+\bm T(\bm y,\bm e_j,\bm y) + \bm T(\bm y,\bm y, \bm e_j)] = \sum_{i=1}^d (\bm T_{iij}+\bm T_{iji}+\bm T_{jii}).$$
Omitted Proof in Section \[sec:simplealg\] {#app:theorem3}
------------------------------------------
When $\tau \ge Cn^{3/4}\log n$ for a large enough constant $C$, under the Independence Condition (Condition \[assump\]), Algorithm \[algo:simple\] finds a vector $\bm x^m$ such that ${\langle \bm x^m, \bm v \rangle} \geq 0.8$ in $O(\log \log n)$ iterations.
$m = O(\log \log n)$;\
$\forall~j, \bm x^0_j = \sum_i \bm T_{iij} + \bm T_{iji} + \bm T_{jii}$;\
$\bm x^0 = \bm x^0 / \|\bm x^0\|$; //$\bm x^0 = \bm x^\dag$\
;
We first show the initial maximizer $\bm x^0$ already has a nontrivial correlation with $\bm v$. Recall $\bm x^0 = \frac{3\tau \bm v+\bm u}{\|3\tau \bm v+\bm u\|_2}$. Note that if $\tau$ is very large such that $\|3\tau \bm v\|_2 \ge 10 \|u\|_2$, then we already have ${\langle \bm x^0, \bm v \rangle} \ge 0.8$. Later we will show that whenever ${\langle \bm x^i, \bm v \rangle} \ge 0.8$ all later iterations have the same property.
Therefore, we are left with the case when $\|\bm u\|_2 \ge 0.1 \|3\tau \bm v\|_2$ (this implies $\tau \le O(n)$). In this case, by Condition \[assump\] we know $|{\langle \bm u, \bm v \rangle}| = O(\sqrt{nm\log n})$ and $\|\bm u\|_2 = O(n\sqrt{m})$, therefore $$\begin{aligned}
\|3 \tau \bm v + \bm u\|_2 \in \left[\sqrt{\|\bm u\|_2^2+\|3\tau \bm v\|_2^2 - O(\tau \sqrt{nm\log n})}, \sqrt{\|\bm u\|_2^2+\|3\tau \bm v\|_2^2 + O(\tau \sqrt{nm\log n})}\right]
\end{aligned}$$ Therefore, $\|3 \tau \bm v + \bm u\|_2 = \Theta(n\sqrt{m})$. Assume $\tau \ge Cn^{3/4}\log^c n$ for large enough $C$ (where we will later show $c = 1$ suffices) $$\begin{aligned}
\langle \bm x^0, \bm v \rangle = \frac{3 \tau + \langle \bm u, \bm v \rangle} {\|3 \tau \bm v + \bm u\|_2} = \frac 1 {O(n \sqrt m)} \Theta(n^{\frac 3 4} \cdot \log^c n) \ge \frac{n^{-\frac 1 4} \cdot \log^c n} {\sqrt m}.
\end{aligned}$$
Now let us consider the first step of power method. Let $\hat{\bm x}^1$ be the vector before normalization. Observe that $\hat{\bm x}^1 = 3 \tau \langle \bm v, \bm x^0 \rangle^2 \bm v + \delta(\bm x^0)$. By Condition \[assump\] we have bounds on $\|\delta(\bm x^0)\|$ and $|{\langle \delta(\bm x^0),v \rangle}|$, therefore we have $$\begin{aligned}
\langle \hat{\bm x}^{1}, \bm v \rangle = 3 \tau \langle \bm v, \bm x^0 \rangle^2 + {\langle \delta(\bm x^0),\bm v \rangle} \in \left[3 \tau \langle \bm v, \bm x^0 \rangle^2 - O(\sqrt{m \log n}), 3 \tau \langle \bm v, \bm x^0 \rangle^2 + O(\sqrt{m \log n})\right].
\end{aligned}$$ Note that when $\tau \ge Cn^{3/4}\log^c n$ and $\log^c n \gg m$, the first term is much larger than $\sqrt{m\log n}$. Hence for the first iteration, we have $\langle \hat{\bm x}^{1}, \bm v \rangle \ge (3-o(1))\tau {\langle \bm v,\bm x^0 \rangle}^2\ge 2Cn^{\frac 1 4} \cdot \log^{3c} n / m$.
Similar as before, when $\|\delta(\bm x^0)\|_2 \le 0.1 \|3\tau {\langle v,\bm x^0 \rangle}^2 v\|_2$, we already have ${\langle \bm x^1, \bm v \rangle} \ge 0.8$. On the other hand, if $\|\delta(\bm x^0)\|_2 \ge 0.1\|3\tau {\langle v,\bm x^0 \rangle}^2 v\|_2$, in this case, by Condition \[assump\] we know $\|\delta(\hat{\bm x}^0)\| = O(\sqrt{nm})$. We again have $\|\hat{\bm x}^1\|_2 \in \sqrt{\|\delta(\bm x^0)\|_2^2+\|3\tau {\langle \bm v, \bm x^0 \rangle}^2 \bm v\|_2^2 \pm O(\tau {\langle \bm v, \bm x^0 \rangle}^2 \sqrt{nm})}$. Therefore, $\|\hat{\bm x}^1\|_2 = O(\sqrt {n m})$. Combining the bounds for the norm of $\hat{\bm x}^1$ and its correlation with $\bm v$, $$\langle \frac{\hat{\bm x}^{1}}{\|\hat{\bm x}^{1}\|}, \bm v \rangle \ge n^{-\frac 1 4} \cdot \log^{3c} n / m^{\frac 3 2}.$$ Therefore, when $\log^c n \gg m$, the correlation between $\bm x^1$ and $\bm v$ is larger than the correlation between $\bm x^0$ and $\bm v$. This shows the first step makes an improvement.
In order to show this for the future steps, we do induction over $p$. The induction hypothesis is for every $p$, either ${\langle \bm x^p, \bm v \rangle} \ge 0.8$ or $$\langle \bm x^p, \bm v \rangle \ge n^{-\frac 1 4} \log^{3^p c} n / m^{2^p - \frac 1 2}.$$ Initially, for $p = 0$, we have already proved the induction hypothesis.
Now assume the induction hypothesis is true for $p$. In the next iteration, let $\hat{\bm x}^{p+1}$ be the vector before normalization. Similar as before we have $\hat{\bm x}^{p+1} = 3 \tau \langle \bm v, \bm x^p \rangle^2 \bm v + \delta(\bm x^p)$.
When ${\langle \bm x^p,\bm v \rangle} \ge 0.8$, by Condition \[assump\] we know the norm of $\|\delta(\bm x^p)\|$ is much smaller than $3\tau {\langle \bm x^p,\bm v \rangle}^2$. Therefore we still have ${\langle \bm x^{p+1},\bm v \rangle} \ge 0.8$.
In the other case, we follow the same strategy as the first step. By Condition \[assump\] we can compute the correlation between $\hat{\bm x}^{p+1}$ and $\bm v$: $$\begin{aligned}
\langle \hat{\bm x}^{p+1}, \bm v \rangle &= 3 \tau \langle \bm v, \bm x^p \rangle^2 \pm O(\sqrt{m \log n}) \\ &\ge 2Cn^{\frac 1 4} \log^{3^{p+1}c} n / m^{2^{p+1} - 1}.
\end{aligned}$$ For the norm of $\hat{\bm x}^{p+1}$, notice that the first term $3 \tau \langle \bm v, \bm x^p \rangle^2 \bm v$ has norm $3 \tau \langle \bm v, \bm x^p \rangle^2$, and the second term $\delta(\bm x^p)$ has norm $\Theta(\sqrt{nm})$. Note that these two terms are almost orthogonal by Independence Condition, therefore $$\|\hat{\bm x}^{p+1}\|_2 = \Theta(\tau \langle \bm v, \bm x^p \rangle^2) + O(\sqrt {n m})$$ If $3\tau \langle \bm v, \bm x^p \rangle^2 \ge \Delta \sqrt {n m}$, then $\|\hat{\bm x}^{p+1}\|_2 \le (3+\Delta')\tau \langle \bm v, \bm x^p \rangle^2$, where $\Delta'$ is a constant that is smaller than $0.1$ when $\Delta$ is large enough. Therefore in this case $\langle \frac{\hat{\bm x}^{p+1}}{\|\hat{\bm x}^{p+1}\|_2}, \bm v \rangle \ge 0.8$. Thus we successfully recover $\bm v$ in the next step.
Otherwise, we know $\|\hat{\bm x}^{p+1}\|_2 = O(\sqrt {n m})$. Then, $$\langle \frac{\hat{\bm x}^{p+1}}{\|\hat{\bm x}^{p+1}\|_2}, \bm v \rangle \ge n^{-\frac 1 4} \cdot \log^{3^{p+1}c} n / m^{2^{p+1} - \frac 1 2}$$ If we select $c = 1$, after $m = O(\log \log n)$ rounds, we have ${\langle \bm x^p, \bm v \rangle} \ge n^{-\frac 1 4} \log^{3^p c} n / m^{2^p - \frac 1 2} \ge 0.8$, therefore we must always be in the first case. As a result $\langle \bm x^m, \bm v \rangle \ge 0.8$.
Let the sequence $\bm T^0, \cdots, \bm T^{m-1}$ be generated according to Section \[sec:resampling\]. Let $\bm Q^i = \tau \bm v^{\otimes 3}+\bm C^i$, where $\bm C^i$’s are tensors with independent Gaussian entries. Each entry in $\bm C^i$ is distributed as $N(0,m)$. The two sets of variables $\{\bm T^i\}$ and $\{\bm Q^i\}$ has the same distribution.
Note that both distributions are multivariate Gaussians. Therefore we only need to show that they have the same first and second moments.
For the first moment, this is easy, we have $\mathbb E [\bm T^p] = \tau \cdot v^{\otimes 3}$ and $\mathbb E [\bm Q^p] = \tau \cdot v^{\otimes 3}$ for all $p$.
For the second moment (covariance), we consider the covariance between $T^p_{ijk}$ and $T^q_{i'j'k'}$. Note that for the distribution $Q$, as long as the 4 tuple $(p,i,j,k)\ne (q,i',j',k')$ the correlation is 0. We first show when $(i,j,k) \neq (i',j',k')$ we have $$\begin{aligned}
\mathrm{Cov}(\bm T^p_{ijk}, \bm T^q_{i'j'k'})
= & ~\mathbb E [(\bm T^p_{ijk} - \tau \bm v_i \bm v_j \bm v_k) (\bm T^q_{i'j'k'} - \tau \bm v_{i'} \bm v_{j'} \bm v_{k'})] \\
= & ~\mathbb E [(\bm B^p_{ijk} - \overline{\bm B_{ijk}} + \bm A_{ijk}) (\bm B^q_{i'j'k'} - \overline{\bm B_{i'j'k'}} + \bm A_{i'j'k'})] \\
= & ~\mathbb E [\bm B^p_{ijk} - \overline{\bm B_{ijk}} + \bm A_{ijk}] \mathbb E [\bm B^q_{i'j'k'} - \overline{\bm B_{i'j'k'}} + \bm A_{i'j'k'}] \\
= & ~0 \\
\end{aligned}$$ Hence for these variables the two distributions have the same covariance.
Next we consider the case $p \neq q$, $$\begin{aligned}
\mathrm{Cov}(\bm T^p_{ijk}, \bm T^q_{ijk})
= & ~\mathbb E [(\bm T^p_{ijk} - \tau \bm v_i \bm v_j \bm v_k) (\bm T^q_{ijk} - \tau \bm v_i \bm v_j \bm v_k)] \\
= & ~\mathbb E [(\bm B^p_{ijk} - \overline{\bm B_{ijk}} + \bm A_{ijk}) (\bm B^q_{ijk} - \overline{\bm B_{ijk}} + \bm A_{ijk})] \\
= & ~-\frac {m - 1} {m^2} \mathbb E[(\bm B^p_{ijk})^2 + (\bm B^q_{ijk})^2] + \sum_{l \neq p, q} \frac {1} {m^2} \mathbb E[(\bm B^l_{ijk})^2] + \mathbb E[\bm A_{ijk}^2] \\
= & ~-\frac {2(m - 1)} {m^2} \cdot m + \frac {m - 2} {m^2} \cdot m + 1 = 0
\end{aligned}$$ The covariance for these entries also match.
Finally we need to consider the variance for each entry of $\bm T^p$ and $\bm Q^p$. To do that we compute the Variance of $\bm T^p_{ijk}$ $$\begin{aligned}
\mathrm{Var}(\bm T^p_{ijk})
= & ~\mathbb E [(\bm T^p_{ijk} - \tau \bm v_i \bm v_j \bm v_k) (\bm T^p_{ijk} - \tau \bm v_i \bm v_j \bm v_k)] \\
= & ~\mathbb E [(\bm B^p_{ijk} - \overline{\bm B_{ijk}} + \bm A_{ijk}) (\bm B^p_{ijk} - \overline{\bm B_{ijk}} + \bm A_{ijk})] \\
= & ~\frac {(m - 1)^2} {m^2} \mathbb E[(\bm B^p_{ijk})^2] + \sum_{l \neq p} \frac {1} {m^2} \mathbb E[(\bm B^l_{ijk})^2] + \mathbb E[\bm A_{ijk}^2] \\
= & ~\frac {(m - 1)^2} {m^2} \cdot m + \frac {m - 1} {m^2} \cdot m + 1 = m
\end{aligned}$$ This is also the same as the variance of $Q^p_{ijk}$. Therefore the two multivariate Gaussians have the same mean and covariance, and must be the same distribution.
Let $\bm T^p$ be generated according to Algorithm \[algo:resample\] and $\bm A^p = \bm T^p - \tau \bm v^{\otimes 3}$. Let $\bm u^0$ be a vector such that $\bm u^0_j = \sum_i \bm A_{iij}^0 + \bm A_{iji}^0 + \bm A_{jii}^0$, and $\delta^p(\bm x^p) = \bm A^p(\bm x^p, \bm x^p, :)+\bm A^p(\bm x^p, :, \bm x^p)+\bm A^p(:, \bm x^p, \bm x^p)$. With high probability, (1) $\|\bm u^0\|_2 = \Theta(n \sqrt{m})$ and $|\langle \bm u^0, \bm v \rangle| = O(\sqrt {n m \log n})$; (2) for the sequence computed by Algorithm \[algo:resample\], $\bm x^0, \bm x^1, \cdots, \bm x^{m-1}$, $\forall~ 0 \leq p \leq m-1$, $\|\delta^p(\bm x^p)\|_2 = \Theta(\sqrt{n m}) \|\bm x^p\|_2^2$ and $|\langle \delta^p(\bm x^p), \bm v \rangle| = O(\sqrt{m \log n}) \|\bm x^p\|_2^2$. As a result Condition \[assump\] is satisfied.
Since by Lemma \[lem:resample\], we know the noise tensors $\bm A^p$ used in $p$-th step behave exactly the same as independent Gaussian tensors. The vectors $\bm u^0$ and $\delta(x^p)$ are therefore spherical Gaussian random variables conditioned on any value of $\bm x^i$. Therefore we can prove this lemma by standard Gaussian concentration results.
\[clm:gaussian\][@laurent2000adaptive] Suppose $\bm x$ is a $d$-dimensional spherical Gaussian, then $$\Pr[|\|\bm x\|^2 - {\mathbb{E}}[\|\bm x\|^2]| \ge \frac{1}{2}{\mathbb{E}}[\|\bm x\|^2]] \le e^{-\Omega(d)}.$$ Also, for any fixed vector $\bm v$, ${\langle \bm x, \bm v \rangle}$ is also a Gaussian distribution that satisfies $$\Pr[|{\langle \bm x,\bm v \rangle}| \ge t\sqrt{{\mathbb{E}}[{\langle \bm x,\bm v \rangle}^2]}] \le e^{-\Omega(t^2)}.$$
For terms like $\|\bm u^p\|$ and $\|\delta(\bm x^p)\|$, we know the norm of a Gaussian random variable obeys the $\chi^2$ distribution and is highly concentrated to its expectation. For terms like ${\langle \bm u^p,\bm v \rangle}$ and ${\langle \delta(\bm x^p), \bm v \rangle}$, we know they are just Gaussian distributions and is always bounded by $O(\sigma\sqrt{\log n})$ with high probability. Therefore we only need to compute the expected norms of these vectors.
$$\begin{aligned}
\mathbb E[\|\bm u^p\|_2^2] &~= \mathbb E[\sum_j (\sum_i \bm A^p_{iij} + \bm A^p_{iji} + \bm A^p_{jii})^2] \\
&~= \mathbb E[\sum_{j} (\sum_{i \neq j} (\bm A^p_{iij})^2 + (\bm A^p_{iji})^2 + (\bm A^p_{jii})^2) + 9 (\bm A^p_{jjj})^2] \\
&~= 3 n (n-1) m + 9 n m \\
&~= \Theta(n^2 m)
\end{aligned}$$
Therefore by Claim \[clm:gaussian\] we have $\|u\|_2 = \Theta(n\sqrt{m})$ with high probability.
$$\begin{aligned}
\mathbb E[\langle \bm u^p, \bm v \rangle] = \mathbb E[\sum_j (\sum_i \bm A^p_{iij} + \bm A^p_{iji} + \bm A^p_{jii}) \bm v_j] = 0
\end{aligned}$$
$$\begin{aligned}
\mathbb E[\langle \bm u^p, \bm v \rangle^2] &~= \mathbb E[\sum_j ((\sum_i \bm A^p_{iij} + \bm A^p_{iji} + \bm A^p_{jii}) \bm v_j)^2] \\
&~= \mathbb E[\sum_j \bm v_j^2 (9 (\bm A^p_{jjj})^2 + \sum_{i \neq j} (\bm A^p_{iij})^2 + (\bm A^p_{iji})^2 + (\bm A^p_{jii})^2)] \\
&~=9 m + 3(n - 1) m \\
&~=\Theta(n m)
\end{aligned}$$
This means ${\langle u,v \rangle}$ is a Gaussian random variable with variance $\sigma^2 = \Theta(nm)$, therefore for any constant $C'$, with probability at least $1-n^{-C'}$ we know $|{\langle u,v \rangle}| \le O(\sqrt{nm \log n})$. We can apply union bound over all $p$ and get the desired result.
Similarly we can compute the expected square norm of $\delta(\bm x^p)$ as below
$$\begin{aligned}
\mathbb E[\|\delta(\bm x^p)\|_2^2] &~= \Theta(1) \mathbb E[\|\bm A^p(\bm x^p, \bm x^p, :)\|_2^2] \\
&~= \Theta(1) \mathbb E[\sum_k (\sum_{i,j} \bm A^p_{ijk} \bm x^p_i \bm x^p_j)^2] \\
&~= \Theta(1) \mathbb E[\sum_k (\sum_{i,j} (\bm A^p_{ijk})^2 (\bm x^p_i)^2 (\bm x^p_j)^2)] \\
&~= \Theta(1) n m \|x^p\|_2^4
\end{aligned}$$
$$\begin{aligned}
& \mathbb E[\langle \delta(\bm x^p), \bm v \rangle] = \sum_{i,j,k} \mathbb E[\bm A^p_{ijk} (\bm x^p_i \bm x^p_j \bm v_k + \bm x^p_i \bm v_j \bm x^p_k + \bm v_i \bm x^p_j \bm x^p_k)] = 0
\end{aligned}$$
$$\begin{aligned}
\mathbb E[\langle \delta(\bm x^p), \bm v \rangle^2] &~= \sum_{i,j,k} \mathbb E[(\bm A^p_{ijk})^2 (\bm x^p_i \bm x^p_j \bm v_k + \bm x^p_i \bm v_j \bm x^p_k + \bm v_i \bm x^p_j \bm x^p_k)^2] \\
&~= 3 m \sum_k \bm v_k^2 \sum_{i,j} (\bm x_i^p)^2 (\bm x_j^p)^2 + 6 m \sum_i (\bm x_i^p)^2 \sum_{j,k} \bm v_i \bm v_j \bm x_j^p \bm x_k^p \\
&~= 3 m \|\bm x^p\|_2^4 + 6 \|\bm x^p\|_2^2 \langle \bm v, \bm x^p \rangle^2 \\
&~\leq 9 m \|\bm x^p\|_2^4
\end{aligned}$$
The bounds on $\|\delta(\bm x^p)\|$ and ${\langle \delta(\bm x^p),v \rangle}$ follows immediately from these expectations.
Omitted Proof in Section \[sec:path\] {#app:path}
-------------------------------------
The smoothed version of the alternative objective is $$\begin{aligned}
g_r(\bm x, t) = \tau \langle \bm v, \bm x \rangle^3 + t^2 \langle 3 \tau \bm v + \bm u, \bm x \rangle + \bm A(\bm x,\bm x,\bm x) - \frac {3 \tau} 4 \left(\|x\|_2^4 + 2 t^2 (n+2) \|x\|_2^2 + t^4 (n^2+2n)\right)
\end{aligned}$$ Its gradient and Hessian are equal to $$\begin{aligned}
\nabla g_r(\bm x, t) = 3 \tau \langle \bm v, \bm x \rangle^2 \bm v + t^2 (3 \tau \bm v + \bm u)
+ \delta(\bm x) - 3 \tau (\|\bm x\|_2^2 \bm x + t^2 (n+2) \bm x).
\end{aligned}$$ and $$\begin{aligned}
\nabla^2 g_r(\bm x, t) = -3 \tau ((\|\bm x\|_2^2 + t^2 (n+2)) \bm I - 2 \langle \bm v, \bm x \rangle \bm v \bm v^T + 2 \bm x \bm x^T) + P_{sym}[\bm A(\bm x, :, :) + \bm A(:, \bm x, :) + \bm A(:, :, \bm x)].
\end{aligned}$$
Similar to Lemma \[lem:smooth\], we can write the smoothing operation as an expectation. By linearity of expectation we know $$g_r(\bm x,t) = g(\bm x, t)+\mathbb E[\|\bm x + t \bm y\|_2^4]$$
We can compute the new terms by the moments of Gaussians: $$\begin{aligned}
\mathbb E[\|\bm x + t \bm y\|_2^4] & = {\mathbb{E}}[(\|\bm x\|_2^2+2t{\langle \bm x, \bm y \rangle} + t^2\|\bm y\|_2^2)^2] \\
& = {\mathbb{E}}[\|\bm x\|_2^4+4t^2{\langle \bm x, \bm y \rangle}^2 + t^4\|\bm y\|_2^4 + 2t^2\|x\|^2\|y\|^2] \\
& = \|x\|_2^4 + t^2 (2 n + 4) \|x\|_2^2 + t^4 (n^2+2n) = \|x\|_2^4 + 2 t^2 (n+2) \|x\|_2^2 + t^4 (n^2+2n).
\end{aligned}$$
Here in the second equation we omitted all the odd order terms for $\bm y$ because those terms have expectation 0. The final step uses the moments of Gaussians.
The equation for $g_r(\bm x,t)$ follows immediately, and since it is a polynomial it is easy to compute its gradient and Hessian.
Before trying to characterize the local maxima on the homotopy path, let us first prove the following property for the matrix $P_{sym}[\bm A(\bm x, :, :) + \bm A(:, \bm x, :) + \bm A(:, :, \bm x)]$.
Let $H(\bm x) = P_{sym}[\bm A(\bm x, :, :) + \bm A(:, \bm x, :) + \bm A(:, :, \bm x)]$, there exists constants $c^-,c^+$ such that with probability at least $1-\exp(-\Omega(n))$, for any unit vector $\bm x$ we have $$c^-\sqrt{n} \le \lambda_{max} \le c^+\sqrt{n}.$$ \[lem:hessianprop\]
For the upperbound, we use the bound on tensor spectral norm. @tomioka2014spectral proved that for a random Gaussian tensor $\bm A$, with probability at least $1-\exp(-\Omega(n))$ we know for any vectors $\bm x, \bm y, \bm z$, $|\bm A(\bm x, \bm y, \bm z)| \le O(\sqrt{n})$. Therefore for any unit vector $\bm y$, $|\bm y^\top H(\bm x) \bm y| = |\bm A(\bm x, \bm y, \bm y) + \bm A(\bm y, \bm x, \bm y) + \bm A(\bm y, \bm y, \bm x)| \le O(\sqrt{n})$.
For the lowerbound, we use the distribution of the largest eigenvalue of Gaussian Orthogonal Ensemble. Suppose $\bm M$ is a random matrix whose entries are i.i.d. standard Gaussians, then the symmetric matrix $\frac{\bm M + \bm M^\top}{\sqrt{2}}$ is distributed according to the Gaussian Orthogonal Ensemble. Let $P_{\bm x^\perp}$ be the projection operator to the orthogonal subspace of $\bm x$, then the key observation is $P_{\bm x^\perp} H(\bm x) P_{\bm x^\perp}$ is (up to a constant scaling) distributed as a Gaussian Orthogonal Ensemble of dimension $(n-1)\times (n-1)$. To see this, the easiest way is to observe that Gaussians are invariant under rotation, so we can take $\bm x = \bm e_1$. Now for $i,j = \{2,3,...,n\}$, $[P_{\bm x^\perp} H(\bm x) P_{\bm x^\perp}]_{i,j} = \bm A_{1ij}+\bm A_{1ji}+\bm A_{i1j}+\bm A_{j1i}+\bm A_{ij1}+\bm A_{ji1}$. The random entries $1ij,i1j,ij1$ do not overlap because $i,j\ne 1$. Therefore the matrix is the sum of three Gaussian Orthogonal Ensembles, and by property of Gaussians that is equivalent to $\sqrt{3}$ times a Gaussian Orthogonal Ensemble. Now, using the result in [@ledoux2007deviation], we know for any fixed $\bm x$, $\Pr[\lambda_{max}(P_{\bm x^\perp} H(\bm x) P_{\bm x^\perp}) \le \sqrt{n}/2] \le 1-\exp(-\Omega(n^2))$. By standard covering argument (the $\epsilon$-net for $n$ dimensional vectors have size $(n/\epsilon)^{O(n)}$ which is much smaller than $\exp(-\Omega(n^2))$), we know with high probability for all $\bm x$ $\lambda_{max}(P_{\bm x^\perp} H(\bm x) P_{\bm x^\perp}) \ge \sqrt{n}/2$. The lemma follows immediately because $\lambda_{max}(H(\bm x)) \ge \lambda_{max}(P_{\bm x^\perp} H(\bm x) P_{\bm x^\perp})$.
Now we are ready to prove Theorem \[thm:homotopypath\]. To capture the properties of the homotopy path, we break it into three lemmas.
When $\tau = n^{3/4}\log n$, $t \ge Cn^{-1}$ for large enough constant $C$, there exists a local maximizer $\bm x^t$ of $g_r(\bm x, t)$ such that $\|\bm x^t - \bm x^\dag\|_2 = o(1) \|\bm x^\dag\|_2$.
Recall according to the objective we chose, the maximizer at infinity $\bm x^\dag$ can be computed explicitly and we know $\bm x^\dag = \frac{3\tau \bm v+\bm u}{3\tau(n+2)}$. By Conjecture \[assump2\], we can estimate the norm and correlation with $\bm v$: $$\|\bm x^\dag\|_2 = \Theta(n^{-3/4}\log^{-1} n), \quad {\langle \bm x^\dag, \bm v \rangle} = (1\pm o(1))/n.$$
We shall first prove in the region $\mathcal{B} = \{\bm x: \|\bm x - \bm x^\dag\|_2 \le \frac 1 2 \|\bm x^\dag\|_2, {\langle \bm x, \bm v \rangle} \le 10/n\}$, the Hessian of the objective function is always negative definite. By standard analysis in convex optimization, this in particular implies two things: 1. There can be at most one local maximizer in this region; 2. If the function is $\mu$-strongly-concave ($\nabla^2 g(\bm x,t)\succeq -\mu I$), and a point $\bm x$ has $\|\nabla g(\bm x,t)\| \le \epsilon$, then there is a local maximizer within $\epsilon/\mu$. This particular implies if there is a point $\bm x$ in the interior $\mathcal{B}' = \{\bm x: \|\bm x - \bm x^\dag\|_2 \le \frac 1 4 \|\bm x^\dag\|_2, {\langle \bm x, \bm v \rangle} \le 2/n\}$ such that $\nabla g(\bm x,t)$ is very small, then there must exist a local maximizer in $\mathcal{B}$.
By Lemma \[lem:smoothalternative\], we know the Hessian is equal to: $$\nabla^2 g_r(\bm x, t) = -3 \tau ((\|\bm x\|_2^2 + t^2 (n+2)) \bm I - 2 \langle \bm v, \bm x \rangle \bm v \bm v^T + 2 \bm x \bm x^T) + P_{sym}[\bm A(\bm x, :, :) + \bm A(:, \bm x, :) + \bm A(:, :, \bm x)].$$
In the region we are interested in, since ${\langle \bm v,\bm x \rangle} \le 10/n \le t^2(n+2)/2$ when $C$ is large enough, we have the first term $$-3\tau((\|\bm x\|_2^2 + t^2 (n+2)) \bm I - 2 \langle \bm v, \bm x \rangle \bm v \bm v^T + 2 \bm x \bm x^T) \preceq -1.5\tau t^2(n+2) \bm I.$$
On the other hand, for the second part we know by Lemma \[lem:hessianprop\] $$P_{sym}[\bm A(\bm x, :, :) + \bm A(:, \bm x, :) + \bm A(:, :, \bm x)] \preceq \frac{1}{2}c^+\sqrt{n}\|\bm x^\dag\|_2 I.$$
By our choice of parameters, $ \tau t^2(n+2) = \Omega(n^{-1/4}\log n)$, and $\sqrt{n}\|\bm x^\dag\|_2 = \Theta(n^{-1/4}\log^{-1}n)$, therefore the first term dominates and we know the Hessian $\nabla^2 g_r(\bm x, t) \preceq -\tau t^2(n+2) \bm I$.
When $t$ is a large polynomial of $n$ (e.g. $t = n^{10}$), simple calculation shows the optima $\bm x^t$ is very close to $\bm x^\dag$, and we have $\bm x^t \in \mathcal{B'}$. When $C/n \le t < n^{10}$, let $t_0 = n^{10}$, select $t_1,t_2,...,t_q$ such that $t_q = t$, and $t_i,t_{i+1}$ are close enough that if $\bm x^{t_i} \in \mathcal{B'}$, by strong concavity we can get $\bm x^{t_{i+1}}$ exists and $\bm x^{t_{i+1}} \in \mathcal{B}$. We will prove $\bm x^{t_i}\in \mathcal{B}'$ by induction. The base case is already done.
Suppose $\bm x^{t_{i-1}}\in \mathcal{B}'$, we know that $\bm x^{t_i} \in \mathcal{B}$. We will use the first order condition to refine our knowledge about $\bm x^{t_i}$ and show $\bm x^{t_i} \in \mathcal{B}'$. From (\[eq:gradient\]), we can derive the expression of stationary points, $$\label{eq:stationary}
\bm x^{t_i} = \frac{3 \tau \langle \bm v, \bm x^{t_i} \rangle^2 \bm v + t^2 (3 \tau \bm v + \bm u) + \delta(\bm x^{t_i})}{3 \tau (\|\bm x^{t_i}\|_2^2 + t^2 (n+2))}$$
Note that $\bm x^{t_i}$ is a stationary point on homotopy path, so it should satisfy Conjecture \[assump2\]. We also know it is in $\mathcal{B}$.
Since $t \ge Cn^{-1}$, $\|\tau \langle \bm v, \bm x^{t_i} \rangle^2 \bm v\|_2 = \Theta(n^{-\frac 5 4} \log n)$, $\|t^2 (3 \tau \bm v + \bm u)\|_2 \ge \Omega(n^{-1})$ and $\|\delta(\bm x^{t_i})\|_2 = \Theta(n^{-1}\log^{-2}n)$. Therefore, if we let $\bm w = 3 \tau \langle \bm v, \bm x^{t_i} \rangle^2 \bm v + \delta(\bm x^{t_i})$ we know $\|\bm w\|_2 \le o(1) \|t^2 (3 \tau \bm v + \bm u)\|_2$. The middle term dominates the numerator. Moreover, $t^2 (n+2) \geq \Omega(n^{-1})$ and $\|\bm x^{t_i}\|_2^2 = \Theta(n^{-3 / 2} \log^{-2} n)$, and thus, $t^2 n$ dominates the denominator. Now we have $$\begin{aligned}
\bm x^{t_i} &~= \frac{3 \tau \langle \bm v, \bm x^{t_i} \rangle^2 \bm v + t^2 (3 \tau \bm v + \bm u) + \delta(\bm x^{t_i})}{3 \tau (\|\bm x^{t_i}\|_2^2 + t^2 (n+2))} \\&~= \frac{t^2 (3 \tau \bm v + \bm u)+\bm w}{3 \tau t^2 (n+2)(1+\epsilon)} \\
&~= \frac{3\tau \bm v+\bm u}{3\tau (n+2)} \cdot \frac{1}{1+\epsilon} + \frac{\bm w}{3\tau t^2(n+2)(1+\epsilon)}. \\
&~= \bm x^\dag + \bm x^\dag (\frac{1}{1+\epsilon} - 1) + \frac{\bm w}{3\tau t^2(n+2)(1+\epsilon)}.
\end{aligned}$$ Since $\epsilon = o(1)$ and $\|w\|_2 \le o(1) \|t^2 (3 \tau \bm v + \bm u)\|_2$, we know the two additional term has norm $o(1)\|\bm x^\dag\|_2$, therefore $\bm x^{t_i}$ is very close to $\bm x^\dag$.
Next we bound the correlation with $\bm v$. We know ${\langle \bm v, \bm x^{t_i} \rangle} \le 10/n$ because $\bm x^{t_i} \in \mathcal{B}$. Also, the correlation between $|{\langle \bm u,\bm v \rangle}| = O(\sqrt{n \log n})$ and $|{\langle \delta(\bm x^{t_i}), \bm v \rangle}| = O(n^{-3/2} \log^{-3/2} n) $ are negligible due to Conjecture \[assump2\], therefore we have $${\langle \bm x^{t_i}, \bm v \rangle} \le \frac{3\tau(10/n)^2 + 3\tau t^2}{3\tau t^2(n+2)(1+\epsilon)} \approx \frac{10^2 + C^2}{C^2 n} \le 2/n.$$
Here the inequality holds as long as $C$ is large enough. Therefore $x^{t_i} \in \mathcal{B}'$ and we finish the induction.
Next lemma shows what happens after the phase transition, when $t$ is small.
When $\tau = n^{3/4}\log n$, $t = n^{-1} \varepsilon(n)$, where $\varepsilon(n) = O(\log^{-2} n)$, the local maximizers (excluding saddle points) $\bm x^t$ of $g_r(\bm x, t)$ are of the following types:
- good maximizers: $\|\bm x^t\|_2^2 = \Theta(1)$ and $\langle \bm v, \bm x^t \rangle = \Theta(1)$;
- bad maximizers: $\|\bm x^t\|_2^2 = \Theta(n^{-\frac 1 2} \log^{-2} n)$ and $\langle \bm v, \bm x^t \rangle \leq O(n^{-\frac 1 2} \log^{-2} n)$;
Now we use the second order necessary conditions. For all local maximizer, their gradient should be $\bm 0$ and their Hessian should be negative semidefinite.
First, from (\[eq:stationary\]), we can compute the inner product between $\bm v$ and $\bm x^t$: $$\langle \bm v, \bm x^t \rangle = \frac{3 \tau \langle \bm v, \bm x^t \rangle^2 + 3 \tau t^2 + t^2 \langle \bm u, \bm v \rangle + \langle \delta(\bm x^t), \bm v \rangle}{3 \tau (\|\bm x^t\|_2^2 + t^2 (n + 2))}$$ Note that $\bm x^t$ should satisfy the conditions in Conjecture \[assump2\], in particular $|{\langle \delta(\bm x^t, \bm v \rangle}| \le O(1)\|\bm x^t\|_2^2$. Also, by Conjecture \[assump2\] we know $t^2 \langle \bm u, \bm v \rangle = t^2 O(\sqrt{n \log n}) \ll 3 \tau t^2$, so it is negligible in scale analysis. Therefore, $$\label{eq:inner}
\langle \bm v, \bm x^t \rangle = \frac{3 \tau \langle \bm v, \bm x^t \rangle^2 + (3 \pm o(1)) \tau t^2 \pm O(\sqrt{\log n}) \|\bm x^t\|_2^2}{3 \tau (\|\bm x^t\|_2^2 + t^2 n)}$$ From (\[eq:stationary\]), we can also compute the square of the norm of $\bm x$: $$\begin{aligned}
\|\bm x^t\|_2^2 = \frac{9 \tau^2 \langle \bm v, \bm x^t \rangle^4 + t^4 \|3 \tau \bm v + \bm u\|_2^2 + \|\delta(\bm x^t)\|_2^2 + \eta(\bm x^t)}{9 \tau^2 (\|\bm x^t\|_2^2 + t^2 (n + 2))^2}
\end{aligned}$$ where the cross term $\eta(\bm x^t)$ $$\eta(\bm x^t) = 6 \tau \langle \bm v, \bm x^t \rangle^2 \langle \bm v, \delta(\bm x^t) \rangle + 6 \tau t^2 \langle \bm v, \bm x^t \rangle^2 (3 \tau + \langle \bm v, \bm u \rangle) + 2 t^2 \langle 3 \tau \bm v + \bm u, \delta(\bm x^t) \rangle$$ is negligible compared to the other terms. We again have the bound on $\|\delta(\bm x^t)\|_2$ from Conjecture \[assump2\] and therefore $$\label{eq:norm}
\|\bm x^t\|_2^2 = \frac{9 \tau^2 \langle \bm v, \bm x^t \rangle^4 + t^4 \Theta(n^2) + \Theta(n \log n) \|\bm x^t\|_2^4}{9 \tau^2 (\|\bm x^t\|_2^2 + t^2 n)^2}$$ We proceed the proof via a case analysis on the relative order between $\|\bm x^t\|_2^2$ and $t^2 n$.
#### Case 1: $\|\bm x^t\|_2^2 \ge t^2 n$:
\
First, recall that the Hessian at $\bm x^t$ must be a negative semidefinite. Therefore, $\tau \|\bm x^t\|_2^2$ must be larger than $\lambda_{max}(P_{sym}(\bm A(\bm x^t, :, :) + \bm A(:, \bm x^t, :) + \bm A(:, :, \bm x^t)))$. By Lemma \[lem:hessianprop\] we have $\tau \|\bm x^t\|_2^2 > \Theta(\sqrt n) \|\bm x^t\|_2$, which implies $\|\bm x^t\|_2 = \Omega(n^{-\frac 1 4} \log^{-1} n)$. As a result, $\Theta(n) \|\bm x^t\|_2^4$ dominates $t^4 \Theta(n^2)$ in the nominator of (\[eq:norm\]). Henceforth, we have $$\|\bm x^t\|_2^2 = \Theta(1)\frac{\langle \bm v, \bm x^t \rangle^4}{\|\bm x^t\|_2^4} + \Theta(n^{-\frac 1 2} \log^{-1} n)$$
We know $\|\bm x^t\|_2^2$ must be within constant factor to either $\frac{\langle \bm v, \bm x^t \rangle^4}{\|\bm x^t\|_2^4}$ or $n^{-\frac 1 2} \log^{-1} n$. These two cases are discussed below
\(1) If $\|\bm x^t\|_2^2 = \Theta(n^{-\frac 1 2} \log^{-1} n)$, plug it into (\[eq:inner\]), we have $$\langle \bm v, \bm x^t \rangle = \Theta(1)\frac{\langle \bm v, \bm x^t \rangle^2}{\|\bm x^t\|_2^2} + \Theta(1)\frac{t^2}{\|\bm x^t\|_2^2} \pm \frac{O(\sqrt{\log n})}{\tau}$$ Therefore, the largest possible $\langle \bm v, \bm x^t \rangle$ is $\Theta(n^{-\frac 1 2} \log^{-1} n)$.
\(2) If $\|\bm x^t\|_2^3 = \Theta(1)\langle \bm v, \bm x^t \rangle^2$, plug it into (\[eq:inner\]): $$\langle \bm v, \bm x^t \rangle = \Theta(1)\|\bm x^t\|_2^2 + \Theta(1)\frac{t^2}{\|\bm x^t\|_2^2} \pm \frac{O(\sqrt{\log n})}{\tau} = \Theta(1)\|\bm x^t\|_2^2$$ Thus, we can conclude both $\|\bm x^t\|_2$ and $\langle \bm v, \bm x^t \rangle$ are bounded by absolute constants.
#### Case 2: $t^2 n\ge \|\bm x^t\|_2^2$
\
We will show this case cannot happen. Recall that the Hessian at $\bm x^t$ must be a negative semidefinite. Therefore, $\tau t^2 n$ must be larger than $\lambda_{max}(P_{sym}(\bm A(\bm x^t, :, :) + \bm A(:, \bm x^t, :) + \bm A(:, :, \bm x^t)))$. By Lemma \[lem:hessianprop\], we have $\tau t^2 n > \Theta(\sqrt n) \|\bm x^t\|_2$, which implies $\|\bm x^t\|_2 = t^2 \tau O(n^{1/2})$. As a result, $3 \tau t^2$ dominates $O(\sqrt{\log n})\|\bm x^t\|_2^2$ in the nominator of (\[eq:inner\]). Henceforth, we have $$\langle \bm v, \bm x^t \rangle = C_1\frac{\langle \bm v, \bm x^t \rangle^2}{t^2 n} + C_2\frac 1 n,$$ where both $C_1,C_2$ are constants within $1\pm 2/3$. Notice that if $\frac{\langle \bm v, \bm x^t \rangle^2}{t^2 n} \ge n^{-1}$, then $\langle \bm v, \bm x^t \rangle = \Theta(t^2 n)$, implying $\frac{\langle \bm v, \bm x^t \rangle^2}{t^2 n} = \Theta(t^2 n) =\Theta( \frac{\varepsilon^2(n)}{n}) \ll n^{-1}$. This is a contradiction, so we know, $\langle \bm v, \bm x^t \rangle$ can only be $\Theta(n^{-1})$.
Moreover, notice that $t^4 \Theta(n^2) \gg \Theta(n \log n) \|\bm x^t\|_2^4 = t^8 \tau^4 O(n^3 \log n)$. Therefore, from (\[eq:norm\]), $$\|\bm x^t\|_2^2 = \frac {1}{t^4} \Theta(n^{-6}) + \Theta(\frac 1 {\tau^2}) \Rightarrow \|\bm x^t\|_2 = \Theta(n^{-\frac 3 4} \log^{-1} n)$$ This contradicts with $\|\bm x^t\|_2 = t^2 \tau O(\sqrt n) = O(n^{-\frac 3 4} \log^{-3} n)$. There cannot be a local maximizer in this case.
Finally we show that the Hessian is correlated with the correct vector $\bm v$ near the threshold.
For $t = n^{-1} \varepsilon(n)$, where $\varepsilon(n) = O(\log^{-2} n)$, let $\bm b$ be the top eigenvector of $\nabla^2 (g_r(\bm x^\dag, t))$, we know $\sin\theta(\bm b, \bm v) \le 1/\log^2 n$.
Recall the formula for the Hessian (\[eq:hessian\]), $$\nabla^2 g_r(\bm x^\dag, t) = -3 \tau ((\|\bm x^\dag\|_2^2 + t^2 (n+2)) \bm I - 2 \langle \bm v, \bm x^\dag \rangle \bm v \bm v^T + 2 \bm x^\dag {\bm x^\dag}^T) + P_{sym}(\bm A(\bm x^\dag, :, :) + \bm A(:, \bm x^\dag, :) + \bm A(:, :, \bm x^\dag)),$$ and $\bm x^\dag = \frac{\bm v} n + \frac{\bm u}{3 \tau n}$ with norm $\Theta(\frac 1 \tau)$ and correlation $\langle \bm x^\dag, \bm v \rangle = \Theta(\frac 1 n)$. Therefore, we have $\|\bm x^\dag\|_2^2 + t^2 n = O(n^{-1} \log^{-4} n)$. By Lemma \[lem:hessianprop\] the spectral norm of $P_{sym}(\bm A(\bm x^\dag, :, :) + \bm A(:, \bm x^\dag, :) + \bm A(:, :, \bm x^\dag))$ is $\Theta(n^{-\frac 1 4} \log^{-1} n)$. Thus, we can write the Hessian as $$\nabla^2 g_r(\bm x^\dag, t) = 6\tau \langle \bm v, \bm x^\dag \rangle \bm v \bm v^\top + \bm E,$$ where the main term $\bm v \bm v^T$ has coefficient $6 \tau \langle \bm v, \bm x^\dag \rangle = \Theta(n^{-\frac 1 4} \log n)$, and the spectral norm of $E$ is bounded by $O(n^{-1/4}\log^{-1}n)$. By Davis Kahan theorem we know $\sin\theta(\bm b, \bm v) \le O(1/\log^2 n)$, that is, the top eigenvector of the Hessian is $O(1/\log^2 n)$ close to $\bm v$.
[^1]: University of California, Irvine. Email:<[email protected]>
[^2]: Duke University. Email:<[email protected]>
[^3]: Duke University. Email:<[email protected]>
[^4]: Google Research. Email:<[email protected]>
[^5]: Throughout this paper by “with high probability” we mean the probability is at least $1-1/n^C$ for a large constant $C$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Kazuo Aoyama, Kazumi Saito, and Tetsuo Ikeda'
title: |
Inverted-File k-Means Clustering:\
Performance Analysis
---
Introduction {#sec:intro}
============
Based on the rapid growth in the ability of various systems to collect vast amounts of data, machine learning is utilizing large-scale data sets for many applications [@jordan]. In this situation, machine learning algorithms are required to efficiently process such large-scale data sets to withstand practical use. A leading trend for managing data sets is to employ large-scale parallel and distributed computing platforms [@jordan]. To execute algorithms in the platform, modifying and adapting them to the platform is necessary. By contrast, we must develop a novel algorithm that efficiently operates even in a single thread by a single process in a modern computer system, which maintains adaptability to parallel and distributed platforms.
We deal with a Lloyd-type $k$-means clustering algorithm [@wu] for operating in a modern computer system. A standard Lloyd’s algorithm [@lloyd; @macqueen], which is an iterative heuristic algorithm, partitions a given object data set into $k$ subsets (clusters) with given positive integer $k$. Repeating two steps of an assignment and an update step until a convergence is achieved from a given initial state, it locally minimizes an objective function defined by the sum of the squared Euclidean distances between all pairs of an object feature vector and a mean feature vector of the cluster to which the object is assigned. The acceleration algorithms, e.g., those in previous work [@elkan; @hamerly], have also been reported and maintain the same solution as the Lloyd’s algorithm if they start with an identical initial state. These general algorithms are independent of a type of object data sets.
A large-scale data set like document collection often consists of high-dimensional sparse object feature vectors, each of which has a small number of non-zero elements. A spherical $k$-means algorithm [@dhillon] is a Lloyd-type algorithm for such a document data set consisting of texts. Unlike general ones, the spherical $k$-means uses feature vectors normalized by their $L_2$ norms, i.e., points on a unit hypersphere, as an input data set and adopts a cosine similarity for a similarity measure between a pair of points. A mean vector of each cluster is also normalized by its $L_2$ norm. An objective function is defined by the sum of the cosine similarities between all the pairs of an object feature vector and a mean feature vector of the cluster to which the object is assigned. By this procedure, a solution by the spherical $k$-means coincides with that by the standard Lloyd’s algorithm.
It is not trivial what data structures the spherical $k$-means should employ for a large-scale sparse data set to achieve high performance, i.e., to operate at high-speed and with low memory consumption. Our challenge is to develop a high-performance Lloyd-type $k$-means clustering algorithm for a large-scale data set with the low sparsity of a few non-zero elements and potentially numerous classes in the same settings as the spherical $k$-means. We also identify the main factors that determine the performance of our newly developed algorithm by analyzing its operation in a modern computer system.
A modern computer system contains two main components: processors and a hierarchical memory system. A processor has several operating units each of which has deep pipelines with superscalar out-of-order execution and multilevel cache hierarchy [@hennepat]. The memory system consists of registers and caches in a processor and external memories, such as a main memory and flash storages. To efficiently operate an algorithm at high throughput in such a system, we must prevent pipeline hazards, which cause the pipeline stalls, as well as reduce the number of instructions. One serious hazard is a control hazard induced by branch mispredictions [@evers; @eyerman]. Another type is data hazards that can occur when data dependence exists between instructions and degrades the pipeline performance [@hennepat]. In the case of cache misses that result in access to external memories, the degradation becomes conspicuous. For designing an efficient algorithm, the numbers of both branch mispredictions and cache misses must be reduced.
We propose an inverted-file $k$-means clustering algorithm: [*IVF*]{}. [*IVF*]{} utilizes sparse expressions for both the sets of given object feature vectors and the mean feature vectors for low memory consumption. In particular, it exploits an inverted-file data structure for the mean feature vectors. An inverted-file data structure is often adopted in search algorithms for a document data set [@samet; @harman; @knuth; @zobel; @buttcher]. In search algorithms, a set of object feature vectors corresponding to an [*invariant*]{} database is structured with an inverted-file format. Given a query, a search algorithm can find preferable documents quickly from an inverted-file database. Our [*IVF*]{} applies the inverted-file data structure to [*variable*]{} mean feature vectors by varying every iteration instead of [*invariant*]{} object feature vectors.
Our contributions are threefold:
1. We present a novel $k$-means clustering algorithm, an inverted-file $k$-means clustering algorithm referred to as [*IVF*]{}, for a large-scale and high-dimensional sparse data set with potentially numerous classes in Section \[sec:prop\]. Our proposed [*IVF*]{} exploits an inverted-file data structure for a set of mean feature vectors, while the search algorithms employ the data structure for an [*invariant*]{} sparse data set like document collection [@zobel; @buttcher; @knuth].
2. We propose a simple yet practical clock-cycle per instruction (CPI) model for analyzing the factors of computational cost. To identify them based on the CPI model, we prepare different data structures for a set of mean feature vectors and compare [*IVF*]{} to the algorithms using those data structures.
3. We experimentally demonstrate that [*IVF*]{} achieves superior high-speed and low memory consumption performance when it is applied to large-scale and high-dimensional real document data sets with large $k$ values. The low memory consumption is caused by the data structure with sparse expressions of both data object and mean feature vectors. By analyzing the results obtained with the [*perf tool*]{} [@perf] based on the CPI model, [*IVF*]{}’s high speed is clearly attributed to three factors: fewer cache misses, fewer branch mispredictions, and fewer instructions. They are detailed in Sections \[sec:exp\] and \[sec:analy\].
The remainder of this paper consists of the following seven sections. Section \[sec:relate\] briefly reviews related work from viewpoints that clarify the distinct aspects of our work. Section \[sec:prop\] explains our proposed [*IVF*]{}. Section \[sec:algos\] describes the designed algorithms for comparison. Section \[sec:exp\] shows our experimental settings and demonstrates the results. Section \[sec:analy\] determines why [*IVF*]{} achieves high performance with a simple yet practical CPI model. Section \[sec:disc\] discusses [*IVF*]{}’s performance and compares it to other similar algorithms. The final section provides our conclusion and future work.
Related work {#sec:relate}
============
This section reviews four distinct topics: Lloyd-type $k$-means clustering algorithms, a spherical $k$-means for document data sets, which is a variant of Lloyd-type algorithms, an inverted-file data structure for sparse data sets, and design guidelines for efficient algorithms suitable for modern computer systems.
Lloyd-Type k-Means Clustering Algorithm {#subsec:lloyd}
---------------------------------------
We begin by defining a $k$-means clustering problem. Given a set of object feature vectors that are points in a $D$-dimensional Euclidean space, ${\cal X}\!=\!\{ \bm{x}_1, \bm{x}_2,\cdots,\bm{x}_N \}$, $|{\cal X}|\! =\! N$, $\bm{x}_i\! \in\! \mathbb{R}^D$, and a positive integer of $k$, a $k$-means clustering problem finds a set of $k$ clusters, ${\cal C}^*\!=\!\{ C^*_1,C^*_2,\cdots,C^*_k \}$: $${\cal C}^* = {\operatornamewithlimits{\mathrm{arg\,min}}}_{{\cal C}=\{ C_1,\cdots, C_k\}}
\left(\: \sum_{C_j\in{\cal C}}\:\sum_{\bm{x}_i\in C_j}
\|\bm{x}_i-\bm{\mu}_j \|_2^2 \:\right)\:,
\label{eq:obj_funct}$$ where $\|\!\star\!\|_2$ denotes the $L_2$ norm of a vector, ${\cal C}$ is the set of $k$ clusters, and $\bm{\mu}_j\! \in\! \mathbb{R}^D$ is the mean feature vector of cluster $C_j$. Solving the $k$-means clustering problem expressed by Eq. (\[eq:obj\_funct\]) is difficult in practical use due to a high computational cost [@aloise].
Instead of a precise solution to the problem, a standard Lloyd’s algorithm [@lloyd; @macqueen] finds a local minimum in an iterative heuristic manner. The algorithm repeats two steps of an assignment and an update step until the convergence or a predetermined termination condition is satisfied.
[**Algorithm**]{} \[algo:lloyd\] shows an overview of a standard Lloyd’s algorithm at the $r$-th iteration. The assignment step at lines 5–13 assigns a point represented by object feature vector $\bm{x}_i$ to cluster $C_j$ whose centroid (mean at the previous iteration $\bm{\mu}_j^{[r-1]}$) is closest to $\bm{x}_i$. At line 9, $d_{min}$ denotes the tentative minimum distance from $\bm{x}_i$ to the centroids and $a(\bm{x}_i)$ is a function of $\bm{x}_i$ that returns closest centroid ID $j$. The update step at line 15 calculates mean feature vector $\bm{\mu}_j^{[r]}\!\in\!{\cal M}^{[r]}$ at the $r$-th iteration using object feature vectors $\bm{x}_{i}\!\in\! C_j^{[r]}$.
\
\[algo:lloyd-start\] \[algo:lloyd-end\]\
\[algo:lloyd-update\]
Acceleration algorithms have also been reported [@elkan; @hamerly; @drake; @ding; @newling; @hattori; @aoyama], which find the same local minimum as the standard Lloyd’s algorithm if they start at the identical initial state. To eliminate the costly distance calculations at line 8, they exploit the inexpensive lower bound on the exact distance. Since the lower bound is calculated based on the triangle inequality in a metric space, the acceleration strategy is a general one independent of the type of given object feature vectors.
Spherical k-Means Clustering Algorithm {#subsec:sph-kmeans}
--------------------------------------
A spherical $k$-means algorithm [@dhillon] is a special type for document data sets where each object is a text that consists of terms, such as words and phrases. The object is represented by a sparse feature vector, where the dimensionality of a feature space containing all the feature vectors is the number of distinct terms and an element of a feature vector is a feature value given to a term such as [*tf-idf*]{} (term-frequency inverse-document-frequency) [@buttcher]. Define [*sparsity*]{} $\eta(\bm{x}_i)$ of feature vector $\bm{x}_i\!\in\!\mathbb{R}^D$ and [*average sparsity*]{} $\bar{\eta}({\cal X})$ of set ${\cal X}\!=\!\{\bm{x}_1,\bm{x}_2,\cdots,\bm{x}_N\}$: $$\begin{aligned}
\eta(\bm{x}_i) &=& \|\bm{x}_i\|_0 / D\: ,\\
\label{eq:sparsity}
\bar{\eta}({\cal X}) &=& \textstyle{\sum_{i=1}^N \eta(\bm{x}_i) / N}\: ,
\label{eq:avg_sparsity}\end{aligned}$$ where $\|\bm{x}_i\|_0$ denotes the $L_0$ norm of $\bm{x}_i$.
\
\[algo:sph-start\] \[algo:sph-end\]\
\[algo:sph-update\]
The spherical $k$-means assumes that object feature vector $\bm{x}_i\!\in\!\mathbb{R}^D$ is normalized by its $L_2$ norm as $\|\bm{x}_i\|_2\!=\!1$, i.e., a point on a unit hypersphere. Instead of a Euclidean distance used by the standard $k$-means algorithm, the spherical $k$-means algorithm employs a cosine similarity between $\bm{x}_i$ and $\bm{\mu}_j$, expressed by $$\rho(\bm{x}_i,\bm{\mu}_j)\!=\! \bm{x}_i\cdot \bm{\mu}_j~ ,$$ where $\bm{x}_i\!\cdot\! \bm{\mu}_j$ denotes the inner product of $\bm{x}_i$ and $\bm{\mu}_j$ and $\|\bm{\mu}_j\|_2\!=\!1$, i.e., $\bm{\mu}_j$ is a point on the unit hypersphere. Then the spherical $k$-means clustering problem is formulated as $${\cal C}^* = {\operatornamewithlimits{\mathrm{arg\,max}}}_{{\cal C}=\{ C_1,\cdots, C_k\}}
\left(\: \sum_{C_j\in{\cal C}}\:\sum_{\bm{x}_i\in C_j}
\bm{x}_i\cdot\bm{\mu}_j \:\right)\:.
\label{eq:sph_obj_funct}$$ Under the condition of $\|\bm{x}_i\|_2\!=\!\|\bm{\mu}_j\|_2\!=\!1$, Eqs. (\[eq:obj\_funct\]) and (\[eq:sph\_obj\_funct\]) are equivalent [^1] because $\|\bm{x}_i-\bm{\mu}_j\|_2^2\!= 2(1 -\!\bm{x}_i\cdot \bm{\mu}_j)$. The spherical $k$-means algorithm based on the same iterative heuristics as the standard one is shown in [**Algorithm**]{} \[algo:sph\]. Thus the spherical $k$-means algorithm [@dhillon] corresponds to the standard $k$-means algorithm for a general data set in the Euclidean space. In the previous work [@dhillon], neither its acceleration algorithms nor how to leverage sparseness of a data (object) set is disclosed. Our work is based on the same settings as the spherical $k$-means and provides an efficient algorithm that exploits the sparseness of a data set.
Inverted-File Data Structure {#subsec:ivfstr}
----------------------------
\
An inverted file is a type of data structures that is often employed for a data set of texts consisting of sparse feature vectors [@samet]. Instead of listing the feature elements of a given object, we list the objects with a given feature element for the inverted file [@knuth].
Figure \[fig:str\] shows a full and a sparse expression of the object feature vectors in (a) a standard structure and (b) an inverted-file data structure. In the standard structure, a feature vector with a full expression is represented by $\bm{x}_i\!\in\!\mathbb{R}^D$, $i\!=\!1,\cdots,N$, where element $x_{(i,t)}$ whose term does not appear in the $i$-th object is padded by zero. Define a set of non-zero elements in $\bm{x}_i$ as $\hat{\cal X}_i\!=\!\{ (t,x_{(i,t)})| x_{(i,t)}\!\neq\!0 \}$. Then a feature vector with sparse expression $\hat{\bm x}_i$ is represented by $\hat{\bm x}_i\!=\!(t_{(i,h)},v_{(i,h)})$, $h\!=\!1,2,\cdots,|\hat{\cal X}_i|\!=\!(nt)_i$, where $v_{(i,h)}\!=\!x_{(i,t_{(i,h)})}$. Assume that the $t$-th elements in each $\bm{x}_i$ are picked up as a vector like Fig. \[fig:str\](b), left. The transpose of the vector is a feature vector with a full expression in the inverted-file data structure, which is represented by $\bm{y}_t\!=\!(y_{(t,1)},\cdots,y_{(t,N)})$, $y_{(t,q)}\!=\!x_{(q,t)}$. We define a set of $\bm{y}_t$ as $\bar{\cal X}\!=\!\{\bm{y}_1,\bm{y}_2,\cdots,\bm{y}_D\}$. Similar to the standard structure, a feature vector with sparse expression $\hat{\bm y}_t$ is represented by $\breve{\bm y}_t\!=\!(i_{(t,q)},u_{(t,q)})$, $q\!=\!1,2,\cdots,(ni)_t$, where $u_{(t,q)}\!=\!y_{(t,i_{(t,q)})}\!=\!x_{(i_{(t,q)},t)}$. Besides, $\breve{\cal X}\!=\!\{\breve{\bm y}_1,\breve{\bm y}_2,\cdots,\breve{\bm y}_D\}$. Note that we adopt a simple array among several sparse expressions.
By applying the sparse expression to a given object set with low sparsity, we can conserve the memory size although the sparse expression needs extra memory capacitance for storing term IDs as $t_{(i,h)}$ or object IDs as $i_{(t,q)}$. Most text-search algorithms utilize an inverted file (or an inverted index) prebuilt from a text data set as a database. Given a query that is often a set of terms such as words or phrases, the search algorithms find relevant texts to the query from the text data set using terms in the query as search keys [@harman; @knuth; @buttcher; @zobel].
As well as text search, an inverted-file structure has also been employed for image search [@sivic]. A search algorithm for object retrieval in videos employed visual words for a feature of an image (a video frame) [@sivic]. The visual words are generated by the vector quantization of local descriptors extracted from images. Consequently, each image is represented as a sparse feature vector, each element of which is a tuple of a visual word ID and a feature value ([*tf-idf*]{}). Based on this representation, similar to a text search, the inverted-file data structure for an image database is utilized to perform a fast search.
So far, the data structure is based on the relationship between an object and the terms contained by it. By using not the foregoing relationship but the relationship between an object and the clusters to which the object belongs, the concept of the inverted-file structure is extended in an image search [@jegou; @babenko]. In this case [@jegou; @babenko], each object is assigned to a disjoint cluster by vector quantization based on $k$-means clustering. In the inverted-file structure, objects are listed for each cluster that contains the objects as its members. Given an image as a query, product quantization [@jegou] narrows down a search space to a subset (cluster) to which the query belongs. This extended inverted-file structure resembles a hash table employed in a local-sensitive hashing (LSH) approach [@indyk; @andoni; @charikar].
There is a $k$-means clustering algorithm that directly exploits a search algorithm using an inverted-file structure at its assignment step [@broder]. A Lloyd-type algorithm uses a linear scan (brute-force) search at the assignment step to find the most similar centroid to each object. Similar to text-search algorithms, the reported algorithm called wand-k-means [@broder] applies an inverted-file structure to a set of [*invariant*]{} data objects. The wand-k-means regards a set of centroids as queries and finds similar objects to each of the queries by a heuristic search algorithm called WAND at the assignment step. Except for the search algorithm, an important difference between wand-k-means and our [*IVF*]{} is in feature vectors represented with an inverted-file structure: [*invariant*]{} data object feature vectors and [*variable*]{} mean feature vectors at every iteration. This difference prompts the question: which can better achieve high performance? We discuss this issue in Section \[subsec:ivfd\].
Design Guidelines of Efficient Algorithms {#subsec:design}
-----------------------------------------
For efficiently processing a large-scale high-dimensional data set in a modern computer system, parallel processing is effective. There are several levels in parallel processing: instruction-level parallelism (ILP), data-level parallelism (DLP), and thread-level parallelism (TLP) [@hennepat]. We focus on ILP and design an efficient algorithm for a single thread by a single process. A Lloyd-type $k$-means clustering algorithm operating at high throughput in ILP achieves high performance in other parallelisms. This is because its procedure is suitable for explicit parallelisms at their costly assignment step, where a linear scan search for each object independently identifies the most similar centroid (mean) to the object in all the $k$ centroids [@jian; @bhimani].
To completely exploit ILP in a modern computer system, which has deep pipelines with superscalar out-of-order execution in a CPU core and a deep memory hierarchy from registers to external storages, pipeline hazards that cause stalls must be reduced. Among them, control hazards caused by branch mispredictions and data hazards arising from the dependency of instructions on the results of previously executed instructions are critical to increase the performance of the algorithms and their implementations.
The impact of branch mispredictions on algorithm performance has been analyzed, and algorithms that reduce the branch mispredictions have been developed [@kaligosi; @edelkamp; @green]. For a classical quicksort, which is a well-known sort algorithm, a counterintuitive observation of selecting as a pivot not a median of a partitioned array but a skewed pivot (an entry distant from the median) leads to high performance is analyzed and explained based on the balance of the number of comparison operations and branch mispredictions [@kaligosi]. BlockQuicksort [@edelkamp], which is a kind of the dual-pivot quicksort, suppresses the branch mispredictions incurred by conditional branches. Besides sort algorithms, the classic Shiloach-Vishkin algorithm that finds connected components, which is one graph algorithm, was improved in terms of speed performance by avoiding branch mispredictions [@green].
Data hazards accompanied by access to external memories like DRAMs seriously affect the speed performance because of high memory latency. To prevent this performance degradation, algorithms and their implementations have been studied, which efficiently exploit caches in a CPU core for reducing expensive access to external memories [@kowarschik; @frigo]. Cache-aware (-conscious) algorithms [@kowarschik] are optimized based on such actual parameters as capacity, block size, and associativity for increasing the cache hit rate, while cache-oblivious algorithms [@frigo] are designed and tuned with cache consideration and without variables that are dependent on the actual parameters. Frequent pattern mining algorithms [@ghoting] reduce cache misses by improving spatial and temporal locality in data access with cache-conscious methods, resulting in high performance. A similarity join algorithm [@perdacher] achieves high-speed performance by transforming a conventional loop iteration into a cache-oblivious one.
Thus, preventing pipeline hazards is important for designing a high-performance algorithm suitable for a modern computer system. Although our proposed [*IVF*]{} is not a cache-aware algorithm, its structure suppresses the pipeline hazards shown in Sections \[sec:prop\] and \[sec:exp\]. The algorithm is analyzed from the viewpoint of the foregoing performance degradation factors that cause pipeline hazards in Section \[sec:analy\].
Proposed algorithm: IVF {#sec:prop}
=======================
\
\[algo:prop-start\]\
We propose an inverted-file $k$-means clustering algorithm ([*IVF*]{}) for a large-scale and high-dimensional sparse data set with potentially numerous classes. [*IVF*]{} is a Lloyd-type algorithm, i.e., an iterative heuristic algorithm, which keeps the same solution as a standard Lloyd algorithm [@lloyd; @macqueen] under an identical initial state. Due to this property, we do not discuss accuracy (or an objective function value) as performance. We evaluate both the maximum memory capacitance and the CPU time (or the clock cycles) required by the algorithm through iterations.
[**Algorithm**]{} \[algo:ours\] shows the [*IVF*]{} pseudocode at the $r$-th iteration. [*IVF*]{} receives a centroid set, which is the mean set at the previous iteration, with inverted-file sparse expression $\breve{\cal M}^{[r-1]}$ and uses an invariant object set with standard sparse expression $\hat{\cal X}$ and returns cluster set ${\cal C}^{[r]}$ consisting of $k$ clusters and $\breve{\cal M}^{[r]}$.
[*IVF*]{} has two steps; assignment and update. The assignment step at lines 5–19 executes a linear-scan search in the triple loop, where an object feature vector is regarded as a query. The $i$-th object feature vector ($\hat{\bm{x}}_i$) consists of $(nt)_i$ tuples $(t_{(i,h)},v_{(i,h)})$, $h\!=\!1,2,\cdots,(nt)_i$, where $(nt)_i$ denotes the $L_0$ norm of $\bm{x}_i$ ($(nt)_i\!=\!\| \bm{x}_i\|_0$), $h$ is the local counter, $t_{(i,h)}$ is the global (serial) term ID from 1 to $D$, and $v_{(i,h)}$ is a corresponding value such as [*tf-idf*]{}. For each term with term ID $t_{(i,h)}$ ($s$ for simplicity), inverted-file centroid array $\breve{\bm{\xi}}_s^{[r-1]}$ is selected. This array consists of $(nc)_s$ tuples $(c_{(s,q)},u_{(s,q)})^{[r-1]}$, $q\!=\!1,2,\cdots,(nc)_s$, where $c_{(s,q)}$ denotes the global centroid ID from 1 to $k$, $u_{(s,q)}$ is the corresponding value, and $(nc)_s$ denotes the centroid (mean) frequency of term ID $s$. Then the partial similarity (inner product) between the $i$-th object and the $c_{(s,q)}$-th centroid is calculated and stored at $\rho_{c_{(s,q)}}$. Just after the inner double loop has been completed, the $i$-th object is assigned to the $a(\hat{\bm{x}}_i)$-th cluster whose centroid most closely resembles.
The update step at lines 21–33 calculates each mean of $k$ clusters based on the object assignment. For cluster $C_j^{[r]}$ whose members $\hat{\bm x}_i\!\in\! C_j^{[r]}$ are determined at the assignment step, each feature value $v_{(i,h)}$ is added to $w_s$, where $s$ denotes global term ID $t_{(i,h)}$ from 1 to $D$. After the addition for all the members, each value $w_p$ ($1\!\leq\! p\!\leq\! D$) is divided by cluster size $|C_j^{[r]}|$ and $\|\bm{w}\|_2$ is calculated. To represent the mean of $C_j^{[r]}$ with the inverted-file sparse expression, we perform the procedure at lines 28–32, where $p$ denotes the global term ID and $q_p$ is the local counter of $p$. Then the mean of $C_j^{[r]}$ is expressed by a set of tuples $(c_{(p,q_p)},u_{(p,q_p)})$ where $c_{(p,q_p)}$ denotes cluster ID $j$ and $u_{(p,q_p)}$ is the corresponding feature value. Thus the tuple $(c_{(p,q_p)},u_{(p,q_p)})$, which is the $q_p$-th element of $\breve{\bm \xi}_p^{[r]}$, is obtained.
[*IVF*]{} simultaneously satisfies the two requirements of low memory consumption and high speed. The sparse expressions for both object set $\hat{\cal X}$ and mean set $\breve{\cal M}$ suppress memory consumption. The inverted-file data structure for the mean (centroid) set achieves high-speed performance. To qualitatively evaluate the [*IVF*]{} performance, we design three algorithms in Section \[sec:algos\] and compare [*IVF*]{} with them in two distinct real document data sets in Section \[sec:exp\]. Furthermore, we analyze the speed performance to identify factors that determine the performance in Section \[sec:analy\].
Compared algorithms {#sec:algos}
===================
[1.08]{}
----------- -------------------- --------------- ---------------
Data
structure Sparse
Two-way merge Non-branch Branch
[*TWM*]{} [@knuth] [***MFN***]{} [*MFB*]{}
Proposed Non-branch Branch
[***IVF***]{} [***IFN***]{} [***IFB***]{}
----------- -------------------- --------------- ---------------
: Classification of compared algorithms
\[table:class\]
To shed light on the characteristics of [*IVF*]{}, we designed three algorithms, which may be not suitable for practical use due to their required memory capacitances. One is called a mean full-expression algorithm with a non-branch ([*MFN*]{}). The others are an inverted-file full-expression algorithm with branch ([*IFB*]{}) and non-branch ([*IFN*]{}). Similar to [*IVF*]{}, all three algorithms represent a given object set with a standard-structure sparse expression in Fig. \[fig:str\](a) bottom. The difference is in their data structures and expressions for a mean set. Table \[table:class\] shows the classification of the three algorithms and [*IVF*]{}.
[*MFN*]{} employs a standard data structure with a full expression for a mean (centroid) set shown in Fig. \[fig:str\](a) top, where subscript $i$ is replaced with $j$ for the means, $j\!=\! 1,2,\cdots,k$. Mean sets $\breve{\cal M}^{[r-1]}$ and $\breve{\cal M}^{[r]}$ at lines 1 (I), 2 (I’), and 34 (I’) in [**Algorithm**]{} \[algo:ours\] are replaced with ${\cal M}^{[r-1]}$ and ${\cal M}^{[r]}$. When mean feature vector $\bm{\mu}_j$ is represented with a full expression, values of entries for some global term IDs may be undefined. Then each of the entries is padded with zero. The similarity between object feature vector $\hat{\bm x}_i\!=\!((t_{(i,1)},v_{(i,1)}),\cdots,(t_{(i,(nt)_i)},v_{(i,(nt)_i)}))$ and centroid (mean) feature vector $\bm{\mu}_j$ is calculated by $$\rho_j = \sum_{h=1}^{(nt)_i} v_{(i,h)}\!\times\! \mu_{(j,t_{(i,h)})}\: ,
\label{eq:mfn_sim}$$ where $\mu_{(j,t_{(i,h)})}$ denotes the element with the global term ID of $t_{(i,h)}$ in $\bm{\mu}_j$. When $\mu_{(j,t_{(i,h)})}\!=\!0$ in Eq. (\[eq:mfn\_sim\]), there are two approaches: the execution of zero multiplication and the insertion of the conditional branch for skipping the zero multiplication. [*MFN*]{} employs the former approach. We call the former approach [*non-branch*]{} and the latter [*branch*]{}. From the algorithmic point of view, lines 9 (II) and 10 (II’) in [**Algorithm**]{} \[algo:ours\] are replaced as follows.\
[**do**]{}
\
The update step is modified from that of the spherical $k$-means algorithm shown in [**Algorithm**]{} \[algo:sph\] for the use of object feature vectors with sparse expression. We can evaluate the effect of the inverted-file data structure [*itself*]{} on the speed performance by comparing [*MFN*]{} with the following [*IFN*]{}.
Both [*IFN*]{} and [*IFB*]{} utilize an inverted-file data structure with full expressions for the means, which resembles that in Fig. \[fig:str\](b) right top. The inverted file has all the $k$ entries for each term while [*IVF*]{} has $(nc)_s\!\leq\!k$ entries for a term whose global term ID is $s$. Then lines 9 (II), 10 (II’), and 12 (IV) in [**Algorithm \[algo:ours\]**]{} are replaced with\
[**do**]{}
,\
where $\bar{\cal M}$ indicates the set of the mean feature vectors represented by the inverted-file data structure with all the $k$ entries for each of the $D$ terms and $\bar{\bm{\xi}}_s$ denotes the value array of the $s$-th term. Mean sets $\breve{\cal M}^{[r-1]}$ and $\breve{\cal M}^{[r]}$ in [**Algorithm**]{} \[algo:ours\] are replaced with $\bar{\cal M}^{[r-1]}$ and $\bar{\cal M}^{[r]}$. The undefined values in $\bar{\bm{\xi}}_s$ are padded with zeros. Then the similarity between $\hat{\bm x}_i$ and the $j$-th centroid (mean) is expressed by $$\rho_j = \sum_{h=1}^{(nt)_i} v_{(i,h)}\!\times\! u_{(s,j)},
\hspace{3mm} s = t_{(i,h)}\: .$$ The difference between [*IFB*]{} and [*IFN*]{} is concerned with whether the zero multiplications in the partial similarity calculations are skipped, based on the conditional branch statement of\
,\
which is inserted at line 11 (III) in [**Algorithm \[algo:ours\]**]{}. The algorithm with the conditional branch is [*IFB*]{} and the other is [*IFN*]{}.
Using the conditional branch at (III) in [**Algorithm \[algo:ours\]**]{} has an advantage and a disadvantage. The advantage is the decrease of the number of costly operations related to floating-point multiplications and additions at line 12 in [**Algorithm \[algo:ours\]**]{}. The disadvantage is the increase of the numbers of both instructions and branch mispredictions. Comparing [*IFB*]{} and [*IFN*]{} in Section \[sec:exp\] explains the impact of branch mispredictions on speed performance.
Let us briefly review the relationship among the four algorithms. Consider [*IFN*]{} as a baseline algorithm. The difference of [*MFN*]{} and [*IFN*]{} is only in their standard and inverted-file data structures. The difference of [*IFB*]{} and [*IFN*]{} is only in how to process the zero multiplications, whether they are skipped by the inserted conditional branch or calculated without the conditional-branch insertion. The difference of [*IVF*]{} and [*IFN*]{} is only in their mean expressions: sparse or full.
Note that [*MFB*]{} and [*TWM*]{} in Table \[table:class\] were not compared. The [*MFB*]{} performance can be estimated by the comparison results of [*IFB*]{} and [*IFN*]{}. [*TWM*]{} was prepared as an algorithm for both the object and the mean feature vectors represented by the standard data structure with a sparse expression. To calculate the similarity of $\hat{\bm x}_i$ and centroid feature vector $\hat{\bm \mu}_j$, the feature values with identical global term IDs have to be detected in both the vectors, i.e., the set-intersection operation in terms of global term ID has to be executed. [*TWM*]{} uses a [*two-way merge*]{} algorithm for the set-intersection operation [@knuth]. [*TWM*]{}, which has many conditional branches that induce cache misses, operated very slowly in our preliminary experiments based on identical settings as the others.
Experiments {#sec:exp}
===========
We describe data sets, a platform for executing the algorithms, and the performance of the four algorithms, our proposed [*IVF*]{} and three others in Section \[sec:algos\].
------------ ---------
(a) PubMed (b) NYT
------------ ---------
Data Sets {#subsec:data}
---------
We employed two different types of large-scale and high-dimensional sparse real document data sets: [*PubMed Abstracts*]{} (PubMed for short) [@pubmed] and [*The New York Times Articles*]{} (NYT).
The PubMed data set contains 8,200,000 documents (texts) represented by the term (distinct word) counts in each. We made a feature vector normalized by its $L_2$ norm from each document, each of which consisted of the [*tf-idf*]{} values of the corresponding terms. Each feature vector was regarded as a point on a unit hypersphere. We chose 1,000,000 feature vectors at random without duplication from all of the vectors as our 1M-sized experimental data sets. The number of distinct terms in the data set (dimensionality) was 140,914. Their average frequency in the documents, i.e., the average number of non-zero elements in the feature vectors, was 58.95, and the average sparsity in Eq. (\[eq:avg\_sparsity\]) was $3.93\!\times\!10^{-4}$.
We extracted 1,285,944 articles from [*The New York Times Articles*]{} from 1994 to 2006 and counted the frequency of the term occurrences after stemming and stop word removal. In the same manner as PubMed, we made a set of feature vectors with 495,714 dimensionality. The average number of non-zero elements in the feature vectors was 225.76, corresponding to an average sparsity of $4.56\!\times\!10^{-4}$.
Platform and Measures {#subsec:platform}
---------------------
All the algorithms were executed on a computer system, which was equipped with two Xeon E5-2697v3 2.6-GHz CPUs with three-level caches from levels 1 to level 3 [@hammarlund] and a 256-GB main memory, by a single thread on a single process within the memory capacity. When the algorithms were executed, two hardware prefetchers related to the level-2 caches in the CPU were disabled by BIOS control [@intel] to measure the effect of the cache misses themselves. The algorithms were implemented in C and compiled with the GNU compiler collection (gcc) version 8.2.0 on the optimization level of [-O0]{}. The performances of the algorithms were evaluated with CPU time (or clock cycles) until convergence and the maximum physical memory size occupied through the iterations.
--------------------------- ----------------------
(a) Avg. \#terms in means (b) \#terms in means
--------------------------- ----------------------
\[c\]\[c\]\[0.9\][Number of clusters: $k$]{} \[c\]\[c\]\[0.9\]
(0,0) (0,0)[(0,22)\[c\][Max. memory size (MB)]{}]{}
\[c\]\[c\]\[0.82\][$10^2$]{} \[c\]\[c\]\[0.82\][$10^3$]{} \[c\]\[c\]\[0.82\][$10^4$]{} \[r\]\[r\]\[0.82\][$10^1$]{} \[r\]\[r\]\[0.82\][$10^2$]{} \[r\]\[r\]\[0.82\][$10^3$]{} ![Maximum memory capacitance for mean tuples required by [*IVF*]{} through iteration in PubMed. []{data-label="fig:mmem"}](./mmem_pubmed1M.eps "fig:"){width="50mm"}
Performance {#subsec:perform}
-----------
### Required Maximum Memory Size
We measured the maximum memory size required by the algorithms through the iterations until the convergence (Fig. \[fig:mem\]). The four algorithms represented the object data set with the sparse expression of the tuple $(t_{(i,h)},v_{(i,h)})$ shown in Sections \[sec:prop\] and \[sec:algos\]. As types of elements $t_{(i,h)}$ and $v_{(i,h)}$, an integer (int) and a 64-bit floating point (double) were used[^2]. The memory capacitance occupied by the object set is expressed by $(\sum_{i=1}^N (nt)_i)\!\times\!(\mbox{sizeof(int + double)})$. Those of PubMed and NYT were 706.8 MB and 3,484 MB.
\[c\]\[c\]\[0.9\][Number of clusters: $k$]{} \[c\]\[c\]\[0.9\]
(0,0) (0,0)[(0,22)\[c\][Avg. CPU time (sec)]{}]{}
\[c\]\[c\]\[0.82\][$10^2$]{} \[c\]\[c\]\[0.82\][$10^3$]{} \[c\]\[c\]\[0.82\][$10^4$]{} \[c\]\[c\]\[0.82\] \[r\]\[r\]\[0.82\][$10^1$]{} \[r\]\[r\]\[0.82\][$10^2$]{} \[r\]\[r\]\[0.82\][$10^3$]{} \[r\]\[r\]\[0.82\][$10^4$]{} \[r\]\[r\]\[0.82\][$10^5$]{} \[r\]\[r\]\[0.7\][*IVF*]{} \[r\]\[r\]\[0.7\][*IFN*]{} \[r\]\[r\]\[0.7\][*IFB*]{} \[r\]\[r\]\[0.7\][*MFN*]{} ![Average CPU time per iteration required by each of four algorithms. CPU time was plotted along $k$ with log-log scale when each algorithm was applied to (a) PubMed and (b) NYT. []{data-label="fig:time"}](./time_pubmed1M.eps "fig:"){width="53mm"}\
(a) PubMed
\
\[c\]\[c\]\[0.9\][Number of clusters: $k$]{} \[c\]\[c\]\[0.9\]
(0,0) (0,0)[(0,22)\[c\][Avg. CPU time (sec)]{}]{}
\[c\]\[c\]\[0.82\][$10^2$]{} \[c\]\[c\]\[0.82\][$10^3$]{} \[c\]\[c\]\[0.82\][$10^4$]{} \[c\]\[c\]\[0.82\] \[r\]\[r\]\[0.82\][$10$]{} \[r\]\[r\]\[0.82\][$10^2$]{} \[r\]\[r\]\[0.82\][$10^3$]{} \[r\]\[r\]\[0.82\][$10^4$]{} \[r\]\[r\]\[0.82\][$10^5$]{} \[r\]\[r\]\[0.7\][*IVF*]{} \[r\]\[r\]\[0.7\][*IFN*]{} \[r\]\[r\]\[0.7\][*IFB*]{} \[r\]\[r\]\[0.7\][*MFN*]{} ![Average CPU time per iteration required by each of four algorithms. CPU time was plotted along $k$ with log-log scale when each algorithm was applied to (a) PubMed and (b) NYT. []{data-label="fig:time"}](./time_nyt1M.eps "fig:"){width="53mm"}\
(b) NYT
By contrast, the memory capacitance for the mean set depends on the algorithms and the number of means $k$. The three algorithms with full expressions ([*MFN*]{}, [*IFB*]{}, and [*IFN*]{}) need an identical memory capacitance expressed by $k\!\times\! D\!\times\! (\mbox{sizeof(double)})$, where $D$ denotes not the number of distinct terms in the mean feature vectors but the dimensionality, including zero padding. The memory capacitances for PubMed and NYT were $1.20\!\times\!k$ MB and $3.96\!\times\!k$ MB and reached 24.0 GB and 79.2 GB at $k\!=\!20,000$.
The memory capacitance required by [*IVF*]{} for the mean feature vectors is expressed by $(\sum_{p=1}^{D} (nc)_p)\!\times\! (\mbox{sizeof(int + double)})$, which is equivalent to $(\sum_{j=1}^k (ntm)_j)\!\times\! (\mbox{sizeof(int + double)})$, where $(ntm)_j$ denotes the number of distinct terms in the $j$-th mean feature vector. Figures \[fig:mterms\](a) and (b) show $(\sum_{j=1}^k (ntm)_j)/k$ for each iteration when [*IVF*]{} started at the initial state chosen randomly in PubMed and both the maximum $(\sum_{j=1}^k (ntm)_j)/k$ through iterations and that at the convergence. As shown in Fig. \[fig:mterms\](a), the average number of mean terms became stable after several iterations for each $k$ value. Figure \[fig:mterms\](b) indicates that the maximum average number almost coincided with the average number at the convergence, and both numbers decreased as a power-law function of $k$. Using the maximum average number of mean terms in Fig. \[fig:mterms\], the maximum memory capacitance that [*IVF*]{} needed was calculated with various $k$ for PubMed. Figure \[fig:mmem\] shows that the memory size increased as a sublinear function of $k$, and even when $k\!=\!20,000$, the memory size was only 345.7 MB.
Thus by applying the sparse expressions to a sparse data set we significantly reduced the memory capacitance occupied by the object and mean feature vectors.
### CPU Time
Figures \[fig:time\](a) and (b) show the average CPU times per iteration with $k$ in the log-log scale required by the four algorithms until the convergence, when they were applied to PubMed and NYT. Regarding the speed performance in the two distinct data sets, the relationships among the algorithms were almost the same. [*IVF*]{} achieved the best performance in the range of large $k$ values. When $k\!=\!20,000$ in PubMed shown in Fig. \[fig:time\](a), the CPU time of [*IVF*]{} was only $33.7\%$ of [*IFN*]{} (the second best). By contrast, both algorithms were competitive in the small $k$ range. These performances are analyzed in Section \[sec:analy\] and scrutinized in Section \[subsec:ifn\].
[*MFN*]{} needed much more CPU time than the others that employed the inverted-file data structure. The CPU time for PubMed reached $4.89$ times more than that of [*IFN*]{}, which only differs from [*MFN*]{} in the mean data structure, whether it is the inverted-file or the standard, as described in Section \[sec:algos\]. This actually indicates that the inverted-file data structure is useful for a large-scale sparse data set.
Our comparison of [*IFB*]{} and [*IFN*]{} was interesting. It intuitively seems that [*IFB*]{}, which skips costly unnecessary floating-point multiplications using the conditional branch, operates faster than [*IFN*]{} that directly executes the multiplications. Surprisingly, [*IFN*]{} was faster than [*IFB*]{} in every range of $k$ in both data sets. [*IFB*]{} required $1.28$ to $1.49$ times more CPU time than [*IFN*]{}. Executing the conditional branch many times, e.g., in the innermost loop of the triple loop, risks degrading the speed performance.
Performance Analysis {#sec:analy}
====================
------------ ---------
(a) PubMed (b) NYT
------------ ---------
When executing the four algorithms in PubMed and NYT, we measured the number of completed (retired) instructions, cache misses, and branch mispredictions with the [*perf tool*]{} (Linux profiling with performance counters) [@perf]. Hereinafter, we label the four numbers as follows: the instructions, the level-1 (L1) data cache misses, the last-level (LL) cache misses, and the branch mispredictions as Inst, L1CM, LLCM, and BM. These four numbers are collectively called [*performance degradation factors*]{} (DFs). As they increase, the speed performance worsens. To estimate the effects of each DF on the total clock cycles (or the CPU time), we introduced a simple yet practical clock-cycle per instruction (CPI) model and analyzed the four algorithms based on it.
Performance Degradation Factor Characteristics {#subsec:dfs}
----------------------------------------------
Figures \[fig:inst\](a) and (b) show the average number of completed instructions through iterations until convergence when the four algorithms were executed in PubMed and NYT. The algorithms had almost the same characteristics, and their relationships were similar in the distinct data sets and shared three characteristic points:
1. The number of instructions of [*MFN*]{} coincided with that of [*IFN*]{}. The rate expressed by $| (\mbox{Inst}_{[\mbox{\tiny\em IFN}]}/\mbox{Inst}_{[\mbox{\tiny\em MFN}]})\!-\! 1|$ was within 1.1%. This fact is adopted as the assumption of the parameter optimization in Section \[subsec:cpimodel\].
2. $\mbox{Inst}_{[\mbox{\tiny\em IFB}]}$ started at a larger value than $\mbox{Inst}_{[\mbox{\tiny\em IFN}]}$ and ended at a smaller value at $k\!=\!20,000$. This is related to the sparsity of the mean feature vectors. Comparing [*IFB*]{} with [*IFN*]{}, $\mbox{Inst}_{[\mbox{\tiny\em IFB}]}$ increased by the insertion of a conditional branch to avoid unnecessary operations of both zero-multiplications and additions at line 12 in [**Algorithm \[algo:ours\]**]{}. As the sparsity is lowered, i.e., fewer terms appeared in the mean feature vectors, more instructions related to the multiplications and additions are skipped. The sparsity became lower with $k$, as shown in Fig. \[fig:mterms\](b). Thus $\mbox{Inst}_{[\mbox{\tiny\em IFB}]}$ and $\mbox{Inst}_{[\mbox{\tiny\em IFN}]}$ intersected at a large $k$ value.
3. $\mbox{Inst}_{[\mbox{\tiny\em IVF}]}$ had remarkable characteristics to $\mbox{Inst}_{[\mbox{\tiny\em IFN}]}$, similar to $\mbox{Inst}_{[\mbox{\tiny\em IFB}]}$. This is discussed in connection with the CPU time in Section \[subsec:ifn\].
------------ ---------
(a) PubMed (b) NYT
------------ ---------
------------ ---------
(a) PubMed (b) NYT
------------ ---------
------------ ---------
(a) PubMed (b) NYT
------------ ---------
------------ ---------
(a) PubMed (b) NYT
------------ ---------
To analyze performance based on CPI, we introduced performance degradation factors per instruction defined by $$\phi_1 = \frac{(\mbox{L1CM}\!-\!\mbox{LLCM})}{\mbox{Inst}},\:\:
\phi_2 = \frac{\mbox{LLCM}}{\mbox{Inst}},\:\:
\phi_3 = \frac{\mbox{BM}}{\mbox{Inst}}\: ,
\label{eq:model_funct}$$ in addition to L1CM, LLCM, and BM. Figures \[fig:clk-inst\], \[fig:l1cm-inst\], \[fig:llcm-inst\], and \[fig:bm-inst\] show the actual CPI, $\phi_1$, $\phi_2$, and $\phi_3$ with the number of clusters $k$, where L1CM$'$ in Fig. \[fig:l1cm-inst\] denotes $(\mbox{L1CM}\!-\!\mbox{LLCM})$ and $k$ is omitted from $\phi_i(k)$ for simplicity. From all the figures, each of the algorithms indicated the same tendencies on the characteristics when applied to PubMed and NYT.
Figure \[fig:clk-inst\] shows that the inverted-file data structure was effective for lowering CPI. The three algorithms with an inverted-file data structure operated at CPIs from 0.26 to 0.61 through all $k$ values in both data sets while [*MFN*]{} ranged from 0.65 to 1.94. [*MFN*]{} whose CPI exceeded 1.0 in the large $k$ range lost the effect of superscalar execution. The others’ CPIs were reasonable because the CPU core had eight units, including four ALUs [@hammarlund]. In the large $k$ range, we arranged the four algorithms in ascending order of CPI: [*IVF*]{}, [*IFN*]{}, [*IFB*]{}, and [*MFN*]{}.
Figures \[fig:l1cm-inst\] and \[fig:llcm-inst\] show L1CM$'$ per instruction ($\phi_1$) and LLCM per instruction ($\phi_2$) for the algorithms along $k$. These figures indicate that the L1CM$'$/Inst and LLCM/Inst of [*MFN*]{} were conspicuously large. The decrease of L1CM$'$/Inst in the large $k$ range was attributed to the high joint probability at which the L1 and LL cache misses occurred. [*IFB*]{} and [*IFN*]{} had identical characteristics in terms of L1CM$'$/Inst and LLCM/Inst in the $k$ range. This fact is used for the assumption of the parameter optimization in Section \[subsec:cpimodel\]. Regarding LLCM/Inst, [*IVF*]{} achieved the lowest values as a whole.
Figure \[fig:bm-inst\] shows BM per instruction ($\phi_3$) with $k$ in the log-log scale. [*IFB*]{} showed different characteristics from the others. Its conditional branch induced many branch mispredictions because the branch predictor in the CPU core often failed to select the next true instruction due to the zeros’ irregular positions in the inverted file. This characteristic negatively impacted the speed performance of [*IFB*]{}, as shown in Section \[subsec:cpimodel\].
Clock-Cycle per Instruction (CPI) Model {#subsec:cpimodel}
---------------------------------------
We introduce a clock-cycle per instruction (CPI) model, which is a simple linear function of $k$, expressed by $$\mbox{CPI}(k) = w_0 +\sum_{i=1}^{3} w_i \cdot \phi_i(k) \: ,
\label{eq:model}$$ where $w_0$ denotes the expected clock cycles per instruction when cache misses and branch mispredictions do not occur, $w_1$ is the overall penalty per L1CM$'$/Inst when a level-1 data cache miss occurs and a last-level cache hit occurs at the worst case, $w_2$ is the expected memory stall cycles per LLCM/Inst, and $w_3$ is the expected branch misprediction penalty per BM/Inst including the penalty of the number of wasted instructions. Note that $w_2$ does not mean the expected memory latency per instruction due to the out-of-order execution [@hennepat].
[1.1]{}
--------------- --------------- --------------- --------------- -------------- --------------- --------------
[Avg. err.]{} [Max. err.]{}
$w_0$ $w_1$ $w_2$ $w_3$ (%) (%)
[*MFN*]{} 0.255 7.52 56.1 23.8 5.96 9.32
[*IFB*]{} 0.262 5.52 30.8 23.8 0.969 3.93
[*IFN*]{} 0.255 5.52 30.8 23.8 0.617 4.55
[***IVF***]{} [**0.243**]{} [**3.13**]{} [**13.5**]{} [**23.8**]{} [**0.461**]{} [**3.19**]{}
--------------- --------------- --------------- --------------- -------------- --------------- --------------
: Optimized CPI model parameters and errors on actual CPIs
\[table:params\]
------------ ---------
(a) PubMed (b) NYT
------------ ---------
For the optimization of parameters $w_i$, we assumed that they are independent of the data sets and dependent on the algorithms. Based on the relationship between the pairs of algorithms, we also made the following three assumptions. The first is that [*MFN*]{} and [*IFN*]{} share $w_0$ because the algorithms have an identical triple loop[^3] at their assignment step, except the accessed data arrays whose data structure is either standard or inverted-file. The second is that [*IFB*]{} and [*IFN*]{} share $w_1$ and $w_2$ since these algorithms only differ over whether the conditional branch in the innermost loop in the triple loop is set. The last is that all the algorithms share $w_3$. Under these assumptions, we optimized the parameters so that the squared error between the actual and model CPI in Eq. (\[eq:model\]) was minimized:
1. [**Target**]{}: $w_1$ and $w_2$ of [*MFN*]{}\
[**Data**]{}: Differences of CPI, $\phi_1$ and $\phi_2$ values\
of [*MFN*]{} and [*IFN*]{}\
[**Condition**]{}: $w_3\!=\!0$.
2. [**Target**]{}: $w_0$ of [*MFN*]{}\
[**Data**]{}: [*MFN*]{}’s CPI data\
[**Condition**]{}: $w_1$ and $w_2$ are fixed at the optimized values\
and $w_3\!=\!0$.
3. [**Target**]{}: $w_1$ and $w_2$ of [*IFN*]{}\
[**Data**]{}: [*IFN*]{}’s CPI data\
[**Condition**]{}: $w_0$ is fixed at the [*MFN*]{}’s value\
and $w_3\!=\!0$.
4. [**Target**]{}: $w_0$ and $w_3$ of [*IFB*]{}\
[**Data**]{}: [*IFB*]{}’s CPI data\
[**Condition**]{}: $w_1$ and $w_2$ are fixed at the [*IFN*]{}’s values.
5. [**Target**]{}: $w_0$, $w_2$, and $w_3$ of [*IVF*]{}\
[**Data**]{}: [*IVF*]{}’s CPI data\
[**Condition**]{}: $w_3$ is fixed at the [*MFN*]{}’s value.
We obtained the parameters for each algorithm by this procedure and evaluated the accuracy of the CPI model by two measures. One is an average error (Avg. err.): $$\mbox{Avg. err.} = \left\{ \frac{1}{|K|} \sum_{k\in K} \left(
\mbox{CPI}_a(k) -\mbox{CPI}_m(k) \right)^2
\right\}^{\frac{1}{2}}\: ,
\label{eq:avgerr}$$ where $K$ is the set of $k$s in the experiments, i.e., $K\!=\!\{200,500,\cdots,20000\}$, and $\mbox{CPI}_a(k)$ and $\mbox{CPI}_m(k)$ denote the actual and model CPIs when the number of clusters is $k$. The other is a maximum error (Max. err.): $$\mbox{Max. err.} = \max_{k\in K}
\left|\frac{\mbox{CPI}_m(k)}{\mbox{CPI}_a(k)} -1\right|\: .
\label{eq:maxerr}$$
Table \[table:params\] shows the optimized parameters and the evaluation results. The parameters were reasonable values based on the computer architecture [@hammarlund] in our experiments. The errors were also below 10% in the range of all the $k$ values. [*IVF*]{}, in particular, reduced the wasted clock cycles that were caused by the cache misses. Figures \[fig:cpi\](a) and (b) show the actual and model CPIs of the four algorithms. We confirmed the model CPIs agree well with the actual CPIs of all the algorithms.
[1.2]{}
\#
---- --------------------------------------------------------------------- ----------------------------------------------------------------------------------------
[ $\hat{\bm{x}}_i\!\in\!\hat{\cal X}$ [**do**]{} ]{} [ $\hat{\bm{x}}_i\!\in\!\hat{\cal X}$ [**do**]{} ]{}
[$/\!/~|\hat{\cal X}|\!=\!N$ repeats.]{}
[$/\!/~\hat{\bm{x}_i}\!=\!(t_{(i,h)},v_{(i,h)})_{h=1}^{(nt)_i}$ ]{}
[ $s\!\leftarrow\!t_{(i,h)}\!\in\!S_i$ [**do**]{} ]{} [ $s\!\leftarrow\!t_{(i,h)}\!\in\!S_i$ [**do**]{} ]{}
[$/\!/~|S_i|\!=\!(nt)_i$ repeats.]{} [$\Leftrightarrow$ Identical to [*IFN*]{} ]{}
[ $u_{(s,j)}\!\in \bar{\bm{\xi}}_s$ [**do**]{} ]{} [ $(c_{(s,q)},u_{(s,q)})\!\in \breve{\bm{\xi}}_s$ [**do**]{} ]{}
[$/\!/~\bar{\bm{\xi}_s} \in \bar{\cal M}$ ]{} [$/\!/~\breve{\bm{\xi}_s} \in \breve{\cal M}$ ]{}
[$/\!/~\bar{\bm{\xi}_s}\!=\!(u_{(s,j)})_{j=1}^{k}$ ]{} [$/\!/~\breve{\bm{\xi}_s}\!=\!(c_{(s,q)},u_{(s,q)})_{q=1}^{(nc)_s}$ ]{}
[$\rho_j\!\leftarrow\!\rho_j\!+\!v_{(i,h)}\!\times\!u_{(s,j)}$ ]{} [$\rho_{c_{(s,q)}}\!\leftarrow\!\rho_{c_{(s,q)}}\!+\!v_{(i,h)}\!\times\!u_{(s,q)}$ ]{}
[$/\!/~\bm{k}$ repeats.]{} [$/\!/~\bm{(nc)_s}$ repeats.]{}
: Comparison of the triple loops in [*IFN*]{} and [*IVF*]{}
\[table:ifn\]
--------- ---------
(a) IFN (b) IVF
--------- ---------
Discussion {#sec:disc}
==========
We compare [*IVF*]{} with two similar algorithms, [*IFN*]{} in Section \[sec:algos\] and [*IVFD*]{} that is related to wand-k-means [@broder] in Section \[subsec:ivfstr\], and discuss their performances.
IFN and IVF {#subsec:ifn}
-----------
[*IFN*]{} operated in less CPU time than [*IVF*]{} in the small $k$ range in Figs. \[fig:time\](a) and (b). From the viewpoints of the performance degradation factors, [*IVF*]{} was inferior in this range to [*IFN*]{} based on the number of instructions in Figs. \[fig:inst\](a) and (b). We focus on the number of instructions needed by each algorithm, especially in the triple loop at the assignment step because most of the CPU time was spent in the triple loop based on our preliminary analyses.
Table \[table:ifn\] shows an overview of the triple loops in [*IFN*]{} and [*IVF*]{}. The two algorithms only differ in the innermost loop labeled as 3. [*IFN*]{} loads feature value $u_{(s,j)}$ in the $j$-th entry in array $\bar{\bm \xi}_s$ from an external memory or a cache, multiplies $u_{(s,j)}$ with $v_{(i,h)}$, and adds a multiplication value to partial similarity $\rho_j$. This procedure is repeated by $$k\cdot \sum_{i=1}^N (nt)_i = \sum_{i=1}^N \sum_{h=1}^{(nt)_i} k\: ,
\label{eq:ifn_mult}$$ where the number of repetitions corresponds to the number of multiplications. By contrast, [*IVF*]{} loads the tuple of mean ID $c_{(s,q)}$ and feature value $u_{(s,q)}$ in the $q$-th entry in array $\breve{\bm \xi}_s$. The number of repetitions of the foregoing procedure is expressed by $$\sum_{i=1}^N \sum_{h=1}^{(nt)_i} (nc)_s,\:~~~\: s = t_{(i,h)} \: .
\label{eq:ivf_mult}$$
Figures \[fig:diag\](a) and (b) intuitively clarify the number of multiplications. This shows a conceptual diagram[^4] of the number of multiplications executed in the triple loops by [*IFN*]{} and [*IVF*]{}. The number of multiplications is represented as the volume surrounded by the curves in the rectangle. The curve in the (Term axis)-(Object axis) plane, which is shared by the two algorithms, depicts a distribution of objects each of whose feature vectors contains a value of the corresponding term. The area surrounded by the curve in Fig. \[fig:diag\](a) is $\sum_{i=1}^{N} (nt)_i$, and the volume is expressed by Eq. (\[eq:ifn\_mult\]) for [*IFN*]{}. The curve in the (Term axis)-(Mean axis) plane in Fig. \[fig:diag\](b) illustrates the distribution of means, each of whose feature vectors contains a value of the corresponding term. The volume of [*IVF*]{} is expressed by Eq. (\[eq:ivf\_mult\]). Figures \[fig:triple\_mult\](a) and (b) show the numbers of multiplications executed by [*IFN*]{} and [*IVF*]{} in their triple loops. The number of multiplications by [*IVF*]{} is smaller than that by [*IFN*]{} in every $k$ range, and such differences gradually increase with $k$, i.e., where the increase of [*IVF*]{}’s curve is suppressed. This is because the average sparsity of the mean feature vectors decreases with $k$ (Fig. \[fig:mterms\](b)).
------------ ---------
(a) PubMed (b) NYT
------------ ---------
------------ ---------
(a) PubMed (b) NYT
------------ ---------
Assume that when a procedure for an entry in an array ($\bar{\bm \xi}_s$ or $\breve{\bm \xi}_s$) in the innermost loop is performed once, the numbers of instructions executed by [*IFN*]{} and [*IVF*]{} are $\alpha$ and $\beta$. Note that $\beta$ is larger than $\alpha$ by the number of instructions by which [*IVF*]{} loads the mean IDs, $c_{(s,q)}$, $q\!\in\!\{ 1,2,\cdots,(nc)_s \}$. We ignore the instructions for loading $(nc)_s$ itself due to their smaller numbers. Then the numbers of instructions are expressed by $$\left\{ \begin{array}{lll}
\alpha\cdot \sum_{i=1}^N \sum_{h=1}^{(nt)_i} k & \mbox{for}
& \mbox{\em IFN} \\
\beta\cdot \sum_{i=1}^N \sum_{h=1}^{(nt)_i} (nc)_s & \mbox{for}
& \mbox{\em IVF}
\end{array}\right. \: .
\label{eq:vol}$$ Both $\alpha$ and $\beta$ depend on the computer architecture on which the algorithms operate. The number of multiplications depends on the sparsity of the object feature vectors, and in [*IVF*]{} it furthermore depends on the sparsity of the mean feature vectors.
We obtained $\alpha\!=\!28$ and $\beta\!=\!40$ in our preliminary analysis of the assembly codes generated from the source codes of the algorithms and applied them to Eq. (\[eq:vol\]). Figures \[fig:triple\_inst\](a) and (b) show the results, which are compared to the average numbers of instructions per iteration in Figs. \[fig:inst\](a) and (b). The cross points of the two curves of [*IFN*]{} and [*IVF*]{} appeared at almost the same $k$ values in Figs. \[fig:triple\_inst\] and \[fig:inst\]. We believe that the difference in the speed performance of [*IFN*]{} and [*IVF*]{} is mainly caused by the difference of the number of instructions in the triple loop.
We provide the condition that [*IVF*]{} achieves better performance than [*IFN*]{} as follows: $$\frac{\alpha}{\beta} >
\frac{\sum_{i=1}^N\sum_{h=1}^{(nt)_i}(nc)_s}{k\cdot\sum_{i=1}^N (nt)_i}
=\frac{1}{k}\cdot
\frac{\sum_{p=1}^D\{ (nc)_p\!\times\!(no)_p\}}{\sum_{p=1}^D (no)_p}\:,
\label{eq:ifn-ivf_cond}$$ where $(nc)_p$ and $(no)_p$ denote the numbers of centroids (means) and objects that contain a term with global term ID $p$, i.e., $(nc)_p$ and $(no)_p$ are the centroid and document frequencies of the $p$-th term. Thus [*IVF*]{} operates faster than [*IFN*]{} when value ($\alpha/\beta$) that is determined by a computer architecture is larger than a right-hand side value in Eq. (\[eq:ifn-ivf\_cond\]) that is determined by given data objects and generated $k$ means.
IVFD and IVF {#subsec:ivfd}
------------
Assume that both the object and mean feature vectors are represented with a sparse expression. This presents a problem: which feature vectors should be inverted to achieve high performance?
### Inverted-File for Data Object Feature Vectors
To address the foregoing problem, we designed a Lloyd-type algorithm [*IVFD*]{} that applies the inverted-file data structure to the data object feature vectors described in Section \[subsec:ivfstr\]. This approach is the same as that of wand-k-means [@broder], although it employs a heuristic search instead of a linear-scan search for determining the objects’ assignments to clusters. To focus on only the basic data structure, [*IVFD*]{} adopts a linear-scan search to find the most similar centroid (mean) when each mean feature vector is given as a query. To reduce the computational cost for updating the mean feature vectors, [*IVFD*]{} utilizes not only the inverted-file data structure but also the standard data structure for the object feature vectors at the expense of consuming double memory capacitance[^5].
, ($k$)\
\[algo:ivfd-start\]\
[**Algorithm \[algo:ivfd\]**]{} shows the [*IVFD*]{} pseudocode at the $r$-th iteration. [*IVFD*]{} receives a set of the mean feature vectors represented by a standard data structure with sparse expression $\hat{\cal M}^{[r-1]}$ and uses two invariant object sets of the feature vectors with inverted-file sparse expression $\breve{\cal X}$ and standard sparse expression $\hat{\cal X}$ and returns cluster set ${\cal C}^{[r]}$ consisting of $k$ clusters and $\hat{\cal M}^{[r]}$. At the assignment step, similarities $\rho_i$, $i\!=\!1,\cdots,N$, between mean feature vector $\hat{\bm{\mu}}_j\!=\!(\tau_{(j,h)},v'_{(j,h)})$ and every object feature vectors are calculated and stored, using inverted-file $\breve{\cal X}$ for the object features. Note that $\tau_{(j,h)}$ and $v'_{(j,h)}$ denote the global term ID accessed by the tuple of mean ID $j$ and local counter $h$ and the corresponding feature value. The inverted-file $\breve{\cal X}$ consists of $D$ arrays $\breve{\bm{\zeta}}_s$ with $(no)_s$ entries, where $s\!=\!\tau_{(j,h)}$. The $q$-th entry in $\breve{\bm{\zeta}}_s$ is tuple $(o_{(s,q)},u_{(s,q)})$, where $q\!=\!1,2,\cdots,(no)_s$ and $o_{(s,q)}$ and $u_{(s,q)}$ denote the object ID $i$ ($o_{(s,q)}\!=\!i$) and the corresponding feature value. At the update step, mean feature vector with standard sparse expression $\hat{\bm{\mu}}_j^{[r]}$ is calculated using object feature vectors with standard sparse expression $\hat{\bm{x}}_i\!\in\!C_j^{[r]}$. This algorithm differs from that in [**Algorithm \[algo:ours\]**]{}. However, the main difference between them is only the order of the triple loop and the data structures for the object feature vectors and the mean feature vectors[^6].
### Performance Comparison
[*IVFD*]{} and [*IVF*]{} were applied to PubMed for evaluating their performance. Figures \[fig:ivfd\_perform\](a) and (b) show the performance-comparison results in terms of the maximum memory capacitance required by the algorithms through iterations until the convergence and the average CPU time per iteration. The horizontal lines labeled 0.707 and 1.414 in Fig. \[fig:ivfd\_perform\](a) denote the memory capacitances occupied by the object feature vectors and the double capacitance. [*IVFD*]{} used double capacitance for the object feature vectors as designed. Regarding speed performance, [*IVFD*]{} needed more CPU time than [*IVF*]{} in all the $k$ ranges. The maximum and minimum rates of the [*IVFD*]{}’s CPU time to the [*IVF*]{}’s were 1.82 at $k\!=\!1,000$ and 1.55 at $k\!=\!20,000$. Although [*IVF*]{} employed the inverted-file data structure for the [*variable*]{} mean feature vectors at the update step, it operated faster than [*IVFD*]{}. This is because constructing the inverted-file mean feature vectors is not costly. Importantly, most CPU time is spent at not the update step but the assignment step. In particular, both the algorithms spent at least 92% of their CPU time for the triple loop in their assignment steps in all the $k$ ranges.
-------------------------- --------------
(a) Required memory size (b) CPU time
-------------------------- --------------
------------------------- ----------------------------
(a) Avg. \#instructions (b) Avg. \#multiplications
------------------------- ----------------------------
Figure \[fig:ivfd\_inst\](a) shows the average number of instructions executed in the triple loop at the assignment step per iteration, and Fig. \[fig:ivfd\_inst\](b) shows the average number of multiplications operated in the triple loop per iteration. The numbers of instructions executed by [*IVFD*]{} and [*IVF*]{} were almost equal. The absolute difference of the numbers of instructions of the algorithms at each $k$ value was at most 0.7%. Both used identical number of multiplications at each iteration. This can be confirmed by comparing [**Algorithm \[algo:ivfd\]**]{} with [**Algorithm \[algo:ours\]**]{}. Both performed the multiplications illustrated as the volume in Fig. \[fig:diag\](b).
----------------------- -----------------------------
(a) CPI (b) \#L1-cache misses
(c) \#LL-cache misses (d) \#branch mispredictions
----------------------- -----------------------------
### Analysis Based on CPI Model
To identify why [*IVFD*]{} needed more CPU time despite executing almost the same number of instructions as [*IVF*]{}, we analyzed [*IVFD*]{} from the viewpoint of performance degradation factors (DFs) and compared it with [*IVF*]{}. Figures \[fig:ivfd\_dfs\](a), (b), (c), and (d) show CPI, the number of L1-data cache misses excluding the LL-cache misses (LLCM$'$) per instruction, the number of LL-cache misses (LLCM) per instruction, and the number of branch mispredictions (BM) per instruction, respectively.
The difference in the CPIs in Fig. \[fig:ivfd\_dfs\](a) corresponds to the CPU time in Fig. \[fig:ivfd\_perform\](b) since the numbers of instructions executed by both algorithms were almost identical. Actually, the [*IVF*]{}’s CPI ranged from 0.27 to 0.31, and [*IVFD*]{}’s ranged from 0.47 to 0.48. The rates of the [*IVFD*]{}’s CPIs to the [*IVF*]{}’s at $k\!=\!1,000,~20,000$ were 1.82 and 1.54, nearly equal to the CPU time rates. In Fig. \[fig:ivfd\_dfs\](d), [*IVF*]{} had more branch mispredictions than [*IVFD*]{}. However, the number was too small, compared with those of the other DFs; its contribution to the CPU time can be ignored, as shown in Fig. \[fig:contrib\](b). The difference in the CPU times (the clock cycles) came from the number of cache misses in Figs. \[fig:ivfd\_dfs\](b) and (c). [*IVFD*]{}’s L1CM$'$ and LLCM per instruction were constant high values. By contrast, [*IVF*]{}’s L1CM$'$ and LLCM per instruction increased with $k$. These characteristics of the LLCMs are explained based on our cache-miss models in Section \[subsubsec:LLCM\_model\].
[1.1]{}
------------ ------- --------------- --------------- ------- ------- ------
[Avg. err.]{} [Max. err.]{}
$w_0$ $w_1$ $w_2$ $w_3$ (%) (%)
[*IVFD*]{} 0.243 [**8.94**]{} [**16.8**]{} 23.8 0.445 1.52
[*IVF*]{} 0.243 [**3.13**]{} [**13.5**]{} 23.8 0.461 3.19
------------ ------- --------------- --------------- ------- ------- ------
: Optimized CPI model parameters and errors on actual CPIs
\[table:ivfd\_params\]
\[c\]\[c\]\[1.0\]
(0,0) (0,0)[(0,-6)\[c\][Number of clusters: $k$]{} ]{}
\[c\]\[c\]\[1.1\]
(0,0) (0,0)[(0,20)\[c\][CPI]{} ]{}
\[c\]\[r\]\[0.9\][$10^2$]{} \[c\]\[c\]\[0.9\][$10^3$]{} \[c\]\[c\]\[0.9\][$10^4$]{} \[r\]\[r\]\[0.9\][0]{} \[r\]\[r\]\[0.9\][0.1]{} \[r\]\[r\]\[0.9\][0.2]{} \[r\]\[r\]\[0.9\][0.3]{} \[r\]\[r\]\[0.9\][0.4]{} \[r\]\[r\]\[0.9\][0.5]{} \[r\]\[r\]\[0.75\][*IVF*]{} \[r\]\[r\]\[0.75\][*IVFD*]{} \[r\]\[r\]\[0.75\][Model]{} ![Actual and model CPI of [*IVFD*]{} and [*IVF*]{} along $k$ in PubMed[]{data-label="fig:ivfd_fit"}](./fit_ivfd_pubmed1M.eps "fig:"){width="55mm"}
We optimized the [*IVFD*]{} parameters by referring to the procedure in Section \[subsec:cpimodel\] to determine the contribution rates of the DFs to the CPU time. We assumed for the optimization that parameters $w_0$ and $w_3$ of [*IVFD*]{} in Eq. (\[eq:model\]) were fixed at the same values as those of [*IVF*]{}. The optimized parameters and results are shown in Table \[table:ivfd\_params\] and Fig. \[fig:ivfd\_fit\]. The optimized CPI model agrees well with the actual CPIs since the average error and the maximum error in [*IVFD*]{} were 0.445% and 1.52%. Parameters $w_1$ and $w_2$ were larger than those of [*IVF*]{}; the stall clock cycles per cache miss were longer. Thus [*IVFD*]{} had more cache misses, each of which induced longer stall clock cycles.
---------------- ---------------
(a) [*IVFD*]{} (b) [*IVF*]{}
---------------- ---------------
Figures \[fig:contrib\](a) and (b) show the contribution rates of each DF to the CPU times in [*IVFD*]{} and [*IVF*]{}. The rates of L1CM$'$ and LLCM were high in [*IVFD*]{}, and the rate of Inst occupied much of the whole of contribution rate in [*IVF*]{}. In terms of branch misprediction (BM), its contribution rates in [*IVFD*]{} and [*IVF*]{} were very small, and we can ignore its values. Since the number of instructions and parameter $w_0$ were equal in [*IVFD*]{} and [*IVF*]{}, the [*IVFD*]{}’s performance degradation was caused by cache misses, more of which were caused by the long arrays $\bm{\zeta}_s$ of object inverted-file $\breve{\cal X}$ in the innermost loop in the triple loop in [**Algorithm \[algo:ivfd\]**]{}.
### LL-Cache-Miss Models for IVFD and IVF {#subsubsec:LLCM_model}
Figure \[fig:ivfd\_dfs\](c) shows that the number of last-level cache misses (LLCM) of [*IVF*]{} increased with $k$ while that of [*IVFD*]{} was almost constant in the $k$ range. We analyzed these characteristics.
The last-level (LL) cache used in our experiments contained 36,700,160 (35 M) bytes in 64-byte blocks with 20-way set associative placement and least-recently used (LRU) replacement. Instead of the actual set associative replacement, we assumed fully associative one in our analysis. Both [*IVFD*]{} and [*IVF*]{} used an inverted-file data structure that consisted of two arrays for 4-byte IDs of objects or centroids and 8-byte feature values.
[*IVF*]{} calculates similarities (inner products) between an object and all centroids (means) in the middle and innermost loop in the triple loop at its assignment step. A probability that a term with global term ID $p$ is used for a similarity calculation is $(no)_p/N$, where $(no)_p$ denotes the number of objects that contain the $p$-th term, i.e., the document frequency of the term. When the array related to the $p$-th term, $\breve{\bm{\xi}}_p\!=\!(c_{(p,q)},u_{(p,q)})_{q=1}^{(nc)_p}\!\in\!\breve{\cal M}$, is accessed, the number of blocks ([*NB*]{}$_{[\mbox{\tiny\em IVF}]}$) that are placed into the LL cache from the main memory is given by $$\mbox{\em NB}_{[\mbox{\tiny\em IVF}]} =
\left\lceil (nc)_p \times
\frac{\mbox{sizeof(int + double)}}{\mbox{block size}}
\right\rceil \: ,
\label{eq:nb_ivf}$$ where $(nc)_p$ denotes the number of centroids that contain the $p$-th term and depends on $k$. Then the expected number of blocks that are placed into the LL cache is expressed by $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}]=
\sum_{p=1} ^D \frac{(no)_p}{N}\cdot
\left\lceil (nc)_p\cdot \gamma \right\rceil\: ,
\label{eq:E_nb_ivf}$$ where $\gamma$ denotes (sizeof(int + double))/(block size). By contrast, when [*IVFD*]{} calculates similarities between a centroid and all objects, the expected number of blocks ([*NB*]{}$_{[\mbox{\tiny\em IVFD}]}$) is expressed by $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}]=
\sum_{p=1} ^D \frac{(nc)_p}{k}\cdot
\left\lceil (no)_p\cdot \gamma \right\rceil\: .
\label{eq:E_nb_ivfd}$$ Assume that $(nc)_p\cdot \gamma$ and $(no)_p\cdot \gamma$ are integers. Then Eqs. (\[eq:E\_nb\_ivf\]) and (\[eq:E\_nb\_ivfd\]) are simplified: $$\begin{aligned}
\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}] &=&
\frac{1}{N}\left(\gamma \textstyle{\sum_{p=1}^D (no)_p (nc)_p}\right) \\
\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}] &=&
\frac{1}{k}\left(\gamma \textstyle{\sum_{p=1}^D (no)_p (nc)_p}\right) \: ,\end{aligned}$$ where $\sum_{p=1}^D (no)_p (nc)_p$ is the number of multiplications that is illustrated as the volume in Fig. \[fig:diag\](b). It is clear that $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}] \ll
\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}]~~~~(N\gg k)\: .
\label{eq:comp_nb0}$$
We compare $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}]$ and $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}]$ with the number of blocks in the actual LL cache when [*IVF*]{} and [*IVFD*]{} are applied to PubMed ($N\!=\!1\!\times\! 10^6$), given $k\!=\!1\!\times\!10^4$. In this comparison, we assume that the number of available blocks ($\mbox{\em NB}_{[\mbox{\tiny\em LLC}]}$) is $5\!\times\! 10^5$ that corresponds to 32 MB. The number of multiplications executed by [*IVF*]{} was $2.21\!\times\! 10^{11}$ shown in Fig. \[fig:triple\_mult\](a) and $\gamma\!=\!(4+8)/64\!=\!3/16$. Then $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}]\!\sim\!4\!\times\! 10^4$ and $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}]\!\sim\!4\!\times\! 10^6$. The inequality in Eq. (\[eq:comp\_nb0\]) is rewritten as $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}] \ll
\mbox{\em NB}_{[\mbox{\tiny\em LLC}]} \ll
\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}]\:.
\label{eq:comp_nb}$$ This inequality held in the $k$ range from 200 to 20,000 when the algorithms were applied to PubMed.
The fact of $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVFD}]}] \!\gg\!
\mbox{\em NB}_{[\mbox{\tiny\em LLC}]}$ means that [*IVFD*]{} almost always fails to use feature values in the LL cache like [*cold-start misses*]{}. Based on this, we assume that the blocks required by [*IVFD*]{} must be always brought into the LL cache from the main memory. Then the number of LL-cache misses (LLCM$_{[\mbox{\tiny\em IVFD}]}$) is given by $$\begin{aligned}
\mbox{LLCM}_{[\mbox{\tiny\em IVFD}]}
&=& \textstyle{\sum_{p=1}^D (nc)_p\lceil (no)_p\cdot\gamma \rceil}
\label{eq:exact_llcm_ivfd} \\
&\sim& \gamma\textstyle{\sum_{p=1}^D (no)_p (nc)_p} \: .
\label{eq:approx_llcm_ivfd}\end{aligned}$$ We show the rate of LLCM$_{[\mbox{\tiny\em IVFD}]}$ in Eq. (\[eq:exact\_llcm\_ivfd\]) to the number of instructions (Inst$_{[\mbox{\tiny\em IVFD}]}$) that was obtained in our experiments as Model in Fig. \[fig:append\_llcm\]. The model curve coincided with the actual rate depicted as [*IVFD*]{} in Fig. \[fig:ivfd\_dfs\](c). Furthermore, we approximate the rate as $$\frac{\mbox{LLCM}_{[\mbox{\tiny\em IVFD}]}}{\mbox{Inst}_{[\mbox{\tiny\em IVFD}]}}\sim
\frac{\gamma\sum_{p=1}^D (no)_p (nc)_p}{\beta\times \mbox{(\#multiplications)}}
= \frac{\gamma}{\beta}\: ,
\label{eq:approx_rate_ivfd}$$ where $\beta$ is the same constant value[^7] as that for [*IVF*]{} in Eq. (\[eq:vol\]) and \#multiplications denotes $\sum_{p=1}^D (no)_p (nc)_p$. Since $\beta\!=\!40$ and $\gamma\!=\!3/16$ in our experiments, $\mbox{LLCM}_{[\mbox{\tiny\em IVFD}]}/\mbox{Inst}_{[\mbox{\tiny\em IVFD}]}\!=\!4.7\!\times\!10^{-3}$. This approximate value is not so far from the [*IVFD*]{} values in Fig. \[fig:ivfd\_dfs\](c) and higher than the corresponding values in Fig. \[fig:append\_llcm\] because $\beta\!\times\!\mbox{(\#multiplications)}$ is slightly smaller than actual Inst$_{[\mbox{\tiny\em IVFD}]}$. Thus $\mbox{LLCM}_{[\mbox{\tiny\em IVFD}]}/\mbox{Inst}_{[\mbox{\tiny\em IVFD}]}$ becomes the constant value depending on the computer architecture.
\[c\]\[c\]\[1.0\]
(0,0) (0,0)[(0,-6)\[c\][Number of clusters: $k$]{} ]{}
\[c\]\[c\]\[0.9\]
(0,0) (0,0)[(0,18)\[c\][LLCM / Inst $(\times\! 10^{-3})$]{} ]{}
\[l\]\[r\]\[0.9\][$10^2$]{} \[c\]\[c\]\[0.9\][$10^3$]{} \[c\]\[c\]\[0.9\][$10^4$]{} \[r\]\[r\]\[0.9\][0]{} \[r\]\[r\]\[0.9\][1]{} \[r\]\[r\]\[0.9\][2]{} \[r\]\[r\]\[0.9\][3]{} \[r\]\[r\]\[0.9\][4]{} \[r\]\[r\]\[0.75\][[*IVF*]{}]{} \[r\]\[r\]\[0.75\][[*IVFD*]{}]{} \[r\]\[r\]\[0.75\][[Model]{}]{} ![Actual and model (LLCM/Inst) of [*IVFD*]{} and [*IVF*]{} along $k$ in PubMed[]{data-label="fig:append_llcm"}](./comp_llcm-inst_pubmed1M.eps "fig:"){width="52mm"}
Next, we model the [*IVF*]{} LLCM (LLCM$_{[\mbox{\tiny\em IVF}]}$). When [*IVF*]{} calculates similarities among successive $z$ objects and all centroids, the expected number of blocks that are placed into the LL cache ($\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z)}]$) is given by $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z)}]
=\sum_{p=1}^D\left\{ 1 -\left(1-\frac{(no)_p}{N}\right)^z \right\}
\lceil (nc)_p \gamma \rceil \:.
\label{eq:ivf_nb}$$ Note that $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(1)}]\!=\!
\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}]$ in Eq. (\[eq:E\_nb\_ivf\]). Let $z^{*}$ denote the maximum integer $z$ under the condition that $\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z)}]$ satisfies $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z)}] \leq
\mbox{\em NB}_{[\mbox{\tiny\em LLC}]}\: .
\label{eq:z_cond}$$ When $z\!=\!z^*$, intuitively, the LL cache is fully occupied by arrays $\breve{\bm \xi}_{t_{(i,h)}}\!\in\!\breve{\cal M}$ related to terms that successive $z^*$ objects $\hat{\bm x}_i$ contain. Consider that when the LL cache is at this state, [*IVF*]{} requires array $\breve{\bm \xi}_p$ related to the $p$-th term, which is not placed in the LL cache. Then the expected number of blocks that are placed into the LL cache ($\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z^*,\mbox{\scriptsize miss})}]$) is given by $$\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z^*,\mbox{\scriptsize miss})}]
=\sum_{p=1}^D \frac{(no)_p}{N} \left(1-\!\frac{(no)_p}{N}\right)^{z^*}
\left\lceil (nc)_p \gamma \right\rceil \:.
\label{eq:miss_nb}$$ Using this value, LLCM$_{[\mbox{\tiny\em IVF}]}$ is given and approximated as $$\begin{aligned}
\mbox{LLCM}_{[\mbox{\tiny\em IVF}]}
&=& N\cdot
\mathbb{E}[\mbox{\em NB}_{[\mbox{\tiny\em IVF}]}^{(z^*,\mbox{\scriptsize miss})}]
\label{eq:exact_llcm_ivf} \\
&\sim& \gamma \sum_{p=1}^D (no)_p (nc)_p
\left(1-\!\frac{(no)_p}{N}\right)^{z^*} \: .
\label{eq:llcm_ivf}\end{aligned}$$ The rate of LLCM$_{[\mbox{\tiny\em IVF}]}$ in Eq. (\[eq:exact\_llcm\_ivf\]) to Inst$_{[\mbox{\tiny\em IVF}]}$ in Fig. \[fig:ivfd\_dfs\](c) is shown as Model in Fig. \[fig:append\_llcm\]. The model curve gave close agreement with the values obtained by the experiments and increased with $k$. Furthermore, this rate is approximated as $$\frac{\mbox{LLCM}_{[\mbox{\tiny\em IVF}]}}{\mbox{Inst}_{[\mbox{\tiny\em IVF}]}}
\sim \left( \frac{\gamma}{\beta} \right)
\frac{\sum_{p=1}^D (no)_p (nc)_p \left(1-\!\frac{(no)_p}{N}\right)^{z^*}}
{\sum_{p=1}^D (no)_p (nc)_p} \: .
\label{eq:approx_rate_ivf}$$ When $k$ approaches asymptotically to $N$, $z^*\!\rightarrow\! 0$ and $(nc)_p\!\rightarrow\!(no)_p$ in Eq. (\[eq:approx\_rate\_ivf\]). Then $$\lim_{k\rightarrow N}
\frac{\mbox{LLCM}_{[\mbox{\tiny\em IVF}]}}{\mbox{Inst}_{[\mbox{\tiny\em IVF}]}}
\sim \frac{\gamma}{\beta}
\sim \frac{\mbox{LLCM}_{[\mbox{\tiny\em IVFD}]}}{\mbox{Inst}_{[\mbox{\tiny\em IVFD}]}} \: .
\label{eq:lim}$$ $\mbox{LLCM}_{[\mbox{\tiny\em IVF}]}/\mbox{Inst}_{[\mbox{\tiny\em IVF}]}$ increases with $k$ and approached to $(\gamma/\beta)$ that is the approximate rate of [*IVFD*]{}.
Due to the reasons mentioned above, applying an inverted-file data structure to the mean feature vectors leads to better performance. We should use [*IVF*]{} rather than [*IVFD*]{} to achieve high performance for large-scale sparse data sets.
Conclusion {#sec:conc}
==========
We proposed an inverted-file $k$-means clustering algorithm ([*IVF*]{}) that operated at high speed and with low memory consumption in large-scale high-dimensional sparse document data sets when large $k$ values were given. [*IVF*]{} represents both the given object feature vectors and the mean feature vectors with sparse expression to conserve occupied memory capacitance and exploits the inverted-file data structure for the mean feature vectors to achieve high-speed performance. We analyzed [*IVF*]{} using a newly introduced clock-cycle per instruction (CPI) model to identify factors for high-speed operation in a modern computer system. Consequently, [*IVF*]{} suppressed the three performance degradation factors of the numbers of cache misses, branch mispredictions, and completed instructions.
As future work, we will evaluate [*IVF*]{} in such practical environments as with parallel and distributed modern computer systems.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was partly supported by JSPS KAKENHI Grant Number JP17K00159.
[50]{} \[1\][\#1]{} url@samestyle \[2\][\#2]{} \[2\][[l@\#1=l@\#1\#2]{}]{}
M. I. Jordan and T. M. Mitchell, “Machine learning: rends, perspectives, and prospects,” *Science*, vol. 349, no. 6245, pp. 255–260, 2015.
X. Wu, V. Kumar, J. R. Quinlan, J. Ghosh, Q. Yang, H. Motoda, G. J. McLachlan, A. Ng, B. Liu, P. S. Yu, Z.-H. Zhou, M. Steinbach, D. J. Hand, and D. Steinberg, “Top 10 algorithms in data mining,” *Knowl. Inf. Syst.*, vol. 14, no. 1, pp. 1–37, 2008.
S. P. Lloyd, “Least squares quantization in ,” *IEEE Trans. Information Theory*, vol. 28, no. 2, pp. 129–137, 1982.
J. B. MacQueen, “Some methods for classification and analysis of multivariate observations,” in *Proc. 5th Berkeley Symp. Mathematical Statistics and Probability*, 1967, pp. 281–297.
C. Elkan, “Using the triangle inequality to accelerate k-means,” in *Proc. 20th Int. Conf. Machine Learning (ICML)*, 2003, pp. 147–153.
G. Hamerly, “Making k-means even faster,” in *Proc. SIAM Int. Conf. Data Mining (SDM)*, 2010, pp. 130–140.
I. S. Dhillon and D. S. Modha, “Concept decompositions for large sparse text data using clustering,” *Machine Learning*, vol. 42, no. 1–2, pp. 143–175, 2001.
J. L. Hennessy and D. A. Patterson, Eds., *Computer architecture, sixth edition: A quantitative approach*.1em plus 0.5em minus 0.4emSan Mateo, CA: Morgan Kaufmann, 2017.
M. Evers and T.-Y. Yeh, “Understanding branches and designing branch predictors for high-performance microprocessors,” *Proc. IEEE*, vol. 89, no. 11, pp. 1610–1620, 2001.
S. Eyerman, J. E. Smith, and L. Eeckhout, “Characterizing the branch misprediction penalty,” in *Proc. Int. Symp. Perform. Anal. Syst. Softw. (ISPASS)*, 2006, pp. 48–58.
H. Samet, Ed., *Foundations of multidimensional and metric data structures*.1em plus 0.5em minus 0.4emSan Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 2006.
D. Harman, E. Fox, R. Baeza-Yates, and W. Lee, “Inverted files,” in *Information retrieval: Data structures algorithms*, W. B. Frakes and R. Baeza-Yates, Eds.1em plus 0.5em minus 0.4emNew Jersey: Prentice Hall, 1992, ch. 3, pp. 28–43.
D. E. Knuth, “Retrieval on secondary keys,” in *The art of computer programming: Volume 3: Sorting and searching*.1em plus 0.5em minus 0.4emAddison-Wesley Professinal, 1998, ch. 5.2.4 6.5.
J. Zobel and A. Moffat, “Inverted files for text search,” *ACM Computing Surveys*, vol. 38, no. 2, article 6, 2006.
S. B, C. L. A. Clarke, and G. V. Cormack, Eds., *Information retrieval: Implementing and evaluating search engines*.1em plus 0.5em minus 0.4emCambridge, Massachusetts: The MIT Press, 2010.
Perf, “Linux profiling with performance counters,” 2019. \[Online\]. Available: <https://perf.wiki.kernel.org/index.php>
D. Aloise, A. Deshpande, P. Hansen, and P. Popat, “-hardness of sum-of-squares clustering,” *Machine Learning*, vol. 75, pp. 245–248, 2009.
J. Drake and G. Hamerly, “Accelerated k-means with adaptive distance bounds,” in *Proc. 5th NIPS Workshop on Optimization for Machine Learning*, 2012.
Y. Ding, Y. Zhao, X. Shen, M. Musuvathi, and T. Mytkowicz, “Yinyang k-means: A drop-in replacement of the classic k-means with consistent speedup,” in *Proc. 32nd Int. Conf. Machine Learning (ICML)*, 2015, pp. 579–587.
J. Newling and F. Fleuret, “Fast k-means with accurate bounds,” in *Proc. 33rd Int. Conf. Machine Learning (ICML)*, 2016.
T. Hattori, K. Aoyama, K. Saito, T. Ikeda, and E. Kobayashi, “Pivot-based k-means algorithm for numerous-class data sets,” in *Proc. SIAM Int. Conf. Mata Mining (SDM)*, 2016, pp. 333–341.
K. Aoyama, K. Saito, and T. Ikeda, “Accelerating a loyd-type k-means clustering algorithm with summable lower bounds in a lower-dimensional space,” *IEICE Trans. Inf. Syst.*, vol. E101-D, no. 11, pp. 2773–2782, 2018.
J. Sivic and A. Zisserman, “Video : text retrieval approach to object matching in videos,” in *Proc. IEEE Int. Conf. Computer Vision (ICCV)*, 2003, pp. 1470–1478.
H. Jgou, M. Douze, and C. Schmid, “Product quantization for nearest neighbor search,” *IEEE Trans. Pattern Anal. Mach. Intell.*, vol. 33, no. 1, pp. 117–128, 2011.
A. Babenko and V. Lempitsky, “The inverted multi-index,” *IEEE Trans. Pattern Anal. Mach. Intell.*, vol. 37, no. 6, pp. 1247–1260, 2015.
P. Indyk and R. Motwani, “Approximate nearest neighbors: owards removing the curse of dimensionality,” in *Proc. ACM Symp. Theory of Computing (STOC)*, 1998, pp. 604–613.
A. Andoni, M. Datar, N. Immorlica, P. Indyk, and V. Mirrojni, “Locality-sensitive hashing using stable distributions,” in *Nearest-neighbor methods in learning and vision; heory and practice*, G. Shakhnarovich, T. Darrell, and P. Indyk, Eds.1em plus 0.5em minus 0.4emThe MIT Press, 2005, ch. 3, pp. 61–72.
M. S. Charikar, “Similarity estimation techniques from rounding algorithms,” in *Proc. ACM Symp. Theory of Computing (STOC)*, 2002, pp. 380–388.
A. Broder, L. Garcia-Pueyo, V. Josifovski, S. Vassilvitskii, and S. Venkatesan, “Scalable k-means by ranked retrieval,” in *Proc. ACM Int. Conf. Web Search and Data Mining (WSDM)*, 2014, pp. 233–242.
L. Jian, C. Wang, Y. Liu, S. Liang, W. Yi, and Y. Shi, “Parallel data mining techniques on graphics processing unit with compute unified device architecture (),” *J. Supercomput.*, vol. 64, pp. 942–967, 2013.
J. Bhimani, M. Leeser, and N. Mi, “Accelerating -means clustering with parallel implementations and computing,” in *Proc. IEEE High Performance Extreme Computing Conf. (HPEC)*, 2015, pp. 233–242.
K. Kaligosi and P. Sanders, “How branch mispredictions affect quicksort,” in *Algorithms-ESA2006. Lecture Notes in Computer Science*, Y. Azar and T. Erlebach, Eds.1em plus 0.5em minus 0.4emSpringer, Berlin, Heidelberg, 2006, pp. 780–791.
S. Edelkamp and A. Wei, “: voiding branch mispredictions in quicksort,” *ACM J. Exp. Algorithmics (JEA)*, vol. 24, no. 1, pp. 1.4:1–1.4:22, 2019.
O. Green, M. Dukhan, and R. Vuduc, “Branch-avoiding graph algorithms,” in *Proc. ACM Symp. Parallelism in Algorithms and Architectures (SPAA)*, 2015, pp. 212–223.
M. Kowarschik and C. Wei, “An overview of cache optimization techniques and cache-aware numerical algorithms,” in *Algorithms for memory heirarchies, Lecture Notes in Computer Science*, U. Meyer, P. Sanders, and J. Sibeyn, Eds.1em plus 0.5em minus 0.4em Berlin, Heidelberg: Springer, 2003, ch. 10, pp. 213–232.
M. Frigo, C. Leiserson, H. Prokop, and S. Ramachandran, “Cache-oblivious algorithms,” *ACM Trans. Algorithms*, vol. 8, no. 1, article 4, 2012.
A. Ghoting, G. Buehrer, S. Parthasarathy, D. Kim, A. Nguyen, Y.-K. Chen, and P. Dubey, “Cache-conscious frequent pattern mining on modern and emerging processors,” *The VLDB Journal*, vol. 16, no. 1, pp. 77–96, 2007.
M. Perdacher, C. Plant, and C. Bhm, “Cache-oblivious high-performance similarity join,” in *Proc. Int. Conf. Management of Data (SIGMOD)*, 2019, pp. 87–104.
D. Dua and E. K. Taniskidou, “Bag of words data set ([P]{}ub[M]{}ed abstracts) in [UCI]{} machine learning repository,” 2017. \[Online\]. Available: <http://archive.ics.uci.edu/ml>
P. Hammarlund, A. J. Martinez, A. A. Bajwa, D. L. Hill, E. Hallnor, H. Jiang, M. Dixon, M. Derr, M. Hunsaker, R. Kumar, R. B. Osborne, R. Rajwar, R. Singhal, R. D’Sa, R. Chappell, S. Kaushik, S. Chennupaty, S. Jourdan, S. Gunther, T. Piazza, and T. Burton, “: The fourth-generation ntel core processor,” *IEEE Micro*, vol. 34, issue 2, pp. 6–20, 2014.
, “Disclosure of hardware prefetcher control on some processors,” 2014. \[Online\]. Available: <https://software.intel.com/en-us/> [articles/disclosure-of-hw-prefetcher-control-on-some-intel-processors](articles/disclosure-of-hw-prefetcher-control-on-some-intel-processors)
[^1]: If mean feature vectors are not normalized by their $L_2$ norms, i.e., they are not points on the unit hypersphere, a solution by the spherical $k$-means algorithm does not always coincide with that by the standard $k$-means algorithm.
[^2]: The tuple was not implemented with a [*structure type*]{} consisting of an int-type and a double-type member to avoid unnecessary memory usage caused by an 8-byte memory alignment adopted by 64-bit CPUs.
[^3]: Regarding the two algorithms, the instructions executed in the triple loop were identical in the corresponding assembly codes.
[^4]: Actually the term order sorted on the number of centroids does not always meet that sorted on the number of objects. For this reason, both the numbers of centroids and objects do not decrease monotonically, as shown in Fig. \[fig:diag\](b).
[^5]: A mean-update step using object feature vectors with inverted-file data structure required much more CPU time than that with the standard data structure in our preliminary experiments.
[^6]: Exactly, [*IVFD*]{} differs from [*IVF*]{} in the positions in source codes at which the final assignment of each object to a cluster is executed. [*IVFD*]{} executes the assignment outside the triple loop; [*IVF*]{} does so inside.
[^7]: Analysis of [*IVFD*]{} and [*IVF*]{} assembly codes showed that both algorithms used the identical number of instructions for each multiplication and addition operation.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: '**The spread of perturbative signals in complex networks is governed by the combined effect of the network topology and its intrinsic nonlinear dynamics. Recently, the resulting spreading patterns have been analyzed and predicted, shown to depend on a single scaling relationship, linking a node’s weighted degree $S_i$ to its intrinsic response time $\tau_i$. The relevant scaling exponent $\theta$ can be analytically traced to the system’s nonlinear dynamics. Here we show that $\theta$ can be obtained via two different derivation tracks, leading to seemingly different functions. Analyzing the resulting predictions, we find that, despite their distinct form, they are fully consistent, predicting the exact same scaling relationship under potentially diverse types of dynamics.**'
author:
- 'Chittaranjan Hens$^{1}$, Uzi Harush$^{2}$, Simcha Haber$^{2}$, Reuven Cohen$^{2}$ & Baruch Barzel$^{2,3,*}$'
title: |
Response times of nodes in a complex network environment -\
two potential derivation tracks
---
In Ref. [@Hens2019] we seek the patterns of signal propagation in a complex network environment. We begin with an $N$ node dynamic system, whose activities $x_i(t)$ ($i = 1,\dots,N$) are driven by [@Barzel2013; @Barzel2015; @Harush2017; @Hens2019]
$$\dod{x_i}{t} = M_0(x_i) + \sum_{j = 1}^N A_{ij} M_1\big( x_i(t) \big) M_2\big( x_j(t) \big),
\label{Dynamics}$$
where ${
A_{i j}
}$ represents the $N \times N$ weighted adjacency matrix, and the nonlinear functions $M_0(x), M_1(x)$ and $M_2(x)$ capture the system’s internal nonlinear mechanisms. To track the spread of perturbative signals we seek the response time $\tau_i$ of all nodes $i = 1,\dots,N$ to a directly incoming signal. We find that these response times are characterized by a single universal exponent $\theta$ via
$$\tau_i \sim S_i^{\theta}
\label{TauSi}$$
where $S_i = \sum_{j = 1}^N {
A_{i j}
}$ is node $i$’s weighted degree. Hence $\theta$ links a well-mapped topological feature, degree $S_i$, to its consequent dynamic observable, response time $\tau_i$. Averaging over all nodes with degree around $S$, *i.e*. $S_i \in (S, S + \dif S)$, we rewrite (\[TauSi\]) as
$$\tau(S) \sim S^{\theta},
\label{TauS}$$
predicting the average response times of nodes in function of their weighted degree.
To predict $\theta$ we link it to the dynamics (\[Dynamics\]) via [@Hens2019]
$$Y \big( R^{-1}(x) \big) = \sum_{n = 0}^{\infty} C_n x^{\Gamma(n)},
\label{YRx}$$
where
$$Y(x) = \left( \dod{ \big[ M_1(x) R(x) \big]}{x} \right)^{-1},
\label{Yx}$$
$R(x) = -M_1(x)/M_0(x)$ and $R^{-1}(x)$ is its inverse function. The function (\[YRx\]) is fully determined by the system’s dynamics through $M_0(x)$ and $M_1(x)$. Its leading power $\Gamma(0)$ determines the exponent $\theta$ via
$$\theta = -2 - \Gamma(0).
\label{Theta}$$
The detailed derivation of this result appears in Ref. [@Hens2019].
Recently, an alternative derivation has been proposed, which leads to a seemingly distinct scaling. In this derivation track Eq. (\[Yx\]) is substituted by
$$Y(x) = \left( M_1(x) \dod{R(x)}{x} \right)^{-1},
\label{YxAlt}$$
in which the function $M_1(x)$ is excluded from the $x$-derivative. Consequently, the power series expansion in (\[YRx\]) is potentially different, and hence $\theta$ in (\[Theta\]) may deviate from the original prediction of Ref. [@Hens2019].
Here we show that the two derivation tracks are, in fact, one and the same, exhibiting identical leading power $\Gamma(0)$, and therefore, predicting the exact same scaling exponent $\theta$. This self consistency of the two alternative derivations, provides further verification for the validity and universal nature of Eq. (\[TauS\]). It also presents a solution to determine between the two alternative derivations, stating that non is preferable over the other - as they lead to indistinguishable results.
Comparing the two tracks
========================
To systematically compare the two derivation tracks we refer to their point of bifurcation. Hence, we follow the original derivation, leading to (\[Yx\]), where it is shown that [@Hens2019]
$$\dfrac{1}{\tau(S)} =
\dfrac{1}{\lambda^2}
\left[
c_1 R \big( R^{-1}(\lambda) \big) M_1^{\prime} \big( R^{-1}(\lambda) \big) +
c_2 M_1 \big( R^{-1}(\lambda) \big) R^{\prime} \big( R^{-1}(\lambda) \big)
\right],
\label{TwoTerms}$$
in which $\lambda \sim 1/S$ approaches zero in the limit of large $S$, and $c_1, c_2$ are arbitrary coefficients (see Supp. Eq. 1.34 in Red. [@Hens2019]). In **Track I** the next step is to arbitrarily set $c_1 = c_2 = 1$. This is motivated by the fact that these coefficients do not contribute to the scaling, and hence their specific value is of no importance. Once set to unity we can transform the r.h.s. of (\[TwoTerms\]) into the product derivative of Eq. (\[Yx\]), obtaining
$$\dfrac{1}{\tau(S)} =
\dfrac{1}{\lambda^2} \dfrac{1}{Y \big( R^{-1}(x) \big)},
\label{Yone}$$
and, consequently, the scaling in (\[Theta\]).
The alternative **Track II** sets $c_1 = 0$, preserving only the second term of (\[TwoTerms\]), arriving at the single derivative of Eq. (\[YxAlt\]), in which only $R(x)$ is differentiated. This track, therefore, predicts $\tau(S)$ via (\[Yone\]), but with $Y(x)$ taken from (\[YxAlt\]) instead of from (\[Yx\]). To analyze the relationship between the two derivation tracks, we rewrite Eq. (\[TwoTerms\]) as
$$\dfrac{1}{\tau(S)} =
\dfrac{1}{\lambda^2}
\Big(
c_1 Z_1(\lambda) + c_2 Z_2(\lambda)
\Big),
\label{Z1Z2}$$
where
$$\begin{aligned}
Z_1(\lambda) &=& R \big( R^{-1}(\lambda) \big) M_1^{\prime} \big( R^{-1}(\lambda) \big)
\label{Z1}
\\[3pt]
Z_2(\lambda) &=& M_1 \big( R^{-1}(\lambda) \big) R^{\prime} \big( R^{-1}(\lambda) \big)
\label{Z2}\end{aligned}$$
represent the two terms comprising the product derivative of (\[TwoTerms\]). Using this description, we can summarize the difference between the two proposed derivations as
- **Track I**. As appears in the original derivation [@Hens2019], we take $c_1 \ne 0$, and hence keep both terms of Eq. (\[Z1Z2\]).
- **Track II**. Setting the coefficient $c_1$ in Eq. (\[Z1Z2\]) to zero, remaining only with the second term $Z_2(\lambda)$.
Mathematically speaking, **Track II** is only justifiable if $c_1$ is truly *equal* to zero, while **Track I** is insensitive to the specific value of $c_1$, relevant even if, *e.g.*, $c_1 \approx 0$. It is for this reason that we find **Track I** more general and rigorous. However, the focus here is on the *outcome* of the two tracks, not their rigorousness.
**Is there a difference between the Tracks?** Clearly Eq. (\[Z1Z2\]) provides a different *value* for $\tau(S)$, depending on $c_1$. Yet, our focus here is not on the specific *value*, but rather on the *asymptotic scaling* $\theta$. This is determined only by the power-series expansion of $Z_1(\lambda)$ vs. that of $Z_2(\lambda)$ in the limit $\lambda \rightarrow 0$, *i.e*. $S \rightarrow \infty$. Writing
$$\begin{aligned}
\lim_{\lambda \rightarrow 0} Z_1(\lambda) &\sim& \lambda^{\alpha}
\label{Alpha}
\\[3pt]
\lim_{\lambda \rightarrow 0} Z_2(\lambda) &\sim& \lambda^{\beta},
\label{Beta}\end{aligned}$$
we distinguish between three different scenarios:
- **Case 1**. $\alpha = \beta$. In this case both terms in (\[Z1Z2\]) provide the same scaling, and hence **Track I** agrees with **Track II**.
- **Case 2**. $\alpha > \beta$. Here in the limit $\lambda \rightarrow 0$ we have $Z_1(\lambda) \ll Z_2(\lambda)$, Eq. (\[Z1Z2\]) is dominated by $Z_2(\lambda)$ independent of $c_1$, and therefore, once again, we have **Track I** and **Track II** in full agreement.
- **Case 3**. $\alpha < \beta$. This is the only scenario where $Z_1(\lambda)$ dominates Eq. (\[Z1Z2\]). Here setting $c_1 = 0$ will leave only the $Z_2(\lambda)$ term and its associated exponent $\beta$, while $c_1 \ne 0$ will predict the dominance of $\alpha$, proposing a potential discrepancy between **Tracks I** and **II**.
Fortunately, we find below that **Case 3** is prohibited, and therefore **Track I**, offered in Ref. [@Hens2019], provides the same prediction for $\theta$ as the suggested alternative **Track II**.
The scaling of $Z_1(\lambda)$ and $Z_2(\lambda)$
================================================
To calculate $\alpha$ and $\beta$ in (\[Alpha\]) and (\[Beta\]) let us first assume that $M_1(x)$ is *not* constant. Indeed, in case $M_1(x) = \rm{Const}$ the two tracks trivially coincide as (\[Yx\]) and (\[YxAlt\]) are identical. Hence we can write (to leading power)
$$M_1(x) \sim x^{\Phi},
\label{M1x}$$
with $\Phi \ne 0$. The specific case where the leading power is $\Phi = 0$, and hence the *next* leading power is required, is treated separately below. Following Eq. (\[M1x\]) we have the leading power of $M^{\prime}(x)$ as
$$M_1^{\prime}(x) \sim x^{\Phi - 1}.
\label{M1Primex}$$
Next we extract the power-series (Hahn expansion [@Hahn1995]) of $R^{-1}(x)$ around $x = 0$ as
$$R^{-1}(x) = \sum_{n = 0}^{\infty} C_n x^{\Omega(n)},
\label{Rx}$$
in which $\Omega(n)$ is a sequence of real powers, increasing with $n$. For small $x$, $R^{-1}(x)$ is dominated by the leading, *i.e. smallest*, powers $\Omega(0), \Omega(1), \dots$. To obtain the powers $\alpha$ and $\beta$ in (\[Alpha\]) and (\[Beta\]), we must express the near-zero behavior of each of the terms in (\[Z1\]) and (\[Z2\]): $M_1(R^{-1}(x)), M_1^{\prime}(R^{-1}(x)), R(R^{-1}(x))$ and $R^{\prime}(R^{-1}(x))$. This depends on the exponents $\Phi$ in (\[M1x\]) and $\Omega(n)$ in (\[Rx\]). As we show below, we must first consider the value of $\Omega(0)$, specifically, whether it is positive, negative or zero (Fig. 1).
[ **1. The case where $\Omega(0) > 0$** (Fig. 1a)]{}. If the leading power in (\[Rx\]) is positive, we have in the limit $x \rightarrow 0$, $R^{-1}(x) \sim x^{\Omega(0)}$ (blue), providing $R(x) \sim x^{1/\Omega(0)}$ (red), and consequently $R^{\prime}(x) \sim x^{-1 + 1/\Omega(0)}$. We can therefore calculate each of the terms comprising $Z_1(\lambda)$ and $Z_2(\lambda)$ in (\[Z1\]) and (\[Z2\]) as
$$\begin{aligned}
M_1 \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{\Phi} = \lambda^{\Phi \Omega(0)}
\label{Terms11}
\\[3pt]
M_1^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{\Phi - 1} = \lambda^{\Phi \Omega(0) - \Omega(0)}
\label{Terms12}
\\[3pt]
R^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{-1 + 1/\Omega(0)} = \lambda^{1 - \Omega(0)}.
\label{Terms13}\end{aligned}$$
Collecting all the terms, and using the fact that $R(R^{-1}(\lambda)) = \lambda$ we obtain
$$\begin{aligned}
Z_1(\lambda) &\sim& \lambda^{1 + \Phi \Omega(0) - \Omega(0)}
\label{Z1Case1}
\\[3pt]
Z_2(\lambda) &\sim& \lambda^{1 + \Phi \Omega(0) - \Omega(0)},
\label{Z2Case1}\end{aligned}$$
and hence we observe $\alpha = \beta$ in (\[Alpha\]) and (\[Beta\]). This corresponds to **Case 1** above, in which it is guaranteed that **Tracks I** and **II** yield identical $\theta$.
**Example**. As an example we consider the dynamic model
$$\dod{x_i}{t} = -Bx_i + \sum_{j = 1}^N {
A_{i j}
} x_i^{\frac{3}{2}} f(x_j),
\label{Model1}$$
for which we have $M_1(x) \sim x^{\frac{3}{2}}$, *i.e*. $\Phi = 3/2$ in (\[M1x\]). For this model we have $R(x) \sim x^{1/2}$, providing $\Omega(0) = 1/2 > 0$, and correspondingly $R^{\prime}(x) \sim x^{-1/2}$ and $R^{-1}(x) \sim x^2$. Using **Track I**’s Eq. (\[Yx\]) we write
$$Y\Big( R^{-1}(\lambda) \Big) = \left. \left( \dod{M_1 R}{x} \right)^{-1} \right|_{x = R^{-1}(\lambda)} =
\left. \dfrac{1}{2x} \right|_{x = \lambda^2} \sim \lambda^{-2},
\label{Y1}$$
providing $\Gamma(0) = -2$, and hence, through Eq. (\[Theta\]), predicting $\theta_{\rm{I}} = -2 - \Gamma(0) = 0$.
On the other hand, following **Track II** we use Eq. (\[YxAlt\]) to write
$$Y\Big( R^{-1}(\lambda) \Big) =
\left. \left( M_1(x) \dod{R}{x} \right)^{-1} \right|_{x = R^{-1}(\lambda)} =
\Big( M_1 \big( \lambda^2 \big) R^{\prime} \big( \lambda^2 \big) \Big)^{-1} = \lambda^{-2},
\label{Tau1I}$$
the same leading power, predicting again that $\theta_{\rm{II}} = 0$. Therefore, as predicted, we observe that
$$\theta_{\rm{I}} = \theta_{\rm{II}}$$
namely that both **Tracks I** and **II** are in perfect agreement.
.When the leading power in (\[Rx\]) is negative, we have $R^{-1}(x) \sim x^{\Omega(0)} \rightarrow \infty$ for $x \rightarrow 0$ (blue). This corresponds to $R(x) \sim x^{1/\Omega(0)}$ in the limit $x \rightarrow \infty$ (red). Hence, in this case, the inverse $R^{-1}(x)$ helps characterize the asymptotic behavior of $R(x)$ in the $x \rightarrow \infty$ limit, as opposed to the $x \rightarrow 0$ limit, obtained in the $\Omega(0)$ positive scenario. We therefore obtain the large $x$ limit of the derivative $R^{\prime}(x)$ as $R^{\prime}(x) \sim x^{1/\Omega(0) - 1}$, helping us, once again, construct all the terms of (\[Z1\]) and (\[Z2\]) as
$$\begin{aligned}
M_1 \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{\Phi} = \lambda^{\Phi \Omega(0)}
\label{Terms21}
\\[3pt]
M_1^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{\Phi - 1} = \lambda^{\Phi \Omega(0) - \Omega(0)}
\label{Terms22}
\\[3pt]
R^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{-1 + 1/\Omega(0)} = \lambda^{1 - \Omega(0)}.
\label{Terms23}\end{aligned}$$
The result is identical to (\[Terms11\]) - (\[Terms13\]), with the only difference that now $\Omega(0)$ is negative, and the argument $R^{-1}(\lambda)$ appearing in all the three functions approaches infinity and not zero. Hence (\[Terms21\]) - (\[Terms23\]) examine the behavior of $M(x)$ and $R(x)$ at the $x \rightarrow \infty$ limit. This is consistent with the fact that under $\Omega(0) < 0$ the steady state activities $x(S)$ scale positively with $S$ [@Barzel2013; @Barzel2015], and therefore the limit $S \rightarrow \infty$, relevant for our asymptotic scaling analysis, is associate with the *large* $x$ nodes. In any case, what matters in the present context is that once collecting all the terms, we, again, arrive at
$$\alpha = \beta = 1 + \Phi \Omega(0) - \Omega(0),
\label{AlphaBeta}$$
satisfying **Case 1**, in which **Tracks I** and **II** remain consistent.
**Example**. Here we consider
$$\dod{x_i}{t} = -Bx_i^2 + \sum_{j = 1}^N {
A_{i j}
} x_i^{\frac{3}{2}} f(x_j),
\label{Model2}$$
for which we have $\Phi = 3/2$ in (\[M1x\]), and $R(x) \sim x^{-1/2}$. We now have $R^{\prime}(x) \sim x^{-3/2}$ and $R^{-1}(x) \sim x^{-2}$, *i.e*. $\Omega(0) = -2 < 0$. Via **Track I** we write
$$Y\Big( R^{-1}(\lambda) \Big) = \left. \left( \dod{M_1 R}{x} \right)^{-1} \right|_{x = R^{-1}(\lambda)} =
1 \sim \lambda^0,
\label{Y2}$$
obtaining $\Gamma(0) = 0$, which using (\[Theta\]) provides $\theta_{\rm{I}} = -2 - \Gamma(0) = -2$. Under **Track II** we have
$$Y\Big( R^{-1}(\lambda) \Big) =
\left. \left( M_1(x) \dod{R}{x} \right)^{-1} \right|_{x = R^{-1}(\lambda)} =
M_1 \big( \lambda^{-2} \big) R^{\prime} \big( \lambda^{-2} \big) = \lambda^0,
\label{Tau2I}$$
again providing $\theta_{\rm{II}} = -2 - \Gamma(0) = -2$. Hence, once again, we observe that
$$\theta_{\rm{I}} = \theta_{\rm{II}},$$
with both **Tracks**, **I** and **II**, in full agreement.
.The third case we consider is when the leading power in in the expansion of $R^{-1}(x)$ vanishes. This represents the scenario where Eq. (\[Rx\]) takes the form
$$R^{-1}(x) \sim C_0 + C_1 x^{\Omega(1)} + \dots \sim 1 + x^{\Omega(1)},
\label{Rx0}$$
where in the last step we omitted the coefficients $C_0, C_1$, for simplicity. The second leading power is, by definition, greater than $\Omega(0)$, namely we have $\Omega(1) > 0$. Here, $R^{-1}(x)$ approaches a constant at $x \rightarrow 0$ (blue), expressed in the inverted $R(x)$ by a shifted polynomial of the form $R(x) \sim (x - 1)^{1/\Omega(1)}$ (red). Consequently $R^{\prime}(x) \sim (x - 1)^{1/\Omega(1) - 1}$. The terms to construct $Z_1(\lambda)$ and $Z_2(\lambda)$ become
$$\begin{aligned}
M_1 \big( R^{-1} (\lambda) \big) &\sim& \Big( 1 + \lambda^{\Omega(1)} \Big)^{\Phi} \sim 1 + \Phi \lambda^{\Omega(1)} + \dots \sim \lambda^0
\label{Terms31}
\\[3pt]
M_1^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( 1 + \lambda^{\Omega(1)} \Big)^{\Phi - 1} \sim 1 + (\Phi - 1) \lambda^{\Omega(1)} + \dots \sim \lambda^0
\label{Terms32}
\\[3pt]
R^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( 1 + \lambda^{\Omega(1)} - 1 \Big)^{1/\Omega(1) - 1} \sim \lambda^{1 - \Omega(1)},
\label{Terms33}\end{aligned}$$
and therefore
$$\begin{aligned}
Z_1(\lambda) &\sim& \lambda
\label{Z1Case3}
\\[3pt]
Z_2(\lambda) &\sim& \lambda^{1 - \Omega(1)}.
\label{Z2Case3}\end{aligned}$$
Using this in (\[Alpha\]) and (\[Beta\]) we obtain $\alpha = 1$ and $\beta = 1 - \Omega(1) < 1$ (since $\Omega(1) > 0$). Consequently, we have $\alpha > \beta$, satisfying **Case 2**, which once again predicts an agreement between **Track I** and **Track II**.
**Example**. To examine this class, the natural choice is to set $M_1(x) \sim \rm{Const}$, *i.e*. $\Phi = 0$, however, in this case **Track I** trivially converges to **Track II**, therefore, we *challenge* ourselves by constructing a non-trivial model
$$\dod{x_i}{t} = B - x_i^2 + \sum_{j = 1}^N {
A_{i j}
} (1 - x_i) f(x_j),
\label{Model3}$$
in which $M_0(x) = B - x^2$ and $M_1(x) = 1 - x$. Here
$$\begin{aligned}
R(x) &=& - \dfrac{1 - x}{B - x^2}
\label{Model3R}
\\[3pt]
R^{\prime}(x) &=& \dfrac{x^2 - 2x + B}{x^4 - 2Bx^2 + B^2}
\label{Model3Rprime}\end{aligned}$$
and the inverse function is
$$R^{-1}(x) = \dfrac{1}{2x}
\left(
-1 + \sqrt{1 + 4(Bx^2 + x)}
\right).
\label{Model3Rinverse}$$
In the limit $\lambda \rightarrow 0$ we can approximate (\[Model3Rinverse\]) by
$$R^{-1}(\lambda) \sim 1 + B \lambda + O(\lambda^2),
\label{Model3RinverseApprox}$$
allowing us to evaluate the relevant approximations for $M_1(R^{-1}(\lambda))$ and $R^{\prime}(R^{-1}(\lambda))$, as required to calculate $\theta$ via both tracks. This time, we first begin with **Track II**, writing
$$Y\big( R^{-1}(\lambda) \big) =
\left. \left( M_1(x) \dod{R}{x} \right)^{-1} \right|_{x = R^{-1}(\lambda)} =
\left. \left(
(1 - x)
\left(
\dfrac{x^2 - 2x + B}{x^4 - 2Bx^2 + B^2}
\right) \right)^{-1} \right|_{x = 1 + B\lambda},
\label{Tau3I}$$
where we have taken (\[Model3Rprime\]) and (\[Model3RinverseApprox\]) to express $R^{\prime}(x)$ and $R^{-1}(\lambda)$. Taking the limit where $\lambda \rightarrow 0$, the rational function on the r.h.s. of (\[Tau3I\]) approaches a constant, leaving us with $Y(R^{-1}(\lambda)) \sim (B\lambda)^{-1}$, namely $\Gamma(0) = -1$, and hence
$$\theta_{\rm II} = -2 - \Gamma(0) = -1.
\label{Tau3IFinal}$$
Next we use **Track I**, writing
$$Y(x) = \left( \dod{M_1R}{x} \right)^{-1} =
\left[
\dod{}{x} \left( (1 - x) \left( - \dfrac{1 - x}{B - x^2} \right) \right)
\right]^{-1} = \dfrac{x^4 - 2Bx^2 + B^2}{2x^2 - 2(B + 1)x + 2B}.
\label{Y3}$$
Taking (\[Model3RinverseApprox\]) to approximate $R^{-1}(\lambda)$ we obtain
$$Y\Big( R^{-1}(\lambda) \Big) \approx \dfrac{1 - 2B + B^2}{(2B - 2B^2)\lambda} \sim \lambda^{-1},
\label{Y3Lambda}$$
where we have only kept leading order terms in the limit $\lambda \rightarrow 0$. Equation (\[Y3Lambda\]) provides us with $\Gamma(0) = -1$, leading to $\theta_{\rm{I}} = -2 - \Gamma(0) = -1$. Once again, even in these rather complex dynamics, we observe that
$$\theta_{\rm{I}} = \theta_{\rm{II}},$$
showing the complete agreement between the two derivation tracks.
.To complete our analysis there is one remaining scenario that has not yet been treated, captured by
$$M_1(x) \sim C_0 + C_1 x^{\Phi(1)} + \dots \sim 1 + x^{\Phi(1)},
\label{M1x0}$$
namely that the leading power in the expansion of $M_1(x)$ vanishes, and therefore its *next* power $\Phi(1) > 0$ continues to play a role in determining $\theta$. The derivative is now
$$M_1^{\prime}(x) \sim x^{\Phi(1) - 1}.
\label{M1Primex0}$$
Taking $R^{-1}(\lambda) \sim \lambda^{\Omega(0)}$, we recalculate the terms in (\[Terms11\]) - (\[Terms13\]) as
$$\begin{aligned}
M_1 \big( R^{-1} (\lambda) \big) &\sim& M_1 \Big( \lambda^{\Omega(0)} \Big) \sim
1 + \lambda^{\Phi(1) \Omega(0)}
\label{Terms41}
\\[3pt]
M_1^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{\Phi(1) - 1} =
\lambda^{\Phi(1) \Omega(0) - \Omega(0)}
\label{Terms42}
\\[3pt]
R^{\prime} \big( R^{-1} (\lambda) \big) &\sim& \Big( \lambda^{\Omega(0)} \Big)^{-1 + 1/\Omega(0)} =
\lambda^{1 - \Omega(0)},
\label{Terms43}\end{aligned}$$
valid under both negative or positive $\Omega(0)$. The case $\Omega(0) = 0$ remains the same as in (\[Terms31\]) - (\[Terms33\]), and therefore does not require specific treatment. Note that the only term that has changed due to $M_1(x)$’s vanishing leading power is the expression in Eq. (\[Terms41\]), whose scaling is different compared with Eqs. (\[Terms11\]) and (\[Terms21\]) that were obtained under $M_1(x) \sim x^{\Phi}$. In the relevant limit, *i.e*. $\lambda \rightarrow 0$, Eq. (\[Terms41\]) is dominated either by the unity term ($\lambda^0$) or by the $\lambda^{\Phi(1) \Omega(0)}$ term, depending on which of the two powers is *smaller*. This allows us to express Eq. (\[Terms41\]) asymptotically as
$$M_1 \big( R^{-1} (\lambda) \big) \sim \lambda^{\min(0,\Phi(1) \Omega(0))},
\label{Terms41Min}$$
keeping only the term raised to the minimal power, $0$ or $\Phi(1) \Omega(0)$. We can now collect all the terms, as we did in all previous cases, to construct $Z_1(\lambda)$ and $Z_2(\lambda)$ in (\[Z1\]) and (\[Z2\]), obtaining
$$\begin{aligned}
Z_1(\lambda) &\sim& \lambda^{1 + \Phi(1) \Omega(0) - \Omega(0)}
\label{Z1Case4}
\\[3pt]
Z_2(\lambda) &\sim& \lambda^{1 + \min(0,\Phi(1) \Omega(0)) - \Omega(0)},
\label{Z2Case4}\end{aligned}$$
for which the powers are
$$\begin{aligned}
\alpha &=& 1 + \Phi(1) \Omega(0) - \Omega(0)
\label{AlphaCase4}
\\[3pt]
\beta &=& 1 + \min \big( 0,\Phi(1) \Omega(0) \big) - \Omega(0).
\label{BetaCase4}\end{aligned}$$
Of course, it is guaranteed, by definition, that
$$\min \big( 0, \Phi(1) \Omega(0) \big) \le \Phi(1) \Omega(0)
\label{Minimum}$$
and therefore that $\alpha \ge \beta$. This represents either **Case 1** or **Case 2**, but excludes **Case 3**, thus predicting, once again, the agreement of **Track I** and **Track II**.
**Example**. A classic example for a dynamics with $M_1(x)$ following the form of Eq. (\[M1x0\]) is the SIS model [@Hufnagel2004], for which $M_1(x) = 1 - x$. This dynamics has already been analyzed in Ref. [@Hens2019], and one can readily confirm that both **Tracks** agree that it exhibits $\theta_{\rm I} = \theta_{\rm II} = -1$. Therefore, to expand the range of our validation we consider a *new* type of dynamics of the form
$$\dod{x_i}{t} = -Bx_i + \sum_{j = 1}^N {
A_{i j}
} \Big(1 - x_i^{\frac{1}{2}} \Big) f(x_j),
\label{Model4}$$
capturing a potential generalization of the classic SIS model. Here we have $M_1(x) \sim 1 - x^{1/2}$, $R(x) = (1 - x^{1/2})/Bx$, and therefore
$$R^{-1}(x) = \dfrac{1}{2 B^2 x^2}
\left(
1 + 2Bx - \sqrt{1 + 4Bx}
\right).
\label{Rinv4}$$
Taking the relevant limit we obtain
$$\lim_{\lambda \rightarrow 0} R^{-1}(\lambda) = 1 - 2 B \lambda + O(\lambda^2),$$
an expansion of the form (\[Rx\]) with $\Omega(0) = 0$ and $\Omega(1) = 1$. Under **Track I** we have
$$Y(x) = \left( \dod{M_1 R}{x} \right)^{-1} =
\left( \dod{}{x} \left( \dfrac{\Big(1 - x^{\frac{1}{2}} \Big)^2}{Bx} \right) \right)^{-1} =
\dfrac{B x^2}{x^{\frac{1}{2}} - 1}.
\label{Y4}$$
Using $R^{-1}(\lambda) \sim 1 - 2B\lambda$ we write
$$Y \big( R^{-1}(\lambda) \big) \sim
\dfrac{B (1 - 2B\lambda)^2}{\sqrt{1 - 2B\lambda} - 1} \sim \lambda^{-1},
\label{YR4}$$
providing $\Gamma(0) = -1$, and, as a result $\theta_{\rm I} = -2 - \Gamma(0) = -1$.
For **Track II** we write
$$\begin{aligned}
Y \big( R^{-1}(\lambda) \big) &=&
\left. \left( M_1(x) \dod{R}{x} \right)^{-1} \right|_{x = R^{-1}(\lambda)} =
\left. \left(
\Big(1 - x^{\frac{1}{2}} \Big)
\left(
\dfrac{x^{\frac{1}{2}} - 2}{2Bx^2}
\right) \right)^{-1} \right|_{x = 1 - 2 B \lambda}
\nonumber \\[3pt]
&=&
\left[\left(
1 - \sqrt{1 - 2B\lambda}
\right)
\dfrac{\sqrt{1 - 2B\lambda} - 2}{2B(1 - 2B\lambda)^2}
\right]^{-1},
\label{Tau4I}\end{aligned}$$
which taking $\lambda \rightarrow 0$ provides
$$Y \big( R^{-1}(\lambda) \big) \sim \lambda^{-1},
\label{TauS4I}$$
*i.e*. $\theta_{\rm II} = -1$. Hence, also here $\theta_{\rm I} =\theta_{\rm II}$, reconfirming the consistency of **Track I** and **Track II**.
Taken together, we arrive at an analytical proof, further strengthened by a broad array of dynamic examples, that the universal scaling exponent $\theta$ is independent of the derivation track, whether the original **Track I**, offered in Ref. [@Hens2019], or the alternative **Track II**, suggested recently.
[1]{}
. . , 15:403, 2019.
. . , [9]{}:[673 – 681]{}, 2013.
. . , [6]{}:[7186]{}, 2015.
. . , [8]{}:[2181]{}, 2017.
. . , [Vienna, Austria]{}, 1995.
. . , 101:15124–9, 2004.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Tensor factorizations have become increasingly popular approaches for various learning tasks on structured data. In this work, we extend the [<span style="font-variant:small-caps;">Rescal</span>]{} tensor factorization, which has shown state-of-the-art results for multi-relational learning, to account for the binary nature of adjacency tensors. We study the improvements that can be gained via this approach on various benchmark datasets and show that the logistic extension can improve the prediction results significantly.'
bibliography:
- 'SLG2013.bib'
---
Introduction
============
Tensor factorizations have become increasingly popular for learning on various forms of structured data such as large-scale knowledge bases, time-varying networks or recommendation data [@nickel_factorizing_2012; @bordes_learning_2011; @bader_temporal_2007; @rendle_factorizing_2010]. The success of tensor methods in these fields is strongly related to their ability to efficiently model, analyze and predict data with multiple modalities. Due to their multilinear nature, tensor models overcome limitations of linear models, such as their limited expressiveness, but at the same time remain more scalable and easier to handle then general non-linear approaches.
[<span style="font-variant:small-caps;">Rescal</span>]{} [@nickel_three-way_2011; @nickel_factorizing_2012] is a tensor factorization for dyadic multi-relational data which has been shown to achieve state-of-the-art results for various relational learning tasks such as link prediction, entity resolution or link-based clustering. Briefly, the [<span style="font-variant:small-caps;">Rescal</span>]{}model can be summarized as following: For relational data with $K$ different dyadic relations and $N$ entities, a third-order *adjacency tensor* ${\mathbf{X}}$ of size $N \times N \times K$ is created, where $$x_{ijk} = \begin{cases}
1, & \text{if } Rel_k(Entity_i, Entity_j) \text{ is true}\\
0, & \text{otherwise.}
\end{cases}$$ This adjacency tensor ${\mathbf{X}}$ is then factorized into latent representations of entities and relations, such that $$X_k \approx A R_k A^T$$ where $X_k$ is the $k$-th frontal slice of ${\mathbf{X}}$. After computing the factorization, the matrix $A \in
\mathbb{R}^{N \times r}$ then holds the latent representations for the entities in the data, i.e. the row ${\mathbf{a}}_i$ holds the latent representation of the $i$-th entity. Furthermore, $R_k \in \mathbb{R}^{r \times r}$ can be regarded as the latent representation of the $k$-th predicate, whose entries encode how the latent components interact for a specific relation. Since $R_k$ is a *full, asymmetric* matrix, the factorization can also handle directed relations. When learning the latent representation of an entity, unique global representation allows the model to efficiently access information that is more distant in the relational graph via information propagation through the latent variables. For instance, it has been shown that [<span style="font-variant:small-caps;">Rescal</span>]{}can propagate information about party membership of presidents and vice presidents over multiple relations, such that the correct latent representations are learned even when the party membership is unknown [@nickel_three-way_2011]. Moreover, since the entries of ${\mathbf{X}}$ are mutually independent given the latent factors $A$ and ${\mathbf{R}}_k$, prediction is very fast, as it reduces to simple vector-matrix-vector products.
In its original form, the [<span style="font-variant:small-caps;">Rescal</span>]{}factorization is computed by minimizing the least-squares error between the observed and the predicted entries; in a probabilistic interpretation this implies that the random variation of the data follows a Gaussian distribution, i.e. that $$x_{ijk} {\sim}{{\mathcal{N}}(\theta, \sigma^2)}$$ where $\theta$ are the parameters of the factorization. However, a Bernoulli is more appropriate for binary variables with $$x_{ijk} {\sim}Bernoulli(\theta)$$ where the parameter $\theta$ is again computed via the factorization of the corresponding adjacency tensor. In the following, we will present a learning algorithm based on logistic regression[^1] using the Bernoulli likelihood model and evaluate on benchmark data what gains can be expected from this updated model on relational data.
Methods
=======
In the following, we interpret [<span style="font-variant:small-caps;">Rescal</span>]{}from a probabilistic point of view. Each entry $x_{ijk}$ in ${\mathbf{X}}$ is regarded as a random variable and we seek to compute the MAP estimates of $A$ and ${\mathbf{R}}$ for the joint distribution $$p({\mathbf{X}}|A,{\mathbf{R}}) = \prod_{ijk}p(x_{ijk}|{\mathbf{a}}^T_iR_k{\mathbf{a}}_j).
\label{eq:rescal-probal}$$ also shows the graphical model in plate notation for the factorization. We will also fix the prior distributions of the latent factors to the Normal distribution, i.e. we set $$\begin{aligned}
{\mathbf{a}}_i & {\sim}{{\mathcal{N}}(0, \lambda_A I)}\\
R_k & {\sim}{{\mathcal{N}}(0, \lambda_R I)}\end{aligned}$$ Furthermore, we will maximize the log-likelihood of \[eq:rescal-probal\], such that the general form of the objective function that we seek to optimize is $$\operatorname*{arg\,min}_{A,{\mathbf{R}}} \mathrm{loss}(X; A, {\mathbf{R}}) +
\lambda_A\|A\|_F^2 +
\sum_k \lambda_R\|R_k\|_F^2
\label{eq:gen}$$ The nature of the loss function depends on the distribution that we assume for $x_{ijk}$. In the following we consider the least-squares and the logistic loss function.
= \[inner sep=4pt, font=,thick\] (x) [$x_{ijk}$]{}; (sigA) [$\sigma_A$]{}; (sig) [$\sigma$]{}; (ai) [${\mathbf{a}}_i$]{}; (aj) [${\mathbf{a}}_j$]{}; (R) [$R_k$]{}; (sigR) [$\sigma_R$]{};
; ; ; ; ; ; ;
Least-Squares Regression
------------------------
In its original form, [<span style="font-variant:small-caps;">Rescal</span>]{}sets the loss function to $$\mathrm{loss}(X; A, {\mathbf{R}}) \coloneqq \sum_k \|X_k - AR_kA^T\|_F^2.
\label{eq:rescal-als}$$ In this case, \[eq:gen\] and \[eq:rescal-als\] maximize the log-likelihood of \[eq:rescal-probal\] when $$\begin{aligned}
x_{ijk} & {\sim}{{\mathcal{N}}({\mathbf{a}}_i^TR_k{\mathbf{a}}_j, \sigma^2)}\end{aligned}$$ It should be noted that although the least-squares error does not imply the correct error model, it has the appealing property that it enables a very efficient and scalable implementation. An algorithm based on alternating least-squares updates of the factor matrices, has been shown to scale up to large knowledge bases via exploiting the sparsity of relational data. For instance, it has been used to factorize YAGO, an ontology which consists of around 3 million entities, 40 relations, and 70 million known facts on a single desktop computer [@nickel_factorizing_2012]. In the following we will refer to this implementation as [<span style="font-variant:small-caps;">Rescal</span>]{}-ALS.
Logistic Regression
-------------------
To describe the random variation in the data via a Bernoulli distribution, we set $$\begin{gathered}
\mathrm{loss}({\mathbf{X}}; A, {\mathbf{R}}) \coloneqq \\
-\sum_{ijk} x_{ijk} \log \sigma(\theta_{ijk})
+ (1-x_{ijk}) \log \left(1 - \sigma(\theta_{ijk})\right)
\label{eq:rescal-logit}\end{gathered}$$ where $$\sigma(\theta_{ijk}) = \frac{1}{1 + \mathrm{exp}(-{\mathbf{a}}_i^TR_k{\mathbf{a}}_j)}$$ Now, \[eq:gen\] and \[eq:rescal-als\] maximize the log-likelihood of \[eq:rescal-probal\] when $$\begin{aligned}
x_{ijk} & {\sim}Bernoulli(\sigma(\theta_{ijk}))\end{aligned}$$ Since, there exists no closed form solution to compute \[eq:rescal-logit\], we use a gradient based approach to compute \[eq:rescal-logit\] via quasi-Newton optimization, i.e. via the L-BFGS algorithm. The partial gradients for $A$ and $R_k$ are $$\begin{aligned}
\frac{\partial}{\partial A} = & \sum_k \left[\sigma(AR_kA^T) -
X_k)\right]AR_k^T +\\
& \quad\quad\quad \left[\sigma(AR_kA^T) - X_k)\right]^TAR_k + 2\lambda_AA \\
\frac{\partial}{\partial R_k} = & A^T \left[\sigma(AR_kA^T) - X_k\right] A +
2\lambda_RR_k\end{aligned}$$ where $\sigma(AR_kA^T)$ denotes the elementwise application of $\sigma(\cdot)$ to $AR_kA^T$. Unfortunately, terms of the form $$\left[\sigma\left(AR_kA^T\right) - X_k\right]A$$ can not be reduced to a significantly simpler form, due to the logistic function. Hence, this approach currently requires to compute the dense matrix $AR_kA^T$, what limits its scalability compared to the alternating least-squares approach. In the following, we will refer to this approach as [<span style="font-variant:small-caps;">Rescal</span>]{}-Logit
Experiments
===========
Kinships Nations Pres. Bact.
---------------------------------------------------------------- ----------- --------- ------- -----------
[<span style="font-variant:small-caps;">Rescal</span>]{}-ALS 0.966 0.848 0.805 0.927
[<span style="font-variant:small-caps;">Rescal</span>]{}-Logit **0.981** 0.851 0.800 **0.938**
MLN 0.85 0.75 - -
IRM 0.66 0.75 - -
: Evaluation results of the area under the precision-recall curve on the Kinships, Nations, Presidents, and Bacteriome datasets.
\[tab:results\]
To evaluate the logistic extension of [<span style="font-variant:small-caps;">Rescal</span>]{}, we conducted link-prediction experiments on the following datasets:
Presidents
: Multi-relational data, consisting of presidents of the United States, their vice-presidents as well as the parties of presidents and vice presidents.
Kinships
: Multi-relational data, consisting of several kinship relations within the Alwayarra tribe.
Nations
: Multi-relational, data consisting of relations between nations such as treaties, military actions, immigration etc.
Bacteriome
: Uni-Relational data, consisting of protein-protein and functional interactions within the context of an E. coli knowledgebase
For all datasets we performed 10-fold cross-validation and evaluated the results using the area under the precision-recall curve. In case of the presidents data, the task was to predict the party membership for presidents only based on the party memberships of their vice-presidents (and vice-versa). For all other datasets, cross-validation has been applied over all existing relations. It can be seen from the results in \[tab:results\] that the logistic extension of [<span style="font-variant:small-caps;">Rescal</span>]{}can considerably improve the prediction results. Especially the improvements for Kinships and Bacteriome are noteworthy, considering the already very good results of [<span style="font-variant:small-caps;">Rescal</span>]{}-ALS.
Conclusion
==========
To improve the modeling of multi-relational data, we have presented an extension for [<span style="font-variant:small-caps;">Rescal</span>]{}based on logistic regression. We have shown on several benchmark datasets that the logistic extension can improve the prediction results significantly. While the evaluation results are very encouraging, future work will have to address the scalability of the presented approach, as the scalability of its current implementation is too limited for practical use on larger datasets.
### Note added in proof {#note-added-in-proof .unnumbered}
Independently, a similar logistic extension of the [<span style="font-variant:small-caps;">Rescal</span>]{}factorization has been proposed in [@london_multi-relational_2013].
[^1]: In the theory of the exponential family, the logistic function describes the inverse parameter mapping for the Bernoulli distribution
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'In this paper, we present the results of CARMA continuum and spectral line observations of the dense core Per-Bolo 45. Although this core has previously been classified as starless, we find evidence for an outflow and conclude that Per-Bolo 45 is actually an embedded, low-luminosity protostar. We discuss the impact of newly discovered, low-luminosity, embedded objects in the Perseus molecular cloud on starless core and protostar lifetimes. We estimate that the starless core lifetime has been overestimated by 4-18% and the Class 0/I protostellar lifetime has been underestimated by 5-20%. Given the relatively large systematic uncertainties involved in these calculations, variations on the order of 10% do not significantly change either core lifetimes or the expected protostellar luminosity function. Finally, we suggest that high resolution (sub)millimeter surveys of known cores lacking near-infrared and mid-infrared emission are necessary to make an accurate census of starless cores.'
author:
- 'Scott Schnee, James Di Francesco, Melissa Enoch, Rachel Friesen, Doug Johnstone, Sarah Sadavoy'
title: 'How Starless Are Starless Cores?'
---
Introduction
============
Starless cores are the eventual birthplaces of stars, individually or possibly in small multiples. They represent the transition between a diffuse molecular cloud and the next generation of stars to form therein. Dense cores are often identified by their dust continuum emission, and their status as either protostellar or starless can be classified by the respective presence or absence of an embedded infrared object or molecular outflow. For example, cores in nearby molecular clouds identified by their 1.1mm emission have been classified by @Enoch08, and cores identified by their 850 emission have been classified by @Hatchell07a, @Jorgensen07, and @Sadavoy10. Given high amounts of extinction from dust, near-infrared observations can sometimes miss deeply embedded protostars, leading to a misclassification of dense cores. Reviews of the evolution of low-mass cores can be found in @DiFrancesco07 and @Ward-Thompson07.
The first “starless” core observed with the [*Spitzer Space Telescope*]{} [@Werner04] by the Cores to Disks legacy program [c2d; @Evans03], L1014, turned out to host an embedded infrared point source [@Young04]. Other supposedly starless cores, such as L1448 and L1521F, were also found to harbour Very Low Luminosity Objects (VeLLOs) that are either protostars or proto-brown dwarfs [@Kauffmann05; @Bourke06]. VeLLOs are defined to be objects embedded in dense cores that have luminosities $L_{int} <
0.1$ [@Kauffmann05; @DiFrancesco07]. A search for VeLLOs in the c2d survey identified 50 such objects, and 15-25% of the “starless” cores in the survey turned out to be misclassified [@Dunham08]. Furthermore, @Dunham08 postulated that there is likely to be a population of embedded objects even fainter than the sensitivity limit of the c2d survey, leading to the conclusion that there is still a significant number of hidden VeLLOs.
Despite the progress made with [*Spitzer*]{}, deeply embedded protostars will still be missed by near-infrared large-field surveys. For example, the well-studied protostar IRAS 16293-2422 has not been detected at wavelengths shorter than 12 [@Jorgensen08]. Recent observations of dense cores with (sub)millimeter interferometers and deeper mid-infrared observations towards known cores have found three new VeLLOs or candidate first hydrostatic cores in the Perseus molecular cloud. A first hydrostatic core is a short-lived stage that occurs after collapse has begun, between the prestellar core and Class 0 protostar phases, in which there is hydrostatic balance between the thermal pressure of molecular hydrogen and gravity. Since first hydrostatic cores qualify as VeLLOs, in this paper we will group these two classes of objects together unless there is observational evidence that a particular VeLLO is a first hydrostatic core. @Dunham11 have estimated that there should be between 1-6 first hydrostatic cores in the Perseus molecular cloud, based on the expected lifetime of first hydrostatic cores [$5
\times 10^2$ - $5 \times 10^4$ years; @Boss95; @Omukai07; @Saigo08; @Tomida10], an assumed lifetime of the embedded protostellar phase [$5.4 \times 10^5$ years; @Evans09], and the number of embedded prototstars in Perseus [66; @Enoch09]. Three candidate first hydrostatic cores have been identified so far in the Perseus molecular cloud. The first, L1448 IRS2E, previously believed to be starless based on non-detections with [*Spitzer*]{} from 3.6$\mu$m to 70$\mu$m, was recently found to host a protostar or first hydrostatic core [@Chen10a], based on the detection of a collimated CO (2-1) outflow. The second, Per-Bolo 58 [@Enoch06], previously believed to be starless based on [*Spitzer*]{} observations from the c2d survey [@Evans03], was found to host an embedded protostar or first hydrostatic core based on a deep 70 map and 3mm continuum CARMA observations [@Enoch10; @Schnee10]. The presence of an embedded source in Per-Bolo 58 was then confirmed with SMA observations showing a 1.3mm continuum source and CO (2-1) outflow [@Dunham11]. Third, L1451-mm was also considered starless based on non-detections in the [*Spitzer*]{} near-infrared and mid-infrared images taken as part of the c2d survey. Follow-up observations of L1451-mm with the SMA, CARMA, and VLA found compact 3mm and 1mm continuum emission coinciding with broadened spectral lines (, , and NH$_3$) and a slow, poorly collimated CO (2-1) outflow [@Pineda11]. The discovery that L1448 IRS2E, Per-Bolo 58, and L1451-mm are not starless suggests that other “starless” cores have been misidentified and instead harbour deeply embedded VeLLOs that were not detected in wide-area near-infrared and mid-infrared surveys.
In this paper, we present evidence that the dense core Per-Bolo 45 [@Enoch06], despite non-detections in the 3.6-70$\mu$m range, is also protostellar. Per-Bolo 45 is located in the Perseus molecular cloud, near the center of the cluster NGC 1333 and between IRAS 7 [@Jennings87] to the northeast and SVS 13 [@Strom76] to the southwest. In §\[OBSERVATIONS\] we describe new CARMA spectral line observations of Per-Bolo 45. The evidence that Per-Bolo 45 is protostellar is presented in §\[ANALYSIS\]. Some implications of the discoveries of embedded sources within Per-Bolo 45, Per-Bolo 58, L1448 IRS2E, and L1451-mm are discussed in §\[DISCUSSION\], and our results are summarized in §\[SUMMARY\].
Observations {#OBSERVATIONS}
============
Here we present spectral line maps of the dense core Per-Bolo 45 made with the Combined Array for Research in Millimeter-wave Astronomy (CARMA). We made new observations of the 3mm transitions of HCN, , SiO, , HNC, , and at $\sim$5 resolution, as described in §\[NEWCARMA\]. We discuss these observations in the context of the 3mm continuum data previously published by @Schnee10 and described briefly in §\[OLDCARMA\].
New CARMA Observations {#NEWCARMA}
----------------------
Spectral line observations in the 3mm window were obtained in July 2010 with CARMA, a 15 element interferometer consisting of nine 6.1 meter antennas and six 10.4 meter antennas. The CARMA correlator records signals in eight separate bands, each with an upper and lower sideband. We configured one band for maximum bandwidth (495MHz with 95 channels per band) to observe continuum emission, providing a total continuum bandwidth of approximately 1GHz. The other seven bands were configured to observe spectral lines and have 383 channels covering 7.8MHz per band, providing native spectral resolution of 20.4kHz (0.067), which we smoothed to 0.1 before analysis. The observations were centered around 90.7GHz, and range from 85.9GHz to 97.9GHz. The half-power beam width of the 10.4m antennas is 74 at the observed frequencies. Seven-point mosaics were made toward the dense core Per-Bolo 45 (with the phase center located at J2000 3:29:07.7 +31:17:16.8, and rest velocity 8.483) in the D and E-array configurations, with baselines that range from 11m to 150m. The largest angular scale to which the observations are sensitive is approximately 30. A summary of the CARMA observations is presented in Table \[MMOBSTAB\].
The observing sequence for the CARMA observations was to integrate on a phase calibrator (0336+323) for 3 minutes and Per-Bolo 45 for 21 minutes. The total time spent on this project was about 8 hours in each of the array configurations. In each set of observations, 3C84 was observed for passband calibration, and observations of Uranus were used for absolute flux calibration. Based on the repeatability of the quasar fluxes, the estimated random uncertainty in the measured source fluxes is $\sigma\sim5$%, and the systematic uncertainty is approximately 20%. Radio pointing was done at the beginning of each track and pointing constants were updated at least every two hours thereafter, using either radio or optical pointing routines [@Corder10]. Calibration and imaging were done using the MIRIAD data reduction package [@Sault95].
Due to instrumental problems, the first correlator band was lost, so we have no data on the HCN (1-0) line. We detected no significant emission of the C$^{34}$S (2-1) line, and it will not be discussed further in this paper.
Previous Observations {#OLDCARMA}
---------------------
Per-Bolo 45 was included in a CARMA 3mm continuum survey of 11 starless cores in the Perseus molecular cloud [@Schnee10]. It is one of two cores that were detected, and has an integrated 3mm flux density of 11$\pm$0.5mJy, corresponding to a mass of 0.8 after making reasonable assumptions for the core and dust properties. See @Schnee10 for more details about the 3mm continuum observations and the derived core properties.
Per-Bolo 45 was classified as a starless core by @Enoch06 [@Enoch08]. If a protostar was found within one full-width half-maximum (FWHM) of the center of the 1.1mm Bolocam core (i.e., 31, or 0.04pc), then it was classified as protostellar, otherwise the Bolocam core was classified as starless. Protostars in @Enoch08 were identified by the shape of their near-infrared to far-infrared spectral energy distributions (SEDs), a minimum flux at 24 and the presence of 70 point sources not classified as galaxy candidates. Per-Bolo 45 was identified as a Class 0 protostar by @Hatchell07a based on the colors of a nearby, but (we believe) unrelated, near-infrared source. There is a possible outflow from Per-Bolo 45 reported by @Hatchell07b, but due to confusion from nearby protostars this identification is given as tentative. In @Jorgensen07, cores were identified as protostellar based on either the presence of [*Spitzer*]{} detections and colors within a SCUBA core, or by the concentration of the light profile of the 850 SCUBA flux. Neither @Jorgensen07, @Kirk07, nor @Dunham08 found evidence for an embedded source in Per-Bolo 45, and it is classified as a starless core by @Sadavoy10. In Fig. \[SPITZERMAPS\], we show [*Spitzer*]{} near-infrared and mid-infrared images of Per-Bolo 45 with CARMA 3 mm continuum overlays, and it is clear that there is no embedded point source detected by [*Spitzer*]{} associated within Per-Bolo 45. The area shown in Figures \[SPITZERMAPS\] - \[TINTMAPS\] is approximately the region in which the gain of the mosaic is greater than 0.5. Therefore, the true classification of Per-Bolo 45 is not clear from the literature, but most studies have labeled it “starless.”
Analysis {#ANALYSIS}
========
We study the kinematics and morphology of the dense core Per-Bolo 45 by fitting the line profiles of (1$_{1,1}$ - 1$_{0,1}$), SiO (2-1), (1-0), HNC (1-0), and (1-0) using the MPFIT package in IDL [@Markwardt09]. Single Gaussians were fit to the SiO, , and HNC spectra, and multiple Gaussians were fit to the 7 components of the spectra and 6 components of the spectra. Fits were only made to those profiles that had three independent velocity channels with a signal-to-noise ratio greater than 3.
Molecular Line Emission {#MOLECULES}
-----------------------
Starless cores are kinematically quiescent objects, with line widths broadened approximately equally by thermal and non-thermal motions. The thermal line widths (FWHM) for the molecular lines presented in this paper are $\sim$0.13, assuming a typical temperature of 10K [@Schnee09]. The average non-thermal line width for NH$_3$ in starless cores in the Perseus molecular cloud is greater than the thermal line width of NH$_3$ by a factor of 1.5 [@Foster09]. Broad line widths (FWHM $\sim$ 0.6) in and towards the 3mm continuum peak of L1451-mm led @Pineda11 to hypothesize the presence of an embedded protostar in a core that had previously been considered starless, which was confirmed by detection of a CO (2-1) outflow. Similarly, velocity gradients in starless cores are small, with a typical value of 1 kms$^{-1}$pc$^{-1}$ [@Goodman93]. We would expect a total velocity gradient across Per-Bolo 45, which has a diameter no greater than 0.2pc, to be $\sim$0.2. Furthermore, we would expect the velocity field to be continuous, as seen in interferometric maps of the starless cores L183 [@Kirk09], L694-2 [@Williams06], and L1544 [@Williams06]. The kinematics of Per-Bolo 45, as traced by , , , and HNC, are described in Section \[NITROGEN\] and \[CARBON\]. Maps of the full-width half-maximum (FWHM) of , , , and HNC are shown in Figure \[FWHMMAPS\], and the LSR velocity maps for , , , and HNC are shown in Figure \[VLSRMAPS\]. Example spectra demonstrating the variety of line profiles in Per-Bolo 45 are shown in Figure \[SPECTRA\].
In the dense, cold, and quiescent interstellar medium, SiO is frozen onto the surfaces of dust grains except in regions shocked by outflows that return SiO to the gas phase and excite its emission [@Martin-Pintado92]. The presence of SiO emission is therefore a good indicator of one or more nearby protostars. SiO emission can be grouped into two classes, one that has broad lines ($>$10) and is Doppler-shifted by velocities $\ge$10 from the ambient cloud, and a second component with narrow lines ($\sim$1) found at velocities very close to the cloud material [e.g., @Lefloch98; @Jimenez-Serra04]. The SiO maps of Per-Bolo 45 are discussed in Section \[SIO\] and presented in Figure \[SIOMAP\].
An 850 SCUBA map of the dense core Per-Bolo 45 was produced as part of SCUBA Legacy Catalogues [@DiFrancesco08], and is shown in Figure \[TINTMAPS\] along with the 3mm continuum emission and integrated intensity maps of , , , and HNC. Whereas the 3mm continuum emission detected with CARMA has only one peak, the integrated intensity maps of , , , and HNC all have multiple peaks. The morphology of emission towards Per-Bolo 45 varies significantly between molecular lines, and this will be discussed below and presented in Figure \[TINTMAPS\].
### and {#NITROGEN}
The line width of in Per-Bolo 45 has a bimodal distribution, with narrow lines ($\sim$0.3) at the position of the dust peak and a small region of much broader (0.5-1.2) line width a few arcseconds to the southeast of the dust peak. The line width of is qualitatively similar to that of , with a narrow line width at the dust peak and much broader width a few arcseconds to the southeast. We argue that the broad line widths (up to 1.2) seen in and towards Per-Bolo 45 point towards the presence of a protostar, which must be low luminosity to have not been detected in mid-infrared [*Spitzer*]{} maps (see §\[EMBEDDED\]). The velocity fields of and (as well as and HNC) are shown in Fig.\[VLSRMAPS\]. The range of velocities seen in and is fairly narrow and similar to the systemic velocity of the core as measured in the single dish maps.
Both and are coincident with the 850 dust emission, but extend about 30 to the southeast beyond the extent of the 3mm continuum. In both lines the secondary peaks have narrow line widths typical of quiescent cores, so we suggest that this material is part of the envelope of Per-Bolo 45 that is not strongly affected by any outflow. The secondary peaks of and are coincident with the elongation of the 850 dust emission seen in the SCUBA map, although they are not accompanied by a similar feature in the 3mm continuum map. Discrepancies on small spatial scales between the dust emission and dense gas tracers have been seen before, for instance in the dense core Oph B [@Friesen09; @Friesen10] and in Oph A-N6 [@Pon09]. Furthermore, we show in this paper that the CARMA continuum maps are not sensitive to the emission from starless material.
### and HNC {#CARBON}
The integrated intensity map (see Fig. \[TINTMAPS\]) shows two peaks, one immediately to the southeast of the dust peak and the other 40 further to the southeast. The integrated intensity map of HNC shows a peak immediately to the southeast of the dust peak and also has three more peaks to the east and southeast of the 3mm continuum emission. Two of the HNC emission peaks to the east and southeast have no counterparts in the dust emission or with the other features in molecular line maps. Both and HNC show a broad range of velocities and both have line widths many times greater than the thermal line width, as shown in Figs.\[FWHMMAPS\] and \[VLSRMAPS\]. In the case of , both redshifted and blueshifted emission are seen with a spread of $\pm$1 from the systemic velocity of the core, and no emission is seen at the systemic velocity of Per-Bolo 45. The HNC emission adjacent to the 3mm continuum peak is at the systemic velocity of the core and has broad line widths (up to 2), while the other HNC peaks are blue-shifted and have narrow line widths.
Although the interpretation of the emission is complicated by dissimilarities between the morphologies of the emission, we argue that the broad line widths (up to 2) and velocities Doppler-shifted from the systemic velocity of the core (by $\pm$1 in the case of ) are likely caused by an outflow driven by Per-Bolo 45. Neither kinematic feature would be expected in a starless core, as explained in Section \[MOLECULES\]. The regions of narrow line width seen in HNC and (as well as and ) likely trace the more quiescent gas in a dense envelope. has been seen to trace outflows in the cluster NGC 1333 [e.g., @Walsh07], of which Per-Bolo 45 is a member. Given that HCN has been also seen to trace outflows in NGC 1333 [@Jorgensen04], it is reasonable to suggest that the broad line width HNC in Per-Bolo 45 traces an outflow as well. HNC has been seen to trace dense material around a protostar as well as material associated with the protostellar outflow [@Arce04], so the appearance of both quiescent and turbulent HNC in Per-Bolo 45 is not unexpected if it harbors a protostar.
### SiO {#SIO}
Per-Bolo 45 exhibits SiO emission, with $\sim$1 line widths (see Fig.\[SPECTRA\]) and velocities within $\sim$1 of the ambient medium (see Fig.\[SIOMAP\]). The systemic velocity of Per-Bolo 45 is $\sim$8.5 [@Kirk07; @Rosolowsky08]. The systemic velocity of the ISM around Per-Bolo 45, measured in an FCRAO map of CS (2-1) by the COMPLETE survey [@Ridge05], is 8.6, with a 1.7 line width. The velocity range of the SiO (2-1) emission towards Per-Bolo 45 is therefore entirely consistent with the range of velocities of the ambient material. The morphology of the SiO emission is perhaps suggestive, being extended with the long axis of the emission pointing towards the 3mm continuum emission.
We interpret the detection of SiO (2-1) emission immediately adjacent to Per-Bolo 45 as further evidence for the presence of an embedded protostar. An outflow launched by a protostar in Per-Bolo 45 would liberate SiO from the dust in the ISM, creating the conditions required for the observed emission. Since we only detect the narrow component of SiO emission that traces the ambient cloud material, the velocity map shown in Fig.\[SIOMAP\] is not sufficient to say for certain what portion of the outflow is exciting the emission. The velocity structure seen in Fig.\[SIOMAP\] is likely that of the ISM around Per-Bolo 45, and not that of the outflow. The velocity range covered by the CARMA observations ($-5$ $\le$ VLSR $\le$ 21) was not wide enough to find the high-velocity, broad component of the SiO emission more closely associated with the outflow, if present.
### The case for the protostellar nature of Per-Bolo 45
We have shown that Per-Bolo 45 exhibits several traits of protostellar cores, despite the non-detection of mid-infrared emission in IRAC and MIPS [*Spitzer*]{} maps (see Fig.\[SPITZERMAPS\]). The line widths seen in , , , and HNC are several times larger than the thermal line width, behaviour not found in starless cores but which could be explained by the presence of an embedded protostar. Broad line widths of and in L1451-mm led to the discovery of a protostar in that core even though it had previously been identified as starless [@Pineda11]. The presence of redshifted and blueshifted emission seen about 1 from the systemic velocity of the core would also not be expected in a starless core, but could be explained by an outflow from an embedded protostar. SiO emission located adjacent to Per-Bolo 45 can be explained by the interaction between the ambient material in the molecular cloud and an outflow launched by the core. Finally, the 3mm continuum emission detected towards Per-Bolo 45 is also suggestive of a possible embedded source, given that 3mm continuum emission was detected with CARMA towards Per-Bolo 58 and L1541-mm, two supposedly “starless” cores that were subsequently found to be protostellar [@Enoch10; @Dunham11; @Pineda11]. None of the other starless cores surveyed with CARMA by @Schnee10 exhibited 3mm continuum emission.
Given that Per-Bolo 45 is in the NGC 1333 cluster, it is important to check that the outflow does not originate from another nearby protostar. The elongation of the SiO emission points back towards the HH 7-11 group associated with SVS 13, and even further perhaps towards IRAS 4, but each of these cores have well-known outflows going in directions other than towards Per-Bolo 45. For example, a CO (3-2) map of NGC 1333 shows no outflow directed towards Per-Bolo 45 from any other object [@Curtis11]. Instead, Per-Bolo 45 seems to be adjacent to a compact red lobe emanating from IRAS 7 to the northeast and a very extended red lobe emanating from IRAS 2 in the southwest. These redshifted lobes have velocity ranges of 12 to 18 and their blueshifted lobes range from -5kms to 3, both outside the range of velocities associated with the redshifted and blueshifted emission from Per-Bolo 45. It would also be difficult to explain how an outflow from elsewhere in the NGC 1333 cluster could interact with the dense material in Per-Bolo 45 traced by and in such a way as to increase turbulence immediately to the south-east of the 3mm continuum peak without also affecting the area with nearly thermal line widths around it. We conclude that the molecular line observations presented in this paper are best explained by the presence of a protostar embedded within Per-Bolo 45.
Embedded Source {#EMBEDDED}
---------------
A relationship between the internal luminosity of a young stellar object (YSO) and its 70 flux is given by Equation 2 in @Dunham08. Per-Bolo 45 is not detected at 70, but we can use a 3$\sigma$ upper limit to its 70 flux to estimate an upper limit to the internal luminosity of the embedded source. We find that the internal luminosity of Per-Bolo 45 is less than $10^{-2}$, similar to the upper limit for L1451-mm by @Pineda11 and lower by a factor of about 10 than the luminosities of the embedded protostars observed by [*Spitzer*]{} with published models shown in Table 1 of @Dunham08. The upper limit to the luminosity of the embedded source in Per-Bolo 45 is consistent with the sensitivity limit of the c2d survey. Assuming that Per-Bolo 45 is protostellar, the lack of observed 70 flux would make this source a VeLLO by the definition given in @DiFrancesco07. A protostellar source with a disk viewed edge-on might remain invisible at the observed sensitivity limits, so it is possible that the internal luminosity of Per-Bolo 45 is higher than that of a VeLLO and it appears dim because of the viewing angle. Given its low 70 luminosity, the source embedded in Per-Bolo 45 is also a plausible first hydrostatic core candidate.
Discussion {#DISCUSSION}
==========
In this paper we identify Per-Bolo 45 as a core with an embedded source, changing its classification from starless core [@Enoch08] to VeLLO. Other cores in the Perseus molecular cloud, previously identified as starless due to their lack of near- and mid-infrared emission [e.g. @Enoch08; @Sadavoy10] but subsequently found to show evidence for protostellar activity, such as launching molecular outflows, include Per-Bolo 58 [@Enoch10; @Dunham11] , L1448 IRS2E [@Chen10a], and L1451-mm [@Pineda11].
Out of the 11 “starless” cores in Perseus surveyed by @Schnee10 with interferometric observations of their 3mm continuum, the only two cores detected (Per-Bolo 45 and Per-Bolo 58) have since been shown to harbour embedded objects. This fraction of misidentified starless cores in Perseus, 2/11 or 18%, is probably an upper limit given that the cores in the @Schnee10 sample were chosen to have high surface brightnesses in the 1.1mm continuum map of @Enoch06, and the presence of a protostar would increase the peak flux of a dense core.
A lower limit to the number of cores in Perseus misidentified as starless can be derived from the number of previously identified starless cores and the number of newly identified protostars and VeLLOs. This estimate provides a lower limit because not every core identified in wide-field dust continuum maps of Perseus has been followed up with (sub)millimeter interferometric molecular line observations to find outflows. @Sadavoy10 find 97 starless cores and 46 protostellar cores in Perseus, using SCUBA 850 maps to identify the cores and [*Spitzer*]{} maps between 3.6 and 70 to determine the starless or protostellar status of each core. @Jorgensen07 and @Enoch08 find similar numbers of starless cores and protostars using similar techniques. Although the census of protostellar cores and starless cores detected by [*Herschel*]{} has not yet been released for the Perseus molecular cloud, a 2:1 ratio of prestellar cores to protostellar cores in the Aquila rift was reported by @Konyves10 and @Bontemps10, in good agreement with the pre-[*Herschel*]{} studies of Perseus described above. Considering the four newly-identified VeLLOs in Perseus, a lower limit of 4/97 “starless” cores, or 4%, have been misclassified. Our upper limit of 18% would imply that 17 cores have been misclassified as starless. We note that it is likely that not all of the dense cores in Perseus have been identified, given that the published (sub)millimeter surveys used to identify such cores have had limited sensitivity to faint and extended structures.
The lifetime of starless cores can be estimated from the relative number of starless cores and protostellar cores, given a reasonable estimate of the protostellar lifetime and assuming that the rate of star formation is constant over time. @Enoch08 found that a mean lifetime for starless cores in nearby molecular clouds is $0.5
\pm 0.3$Myr, so the $\le$20% change in the number of starless cores in Perseus resulting from the discovery of embedded sources in previously identified “starless” cores is roughly comparable to the uncertainty coming from other systematics in the estimate of starless core lifetimes.
Similarly, increasing the number of protostars increases the resultant protostellar lifetime. @Evans09 found 87 Class 0/I protostars in the Perseus molecular cloud, so adding 4-17 more sources into the Class 0/I category would increase the Class 0/I lifetime by $\sim$5-20%. This increase in the Class 0/I lifetime would not significantly affect modelling of the expected accretion luminosity averaged over the lifetime of a protostellar core, given the significant uncertainties on both the observational and theoretical sides of the problem, as summarized in @Offner11.
Many studies of nearby molecular clouds [e.g., @Motte98; @Nutter07; @Ward-Thompson07; @Enoch08; @Konyves10] have shown that the starless core mass function (CMF) has a similar shape to the stellar initial mass function (IMF). These studies depend on the ability to distinguish starless dense cores from protostellar dense cores. @Hatchell08, @Enoch08 and @DiFrancesco10 each found that their most massive cores tended to be protostellar, which suggests that either prestellar cores have an upper mass limit, consistent with gravitational stability inequalities (e.g., Jeans analysis), or that the more massive prestellar cores have very short lifetimes. Furthermore, @Sadavoy10b found that the most massive “starless” cores in Orion and Perseus had ambiguous infrared emission towards them and thus, changed their classification to “undetermined.” If the most massive “starless” cores tend to be misclassified VeLLOs, then the CMF will lack the massive core tail found in the IMF. This difference will steepen the starless CMF slope with respect to the IMF slope, i.e., the two mass functions will have different shapes.
Three of the VeLLOs in Perseus (Per-Bolo 45, Per-Bolo 58, and L1448 IRS2e) were included in the @Sadavoy10b investigation into the gravitational stability of starless cores in Perseus. None of these sources were found to be particularly unstable. Instead, they reside, along with the majority of starless cores, in a mass range of 1-3 Jeans masses [see @Sadavoy10b for details on how the stability analysis was performed]. The only possible hint that these cores are unique is that all three have relatively small sizes (compared with other cores at similar level of Jeans instability). That these cores are not clearly distinguishable by a Jeans analysis further suggests that deep mid-infrared and interferometric observations will be required to uncover the majority of VeLLOs. Given that only a few “starless” cores have been proven to be misclassified, and that these cores do not stand out as being especially massive, the shape of the CMF, as discussed above, has not been biased by the presence of low-luminosity protostellar objects.
In a 3mm continuum survey of 11 “starless” cores, @Schnee10 found that only two (Per-Bolo 45 and Per-Bolo 58) were detected and concluded that the density distribution in starless cores must be smooth and not strongly peaked. The subsequent discovery that both cores are actually VeLLOs [@Enoch10; @Dunham11 this paper] strengthens the claim that starless cores in Perseus have smooth and shallow density profiles. A shallow inner density profile in starless cores has been previously reported by several groups, [e.g., @Ward-Thompson94; @Ward-Thompson99; @Shirley00]. Had the starless cores in our Perseus sample been in the process of fragmentation, we would have been able to detect the resultant fragments in the dust emission maps.
Although [*Spitzer*]{} surveys were more sensitive than those conducted with previous instruments and found embedded objects in cores that had been identified as starless [@Young04; @Kauffmann05; @Bourke06], it is also true that deeply embedded protostars can be missed at wavelengths less than 12 [@Jorgensen08] by surveys like c2d [@Evans03]. In addition, @Dunham08 have found that there is likely to be a population of VeLLOs too faint to have been detected in current surveys. New [*Herschel*]{} observations will be able to find fainter protostars than [*Spitzer*]{} was able to detect, and in the Aquila region seven of the $\sim$50 Class 0 protostars detected by [ *Herschel*]{} were missed by [*Spitzer*]{} [@Bontemps10]. Still, @Bontemps10 reported that compact sources in Aquila were identified down to the [*Herschel*]{} 70 detection limit, implying that a population of even fainter sources have yet to be discovered. Given that low-luminosity embedded objects can be found through high-resolution observations of their dust continuum emission and molecular outflows, we suggest that interferometric (sub)millimeter observations are a promising method for determining what fraction of “starless” cores have been misclassified.
Summary {#SUMMARY}
=======
In this paper, we report on CARMA maps of the 3mm continuum and 3mm window spectral lines of , SiO, , HNC, and . Our main results are:
- [Despite non-detections in [*Spitzer*]{} maps at 3.6-24, Per-Bolo 45 is a protostellar core, as inferred from large line widths, Doppler-shifted emission, and the presence of SiO emission.]{}
- [There are at least four cores in Perseus, previously identified as starless, that recent observations have shown to contain low-luminosity embedded objects with molecular outflows [@Chen10a; @Enoch10; @Dunham11; @Pineda11]. If this result can be generalized to other nearby molecular clouds, we estimate that the lifetime of starless cores has been overestimated by 4-18% and the lifetime of Class 0/I protostars has been underestimated by 5-20%. These changes are within the previously published uncertainties of starless and protostellar core lifetimes [@Enoch08; @Evans09].]{}
- [Although recent infrared surveys of nearby molecular clouds have made great progress towards classifying starless and protostellar cores [e.g., @Hatchell07a; @Jorgensen07; @Jorgensen08; @Enoch08; @Dunham08; @Evans09; @Sadavoy10; @Konyves10; @Bontemps10], we suggest that an improved census will require high-resolution (sub)millimeter observations to survey known cores and identify embedded low-luminosity objects and outflows.]{}
We thank our anonymous referee for comments that have significantly strengthed this paper. JDF acknowledges support by the National Research Council of Canada, the Canadian Space Agency (via a SSEP Grant), and the Natural Sciences and Engineering Council of Canada (via a Discovery Grant). Support was provided to ME by NASA through the [*Spitzer Space Telescope*]{} Fellowship Program. DJ is supported by a Natural Sciences and Engineering Research Council of Canada (NSERC) Discovery Grant. We thank the CARMA staff, students and postdocs for their help in making these observations. Support for CARMA construction was derived from the Gordon and Betty Moore Foundation, the Kenneth T. and Eileen L. Norris Foundation, the Associates of the California Institute of Technology, the states of California, Illinois and Maryland, and the National Science Foundation. Ongoing CARMA development and operations are supported by the National Science Foundation under a cooperative agreement, and by the CARMA partner universities.
[*Facilities*]{}: CARMA
Arce, H. G., Borkin, M. A., Goodman, A. A., Pineda, J. E., & Halle, M. W. 2010, , 715, 1170 Arce, H. G., & Sargent, A. I. 2004, , 612, 342 Bontemps, S., Andr[é]{}, P., K[ö]{}nyves, V., et al. 2010, , 518, L85 Boss, A. P., & Yorke, H. W. 1995, , 439, L55 Bourke, T. L., Bourke, Myers, P. C., Evans, N.J., II, et al. 2006, , 649, L37 Chen, X., Arce, H. G., Zhang, Q., et al. 2010, , 715, 1344 Corder, S. A., Wright, M. C. H., & Carpenter, J. M. 2010, , 7733, Curtis, E. I., Richer, J. S., Swift, J. J., & Williams, J. P. 2010, , 408, 1516 Di Francesco, J., Sadavoy, S., Motte, F., et al. 2010, , 518, L91 Di Francesco, J., Johnstone, D., Kirk, H., MacKenzie, T., & Ledwosinska, E. 2008, , 175, 277 Di Francesco, J., Evans, N. J., II, Caselli, P., et al. 2007, Protostars and Planets V, B. Reipurth, D. Jewitt, and K. Keil (eds.), University of Arizona Press, Tucson, p.17-32 Dunham, M. M., Chen, X., Arce, H. G., et al. 2011, , 742, 1 Dunham, M. M., Crapsi, A., Evans, N. J., II, Bourke, T. L., Huard, T. L., Myers, P. C., & Kauffmann, J. 2008, , 179, 249 Enoch, M. L., Lee, J.-E., Harvey, P., Dunham, M. M., & Schnee, S. 2010, , 722, L33 Enoch, M. L., Evans, N. J., II, Sargent, A. I., & Glenn, J. 2009, , 692, 973 Enoch, M. L., Evans, N. J., II, Sargent, A. I., et al. 2008, , 684, 1240 Enoch, M. L., Young, K. E., Glenn, J., et al. 2006, , 638, 293 Evans, N. J., Dunham, M. M., Jørgensen, J. K., et al. 2009, , 181, 321 Evans, N. J., II, Allen, L. E., Blake, G. A., et al. 2003, , 115, 965 Foster, J. B., Rosolowsky, E. W., Kauffmann, J., et al. 2009, , 696, 298 Friesen, R. K., Di Francesco, J., Shirley, Y. L., & Myers, P. C. 2009, , 697, 1457 Friesen, R. K., Di Francesco, J., Shimajiri, Y., & Takakuwa, S. 2010, , 708, 1002 Goodman, A. A., Benson, P. J., Fuller, G. A., & Myers, P. C. 1993, , 406, 528 Hatchell, J., & Fuller, G. A. 2008, , 482, 855 Hatchell, J., Fuller, G. A., Richer, J. S., Harries, T. J., & Ladd, E. F. 2007, , 468, 1009 Hatchell, J., Fuller, G. A., & Richer, J. S. 2007, , 472, 187 Jennings, R. E., Cameron, D. H. M., Cudlip, W., & Hirst, C. J. 1987, , 226, 461 Jessop, N. E., & Ward-Thompson, D. 2000, , 311, 63 Jim[é]{}nez-Serra, I., Mart[í]{}n-Pintado, J., Rodr[í]{}guez-Franco, A., & Marcelino, N. 2004, , 603, L49 J[ø]{}rgensen, J. K., Johnstone, D., Kirk, H., et al. 2008, , 683, 822 J[ø]{}rgensen, J. K., Johnstone, D., Kirk, H., & Myers, P. C. 2007, , 656, 293 J[ø]{}rgensen, J. K., Hogerheijde, M. R., Blake, G. A., et al. 2004, , 415, 1021 Kauffmann, J., Bertoldi, F., Evans, N. J., II, & the C2D Collaboration 2005, Astronomische Nachrichten, 326, 878 Kirk, J. M., Crutcher, R. M., & Ward-Thompson, D. 2009, , 701, 1044 Kirk, H., Johnstone, D., & Tafalla, M. 2007, , 668, 1042 K[ö]{}nyves, V., Andr[é]{}, P., Men’shchikov, A., et al. 2010, , 518, L106 Lefloch, B., Castets, A., Cernicharo, J., & Loinard, L. 1998, , 504, L109 Lovas, F. J. 1992, Journal of Physical and Chemical Reference Data, 21, 181 Markwardt, C. B. 2009, Astronomical Data Analysis Software and Systems XVIII, 411, 251 Martin-Pintado, J., Bachiller, R., & Fuente, A. 1992, , 254, 315 Motte, F., Andre, P., & Neri, R. 1998, , 336, 150 Nutter, D., & Ward-Thompson, D. 2007, , 374, 1413 Offner, S. S. R., & McKee, C. F. 2011, , 736, 53 Omukai, K. 2007, , 59, 589 Pineda, J. E., Arce, H. G., Schnee, S., et al. 2011, arXiv:1109.1207 Pon, A., Plume, R., Friesen, et al. 2009, , 698, 1914 Ridge, N. A., Di Francesco, J., Kirk, H., et al. 2006, , 131, 2921 Rosolowsky, E. W., Pineda, J. E., Foster, J. B., et al. 2008, , 175, 509 Sadavoy, S. I., Di Francesco, J., Bontemps, S., et al. 2010, , 710, 1247 Sadavoy, S. I., Di Francesco, J., & Johnstone, D. 2010, , 718, L32 Saigo, K., Tomisaka, K., & Matsumoto, T. 2008, , 674, 997 Sault, R. J., Teuben, P. J., & Wright, M. C. H. 1995, Astronomical Data Analysis Software and Systems IV, 77, 433 Schnee, S., Enoch, M., Johnstone, D., et al. 2010, , 718, 306 Schnee, S., Rosolowsky, E., Foster, J., Enoch, M., & Sargent, A. 2009, , 691, 1754 Sch[ö]{}ier, F. L., van der Tak, F. F. S., van Dishoeck, E. F., & Black, J. H. 2005, , 432, 369 Shirley, Y. L., Evans, N. J., II, Rawlings, J. M. C., & Gregersen, E. M. 2000, , 131, 249 Strom, S. E., Vrba, F. J., & Strom, K. M. 1976, , 81, 314 Tafalla, M., Myers, P. C., Caselli, P., & Walmsley, C. M. 2004, , 416, 191 Tin[é]{}, S., Roueff, E., Falgarone, E., Gerin, M., & Pineau des For[ê]{}ts, G. 2000, , 356, 1039 Tomida, K., Machida, M. N., Saigo, K., Tomisaka, K., & Matsumoto, T. 2010, , 725, L239 Walsh, A. J., Myers, P. C., Di Francesco, J., et al. 2007, , 655, 958 Ward-Thompson, D., Andr[é]{}, P., Crutcher, R., et al. 2007, Protostars and Planets V, B. Reipurth, D. Jewitt, and K. Keil (eds.), University of Arizona Press, Tucson, p.33-46 Ward-Thompson, D., Motte, F., & Andre, P. 1999, , 305, 143 Ward-Thompson, D., Scott, P. F., Hills, R. E., & Andre, P. 1994, , 268, 276 Werner, M. W., Roellig, T. L., Low, F. J., et al. 2004, , 154, 1 Williams, J. P., Lee, C. W., & Myers, P. C. 2006, , 636, 952 Young, C. H., J[ø]{}rgensen, J.K., Shirley, Y. L., et al. 2004, , 154, 396
[lccc]{} NH$_2$D (1$_{1,1}$ - 1$_{0,1}$ F=2-2) & 85.9262703 & 7.4$\times$5.5 & 0.45\
SiO (2-1) & 86.84696 & 7.6$\times$5.7 & 0.42\
HCO$^+$ (1-0) & 89.188523 & 6.9$\times$5.3 & 0.42\
HNC (1-0) & 90.663568 & 7.2$\times$5.4 & 0.42\
N$_2$H$^+$ (1-0) F$_1$=2-1 F=3-2 & 93.173777 & 7.0$\times$5.3 & 0.46\
C$^{34}$S (2-1) & 96.4129495 & 6.9$\times$4.9 & 0.47
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We prove that to every pair of Lie algebroids $(L,A)$ corresponds a Kapranov dg-manifold structure on $A[1]\oplus L/A$, canonical up to isomorphism. As a consequence, ${\Gamma(\wedge^\bullet A^\vee\otimes L/A)}$ carries a canonical $L_\infty[1]$ algebra structure whose binary bracket is a cocycle representative of the Atiyah class of the pair $(L,A)$. For Lie algebroids over ${\mathbb{R}}$, we conjecture that this Kapranov dg-manifold ought to be considered as the derived formal neighborhood of a certain substack in a double-quotient differentiable stack. The second main purpose of the paper is the construction, for Lie algebroid pairs, of Poincaré–Birkhoff–Witt type isomorphisms described by an iteration formula, which allows explicit computations. Such Poincaré–Birkhoff–Witt isomorphisms constitute simultaneous extensions of the classical PBW map of Lie theory and the inverse of the complete symbol map of differential operators.'
address:
- 'Département de mathématiques, université de Lorraine'
- 'Department of Mathematics, Penn State University'
- 'Department of Mathematics, Penn State University'
author:
- 'Camille Laurent-Gengoux'
- Mathieu Stiénon
- Ping Xu
bibliography:
- 'biblio.bib'
title: |
Poincaré–Birkhoff–Witt isomorphisms\
and Kapranov dg-manifolds
---
=1
Introduction
============
Lately some surprising homotopic algebra structures have arisen in connection with Lie algebroids. This paper aims to solve some puzzles in that direction.
In his work [@Kapranov] on Rozansky–Witten invariants, Kapranov discovered a natural $L_\infty[1]$ algebra structure on the Dolbeault complex ${{\Omega}}^{0,\bullet}(T^{1,0}_X)$ of an arbitrary Kähler manifold $X$. Its unary bracket $\lambda_1$ is the Dolbault differential operator $\overline{\partial}$, its binary bracket $\lambda_2$ is the composition $${{\Omega}}^{0,k}(T^{1,0}_X)\otimes{{\Omega}}^{0,l}(T^{1,0}_X){\xrightarrow{\wedge}}{{\Omega}}^{0,k+l}(T^{1,0}_X\otimes T^{1,0}_X)
{\xrightarrow{\mathcal{R}_2}}{{\Omega}}^{0,k+l+1}(T^{1,0}_X)$$ of the exterior product and the contraction with the Atiyah cocycle $$\mathcal{R}_2\in{{\Omega}}^{0,1}\big(\operatorname{Hom}(T^{1,0}_X\otimes T^{1,0}_X,T^{1,0}_X)\big) ,$$ and the higher brackets are obtained similarly by composing the exterior product and a contraction. This $L_\infty[1]$ algebra plays an important role in derived geometry [@Caldararu; @Markarian; @Ramadoss].
The Atiyah class $\alpha_E$ of a holomorphic vector bundle $E$ over a complex manifold $X$, which is an element of the sheaf cohomology group $H^1\big(X;{{\Omega}}_X^1\otimes\operatorname{End}(E)\big)$, captures the obstruction to the existence of a holomorphic connection on $E$ [@Atiyah]. A Dolbeault representative of $\alpha_E$ can be obtained as follows. Considering $T^{1,0}_X$ as a complex Lie algebroid, choose a $T^{1,0}_X$-connection $\nabla^{1,0}$ on $E$. Being a holomorphic vector bundle, $E$ carries a canonical flat $T^{0,1}_X$-connection $\overline{\partial}$. Adding $\nabla^{1,0}$ and $\overline{\partial}$, we obtain a $T_X\otimes{\mathbb{C}}$-connection $\nabla$ on $E$. The element $\mathcal{R}_2\in{{\Omega}}^{1,1}(\operatorname{End}E)$ defined by $$\mathcal{R}_2(a,b)s=\nabla_a\nabla_b s-\nabla_b\nabla_a s-\nabla_{{[a,b]}}s,
\quad\forall a\in{\Gamma(T^{0,1}_X)}, b\in{\Gamma(T^{1,0}_X)}, s\in{\Gamma(E)}$$ is a Dolbeault 1-cocycle (called Atiyah cocycle of $E$) whose cohomology class (which is independent of the choice of $\nabla^{1,0}$) is the Atiyah class $$\alpha_E\in H^{1,1}(X;\operatorname{End}E)\cong H^1(X;{{\Omega}}_X^1\otimes\operatorname{End}E) .$$
Chen and two of the authors discovered that Atiyah classes can be defined in the much wider context of *pairs of Lie algebroids* [@CSX]. We say that $(L,A)$ is a pair of Lie algebroids or *Lie pair* if $A$ is a subalgebroid of a Lie algebroid $L$ over a common base manifold. The Atiyah class $\alpha_{L/A}$ of the pair, which is a simultaneous generalization of both the classical Atiyah class of holomorphic tangent bundles and the Molino class of foliations [@Atiyah; @Molino; @Molino_topology], captures the obstruction to the existence of an *$A$-compatible* $L$-connection on $L/A$. Furthermore, it was brought to light that the graded vector space ${\Gamma(\wedge^\bullet A^\vee \otimes L/A)}$ admits a puzzling $\operatorname{\textit{Leibniz}}_\infty[1]$ algebra structure — a natural generalization of Stasheff’s $L_\infty[1]$ algebras [@Lada_Stasheff] first introduced by Loday [@Loday] in which the requirement that the multibrackets be (skew-)symmetric is dropped. Its unary bracket $\lambda_1:{\Gamma(\wedge{^{\bullet}}A^\vee \otimes L/A)}\to{\Gamma(\wedge^{\bullet+1} A^\vee \otimes L/A )}$ is the Chevalley–Eilenberg coboundary operator associated to the Bott representation of the Lie algebroid $A$ on $L/A$, its binary bracket $\lambda_2$ is the composition of the wedge product and the contraction with a Chevalley–Eilenberg Lie algebroid 1-cocycle ${\mathcal{R}}_2 \in{\Gamma(A^\vee \otimes S^2 (L/A)^\vee \otimes L/A)}$ representative of the Atiyah class $$\alpha_{L/A}\in H^1_{\operatorname{CE}}(A;S^2 (L/A)^\vee\otimes L/A) ,$$ and all multibrackets $\lambda_k:\otimes^k{\Gamma(\wedge^{\bullet} A^\vee \otimes L/A)}\to
{\Gamma(\wedge^{\bullet+1} A^\vee \otimes L/A)}$ with $k\geqslant 2$ are ${\Gamma(\wedge{^{\bullet}}A^\vee)}$-multilinear. The approach in [@CSX] being rather computational, this $\operatorname{\textit{Leibniz}}_\infty[1]$ algebra remained quite enigmatic.
It was natural to wonder whether this result could be improved.
\[Brussels\] Can the sequence of multibrackets $\{\lambda_k\}_{k\in{\mathbb{N}}}$ be tweaked so as to make ${\Gamma(\wedge^\bullet A^\vee\otimes L/A)}$ an $L_\infty[1]$ algebra rather than a mere $\operatorname{\textit{Leibniz}}_\infty[1]$ algebra? If so, how can this $L_\infty[1]$ algebra be interpreted geometrically?
It is well known that a ${\mathbb{Z}}$-graded vector space $\mathfrak{g}=\bigoplus_{i\in{\mathbb{Z}}}\mathfrak{g}_i$ of finite type (i.e. with each $\mathfrak{g}_i$ finite-dimensional) is an $L_\infty[1]$ algebra if and only if it is a dg-manifold [@Kontsevich_formality; @MR1183483]. By a dg-manifold[^1], we mean a ${\mathbb{Z}}$-graded manifold endowed with a vector field $Q$ of degree $+1$ satisfying ${[Q,Q]}=0$. Such a vector field is said to be homological [@Vaintrob]. Hence we are led to seek a particular type of homological vector field on the ${\mathbb{Z}}$-graded manifold $A[1]\oplus L/A$, whose precise characterization is encapsulated in the new notion of Kapranov dg-manifold, the core concept of the present paper.
A *Kapranov dg-manifold* consists of a pair of smooth vector bundles $A$ and $E$ over a common base manifold and a pair of homological vector fields on the ${\mathbb{Z}}$-graded manifolds $A[1]\oplus E$ and $A[1]$ such that the inclusion $A[1]{\hookrightarrow}A[1]\oplus E$ and the projection $A[1]\oplus E{\twoheadrightarrow}E$ are morphisms of dg-manifolds. Isomorphisms of Kapranov dg-manifolds are defined in the obvious way.
It is worth stressing that the addition of a homological vector field on $A[1]$ turns $A$ into a Lie algebroid. Given a representation of a Lie algebroid $A$ on a vector bundle $E$, the Chevalley–Eilenberg differential provides a pair of homological vector fieds on $A[1]$ and $A[1]\oplus E$ compatible with the inclusion and projection maps. The resulting Kapranov dg-manifold is said to be *linear*.
Our main result is the following
\[Amsterdam\] Given a Lie pair $(L,A)$, there exists a homological vector field $${\Gamma(\wedge{^{\bullet}}A^\vee\otimes\hat{S}((L/A)^\vee))}
{\xrightarrow{{D}}} {\Gamma(\wedge^{\bullet+1} A^\vee\otimes\hat{S}((L/A)^\vee))}$$ on the graded manifold $A[1]\oplus L/A$, compatible with the inclusion $A[1]{\hookrightarrow}A[1]\oplus L/A$ and the projection $A[1]\oplus L/A{\twoheadrightarrow}A[1]$, of the form $${D}=d_A^{\nabla^{\operatorname{Bott}}}+\sum_{k=2}^{\infty}{\mathcal{R}}_k ,$$ where
1. $d_{A}^\nabla$ denotes the Chevalley–Eilenberg differential associated to the Bott representation of the Lie algebroid $A$ on the vector bundle $L/A$;
2. ${\mathcal{R}}_2\in{\Gamma(A^\vee\otimes S^2(L/A)^\vee\otimes L/A)}$ denotes the symmetric Atiyah 1-cocycle associated to a choice of $L$-connection $\nabla$ on $L/A$ extending the Bott representation;
3. the terms ${\mathcal{R}}_k\in{\Gamma(A^\vee\otimes S^k(L/A)^\vee\otimes L/A)}$ with $k\geqslant 3$ are algebraic functions of ${\mathcal{R}}_2$, the torsion and curvature of $\nabla$, the curvature of a splitting of the short exact sequence $0\to A\to L\to L/A\to 0$, and their higher order covariant derivatives;
and, abusing notations, we identify the tensor ${\mathcal{R}}_k$ with the induced contraction operator. This Kapranov dg-manifold structure is natural up to a canonical isomorphism.
We follow a two-step strategy to prove Theorem \[Amsterdam\], which is also aimed at answering the second question raised in Problem \[Brussels\]. First, working within the realm of real Lie algebroids where Lie’s third theorem applies (at least locally), we unveil the geometric meaning of the Kapranov dg-manifold stemming from a Lie pair and translate our geometric intuition into purely algebraic formulas. Later, we show that the restriction to real Lie algebroids is unnecessary and that Theorem \[Amsterdam\] holds for Lie algebroids over any field ${\mathbb{K}}$ of characteristic zero.
Given a Lie pair $(L,A)$ over ${\mathbb{R}}$, Lie’s fundamental theorems assert the existence of a pair $({\mathscr{L}},{\mathscr{A}})$ of local Lie groupoids, which the Lie functor transforms into $L$ and $A$ respectively. Let us assume for the sake of simplicity that both ${\mathscr{L}}$ and ${\mathscr{A}}$ are simply $s$-connected, and ${\mathscr{A}}$ is a closed Lie subgroupoid of ${\mathscr{L}}$. By ${\mathfrak{X}}_{{\mathscr{A}}}$, we denote the differentiable stack presented by the Lie groupoid ${\mathscr{A}}$ [@MR2817778]. The groupoid ${\mathscr{A}}$ acts on ${\mathscr{L}}$ from both the right and the left. Consider the differentiable quotient stack $[{\mathscr{A}}\backslash{\mathscr{L}}/{\mathscr{A}}]$. It contains ${\mathfrak{X}}_{{\mathscr{A}}}$ as a smooth substack because the homogeneous space ${\mathscr{L}}/{\mathscr{A}}$ obviously contains the unit space $M$ of $({\mathscr{L}},{\mathscr{A}})$ as a distinguished embedded submanifold [@LWX:CMP]. Now the Lie groupoid ${\mathscr{A}}$ acts from the left on ${\mathscr{L}}/{\mathscr{A}}$. Therefore the differentiable quotient stack $[{\mathscr{A}}\backslash M]$ is a substack of the differentiable quotient stack $[{\mathscr{A}}\backslash({\mathscr{L}}/{\mathscr{A}})]$, the former being isomorphic to ${\mathfrak{X}}_{{\mathscr{A}}}$ and the latter isomorphic to $[{\mathscr{A}}\backslash{\mathscr{L}}/{\mathscr{A}}]$.
The Kapranov dg-manifold of Theorem \[Amsterdam\] is the derived formal neighborhood of ${\mathfrak{X}}_{{\mathscr{A}}}$ in $[{\mathscr{A}}\backslash{\mathscr{L}}/{\mathscr{A}}]$.
Although this conjecture remains open, we use it as the guideline of our construction of the Kapranov dg-manifold of Theorem \[Amsterdam\]. Here is a succinct two-step argumentation in favor of its plausibility. First, the sheaf of functions on the quotient stack $[{\mathscr{A}}\backslash({\mathscr{L}}/{\mathscr{A}})]$ should be *derived*, i.e. replaced by a differential graded commutative algebra, namely the Chevalley–Eilenberg cochain complex $$\label{Zagreb}
\cdots \to {\Gamma(\wedge^{k} A^\vee)}\otimes_R C^\infty({\mathscr{L}}/{\mathscr{A}})
{\xrightarrow{d_{A}}} {\Gamma(\wedge^{k+1} A^\vee)}\otimes_R C^\infty({\mathscr{L}}/{\mathscr{A}})
\to \cdots$$ stemming from the infinitesimal left action of $A$ on ${\mathscr{L}}/{\mathscr{A}}$, whose cohomology plays the role of the submodule of $A$-invariant functions on ${\mathscr{L}}/{\mathscr{A}}$. Next, we need to replace $C^\infty({\mathscr{L}}/{\mathscr{A}})$ by the functions on the *formal neighborhood* of $M$ in ${\mathscr{L}}/{\mathscr{A}}$. The source map $s:{\mathscr{L}}\to M$ factors through the projection ${\mathscr{L}}{\twoheadrightarrow}{\mathscr{L}}/{\mathscr{A}}$ and, consequently, turns the homogeneous space ${\mathscr{L}}/{\mathscr{A}}$ into a fiber bundle over $M$. Thus $s:{\mathscr{L}}/{\mathscr{A}}\to M$ and $\pi:L/A\to M$ are two fiber bundles over $M$. By definition, the functions on the formal neighborhood of $M$ in ${\mathscr{L}}/{\mathscr{A}}$ are the $s$-fiberwise $\infty$-jets of smooth functions on ${\mathscr{L}}/{\mathscr{A}}$ evaluated along the unit space $M$. Since the vector bundle $\pi:L/A\to M$ is the restriction to the submanifold $M$ of the tangent bundle to the $s$-fibers of ${\mathscr{L}}/{\mathscr{A}}$, $L/A$ may be regarded as the normal bundle of $M$ in ${\mathscr{L}}/{\mathscr{A}}$ and identified to a tubular neighborhood of $M$ in ${\mathscr{L}}/{\mathscr{A}}$. Therefore, the functions on the formal neighborhood of $M$ in ${\mathscr{L}}/{\mathscr{A}}$ may be identified with the functions on the formal neighborhood of the zero section of the vector bundle $\pi:L/A\to M$, i.e. with ${\Gamma(\hat{S}(L/A)^\vee)}$. Substituting ${\Gamma(\hat{S}(L/A)^\vee)}$ for $C^\infty({\mathscr{L}}/{\mathscr{A}})$ in the Chevalley-Eilenberg complex , we obtain the dgca $$\big({\Gamma(\wedge^\bullet A^\vee\otimes\hat{S}(L/A)^\vee)}, {D}\big) ,$$ whose differential ${D}$ ought to be the homological vector field on $A[1]\oplus L/A$ of Theorem \[Amsterdam\].
In order to obtain an explicit formula for the homological vector field, an explicit identification between the fiber bundles $L/A$ and ${\mathscr{L}}/{\mathscr{A}}$ is needed. We show that every choice of (i) a splitting $j$ of the short exact sequence of vector bundles $0\to A\to L\to L/A\to 0$ and (ii) an $L$-connection $\nabla$ on $L/A$ extending the Bott representation of $A$ on $L/A$ determines a geodesic flow and an exponential map $\exp^{\nabla,j}:L/A\to{\mathscr{L}}/{\mathscr{A}}$. When the splitting $j$ identifies $L/A$ with a Lie subalgebroid of $L$, say $B$, the homogeneous space ${\mathscr{L}}/{\mathscr{A}}$ gets identified locally with the corresponding Lie subgroupoid of ${\mathscr{L}}$, say ${\mathscr{B}}$, the $L$-connections $\nabla$ on $L/A$ extending the Bott representation are in one-one correspondence with the $B$-connections on $B$, and the resulting maps $\exp^\nabla:B\to{\mathscr{B}}$ are precisely the exponential maps investigated in [@Landsman; @NWX]. In particular, when the Lie algebroid $B$ is a Lie algebra $\mathfrak{g}$, a $B$-connection on $B$ is simply a linear map $\mathfrak{g}\otimes\mathfrak{g}\to\mathfrak{g}$. The classical Lie-theoretic exponential map $\exp:\mathfrak{g}\to G$ corresponds to the trivial linear map $\mathfrak{g}\otimes\mathfrak{g}{\xrightarrow{0}}\mathfrak{g}$. This is the content of Section \[Hungary\].
Although an exponential map $ \exp^\nabla:L/A\to{\mathscr{L}}/{\mathscr{A}}$ can be defined geometrically on a small neighborhood of the zero section of $L/A$, what we really need in the present paper is a *formal exponential map* identifying the formal neighborhood of the zero section of $L/A$ to the formal neighborhood of $M$ in ${\mathscr{L}}/{\mathscr{A}}$. We prove that this formal exponential map — the $\infty$-jet of $\exp^\nabla:L/A\to{\mathscr{L}}/{\mathscr{A}}$ in the direction of the fibers and along the zero section — admits an alternative and purely algebraic definition, which is valid for pairs of Lie algebroids over ${\mathbb{K}}$. In particular this algebraic definition makes sense for pairs of complex Lie algebroids despite the loss of the groupoid picture in that case. Indeed, we prove that the algebraic incarnation of the formal exponential map is a *Poincaré–Birkhoff–Witt isomorphism*, which we believe to be of independent interest.
At this point, it is useful to recall the relation between the classical exponential map and PBW isomorphism in Lie theory. Let $G$ be a Lie group and let ${\mathfrak{g}}$ be its Lie algebra. The algebra $D'_0({\mathfrak{g}})$ of distributions on ${\mathfrak{g}}$ with support $\{0\}$ is canonically identified with the symmetric algebra $S({\mathfrak{g}})$ while the algebra $D'_e(G)$ of distributions on $G$ with support $\{e\}$ is canonically identified with the universal enveloping algebra ${\mathcal{U}({\mathfrak{g}})}$. The classical Lie-theoretic exponential map $\exp:{\mathfrak{g}}\to G$, which is a local diffeomorphism near $0$, can be used to push forward distributions on the Lie algebra to distributions on the Lie group. The induced isomorphism $D'_0({\mathfrak{g}}){\xrightarrow{\simeq}} D'_e(G)$ is exactly the classical PBW isomorphism given by the symmetrization map $$X_1\odot\cdots\odot X_n\mapsto\frac{1}{n!}\sum_{\sigma\in S_n}X_{\sigma (1)}\cdots X_{\sigma (n)} ,$$ where $S_n$ denotes the symmetric group over $n$ objects.
The classical exponential map ${\mathfrak{g}}\to G$ results from the choice of the zero map ${\mathfrak{g}}\otimes{\mathfrak{g}}{\xrightarrow{0}}{\mathfrak{g}}$, and therefore the symmetrization map is only one of many different exotic PBW isomorphisms.
Let $\mathcal{N}(L/A)$ denote the space of all functions on $L/A$ which, together with their derivatives of all degrees in the direction of the $\pi$-fibers, vanish along the zero section. The space of $\pi$-fiberwise differential operators on $L/A$ along the zero section is canonically identified to the symmetric $R$-algebra ${\Gamma(S(L/A))}$. Therefore, we have the short exact sequence of $R$-algebras $$0 \to \mathcal{N}(L/A) \to {C^{\infty}(L/A)} \to \operatorname{Hom}_R\big({\Gamma(S(L/A))},R\big) \to 0 .$$
Likewise, let $\mathcal{N}({\mathscr{L}}/{\mathscr{A}})$ denote the space of all functions on ${\mathscr{L}}/{\mathscr{A}}$ which, together with their derivatives of all degrees in the direction of the $s$-fibers, vanish along the unit section. The space of $s$-fiberwise differential operators on ${\mathscr{L}}/{\mathscr{A}}$ along the unit section is canonically identified to the quotient of the enveloping algebra $\mathcal{U}(L)$ by the left ideal generated by ${\Gamma(A)}$. Therefore, we have the short exact sequence of $R$-algebras $$0 \to \mathcal{N}({\mathscr{L}}/{\mathscr{A}}) \to {C^{\infty}({\mathscr{L}}/{\mathscr{A}})} \to
\operatorname{Hom}_R\big(\tfrac{\mathcal{U}(L)}{\mathcal{U}(L){\Gamma(A)}},R\big) \to 0 .$$
Since the exponential $\exp^{\nabla,j}$ (or more precisely its dual) is an isomorphism from $\mathcal{N}({\mathscr{L}}/{\mathscr{A}})$ to $\mathcal{N}(L/A)$, it induces an isomorphism of $R$-modules $$\operatorname{Hom}_R\big(\tfrac{\mathcal{U}(L)}{\mathcal{U}(L){\Gamma(A)}},R\big)
{\xrightarrow{\cong}} \operatorname{Hom}_R\big({\Gamma(S(L/A))},R\big) .$$ The latter determines a map ${\Gamma(S(L/A))}\to \tfrac{\mathcal{U}(L)}{\mathcal{U}(L){\Gamma(A)}}$, which is exactly the [*PBW map*]{} we expected. In fact, we prove that such a PBW map can be defined algebraically by an iteration formula for any pair of Lie algebroids over ${\mathbb{K}}$:
\[Copenhagen\] Let $(L, A)$ be any pair of Lie algebroids over ${\mathbb{K}}$.
1. Each choice of a splitting $j: L/A\to L$ of the short exact sequence of vector bundles $0\to A\to L\to L/A\to 0$ and of an $L$-connection $\nabla$ on $L/A$ extending the Bott $A$-representation determines a unique isomorphism of $R$-coalgebras $\operatorname{pbw}^{\nabla,j}:{\Gamma(S(L/A))}\to\tfrac{\mathcal{U}(L)}{\mathcal{U}(L){\Gamma(A)}}$, called the Poincaré–Birkhoff–Witt map, which is also a morphism of filtered $R$-modules.
2. The map $\operatorname{pbw}^{\nabla,j}$ satisfies $$\begin{gathered}
\operatorname{pbw}^{\nabla,j}(1)=1 \\
\operatorname{pbw}^{\nabla,j}(b)=j(b) \\
\operatorname{pbw}^{\nabla,j}(b^{n+1})=j(b)\cdot\operatorname{pbw}^{\nabla,j}(b^n)-\operatorname{pbw}^{\nabla,j}\big(\nabla_{j(b)} (b^n)\big) \end{gathered}$$ for all $b\in{\Gamma(L/A)}$, and $n\in{\mathbb{N}}$. Here $b^k$ stands for the symmetric product $b\odot b\odot \cdots \odot b$ of $k$ copies of $b$.
When $L=T_M$ and $A$ is the trivial Lie subalgebroid of $L$ of rank 0, the $\operatorname{pbw}^{\nabla,j}$ map above is the inverse of the so called ‘complete symbol map,’ which is an isomorphism from the space ${\mathcal{U}(T_M)}$ of differential operators on $M$ to the space ${\Gamma(S(T_M))}$ of fiberwisely polynomial functions on $T{^{\vee}}_M$. The complete symbol map was generalized to arbitrary Lie algebroids over ${\mathbb{R}}$ by Nistor, Weinstein, and one of the authors [@NWX]. It played an important role in quantization theory [@MR706215; @NWX]. However, to the best of our knowledge, even in these classical cases, the iteration formula in Theorem \[Copenhagen\] seems to be new and may have interesting applications.
The $\operatorname{pbw}$-isomorphism allows us to establish an explicit isomorphism of algebras $\operatorname{Hom}_R\big(\tfrac{\mathcal{U}(L)}{\mathcal{U}(L){\Gamma(A)}},R\big)
\simeq \operatorname{Hom}_R\big({\Gamma(S(L/A))},R\big)\cong {\Gamma( \hat{S} (L/A)^\vee)}$. When $(L, A)$ is a pair of Lie algebroids over ${\mathbb{R}}$, $\operatorname{Hom}_R\big(\tfrac{\mathcal{U}(L)}{\mathcal{U}(L){\Gamma(A)}},R\big)$ is exactly the space of functions on the formal neighborhood of $M$ in ${\mathscr{L}}/{\mathscr{A}}$. Theorem \[Amsterdam\] can thus be proved directly using the iteration formula defining the $\operatorname{pbw}$-map. This is discussed in Section \[Paris\].
A Kapranov dg-manifold $A[1]\oplus E$ is said to be [*linearizable*]{} if it is isomorphic to the linear Kapranov dg-manifold associated to an $A$-module structure on $E$.
A natural question is the following
When is the Kapranov dg-manifold associated to a Lie pair as in Theorem \[Amsterdam\] linearizable?
Our next result completely answers this question.
\[Dublin\] Let $(L,A)$ be a Lie pair over a smooth manifold $M$. The following assertions are equivalent.
1. \[Vavin\] The Kapranov dg-manifold stemming from the Lie pair $(L,A)$ (see Theorem \[Ohio\]) is linearizable.
2. The Atiyah class $\alpha_{L/A}\in H_{\operatorname{CE}}^1(A;S^2(L/A){^{\vee}}\otimes L/A)$ of the Lie pair $(L,A)$ vanishes.
3. For every splitting $j:L/A\to L$, there exists an extension of the Bott $A$-connection on $L/A$ to an $L$-connection $\nabla$ on $L/A$ such that the associated Poincaré–Birkhoff–Witt map $\operatorname{pbw}^{\nabla,j}$ intertwines the infinitesimal $A$-actions on ${\Gamma(S(L/A))}$ and $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$.
In the case of Lie algebroids over ${\mathbb{R}}$, the following two assertions may be added to the list above.
1. For every splitting $j:L/A\to L$, there exists an extension of the Bott $A$-connection on $L/A$ to an $L$-connection $\nabla$ on $L/A$ such that the fiberwise $\infty$-jet of the associated exponential map $\exp^{\nabla,j}$ along the zero section of $L/A\to M$ intertwines the infinitesimal actions of $A$ on $L/A$ and ${\mathscr{L}}/{\mathscr{A}}$.
2. \[Duroc\] There exists a fiber-preserving diffeomorphism $\phi$ from a neighborhood of the zero section in $L/A$ to a neigborhood of the identity section of ${\mathscr{L}}/{\mathscr{A}}$, which fixes the submanifold $M$, and whose fiberwise $\infty$-jet along $M$ intertwines the infinitesimal actions of $A$ on $L/A$ and ${\mathscr{L}}/{\mathscr{A}}$.
In fact, a stronger version of the equivalence of (\[Vavin\]) and (\[Duroc\]) in Theorem \[Dublin\] is proved by Voglaire and one of the authors [@Gengoux-Voglaire] extending a theorem of Bordemann [@MR3014184].
As an application, we consider the Lie pair $(L=T_X\otimes {\mathbb{C}},A=T^{0, 1}_X)$ associated to a complex manifold $X$. We obtain a Kapranov dg-manifold on $T^{0,1}_X[1]\oplus T^{1, 0}_X$, whose homological vector field involves the Atiyah cocycle as the second order term. When $X$ is a Kähler manifold, we recover the standard example of Kapranov [@Kapranov]. As is standard [@MR1839580], this Kapranov dg manifold describes the formal neighborhood of the diagonal of $X\times X$. As a consequence of Theorem \[Dublin\], we prove that the Kapranov dg-manifold $T^{0,1}_X[1]\oplus T^{1, 0}_X$ is linearizable if and only if the Atiyah class of $X$ vanishes, which is specific to the diagonal embedding.
We conclude with a few remarks on the relations between the present paper and the exisiting literature. Kapranov dg-manifolds are closely related to Costello’s $L_\infty$-spaces [@arXiv:1112.0816; @arXiv:1404.5426]; the precise relation will be studied somewhere else. However, a Kapranov dg-manifold $A[1]\oplus E$, when considered as a vector bundle over $A[1]$, is not a $Q$-bundle in the sense of Kotov–Strobl [@MR3293862] since the dg-structure is not compatible with the linear structure. After [@MR2989383] appeared, we learned from Bressler [@Bressler] that he had also obtained a similar construction. Finally, we would like to point out that this work is related to [@MR3217749]. However, we warn the reader that ‘PBW isomorphism’ does not have the same meaning here and in [@MR3217749]. In the present paper, the PBW map is an explicit $R$-linear isomorphism which always exists once a connection has been chosen, just as is the case for (the inverse of) the complete symbol map, while the PBW map in [@MR3217749] is an isomorphism of filtered ${\mathcal{U}(A)}$-modules which exists only when the Atiyah class vanishes. Hence existence of the PBW map in [@MR3217749] is in a certain sense equivalent to linearizability of $A[1]\oplus L/A$ (compare [@MR3217749 Theorem 5.4] and Theorem \[Dublin\]).
Terminology and notations {#terminology-and-notations .unnumbered}
-------------------------
### Shuffles {#shuffles .unnumbered}
A $(p,q)$-shuffle is a permutation $\sigma$ of the set $\{1,2,\cdots,p+q\}$ such that $\sigma(1)<\sigma(2)<\cdots<\sigma(p)$ and $\sigma(p+1)<\sigma(p+2)<\cdots<\sigma(p+q)$. The symbol ${\mathfrak{S}_{p}^{q}}$ denotes the set of $(p,q)$-shuffles.
### Graduation shift {#graduation-shift .unnumbered}
Given a graded vector space $V=\bigoplus_{k\in{\mathbb{Z}}}V^{(k)}$, $V[i]$ denotes the graded vector space obtained by shifting the grading on $V$ according to the rule $(V[i])^{(k)}=V^{(i+k)}$. Accordingly, if $E=\bigoplus_{k\in{\mathbb{Z}}}E^{(k)}$ is a graded vector bundle over $M$, $E[i]$ denotes the graded vector bundle obtained by shifting the degree in the fibers of $E$ according to the above rule.
### Koszul sign {#koszul-sign .unnumbered}
The Koszul sign ${\operatorname{sgn}(\sigma; v_1, \cdots, v_{n})}$ of a permutation $\sigma$ of homogeneous vectors $v_1,v_2,\dots,v_n$ of a ${\mathbb{Z}}$-graded vector space $V=\bigoplus_{n\in{\mathbb{Z}}}V_n$ is determined by the equality $$v_{\sigma(1)}\odot v_{\sigma(2)}\odot\cdots\odot v_{\sigma(n)}
= {\operatorname{sgn}(\sigma; v_1, \cdots, v_n)} \cdot v_1\odot v_2\odot\cdots\odot v_n$$ in the graded commutative algebra $S(V)$.
### $L_\infty$ algebra {#l_infty-algebra .unnumbered}
An $L_\infty[1]$-algebra [@Lada_Stasheff; @MR1183483; @Cattaneo-Schaetz] is a ${\mathbb{Z}}$-graded vector space $V=\bigoplus_{n\in{\mathbb{Z}}}V_n$ endowed with a sequence $(\lambda_k)_{k=1}^\infty$ of linear maps $\lambda_k: S^k(V)\to V[1]$ satisfying the generalized Jacobi identities $$\sum_{p+q=n}\sum_{\sigma\in{\mathfrak{S}_{p}^{q}}}
{\operatorname{sgn}(\sigma;v_1,\cdots,v_n)}\
\lambda_{1+q}\big(\lambda_p(v_{\sigma(1)},\cdots,v_{\sigma(p)}),
v_{\sigma(p+1)},\cdots,v_{n}\big)=0$$ for each $n\in{\mathbb{N}}$ and for all homogeneous vectors $v_1,v_2,\dots,v_n\in V$. A ${\mathbb{Z}}$-graded vector space $V$ is an $L_\infty$-algebra if and only if $V[1]$ is an $L_\infty[1]$-algebra.
### Lie algebroid {#lie-algebroid .unnumbered}
We use the symbol ${\mathbb{K}}$ to denote either of the fields ${\mathbb{R}}$ and ${\mathbb{C}}$. A *Lie algebroid* over ${\mathbb{K}}$ is a ${\mathbb{K}}$-vector bundle $L\to M$ together with a bundle map $\rho:L\to TM\otimes_{\mathbb{R}}{\mathbb{K}}$ called the *anchor*, and a bracket $[\cdot,\cdot]$ on sections of $L$, such that $$[X,fY]=f[X,Y]+\big(\rho(X)f\big)Y$$ for all $X,Y\in\Gamma(L)$ and $f\in C^\infty(M,{\mathbb{K}})$. In that case $\rho$ seen as a map on sections of $L$ is a morphism of Lie algebras. Unless specified, in this paper, by a Lie algebroid, we always mean a Lie algebroid over ${\mathbb{K}}$.
### Lie pair {#lie-pair .unnumbered}
By a *Lie pair* $(L,A)$, we mean an inclusion $A\hookrightarrow L$ of Lie algebroids over a smooth manifold $M$.
### Tensor products {#tensor-products .unnumbered}
The symbol $R$ always denotes the algebra of smooth functions on $M$ with values in ${\mathbb{K}}$. For any two $R$-modules $P$ and $Q$, we write $P\otimes_{R} Q$ to denote the tensor product of $P$ and $Q$ as $R$-modules and $P\otimes Q$ to denote the tensor product of $P$ and $Q$ regarded as ${\mathbb{K}}$-modules.
Acknowledgments {#acknowledgments .unnumbered}
---------------
We would like to thank several institutions for their hospitality, which allowed the completion of this project: Penn State University (Laurent-Gengoux), Université de Lorraine (Stiénon), and Université Paris Diderot (Xu). We also wish to thank Paul Bressler, Yannick Voglaire, and Jim Stasheff for useful discussions and comments.
Poincaré–Birkhoff–Witt isomorphisms
===================================
Connections and representations for Lie algebroids {#Edmonton}
--------------------------------------------------
Let $M$ be a smooth manifold, let $L\to M$ be a Lie ${\mathbb{K}}$-algebroid with anchor map ${\rho}:L\to T_M\otimes_{{\mathbb{R}}}{\mathbb{K}}$, and let $E{\xrightarrow{\varpi}}M$ be a vector bundle over ${\mathbb{K}}$. The algebra of smooth functions on $M$ with values in ${\mathbb{K}}$ will be denoted $R$.
The traditional description of a (linear) $L$-connection on $E$ is in terms of a *covariant derivative* $${\Gamma(L)}\times{\Gamma(E)}\to{\Gamma(E)}: (x,e)\mapsto \nabla_x e$$ characterized by the following two properties: $$\begin{gathered}
\nabla_{f\cdot x} e=f\cdot \nabla_x e , \label{Faro} \\
\nabla_x (f\cdot e)=\rho(x)f\cdot e+f\cdot\nabla_x e \label{Glasgow}
,\end{gathered}$$ for all $x\in{\Gamma(L)}$, $e\in{\Gamma(E)}$, and $f\in R$.
A *representation of a Lie algebroid* $A$ on a vector bundle $E\to M$ is a flat $A$-connection $\nabla$ on $E$, i.e. a covariant derivative $\nabla:{\Gamma(A)}\times{\Gamma(E)}\to{\Gamma(E)}$ satisfying $$\label{Helsinki}
\nabla_{a_1}\nabla_{a_2} e-\nabla_{a_2}\nabla_{a_1} e=\nabla_{{[a_1,a_2]}}e ,$$ for all $a_1,a_2\in{\Gamma(A)}$ and $e\in{\Gamma(E)}$. A vector bundle endowed with a representation of the Lie algebroid $A$ is called an *$A$-module*. More generally, given a left $R$-module ${\mathcal{M}}$, by an [*infinitesimal action*]{} of $A$ on ${\mathcal{M}}$, we mean a ${\mathbb{K}}$-bilinear map $\nabla:{\Gamma(A)}\times{\mathcal{M}}\to{\mathcal{M}}$, $(a,e)\mapsto\nabla_a e$ satisfying Equations. , , and . In other words, $\nabla$ is a representation of the Lie–Rinehart algebra $({\Gamma(A)},R)$ [@Rinehart].
\[Istambul\] Let $(L,A)$ be a Lie pair, i.e. an inclusion $A{\hookrightarrow}L$ of Lie algebroids. The *Bott representation* of $A$ on the quotient $L/A$ is the flat connection defined by $$\nabla^{\operatorname{Bott}}_a q(l)=q\big({[a,l]}\big), \quad\forall a\in{\Gamma(A)},l\in{\Gamma(L)} ,$$ where $q$ denotes the canonical projection $L{\twoheadrightarrow}L/A$. Thus the quotient $L/A$ of a Lie pair $(L,A)$ is an $A$-module.
Universal enveloping algebra of a Lie algebroid
-----------------------------------------------
Let $L$ be a Lie ${\mathbb{K}}$-algebroid over a smooth manifold $M$ and let $R$ denote the algebra of smooth functions on $M$ taking values in ${\mathbb{K}}$. The vector space $\mathfrak{g}:=R\oplus{\Gamma(L)}$ admits a natural Lie algebra structure given by the Lie bracket $$(f+X)\otimes(g+Y) \longmapsto X(g)-Y(f)+{\left[ X , Y \right]} ,$$ where $f,g\in R$ and $X,Y\in{\Gamma(L)}$. Its universal enveloping algebra ${\mathcal{U}(\mathfrak{g})}$ is the quotient of the tensor algebra $T(\mathfrak{g})=\bigoplus_{k=0}^{\infty}
\big(\bigotimes_{{\mathbb{K}}}^k (R\oplus{\Gamma(L)})\big)$ by the ideal generated by the subset of all elements of the form $$(f+X)\otimes(g+Y)-(g+Y)\otimes(f+X)-\big(X(g)-Y(f)+{\left[ X , Y \right]}\big)$$ with $f,g\in R$ and $X,Y\in{\Gamma(L)}$. Let $i$ denote the natural inclusion of $\mathfrak{g}$ into ${\mathcal{U}(\mathfrak{g})}$ and let $\mathcal{V}(\mathfrak{g})$ denote the subalgebra of ${\mathcal{U}(\mathfrak{g})}$ generated by $i(\mathfrak{g})$. The *universal enveloping algebra ${\mathcal{U}(L)}$ of the Lie algebroid $L$* is the quotient of $\mathcal{V}(\mathfrak{g})$ by the two-sided ideal generated by the elements of the form $$i(f)\otimes i(g+Y)-i(fg+fY)$$ with $f,g\in R$ and $Y\in{\Gamma(L)}$. Note that we have implicitly used the left $R$-module structure of $\mathfrak{g}$. The graduation of $T(\mathfrak{g})$ induces a natural ascending filtration $$\label{Jakarta}
\cdots {\hookrightarrow}\mathcal{U}^{\leqslant n-1}(L) {\hookrightarrow}\mathcal{U}^{\leqslant n}(L) {\hookrightarrow}\mathcal{U}^{\leqslant n+1}(L) {\hookrightarrow}\cdots$$ on ${\mathcal{U}(L)}$.
When the base $M$ of the Lie algebroid $L$ is the one-point space so that the only fiber is a Lie algebra $\mathfrak{h}$, the universal enveloping algebra of the Lie algebroid is the the universal enveloping algebra of the Lie algebra $\mathfrak{h}$.
When the Lie algebroid $L$ is the tangent bundle $T_M\to M$, its universal enveloping algebra ${\mathcal{U}(L)}$ is the algebra of differential operators on $M$.
In general, the universal enveloping algebra ${\mathcal{U}(L)}$ of the Lie algebroid $L$ associated to a Lie groupoid ${\mathscr{L}}$ is canonically identified with the associative algebra of source-fiberwise differential operators on ${C^{\infty}({\mathscr{L}})}$ invariant under left translations [@Weinstein:book].
The universal enveloping algebra ${\mathcal{U}(L)}$ of the Lie algebroid $L\to M$ is a coalgebra over $R$ [@Xu:quantumgroupoids]. Its comultiplication $$\Delta:{\mathcal{U}(L)}\to{\mathcal{U}(L)}\otimes_R{\mathcal{U}(L)}$$ is compatible with its filtration and characterized by the identities $$\begin{gathered}
\Delta(1)=1\otimes 1; \\
\Delta(x)=1\otimes x+x\otimes 1, \quad \forall x\in {\Gamma(L)}; \\
\Delta(u\cdot v)=\Delta(u)\cdot\Delta(v), \quad \forall u,v\in{\mathcal{U}(L)} ,\end{gathered}$$ where $1\in R$ denotes the constant function on $M$ with value $1$ while the symbol $\cdot$ denotes the multiplication in ${\mathcal{U}(L)}$. We refer the reader to [@Xu:quantumgroupoids] for the precise meaning of the last equation above. Explicitly, we have $$\begin{gathered}
\Delta(b_1\cdot b_2\cdot\cdots\cdot b_n)=
1\otimes(b_1\cdot b_2\cdot\cdots\cdot b_n) \\
+ \sum_{\substack{p+q=n \\ p,q\in{\mathbb{N}}}}\sum_{\sigma\in{\mathfrak{S}_{p}^{q}}} (b_{\sigma(1)}\cdot\cdots\cdot b_{\sigma(p)}) \otimes
(b_{\sigma(p+1)}\cdot\cdots\cdot b_{\sigma(n)}) \\
+ (b_1\cdot b_2\cdot\cdots\cdot b_n)\otimes 1
,\end{gathered}$$ for all $b_1,\dots,b_n\in{\Gamma(L)}$.
The coalgebras ${\Gamma(S(L/A))}$ and $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$
--------------------------------------------------------------------------------------------
Let $(L,A)$ be a pair of Lie algebroids over ${\mathbb{K}}$.
Writing ${\mathcal{U}(L)}{\Gamma(A)}$ for the left ideal of ${\mathcal{U}(L)}$ generated by ${\Gamma(A)}$, the quotient $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ is automatically a filtered $R$-coalgebra since $$\Delta\big({\mathcal{U}(L)}{\Gamma(A)}\big)\subseteq {\mathcal{U}(L)}\otimes_R \big({\mathcal{U}(L)}{\Gamma(A)}\big) + \big({\mathcal{U}(L)}{\Gamma(A)}\big) \otimes_R {\mathcal{U}(L)}$$ and the filtration on ${\mathcal{U}(L)}$ descends to a filtration $$\label{Kolkata}
\cdots {\hookrightarrow}\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}\right)^{\leqslant n-1} {\hookrightarrow}\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}\right)^{\leqslant n} {\hookrightarrow}\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}\right)^{\leqslant n+1} {\hookrightarrow}\cdots$$ of $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$.
Likewise, deconcatenation defines a graded $R$-coalgebra structure on ${\Gamma(S(L/A))}$. The comultiplication $$\Delta:{\Gamma(S(L/A))}\to{\Gamma(S(L/A))}\otimes_R{\Gamma(S(L/A))}$$ is given by $$\begin{gathered}
\label{London}
\Delta(b_1\odot b_2\odot\cdots\odot b_n)=
1\otimes(b_1\odot b_2\odot\cdots\odot b_n) \\
+ \sum_{\substack{p+q=n \\ p,q\in{\mathbb{N}}}}\sum_{\sigma\in{\mathfrak{S}_{p}^{q}}} (b_{\sigma(1)}\odot\cdots\odot b_{\sigma(p)}) \otimes (b_{\sigma(p+1)}\odot\cdots\odot b_{\sigma(n)}) \\
+ (b_1\odot b_2\odot\cdots\odot b_n)\otimes 1
,\end{gathered}$$ for all $b_1,\dots,b_n\in{\Gamma(L/A)}$. The symbol $\odot$ denotes the symmetric product in ${\Gamma(S(L/A))}$.
Poincaré–Birkhoff–Witt isomorphism {#Madrid}
----------------------------------
Our first main result is a generalization of the classical Poincaré–Birkhoff–Witt theorem to Lie pairs.
\[Nairobi\] Let $(L,A)$ be a Lie pair. Given a splitting $j:L/A\to L$ of the short exact sequence $0\to A \to L \to L/A \to 0$ and a $L$-connection $\nabla$ on $L/A$, there exists a unique isomorphism of filtered $R$-coalgebras $$\operatorname{pbw}^{\nabla,j}:{\Gamma(S(L/A))}\to\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$$ satisfying $$\begin{gathered}
\operatorname{pbw}^{\nabla,j}(f)=f, \quad\forall f\in R; \label{Ottawa} \\
\operatorname{pbw}^{\nabla,j}(b)=j(b), \quad\forall b\in{\Gamma(L/A)}; \label{Pittsburgh} \end{gathered}$$ and, for all $n\in{\mathbb{N}}$ and $b\in{\Gamma(L/A)}$, $$\operatorname{pbw}^{\nabla,j}(b^{n+1})=j(b)\cdot\operatorname{pbw}^{\nabla,j}(b^n)-\operatorname{pbw}^{\nabla,j}\big(\nabla_{j(b)}(b^n)\big) \label{Quito}
.$$
Equation is equivalent to $$\begin{gathered}
\label{Rome}
\operatorname{pbw}^{\nabla,j}(b_0 \odot \cdots \odot b_n) \\ =\tfrac{1}{n+1}\sum_{i=0}^{n} \Big( j(b_i) \cdot
\operatorname{pbw}^{\nabla,j}(b_0 \odot \cdots \odot \widehat{b_i} \odot \cdots \odot b_n ) \\ -
\operatorname{pbw}^{\nabla,j}\big(\nabla_{j(b_i)} ( b_0 \odot \cdots \odot \widehat{b_i} \odot \cdots \odot b_n ) \big) \Big) \end{gathered}$$ for all $b_0,\dots,b_n \in {\Gamma(L/A)}$.
It is immediate that Equations , , and together define inductively a unique $R$-linear map $\operatorname{pbw}^{\nabla,j}$. The proof that this Poincaré–Birkhoff–Witt map $\operatorname{pbw}^{\nabla,j}$ is an isomorphism of filtered $R$-coalgebras will only be sketched here because a more enlightening proof exploiting the close relation between $\operatorname{pbw}^{\nabla,j}$ and exponential maps will be given later.
We introduce the functor $\operatorname{Gr}$ which takes a filtered vector space $$\cdots\subset\mathscr{A}^{\leqslant k-1}\subset\mathscr{A}^{\leqslant k}\subset\mathscr{A}^{\leqslant k+1}\subset\cdots$$ to the associated graded vector space $$\operatorname{Gr}\big(\mathscr{A}\big)=\bigoplus_{k} \frac{\mathscr{A}^{\leqslant k}}{\mathscr{A}^{\leqslant k-1}} .$$ Rinehart proved that, for every Lie algebroid $L$, the symmetrization map $$\operatorname{sym}:{\Gamma(S(L))}\to\operatorname{Gr}\big({\mathcal{U}(L)}\big) ,$$ defined by $$l_1\odot \cdots\odot l_n \mapsto \frac{1}{n!}\sum_{\sigma\in S_n}l_{\sigma(1)}\cdots l_{\sigma(n)}
,\quad\forall l_1,\dots,l_n\in{\Gamma(L)} ,$$ is an isomorphism of graded vector spaces [@Rinehart].
\[Stockholm\] For all $n\in{\mathbb{N}}$ and $b_1,\dots,b_n\in{\Gamma(L/A)}$, $$\operatorname{pbw}^{\nabla,j}(b_1\odot\cdots\odot b_n)-\frac{1}{n!}\sum_{\sigma\in S_n}
j(b_{\sigma(1)})\cdot j(b_{\sigma(2)})\cdot \cdots \cdot j(b_{\sigma(n)})$$ is an element of $\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}\right)^{\leqslant n-1}$.
It follows from the inductive relation that $$\begin{gathered}
\operatorname{pbw}^{\nabla,j}(b_1\odot\cdots\odot b_n)-\frac{1}{n}\sum_{k=1}^n j(b_k)\cdot\operatorname{pbw}^{\nabla,j}(b_1\odot\cdots
\odot\widehat{b_k}\odot\cdots\odot b_n) \\
=-\frac{1}{n}\sum_{k=1}^n \operatorname{pbw}^{\nabla,j}\big(\nabla_{j(b_k)}(b_1\odot\cdots
\odot\widehat{b_k}\odot\cdots\odot b_n)\big) \end{gathered}$$ belongs to $\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}\right)^{\leqslant n-1}$ as $$\nabla_{j(b_k)}(b_1\odot\cdots\odot\widehat{b_k}\odot\cdots\odot b_n)\in {\Gamma(S^{n-1}(L/A))}$$ and $\operatorname{pbw}^{\nabla,j}$ respects the filtrations. The result follows by induction on $n$.
Consider the canonical projection $$p:{\mathcal{U}(L)}\to\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$$ and the inclusion $$j:{\Gamma(S(L/A))}\to{\Gamma(S(L))}$$ induced by the splitting $j$.
Lemma \[Stockholm\] asserts the commutativity of the diagram $$\begin{tikzcd}[column sep=huge]
{\Gamma(S(L))} \arrow{r}{\operatorname{sym}} & \operatorname{Gr}\big({\mathcal{U}(L)}\big) \arrow[two heads]{d}{\operatorname{Gr}(p)} \\
{\Gamma(S(L/A))} \arrow[hook]{u}{j} \arrow{r}[swap]{\operatorname{Gr}(\operatorname{pbw}^{\nabla,j})} & \operatorname{Gr}\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}\right)
\end{tikzcd} .$$
The functor $\operatorname{Gr}$ has the following remarkable property: a morphism $\phi$ of filtered vector spaces is invertible if and only if the associated morphism of graded vector spaces $\operatorname{Gr}(\phi)$ is invertible. Consequently, in order to prove Theorem \[Nairobi\], it suffices to prove that $\operatorname{Gr}(\operatorname{pbw}^{\nabla,j})$ is an isomorphism of graded vector spaces. Since $\operatorname{Gr}(\operatorname{pbw}^{\nabla,j})=\operatorname{Gr}(p)\circ\operatorname{sym}\circ j$ where $j$ is injective, $\operatorname{sym}$ is bijective, and $\operatorname{Gr}(p)$ is surjective, we simply need to check that $\operatorname{sym}$ induces an isomorphism between the cokernel of $j:{\Gamma(S^n(L/A))}\to{\Gamma(S^n(L))}$ and the kernel of $\operatorname{Gr}^n(p)$.
The cokernel of $j:{\Gamma(S^n(L/A))}\to{\Gamma(S^n(L))}$ is clearly isomorphic to the subspace ${\Gamma(S^{n-1}(L)\otimes A)}$ of ${\Gamma(S^n(L))}$. Moreover, $\operatorname{Gr}^n\big({\mathcal{U}(L)}\big)=\frac{\mathcal{U}^{\leqslant n}(L)}
{\mathcal{U}^{\leqslant n-1}(L)}$ and $$\operatorname{Gr}^n\left(\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}\cdot{\Gamma(A)}}\right)=
\frac{\mathcal{U}^{\leqslant n}(L)}{\mathcal{U}^{\leqslant n-1}(L)\cdot{\Gamma(A)}
+\mathcal{U}^{\leqslant n-1}(L)} ,$$ imply that $$\ker\big(\operatorname{Gr}^n(p)\big)=\frac{\mathcal{U}^{\leqslant n-1}(L)\cdot{\Gamma(A)}+\mathcal{U}^{\leqslant n-1}(L)}{\mathcal{U}^{\leqslant n-1}(L)} .$$
\[Tokyo\] For all $a\in{\Gamma(A)}$ and $b\in{\Gamma(L)}$, we have $$\operatorname{sym}(\underset{n\text{ factors}}{\underbrace{b\odot\cdots\odot b}}\odot a) = \underset{n\text{ factors}}{\underbrace{b\cdot\cdots\cdot b}}\cdot a \quad\text{in }
\operatorname{Gr}^{n+1}\big(\mathcal{U}(L)\big)=\frac{\mathcal{U}^{\leqslant n+1}(L)}{\mathcal{U}^{\leqslant n}(L)} .$$
First, observe that, for every $k\in\{0,1,2,\dots,n\}$, $$b^k ab^{n-k}-b^n a=b^k{[a,b]}b^{n-k-1}+ b^{k+1}{[a,b]}b^{n-k-2}+\cdots+b^{n-1}{[a,b]}\in
\mathcal{U}^{\leqslant n}(L) .$$ Then, in $\operatorname{Gr}^{n+1}\big(\mathcal{U}(L)\big)$, we have $$\begin{gathered}
\operatorname{sym}(b^n\odot a)-b^n\cdot a \\
=\frac{1}{n+1}(a b^n+b a b^{n-1}+\cdots+b^{n-1}ab+b^n a)-b^n a \\
=\frac{1}{n+1}\sum_{k=0}^n \big(b^k a b^{n-k}-b^n a\big)
=0 .\end{gathered}$$
For every $n\in{\mathbb{N}}$, the graded linear isomorphism $$\operatorname{sym}:{\Gamma(S(L))}\to\operatorname{Gr}\big({\mathcal{U}(L)}\big)$$ maps the subspace ${\Gamma(S^{n-1}(L)\odot A)}$ of ${\Gamma(S^n(L))}$ isomorphically onto the subspace $\frac{\mathcal{U}^{\leqslant n-1}(L)\cdot{\Gamma(A)}+\mathcal{U}^{\leqslant n-1}(L)}
{\mathcal{U}^{\leqslant n-1}(L)}$ of $\operatorname{Gr}^n\big({\mathcal{U}(L)}\big)$.
Since $\operatorname{sym}:{\Gamma(S(L))}\to\operatorname{Gr}\big({\mathcal{U}(L)}\big)$ is an isomorphism of graded $R$-modules, $\mathcal{U}^{\leqslant n-1}(L)$ is spanned by its elements of the form $b^k$ with $b\in{\Gamma(L)}$ and $k\in\{0,1,2,\cdots,n-1\}$. Consequently, $\frac{\mathcal{U}^{\leqslant n-1}(L)\cdot{\Gamma(A)}+\mathcal{U}^{\leqslant n-1}(L)}
{\mathcal{U}^{\leqslant n-1}(L)}$ is spanned by its elements of the form $b^{n-1}\cdot a$ with $b\in{\Gamma(L)}$ and $a\in{\Gamma(A)}$. By Lemma \[Tokyo\], $\operatorname{sym}$ induces an isomorphism of $R$-modules $${\Gamma(S^{n-1}(L)\odot A)} {\xrightarrow{\operatorname{sym}}}
\frac{\mathcal{U}^{\leqslant n-1}(L)\cdot{\Gamma(A)}+\mathcal{U}^{\leqslant n-1}(L)}
{\mathcal{U}^{\leqslant n-1}(L)}$$ mapping $b^{n-1}\odot a$ to $b^{n-1}\cdot a$.
Exponential maps {#Urumqi}
================
Troughout this section, we work over the field ${\mathbb{R}}$ of real numbers exclusively. Thus we restrict ourselves to real Lie algebroids.
For any Lie group, there is a canonical exponential map from the associated Lie algebra to the Lie group itself. The construction of the exponential map generalizes to Lie groupoids though in a noncanonical way as a choice of connection is needed (see Landsman [@Landsman]). In the particular case of a Lie group ${\mathscr{G}}$, one of these connection-induced exponential maps stands out: the one associated to the ‘zero’ connection defined by $ \nabla_X Y =0$ for every left-invariant vector fields $X$ and $Y$ on ${\mathscr{G}}$.
Here, given an inclusion of Lie groupoids ${\mathscr{A}}{\hookrightarrow}{\mathscr{L}}$ (and the corresponding inclusion of Lie algebroids $A{\hookrightarrow}L$), we define an exponential map which takes (a neighborhood of the zero section of) $L/A$ to (a neighborhood of the ‘identity’ section of) ${\mathscr{L}}/{\mathscr{A}}$. The definition requires two choices: (1) a splitting of the short exact sequence of vector bundles $$0\to A \to L \to L/A \to 0$$ and (2) an $L$-connection on $L/A$. When the Lie subalgebroid $A$ is a vector bundle of rank zero, we recover Landsman’s exponential map [@Landsman].
${s}$- and ${\rho}$-paths
-------------------------
Let ${\mathscr{L}}$ be a Lie groupoid over a manifold $M$ with source ${s}:{\mathscr{L}}\to M$, target ${t}:{\mathscr{L}}\to M$, and identity $1:M\to{\mathscr{L}}$. The subbundle $L=\{ X \in 1^* (T_{{\mathscr{L}}}) | s_* X=0 \}$ of the pullback of $T_{{\mathscr{L}}}$ via $1:M\to{\mathscr{L}}$ carries a Lie algebroid structure whose anchor ${\rho}:L\to T_M$ is the (restriction of) the differential ${t}_*:T_{{\mathscr{L}}}\to T_M$ of the target [@Mackenzie:book]. We write $\pi$ to denote the bundle projection $L\to M$.
In what follows, the symbol $I$ always denotes some open interval of the real line containing $0$.
An *${s}$-path* is a smooth curve $\gamma:I\to{\mathscr{L}}$ originating from a point of the unit submanifold of ${\mathscr{L}}$ (i.e. $\gamma(0)=1_m$ for some $m\in M$) and fully contained in one of the ${s}$-fibers (i.e. ${s}\circ\gamma(t)=m$ for all $t\in I$).
A *${\rho}$-path* is a smooth curve $\beta:I\to L$ satisfying $$\pi_*\big(\beta'(t)\big)={\rho}\big(\beta(t)\big), \quad\forall t\in I .$$
\[Vienna\] Every ${s}$-path $\gamma$ determines a unique ${\rho}$-path $\beta$ through the relation $$\beta(t)=\left.\frac{d}{d\tau} \big(\gamma(t)\big)^{-1}\gamma(\tau) \right|_t .$$ Conversely, every ${\rho}$-path $\beta$ determines a unique ${s}$-path $\gamma$. The origins of the two paths satisfy the relation $\gamma(0)=1_{\pi(\beta(0))}$.
Connections as horizontal liftings
----------------------------------
Let $M$ be a smooth manifold, let $L\to M$ be a Lie algebroid with anchor map ${\rho}$, and let $E{\xrightarrow{\varpi}}M$ be a vector bundle.
In Section \[Edmonton\], we recalled the traditional definition of $L$-connections on $E$ in terms of covariant derivatives ${\Gamma(L)}\times{\Gamma(E)}{\xrightarrow{\nabla}}{\Gamma(E)}$. However, $L$-connections on $E$ admit the following alternate description:
A (linear) $L$-connection on $E$ is a map $L\times_M E{\xrightarrow{h}} T_E$, called *horizontal lifting*, such that the diagram $$\label{Warsaw}
\begin{tikzcd}[column sep=small] & & L \arrow{rr}{\rho} \arrow{rd} & & T_M \arrow{ld} \\
L \times_M E \arrow{rr}{h} \arrow{rru} \arrow{rd} & & T_E \arrow{rru}[near end]{\varpi_*}
\arrow{ld} & M & \\
& E \arrow{rru}[swap]{\varpi} & & &
\end{tikzcd}$$ commutes and its faces $$\begin{tikzcd} L \times_M E \arrow{r}{h} \arrow{d} & T_E \arrow{d} \\ E \arrow{r}[swap]{\operatorname{id}} & E
\end{tikzcd} \qquad \text{and} \qquad \begin{tikzcd} L \times_M E \arrow{r}{h} \arrow{d}
& T_E \arrow{d}{\varpi_*} \\ L \arrow{r}[swap]{\rho} & T_M \end{tikzcd}$$ are vector bundle maps.
The covariant derivative and horizontal lift describing a given connection are related to one another by the identity $$\label{XiAn}
e_*\big(\rho(l_m)\big)-h(l_m,e_m)=\tau_{e_m} \big((\nabla_l e)_m\big) ,$$ which holds for all $m\in M$, $l\in{\Gamma(L)}$, and $e\in{\Gamma(E)}$. Here, $\tau_{e_m}$ denotes the canonical linear isomorphism between the fiber $E_m$ and its tangent space at the point $e_m$.
Geodesic vector field
---------------------
Suppose a splitting $j:L/A\to L$ of the short exact sequence $0\to A \to L \to L/A\to 0$ and an $L$-connection $\nabla$ on $L/A$ have been chosen. Together, the connection and the splitting determine a *‘geodesic’ vector field* $\Xi^{\nabla,j}$ on $L/A$: the value of $\Xi$ at a point $x\in L/A$ is the horizontal lift of the vector $j(x)\in L$. More precisely, $$\Xi_x=h\big(j(x),x\big), \quad\forall x\in L/A ,$$ where $h:L\times_M (L/A) \to T_{L/A}$ denotes the horizontal lifting associated to the linear $L$-connection on $L/A$. This vector field $\Xi$ is called geodesic because its integral curves $t\mapsto b(t)$ satisfy the equation $$b'(t)=h\big(j(b(t)),b(t)\big) ,$$ which is similar to the usual geodesic equation.
\[Yerevan\] The flow $t\mapsto\Phi^\Xi_t$ of the geodesic vector field $\Xi$ satisfies the following two properties:
1. $\Phi^\Xi_t(rx) = r \Phi^\Xi_{rt}(x)$, for all $x\in L/A$ and all real numbers $t,r$ in a sufficiently small open interval around $0$;
2. $\Phi^\Xi_{t_1+t_2}=\Phi^\Xi_{t_2}\circ\Phi^\Xi_{t_1}$, for all real numbers $t_1,t_2$ in a sufficiently small open interval around $0$.
Given a section $\sigma$ of the vector bundle $L/A{\xrightarrow{\pi}}M$, consider the map $$\Psi^{\Xi,\sigma}:{\mathbb{R}}\times M\to{\mathbb{R}}\times M$$ defined by $$\Psi^{\Xi,\sigma}(t,m)=\big(t,\pi\circ\Phi^{\Xi}_t\circ\sigma(m)\big) .$$ Since $\pi\circ\Phi^{\Xi}_0\circ\sigma=\operatorname{id}_M$ and $$\label{Belgium}
\left.\frac{d}{d\tau}\pi\circ\Phi^{\Xi}_{\tau}\circ\sigma(m)\right|_0=\pi_*\Xi_{\sigma(m)}
=\pi_*h\big(j(\sigma(m)),\sigma(m)\big)=\rho\big(j(\sigma(m))\big)
,$$ the differential of $\Psi^{\Xi,\sigma}$ at the point $(0,m)$ is represented by the Jacobian matrix $$\left.J_{\Psi^{\Xi,\sigma}}\right|_{(0,m)}=\begin{pmatrix} 1 & 0 \\ \rho\big(j(\sigma(m))\big) & \operatorname{id}_{{(T_M)}_m} \end{pmatrix} ,$$ which is invertible. Hence $\Psi^{\Xi,\sigma}$ is locally diffeomorphic in a small neighborhood of $\{0\}\times M$.
Therefore, for $t$ sufficiently close to $0$, the map $$\psi^{\Xi,\sigma}_t:=\pi\circ\Phi^{\Xi}_t\circ\sigma$$ is locally diffeomorphic and we can define a one-parameter family of sections $\sigma_t$ of $L/A{\xrightarrow{\pi}}M$ by setting $$\label{Canada}
\sigma_t=\Phi^{\Xi}_t\circ\sigma\circ\big(\psi^{\Xi,\sigma}_t\big)^{-1}
.$$ Obviously, we have $\sigma_0=\sigma$.
The following Lemma will be used later in the proof of Lemma \[Oman\].
\[Denmark\] For every section $\sigma$ of the vector bundle $L/A{\xrightarrow{\pi}}M$, we have $$\left.\frac{d\sigma_t}{dt}\right|_0=-\nabla_{j(\sigma)}\sigma .$$
Differentiating $\Phi^{\Xi}_t\circ\sigma(m)=\sigma_t\circ\psi^{\Xi,\sigma}_t(m)$ with respect to the parameter $t$, we obtain $$\begin{aligned}
\left.\frac{d}{dt}\Phi^{\Xi}_t\big(\sigma(m)\big)\right|_0
=&\left.\frac{d}{dt}\sigma_t\circ\psi^{\Xi,\sigma}_0(m)\right|_0+\left.\frac{d}{dt}\sigma_0\circ\psi^{\Xi,\sigma}_t(m)\right|_0
\\ =&\left.\frac{d}{dt}\sigma_t(m)\right|_0+\left.\frac{d}{dt}\sigma\circ\pi\circ\Phi^{\Xi}_t\circ\sigma(m)\right|_0 \end{aligned}$$ and hence, making use of Equations and , $$\begin{aligned}
\left.\frac{d}{dt}\sigma_t(m)\right|_0 =& \Xi_{\sigma(m)}-\sigma_*\pi_*\Xi_{\sigma(m)} \\
=& h\big(j(\sigma(m)),\sigma(m)\big)-\sigma_*\rho\big(j(\sigma(m))\big) \\
=& -\tau_{\sigma(m)}\big((\nabla_{j(\sigma)}\sigma)(m)\big)
.\end{aligned}$$ Therefore, we have established the equality $\left.\frac{d\sigma_t}{dt}\right|_0=-\nabla_{j(\sigma)}\sigma$ in ${\Gamma(L/A)}$.
Exponential map {#Hungary}
---------------
We can now outline the definition of the exponential map $\exp^{\nabla,j}$ from *a neighborood of the zero section of* $L/A$ to *a neighborhood of the identity section of* ${\mathscr{L}}/{\mathscr{A}}$. This map depends on the splitting $j$ and the connection $\nabla$ we have chosen; different choices yielding different exponential maps.
1. Given an element $x\in L/A$, consider the integral curve $t\mapsto b_x(t)$ in $L/A$ of the geodesic vector field $\Xi^{\nabla,j}$ originating from $b_x(0)=x$. According to Lemma \[Yerevan\], provided $x$ lies sufficiently close to the zero section of $L/A$, this curve is defined up to time $t=1$.
2. The lifted curve $t\mapsto j\big(b_x(t)\big)$ is a ${\rho}$-path in the Lie algebroid $L$, for the ‘image’ of the geodesic equation $b'_x(t) = h\big(j(b_x(t)),b_x(t)\big)$ under the projection $\pi_*:T_{L/A}\to T_M$ is precisely the $\rho$-path equation $\pi_* \big((j\circ b_x)'(t)\big) = {\rho}\big( j\circ b_x(t) \big)$ as $\pi_*\big(h(l,e)\big)={\rho}(l)$ per Diagram .
3. The $\rho$-path $t\mapsto j\circ b_x(t)$ determines a unique ${s}$-path $t\mapsto g^x(t)$ in the Lie groupoid ${\mathscr{L}}$. (See Lemma \[Vienna\].)
4. Composing the path $g^x$ with the canonical projection ${\mathscr{L}}\to{\mathscr{L}}/{\mathscr{A}}$, we obtain a path in ${\mathscr{L}}/{\mathscr{A}}$ whose value at $t=1$ is taken to be the image of $x$ by the exponential map $\exp^{\nabla,j}$.
\[Italy\] The exponential $$\exp^{\nabla,j}: L/A\to{\mathscr{L}}/{\mathscr{A}}$$ associated to an $L$-connection $\nabla$ on $L/A$ and a splitting $j:L/A\to L$ is the map which takes a point $x$ of $L/A$ to the projection of $g^x(1)$ in ${\mathscr{L}}/{\mathscr{A}}$, where $t\mapsto g^x(t)$ is the unique path in ${\mathscr{L}}$ satisfying $$\left.\frac{d}{d\tau}\big(g^x(t)\big)^{-1}g^x(\tau)\right|_t = j\big(\Phi^{\Xi^{\nabla,j}}_t(x)\big)
\qquad\text{and}\qquad g^x(0)=1_{\pi(x)} .$$
We reformulate Lemma \[Yerevan\].
\[Japan\]
1. For any $x\in L/A$ and any real numbers $r,t$ in a sufficiently small open interval around $0$, we have $$g^{rx}(t)=g^{x}(rt) .$$
2. For any $x\in L/A$ and any real numbers $t_1,t_2$ in a sufficiently small open interval around $0$, we have $$g^{x}(t_1+t_2)=g^{x}(t_1)\cdot g^{\Phi^\Xi_{t_1}(x)}(t_2) .$$
\[Koweit\] The exponential map $\exp^{\nabla,j}:L/A\to {\mathscr{L}}/{\mathscr{A}}$ is a fiber bundle map over $\operatorname{id}:M\to M$ which maps the zero section of $L/A$ to the identity section of ${\mathscr{L}}/{\mathscr{A}}$ and is locally diffeomorphic around these sections.
The result follows from the following two observations: **(1)** $\exp^{\nabla,j}(0_m)=1_m$ for every $m\in M$ since the geodesic vector field $\Xi$ vanishes along the zero section of $L/A$; **(2)** the differential of $\exp^{\nabla,j}$ in the direction of the fibers and along the zero section is the identity map since, by Lemma \[Japan\], $$\begin{gathered}
\left.\frac{d}{dr}\exp(rx)\right|_0=\left.\frac{d}{dr}\boldsymbol{q}\big(g^{rx}(1)\big)\right|_0
=\boldsymbol{q}_*\left(\left.\frac{d}{dr} g^x(r)\right|_0\right) \\
=\boldsymbol{q}_*\left(\left.\frac{d}{dr} \big(g^x(0)\big)^{-1}\cdot g^x(r)\right|_0\right)
=q\big(j\circ \Phi^\Xi_{0}(x)\big)=\Phi^\Xi_{0}(x)=x, \end{gathered}$$ where $\boldsymbol{q}:{\mathscr{L}}\to{\mathscr{L}}/{\mathscr{A}}$ and $q:L\to L/A$ are the canonical projections.
Consider the special case where $L$ is the tangent bundle to a smooth manifold $M$ and $A$ is its trivial subbundle of rank 0. Here $j$ is necessarily the identity map on $T_M$ while $\nabla$ is simply a connection on $T_M$. If $t\mapsto\gamma(t)$ is the geodesic curve on $M$ with tangent vector $x$ at time $0$, then $t\mapsto\dot{\gamma}(t)$ is the integral curve $b_x$ of $\Xi^{\nabla,j}$ starting from the point $x$ at time $0$ and $t\mapsto \big(\gamma(0),\gamma(t)\big)$ is the corresponding $s$-path $g^x$ in the pair groupoid $M\times M\rightrightarrows M$. Thus we recover the usual exponential map of classical differential geometry: $\exp^{\nabla,j}(x)=\big(x,\gamma(1)\big)$.
When $A$ is the trivial subbundle of rank $0$ of a Lie algebroid $L$, $j$ is necessarily the identity map on $L$ and $\exp^{\nabla,j}$ is the Landsman exponential map corresponding to the $L$-connection $\nabla$ on $L$ [@Landsman; @NWX]. In particular, given a Lie algebra $\mathfrak{g}=\operatorname{Lie}(G)$, consider the Lie pair $(L,A)$ with $L=\mathfrak{g}$ and $A=\{0\}$. The exponential map $\exp^{\nabla,j}$ determined by the trivial connection $\nabla_{\xi}\eta=0, \forall \xi,\eta\in\mathfrak{g}$ is the usual exponential map $\exp:\mathfrak{g}\to G$ of Lie theory.
When the image of the splitting $j:L/A\to L$ is a Lie subalgebroid $B$ of $L$, the $L$-connection $\nabla$ on $L/A$ is the sum of the Bott $A$-connection on $B$ and some $B$-connection on $B$, and $\exp^{\nabla,j}$ is the Landsman exponential map corresponding to this $B$-connection on $B$ [@Landsman].
Fiberwise jets and differential operators along a section
---------------------------------------------------------
Given a surjective submersion $\pi : P \to M$ together with a section $\epsilon: M \to P $, we define an equivalence relation $\sim$ on $C^\infty(P)$. Two functions $f$ and $g$ are equivalent ($f \sim g$) if and only if $f-g$ and all its derivatives vanish on $\epsilon (M)$, or equivalently, if and only if, for every $\pi$-fiberwise differential operator $D$, the function $D(f-g)$ vanishes along $\epsilon(M)$. Thus the equivalence class of a function $f \in C^\infty(P)$ is its *$\pi$-fiberwise infinite jet along $\epsilon(M)$*.
The quotient of $C^\infty(P)$ by the equivalence class $\mathcal{N}(P,M)$ of the zero function is, by construction, the algebra of all such jets. It will be denoted $C^\infty_\sim(P,M)$.
Now, consider the space ${\mathcal{D}}(P,M)$ of all maps $ C^\infty(P) \to C^\infty(M)$ obtained by composition of a $\pi$-fiberwise differential operator on $P$ (seen as an endomorphism of $C^\infty(P)$) with the restriction $\epsilon^* : C^\infty(P) \to C^\infty(M)$. By construction, ${\mathcal{D}}(P,M)$ is a coalgebra.
In the category $R\operatorname{-mod}$ of left modules over $R=C^\infty(M)$, the algebra $C^\infty_\sim(P,M)$ is the dual of the coalgebra ${\mathcal{D}}(P,M)$ : $$\operatorname{Hom}_{R\operatorname{-mod}}\big({\mathcal{D}}(P,M),R\big)\cong C^\infty_\sim(P,M) .$$
Denoting by $F^{k} {\mathcal{D}}(P,M)$ the *$\pi$-fiberwise differential operators evaluated along $\epsilon(M)$ of order at most $k$* and by $F^{k} C^\infty_\sim(P,M)$ the *$\pi$-fiberwise jets (along $\epsilon(M)$) of functions on P which vanish up to order $k$ along $\epsilon(M)$*, we obtain dual filtrations $$\begin{gathered}
\cdots {\hookrightarrow}F^{k-1} {\mathcal{D}}(P,M) {\hookrightarrow}F^{k} {\mathcal{D}}(P,M) {\hookrightarrow}F^{k+1} {\mathcal{D}}(P,M) {\hookrightarrow}\cdots \\
\cdots {\hookrightarrow}F^{k+1} C^\infty_\sim(P,M) {\hookrightarrow}F^{k} C^\infty_\sim(P,M)
{\hookrightarrow}F^{k+1} C^\infty_\sim(P,M) {\hookrightarrow}\cdots \end{gathered}$$ of the coalgebra ${\mathcal{D}}(P,M)$ of *$\pi$-fiberwise differential operators on $P$ evaluated along $\epsilon(M)$* and the algebra $C^\infty_\sim(P,M)$ of *$\pi$-fiberwise infinite jets of functions on $P$ along $\epsilon(M)$*.
Induced exponential map on jets and differential operators
----------------------------------------------------------
Given a pair of surjective submersions $\pi_1 : P_1 \to M_1$ and $\pi_2 : P_2 \to M_2$, a pair of sections $\epsilon_1: M_1 \to P_1$ and $\epsilon_2: M_2 \to P_2$, and a pair of diffeomorphisms $\Psi:P_1 \to P_2$ and $\psi: M_1 \to M_2$ intertwining the submersions $\pi_1$ and $\pi_2 $ and the sections $\epsilon_1$ and $\epsilon_2$ as well, the algebra isomorphism $\Psi^*:C^\infty(P_2)\to C^\infty(P_1)$ maps $\mathcal{N}(P_1,M_1)$ onto $\mathcal{N}(P_2,M_2)$ and, therefore, determines a coalgebra isomorphism $$\Psi_*:{\mathcal{D}}(P_1,M_1)\to{\mathcal{D}}(P_2,M_2) .$$
For instance, given an inclusion of Lie groupoids ${\mathscr{A}}{\hookrightarrow}{\mathscr{L}}$ over a manifold $M$ and the corresponding inclusion of Lie algebroids $A{\hookrightarrow}L$, the exponential map $\exp^{\nabla,j}:L/A\to{\mathscr{L}}/{\mathscr{A}}$ of Definition \[Italy\] induces an isomorphism of coalgebras $$\label{Laos} \exp^{\nabla,j}_*:{\mathcal{D}}(L/A,M)\to{\mathcal{D}}({\mathscr{L}}/{\mathscr{A}},M)
.$$ Indeed, in view of Proposition \[Koweit\], it suffices to take $$\begin{aligned}
\pi_1&=(L/A\to M) & \pi_2&=({\mathscr{L}}/{\mathscr{A}}{\xrightarrow{s}}M) \\
\epsilon_1&=(M{\xrightarrow{0}}L/A) & \epsilon_2&=(M{\xrightarrow{1}}{\mathscr{L}}/{\mathscr{A}}) \\
\Psi&=(L/A{\xrightarrow{\exp^{\nabla,j}}}{\mathscr{L}}/{\mathscr{A}}) & \psi&=(M{\xrightarrow{\operatorname{id}}}M) \end{aligned}$$ in the construction above.
Let $E\to M$ be a vector bundle. The filtered coalgebras ${\mathcal{D}}(E,M) $ and ${\Gamma(S(E))}$ are canonically isomorphic.
To the symmetric product $x_1\odot x_2\odot\cdots\odot x_n$ in ${\Gamma(S(E))}$ of any sections $x_1,x_2,\dots,x_n$ of $E\to M$ we associate the fiberwise differential operator in ${\mathcal{D}}(E,M)$ which takes a function $f\in C^\infty(E)$ to the function $$m\mapsto\left.\frac{d}{dt_1}\right|_0 \left.\frac{d}{dt_2}\right|_0 \cdots \left.\frac{d}{dt_n}\right|_0
f\big(t_1 x_1(m)+t_2 x_2(m)+\cdots+t_n x_n(m)\big)$$ of $C^\infty(M)$.
Let ${\mathscr{A}}\hookrightarrow{\mathscr{L}}$ be an inclusion of Lie groupoids and let $A\hookrightarrow L$ denote the corresponding inclusion of Lie algebroids. The filtered coalgebras ${\mathcal{D}}({\mathscr{L}}/{\mathscr{A}},M) $ and $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ are canonically isomorphic.
To the image in $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ of the product $x_1\cdot x_2\cdot \cdots \cdot x_n\in{\mathcal{U}(L)}$ of any sections $x_1,x_2,\dots,x_n$ of $L$ we associate the fiberwise differential operator in ${\mathcal{D}}({\mathscr{L}}/{\mathscr{A}},M)$ which takes a function $f\in C^\infty({\mathscr{L}}/{\mathscr{A}})$ to the function $$1^*\big(\overrightarrow{x_1}\circ\overrightarrow{x_2}\circ\dots\circ\overrightarrow{x_n}(\boldsymbol{q}^*f)\big)\in C^\infty(M) .$$ Here $\overrightarrow{x}$ stands for the left-invariant vector field on ${\mathscr{L}}$ corresponding to the section $x$ of $L$. As previously, $\boldsymbol{q}$ denotes the canonical projection $\boldsymbol{q}:{\mathscr{L}}\to{\mathscr{L}}/{\mathscr{A}}$ and $1$ denotes the identity section $1:M\to{\mathscr{L}}$.
It turns out that the isomorphism coincides with the $\operatorname{pbw}$ map defined in Section \[Madrid\]. This is our second core result.
\[Madagascar\] Let ${\mathscr{A}}\hookrightarrow{\mathscr{L}}$ be an inclusion of Lie groupoids and let $A\hookrightarrow L$ denote the corresponding inclusion of Lie algebroids. Given a splitting $j:L/A \to L$ of the short exact sequence of vector bundles $0\to A\to L\to L/A\to 0$ and an $L$-connection $\nabla$ on $L/A$, the filtered coalgebra isomorphism $\exp^{\nabla,j}_*:{\mathcal{D}}(L/A,M)\to{\mathcal{D}}({\mathscr{L}}/{\mathscr{A}},M)$ induced by the exponential map coincides with the Poincaré–Birkhoff–Witt isomorphism $\operatorname{pbw}^{\nabla,j}:{\Gamma(S(L/A))}\to\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$.
We will make use of the bundle map $$\begin{tikzcd} L/A \arrow{r}{\boldsymbol{E}} \arrow{d}[swap]{\pi} & {\mathscr{L}} \arrow{d}{s} \\
M \arrow{r}[swap]{\operatorname{id}} & M \end{tikzcd} ,$$ defined by $\boldsymbol{E}(x)=g^x(1)$ for all $x\in L/A$, and the associated map $$\mathscr{E}:{\Gamma(L/A)}\to{\Gamma({\mathscr{L}})} .$$ We remind the reader that the exponential map $\exp^{\nabla,j}$ is simply the composition of $\boldsymbol{E}:L/A\to{\mathscr{L}}$ with the canonical projection $\boldsymbol{q}:{\mathscr{L}}\to{\mathscr{L}}/{\mathscr{A}}$.
\[Niger\]
1. For any $l\in{\mathscr{L}}$ and $x\in{\Gamma(L/A)}$, we have $$\left. \frac{d}{d\tau} l\cdot\mathscr{E}(\tau x) \right|_0=\overrightarrow{j(x)}|_l ,$$ where $\overrightarrow{j(x)}$ denotes the left-invariant vector field on the Lie groupoid ${\mathscr{L}}$ (tangent to the s-fibers) associated to the section $j(x)$ of the Lie algebroid $L$.
2. For any $x\in{\Gamma(L/A)}$ and any real numbers $t_1,t_2$ in a sufficiently small open interval around $0$, we have $$\mathscr{E}\big((t_1+t_2)x\big)=\mathscr{E}(t_1x)\cdot\mathscr{E}\big(t_2 x_{t_1}\big) ,$$ where the r.h.s. is a product of $s$-sections of ${\mathscr{L}}$ and $$x_{t_1}=\Phi^{\Xi}_{t_1}\circ x\circ (\psi^{\Xi,x}_{t_1})^{-1}$$ as in Equation .
The first assertion follows from $$\left.\frac{d}{d\tau}f\big(l\cdot\mathscr{E}(\tau x)\big)\right|_0
=\left.\frac{d}{d\tau}f\big(l\cdot g^{\tau x}(1)\big)\right|_0
=\left.\frac{d}{d\tau}f\big(l\cdot g^{x}(\tau)\big)\right|_0
=\overrightarrow{j(x)}|_l (f)
,$$ in which we have made use of the first part of Lemma \[Japan\]. The second part of Lemma \[Japan\] can be rewritten as $$\boldsymbol{E}\big((t_1+t_2)x\big)=\boldsymbol{E}(t_1 x)\cdot\boldsymbol{E}\big(t_2\Phi^{\Xi}_{t_1}(x)\big) .$$ Passing to sections of $L/A\to M$, we obtain the second assertion above.
Recall that the universal enveloping algebra ${\mathcal{U}(L)}$ of the Lie algebroid $L$ can be regarded as the associative algebra of $s$-fiberwise differential operators on ${C^{\infty}({\mathscr{L}})}$ invariant under left translations. The map $\mathscr{E}:{\Gamma(L/A)}\to{\Gamma({\mathscr{L}})}$ induces a map $$\mathcal{E}:{\Gamma(S(L/A))}\to{\mathcal{U}(L)}$$ as follows. The element $\mathcal{E}[x_1\odot x_2\odot\cdots\odot x_n]$ of ${\mathcal{U}(L)}$ corresponding to $x_1\odot x_2\odot\cdots\odot x_n \in {\Gamma(S^n(L/A))}$ is the left invariant differential operator which takes a function $f\in{C^{\infty}({\mathscr{L}})}$ to the function $${\mathscr{L}}\ni l\mapsto \left.\frac{d}{dt_1}\right|_0 \left.\frac{d}{dt_2}\right|_0 \cdots
\left.\frac{d}{dt_n}\right|_0 f\big( l\cdot\mathscr{E}(t_1 x_1+t_2 x_2 + \cdots + t_n x_n) \big) \in {\mathbb{R}}.$$ In particular, for any $x\in{\Gamma(L/A)}$ and $n\in{\mathbb{N}}$, we have $$\begin{gathered}
\big(\mathcal{E}[x^n] f\big)(l)
=\left.\frac{d}{dt_1}\right|_0 \left.\frac{d}{dt_2}\right|_0 \cdots
\left.\frac{d}{dt_n}\right|_0 f\Big( l\cdot\mathscr{E}\big((t_1+t_2+\cdots+t_n)x\big)\Big) \\
=\left.\frac{d^n}{dt^n}f\big( l\cdot\mathscr{E}(tx)\big)\right|_0 .\end{gathered}$$ We declare that $\mathcal{E}[r]=r$ for every $r\in R=C^\infty(M)$ seen as an element of degree 0 in ${\Gamma(S(L/A))}$.
It is easy to check that $\mathcal{E}$ is a morphism of left $R$-modules.
\[Oman\] For all $x\in{\Gamma(L/A)}$ and $n\in{\mathbb{N}}$, we have $\mathcal{E}[x]=\overrightarrow{j(x)}$ and $$\mathcal{E}[x^{n+1}]=\overrightarrow{j(x)}\circ\mathcal{E}[x^n]-\mathcal{E}[\nabla_{j(x)}(x^n)] .$$
Indeed, it follows from Lemma \[Niger\] that $$\big(\mathcal{E}[x] f\big)(l)=\overrightarrow{j(x)}|_l (f)$$ and $$\begin{aligned}
\big(\mathcal{E}[x^{n+1}] f\big)(l)
=&\left.\frac{d}{dt_0}\right|_0 \left.\frac{d}{dt_1}\right|_0 \cdots \left.\frac{d}{dt_n}\right|_0
f\Big( l\cdot\mathscr{E}\big((t_0+t_1+\cdots+t_n)x\big)\Big) \\
=&\left.\frac{d}{dt_0}\right|_0 \left.\frac{d}{dt_1}\right|_0 \cdots \left.\frac{d}{dt_n}\right|_0
f\Big( l\cdot\mathscr{E}(t_0 x)\cdot\mathscr{E}\big((t_1+t_2+\cdots+t_n)x_{t_0}\big)\Big) \\
=&\left.\frac{d}{dt_0}\right|_0 \Big(\mathcal{E}\big[(x_{t_0})^n\big]f\Big)
\big(l\cdot\mathscr{E}(t_0 x)\big) \\
=&\left.\frac{d}{dt_0}\right|_0 \Big(\mathcal{E}\big[x^n\big]f\Big)
\big(l\cdot\mathscr{E}(t_0 x)\big)
+\left.\frac{d}{dt_0}\right|_0 \Big(\mathcal{E}\big[(x_{t_0})^n\big]f\Big)
\big(l\big)
\\
=&\Big(\overrightarrow{j(x)}\big(\mathcal{E}[x^n]f\big)\Big)(l)
+\Big(\mathcal{E}\Big[\left.\frac{d}{dt_0}(x_{t_0})^n\right|_0\Big] f\Big)(l) \\
=&\Big(\overrightarrow{j(x)}\big(\mathcal{E}[x^n]f\big)\Big)(l)
-\big(\mathcal{E}\big[\nabla_{j(x)}(x^n)\big]f\big)(l)
.\end{aligned}$$ The last equality above is a consequence of Lemma \[Denmark\].
One easily observes that $${\Gamma(S(L/A))}\cong{\mathcal{D}}(L/A,M){\xrightarrow{\exp^{\nabla,j}_*}}
{\mathcal{D}}({\mathscr{L}}/{\mathscr{A}},M)\cong\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$$ is the composition of $\mathcal{E}$ with the projection ${\mathcal{U}(L)}\to\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ the same way $\exp^{\nabla,j}$ is the composition of $\boldsymbol{E}$ with the projection $\boldsymbol{q}:{\mathscr{L}}\to{\mathscr{L}}/{\mathscr{A}}$. Then Lemma \[Oman\] asserts that $$\exp^{\nabla,j}_*:{\Gamma(S(L/A))}\to\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$$ is precisely the map $\operatorname{pbw}^{\nabla,j}$ inductively defined in Theorem \[Nairobi\].
When $L=T_M$ and $A$ is the trivial Lie subalgebroid of $L$ of rank 0, the $\operatorname{pbw}^{\nabla,j}$ map of Theorem \[Madagascar\] is the inverse of the so called ‘complete symbol map,’ which is an isomorphism from the space ${\mathcal{U}(T_M)}$ of differential operators on $M$ to the space ${\Gamma(S(T_M))}$ of fiberwisely polynomial functions on $T{^{\vee}}_M$. The complete symbol map was generalized to arbitrary Lie algebroids over ${\mathbb{R}}$ by Nistor, Weinstein, and one of the authors [@NWX]. It played an important role in quantization theory [@MR706215; @NWX].
Kapranov dg-manifolds
=====================
Definition
----------
Let $A$ be a Lie algebroid over a smooth manifold $M$, and $R$ be the algebra of smooth functions on $M$ valued in ${\mathbb{K}}$. The Chevalley–Eilenberg differential $$d_A:{\Gamma(\wedge^k A{^{\vee}})}\to{\Gamma(\wedge^{k+1} A{^{\vee}})}$$ defined by $$\begin{gathered}
\label{eqCE}
\big(d_A \alpha\big)(a_0,a_1,\cdots,a_k)=
\sum_{i=0}^{n} {\rho}(a_i)\big(\alpha(a_0,\cdots,\widehat{a_i},\cdots,a_k)\big) \\
+\sum_{i<j}\alpha({[a_i,a_j]},a_0,\cdots,\widehat{a_i},\cdots,\widehat{a_j},\cdots,a_k) \end{gathered}$$ and the exterior product make $\bigoplus_{k\geqslant 0}{\Gamma(\wedge^k A{^{\vee}})}$ into a differential graded commutative $R$-algebra.
The Chevalley–Eilenberg covariant differential associated to a representation $\nabla^\natural$ of a Lie algebroid $A\to M$ of rank $n$ on a vector bundle $E\to M$ is the operator $$d_A^{\nabla^\natural}: {\Gamma(\wedge^k A{^{\vee}}\otimes E)}\to{\Gamma(\wedge^{k+1} A{^{\vee}}\otimes E)}$$ that takes a section $\alpha\otimes e$ of $\wedge^k A{^{\vee}}\otimes E$ to $$d_A^{\nabla^\natural}(\alpha\otimes e)=(d_A\alpha)\otimes e
+(-1)^{k}\sum_{j=1}^{n}(\alpha\wedge\beta_j)\otimes \nabla^\natural_{b_j}e ,$$ where $b_1,b_2,\dots,b_n$ and $\beta_1,\beta_2,\dots,\beta_n$ are any pair of dual local frames for the vector bundles $A$ and $A{^{\vee}}$. Because the connection $\nabla$ is flat, $d_A^{\nabla^\natural}$ is a coboundary operator: $d_A^{\nabla^\natural}\circ d_A^{\nabla^\natural}=0$.
Recall that a *dg-manifold* is a ${\mathbb{Z}}$-graded manifold endowed with a vector field $Q$ of degree $+1$ satisfying ${[Q,Q]}=0$. Such a vector field $Q$ is said to be homological [@Vaintrob]. In the literature, dg-manifolds are also called $Q$-manifolds [@MR1432574; @Kontsevich_formality]. We refer the reader to [@MR3293862; @MR1958834] for details. Below are several standard examples of dg-manifolds.
If $A$ is a Lie algebroid, $A[1]$ is a dg-manifold with the Chevalley–Eilenberg differential $d_{A}$ serving as the homological vector field $Q$. In fact, according to Va[ĭ]{}ntrob [@Vaintrob], for any vector bundle $A\to M$, $A[1]$ is a dg-manifold if and only if $A$ is a Lie algebroid over $M$.
If $E\to M$ is a vector bundle and $s\in \Gamma(E)$, then $E[-1]$ is a dg-manifold with the contraction operator $i_s$ serving as the homological vector field $Q$.
Suppose ${\mathfrak{g}}=\sum_{i\in {\mathbb{Z}}}{\mathfrak{g}}_i$ is a ${\mathbb{Z}}$-graded vector space of finite type, i.e. each ${\mathfrak{g}}_i$ is a finite dimensional vector space. Then ${\mathfrak{g}}[1]$ is a dg-manifold if and only if ${\mathfrak{g}}$ is an $L_\infty$-algebra.
We are now ready to introduce the main object of the paper.
A *Kapranov dg-manifold* is a pair of vector bundles $A$ and $E$ over a common base manifold together with a homological vector field on each of the graded manifolds $A[1]$ and $A[1]\oplus E$ such that the inclusion $A[1]{\hookrightarrow}A[1]\oplus E$ and the projection $A[1]\oplus E{\twoheadrightarrow}A[1]$ are morphisms of dg-manifolds.
Note that the requirement of $A[1]$ being a dg-manifold implies that $A$ must be a Lie algebroid. Kapranov dg-manifolds are related to Costello’s $L_\infty$ algebras over the differential graded algebra $({\Gamma(\wedge{^{\bullet}}A{^{\vee}})},d_A)$ (see [@arXiv:1112.0816]).
Given a representation of a Lie algebroid $A$ over a vector bundle $E$, the Chevalley–Eilenberg differentials provide a pair of homological vector fieds on $A[1]$ and $A[1]\oplus E$ compatible with the inclusion and projection maps. The resulting Kapranov dg-manifold $A[1]\oplus E$ is said to be *linear*.
A *morphism of Kapranov dg-manifolds* from $A[1]\oplus E$ to $A'[1]\oplus E'$ consists of a pair $(\Phi,\phi)$ of morphisms of dg-manifolds $\Phi:A[1]\oplus E \to A'[1]\oplus E'$ and $\phi:A[1]\to A'[1]$ such that the following two diagrams commute: $$\begin{array}{cc}
\begin{tikzcd} A[1]\oplus E \arrow{r}{\Phi} & A'[1]\oplus E' \\
A'[1] \arrow{r}[swap]{\phi} \arrow[hook]{u} & A'[1] \arrow[hook]{u} \end{tikzcd} &
\begin{tikzcd} A[1]\oplus E \arrow{r}{\Phi} \arrow[two heads]{d} & A'[1]\oplus E' \arrow[two heads]{d} \\
A[1] \arrow{r}[swap]{\phi} & A'[1] .\end{tikzcd}
\end{array}$$ Such a morphism is said to *fix the dg-submanifold $A[1]$* if $A=A'$ and $\phi=\operatorname{id}_{A[1]}$, in which case it is simply denoted by $\Phi$. An *isomorphism of Kapranov dg-manifolds* is a morphism $(\Phi,\phi)$ of Kapranov dg-manifolds where both $\Phi$ and $\phi$ are dg-manifold isomorphisms.
A Kapranov dg-manifold $(A[1]\oplus E,D)$ is said to be *linearizable* if there exists a representation $\nabla^\natural$ of the Lie algebroid $A$ on the vector bundle $E$ and an isomorphism of Kapranov dg-manifolds from $(A[1]\oplus E,D)$ to $(A[1]\oplus E,d_A^{\nabla^\natural})$ fixing the dg-submanifold $(A[1],d_A)$.
Equivalent characterizations
----------------------------
Given a smooth vector bundle $E$ over the manifold $M$, recall that deconcatenation (see Equation ) defines a comultiplication on ${\Gamma(S(E))}$ while concatenation defines a multiplication on ${\Gamma(S(E^\vee))}$. Let $\mathfrak{e}$ denote the ideal of the graded algebra ${\Gamma(S(E^\vee))}$ generated by ${\Gamma(E^\vee)}$, and $R$ the algebra of smooth functions on $M$. The algebra $\operatorname{Hom}_R\big({\Gamma(S(E))},R\big)$ dual to the coalgebra ${\Gamma(S(E))}$ is the $\mathfrak{e}$-adic completion of ${\Gamma(S(E^\vee))}$. It will be denoted by ${\Gamma(\hat{S}(E^\vee))}$. Equivalently, one can think of the completion $\hat{S}(E^\vee)$ of $S(E^\vee)$ as a bundle of algebras over $M$. Therefore we have a pairing ${\langle \varepsilon|e\rangle}\in R$ for any $\varepsilon\in{\Gamma(\hat{S}(E^\vee))}$ and $e\in {\Gamma(S(E))}$, which can be obtained as follows. Consider $e$ as a fiberwise polynomial on $E^\vee$ and $\varepsilon$ a fiberwise differential operator on $E^\vee$ with fiberwisely constant coefficients. Then ${\langle \varepsilon|e\rangle}$ is the function on $M$ obtained by applying $\varepsilon$ on $e$ fiberwisely. The following lemma is immediate.
\[Portugal\]
1. Let $\varepsilon$ be a section of $\hat{S}(E{^{\vee}})$. Then $\varepsilon\in{\Gamma(\hat{S}^{\geqslant n}(E{^{\vee}}))}$ if and only if ${\langle \varepsilon|e\rangle}=0$ for all $e\in{\Gamma(S^{<n}(E))}$.
2. Let $e$ be a section of $S(E)$. Then $e\in{\Gamma(S^{\leqslant n}(E))}$ if and only if ${\langle \varepsilon|e\rangle}=0$ for all $\varepsilon\in{\Gamma(\hat{S}^{>n}(E{^{\vee}}))}$.
Let $\nabla^\natural:{\Gamma(A)}\times{\Gamma(E)}\to{\Gamma(E)}$ be a representation of $A$ on $E$. Consider the space ${{\mathfrak{X}}^{\operatorname{fv}}}(E)$ of formal vertical vector fields on $E$ along the zero section, which are by definition $R$-linear derivations of the algebra of formal functions on $E$ along the zero section. Since the latter can be identified naturally with ${\Gamma(\hat{S}(E{^{\vee}}))}$, the space ${{\mathfrak{X}}^{\operatorname{fv}}}(E)$ is naturally identified with ${\Gamma(\hat{S}(E{^{\vee}})\otimes E)}$. The Lie algebra structure on ${\Gamma(\hat{S}(E{^{\vee}})\otimes E)}$ inherited from ${{\mathfrak{X}}^{\operatorname{fv}}}(E)$ extends to a graded Lie algebra structure on $\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n(A{^{\vee}})
\otimes\hat{S}(E{^{\vee}})\otimes E)}$: $$\begin{gathered}
\label{Congo}
{[\alpha_1\otimes\varepsilon_1^{n_1}\otimes e_1,\alpha_2\otimes\varepsilon_2^{n_2}\otimes e_2]} \\
= n_2{\langle \varepsilon_2|e_1\rangle}\cdot(\alpha_1\wedge\alpha_2)\otimes
(\varepsilon_1^{n_1}\odot\varepsilon_2^{n_2-1})\otimes e_2 \\
- n_1{\langle \varepsilon_1|e_2\rangle}\cdot(\alpha_1\wedge\alpha_2)\otimes
(\varepsilon_1^{n_1-1}\odot\varepsilon_2^{n_2})\otimes e_1 ,\end{gathered}$$ for all $\alpha_1,\alpha_2\in{\Gamma(\wedge A{^{\vee}})}$, $\varepsilon_1,\varepsilon_2\in{\Gamma(E{^{\vee}})}$, $e_1,e_2\in{\Gamma(E)}$, and $n_1,n_2\in{\mathbb{N}}$. Since $E$ is an $A$-module, so is $\hat{S}(E{^{\vee}}) \otimes E$ and we can consider the corresponding Chevalley–Eilenberg differential $$\label{Rwanda}
d_A^{\nabla^\natural} : {\Gamma( \wedge^\bullet (A{^{\vee}}) \otimes \hat{S}(E{^{\vee}}) \otimes E )} \to
{\Gamma( \wedge^{\bullet+1} (A{^{\vee}}) \otimes \hat{S}(E{^{\vee}}) \otimes E )},$$ The proof of the following result is straightforward.
\[Sweden\] Given a representation $\nabla^\natural:{\Gamma(A)}\times{\Gamma(E)}\to{\Gamma(E)}$ of $A$ on $E$, the differential and the Lie bracket make $$\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n(A{^{\vee}})\otimes\hat{S}(E{^{\vee}})\otimes E)}$$ a differential graded Lie algebra.
\[Tasmania\] Given a Lie algebroid $A$ over a smooth manifold $M$ and a smooth vector bundle $E$ over $M$, the following data are equivalent.
1. \[uno\] A homological vector field $D$ on $A[1]\oplus E$ making it a Kapranov dg-manifold.
2. \[quattro\] A degree $+1$ derivation $D$ of the graded algebra $\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n A{^{\vee}}\otimes\hat{S}(E{^{\vee}}))}$, which preserves the decreasing filtration $$\label{Uruguay}
\cdots{\hookrightarrow}{\Gamma(\wedge A{^{\vee}}\otimes\hat{S}^{\geqslant n+1}(E{^{\vee}}))}
{\hookrightarrow}{\Gamma(\wedge A{^{\vee}}\otimes\hat{S}^{\geqslant n}(E{^{\vee}}))}{\hookrightarrow}\cdots$$ and satisfies $D^2=0$ and $D(\alpha\otimes 1)=d_A(\alpha)\otimes 1$, for all $\alpha\in{\Gamma(\wedge A{^{\vee}})}$.
3. \[due\] An infinitesimal action $\delta:{\Gamma(A)}\times{\Gamma(S(E))}\to
{\Gamma(S(E))}$ of $A$ on the symmetric coalgebra $\Gamma (S(E))$ by coderivations such that $\delta_a(1)=0$, for all $a\in{\Gamma(A)}$.
4. \[tre\] An infinitesimal action $\delta^*:{\Gamma(A)}\times{\Gamma(\hat{S}(E{^{\vee}}))}\to{\Gamma(\hat{S}(E{^{\vee}}))}$ of $A$ on the completed symmetric algebra $\Gamma (\hat{S}(E{^{\vee}}))$ by derivations such that $\delta^*_a(\varepsilon)\in{\Gamma(\hat{S}^{\geqslant 1}(E{^{\vee}}))}$, for all $a\in{\Gamma(A)}$ and $\varepsilon\in{\Gamma(E{^{\vee}})}$.
5. \[cinque\] A representation $\nabla^\natural:{\Gamma(A)}\times{\Gamma(E)}\to{\Gamma(E)}$ of $A$ on $E$ together with a solution ${\mathcal{R}}\in{\Gamma(A{^{\vee}}\otimes\hat{S}^{\geqslant 2}(E{^{\vee}})\otimes E)}$ of the Maurer–Cartan equation $$d_A^{\nabla^\natural}{\mathcal{R}}+\tfrac{1}{2}{[{\mathcal{R}},{\mathcal{R}}]} =0$$ in the dgla $\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n A{^{\vee}}\otimes\hat{S}(E{^{\vee}})\otimes E)}$ of Lemma \[Sweden\].
6. \[sei\] An $L_\infty[1]$ algebra structure on the graded vector space $$\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n A{^{\vee}}\otimes E)}$$ defined by a sequence $(\lambda_k)_{k\in{\mathbb{N}}}$ of multibrackets $$\lambda_k:S^k\big({\Gamma(\wedge A{^{\vee}}\otimes E)}\big)\to{\Gamma(\wedge A{^{\vee}}\otimes E)}[1]$$ such that each $\lambda_k$ with $k\geqslant 2$ is ${\Gamma(\wedge A{^{\vee}})}$-multilinear and $\lambda_1$ is the Chevalley–Eilenberg covariant differential $d_A^{\nabla}$ associated to a representation $\nabla$ of the Lie algebroid $A$ on the vector bundle $E$.
The infinitesimal $A$-actions $\delta$ and $\delta^*$ are related to the homological vector field $D$ through the identities $$\begin{gathered}
\delta^*_a \sigma= i_a D(1\otimes\sigma) \\
\rho(a){\langle \sigma|s\rangle}={\langle \delta^*_a \sigma|s\rangle}+{\langle \sigma|\delta_a s\rangle} \end{gathered}$$ where $a\in{\Gamma(A)}$, $\sigma\in{\Gamma(\hat{S}(E{^{\vee}}))}$, $s\in{\Gamma(S(E))}$. The representation $\nabla^\natural$ of $A$ on $E$ is defined by $$\label{Vietnam} \nabla^\natural_a e=\operatorname{pr}_1(\delta_a e) ,$$ where $a\in{\Gamma(A)}$, $e\in{\Gamma(E)}$, and $\operatorname{pr}_1$ denotes the canonical projection of $\bigoplus_{n=0}^\infty S^n(E)$ onto $S^1(E)$. The solution ${\mathcal{R}}$ of the Maurer–Cartan equation is the difference $$\label{Scandinavie} {\mathcal{R}}=D-d_A^{\nabla^\natural} .$$ Finally, the multibrackets $(\lambda_k)_{k\in{\mathbb{N}}}$ satisfy $$\lambda_k(\xi_1\otimes e_1,\cdots,\xi_k\otimes e_k)=(-1)^{{\left| \xi_1 \right|}+\cdots+{\left| \xi_k \right|}}
\xi_1\wedge\cdots\wedge\xi_k\wedge{\mathcal{R}}_k (e_1,\cdots,e_k)$$ for all $e_1,\dots,e_k\in{\Gamma(E)}$ and all homogeneous elements $\xi_1,\dots,\xi_k$ of ${\Gamma(\wedge A{^{\vee}})}$. Here ${\mathcal{R}}_k$ denotes the component of ${\mathcal{R}}$ in ${\Gamma(A{^{\vee}}\otimes S^k(E{^{\vee}})\otimes E)}$.
(\[uno\])$\Rightarrow$(\[quattro\]) A homological vector field on $A[1]\oplus E$ is, by definition, a degree $+1$ derivation $D$ of the graded algebra $\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n
A{^{\vee}}\otimes\hat{S}(E{^{\vee}}))}$ satisfying $D^2=0$. The projection $A[1]\oplus E{\twoheadrightarrow}A[1]$ is a morphism of dg-manifolds if and only if $$D(\alpha\otimes 1)=d_A \alpha \otimes 1, \quad \forall \alpha \in {\Gamma(\wedge A{^{\vee}})} .$$ The inclusion $A[1]{\hookrightarrow}A[1]\oplus E$ is a morphism of dg-manifolds if and only if $$\pi\circ D=d_A\circ\pi ,$$ where $\pi:{\Gamma(\wedge A{^{\vee}}\otimes \hat{S}(E{^{\vee}}))}
\to {\Gamma(\wedge A{^{\vee}})}$ is defined by $$\pi(\alpha\otimes\sigma)= \begin{cases}
\sigma\cdot\alpha & \text{if }\sigma\in{\Gamma(S^0(E{^{\vee}}))}=R \\ 0 & \text{if } \sigma\in{\Gamma(\hat{S}^{\geqslant 1}(E{^{\vee}}))} .\end{cases}$$ Therefore, for every $\alpha\in{\Gamma(\wedge A{^{\vee}})}$ and $\varepsilon\in{\Gamma(E{^{\vee}})}$, we have $$\pi\big(D(\alpha\otimes\varepsilon)\big)
=d_A\big(\pi(\alpha\otimes\varepsilon)\big)
=d_A(0)=0$$ and thus $D(\alpha\otimes\varepsilon)\in{\Gamma(\wedge A{^{\vee}}\otimes \hat{S}^{\geqslant 1}(E{^{\vee}}))}$. Hence, since $D$ is a derivation, $D$ preserves the filtration .
(\[quattro\])$\Rightarrow$(\[uno\]) For all $\alpha\in{\Gamma(\wedge A{^{\vee}})}$ and $\sigma\in{\Gamma(\hat{S}^{\geqslant 1}(E{^{\vee}}))}$, we have $D(\alpha\otimes\sigma)\in{\Gamma(
\wedge A{^{\vee}}\otimes\hat{S}^{\geqslant 1}(E{^{\vee}}))}$ since $D$ preserves the filtration . Therefore, for all $\alpha\in{\Gamma(\wedge A{^{\vee}})}$ and $\sigma\in{\Gamma(\hat{S}^{\geqslant 1}(E{^{\vee}}))}$, we have $$\pi\big(D(\alpha\otimes\sigma)\big)=0=
d_A\big(\pi(\alpha\otimes\sigma)\big) .$$ Moreover, for all $\alpha\in{\Gamma(\wedge A{^{\vee}})}$ and $f\in{\Gamma(S^0(E{^{\vee}}))}=R$, we have $$\pi\big(D(\alpha\otimes f)\big)
=\pi\big(D(f\alpha\otimes 1)\big)
=\pi\big(d_A(f\alpha)\otimes 1\big)
=d_A\big(\pi(\alpha\otimes f)\big) .$$ Hence $\pi\circ D=d_A\circ\pi$.
(\[due\])$\Leftrightarrow$(\[tre\]) The infinitesimal $A$-actions $\delta$ and $\delta^*$ determine each other through the identity $$\label{Poland}
{\rho}(a){\langle \varepsilon|e\rangle}={\langle \delta^*_a(\varepsilon)|e\rangle}+
{\langle \varepsilon|\delta_a(e)\rangle},$$ where $a\in{\Gamma(A)}$, $\varepsilon\in{\Gamma(\hat{S}(E^\vee))}$, and $e\in{\Gamma(S(E))}$. Obviously, $\delta_a$ is a coderivation if and only if $\delta^*_a$ is a derivation. Since $\delta_a$ is a coderivation for all $a\in{\Gamma(A)}$, Theorem \[shark\] asserts that $\delta_a(1)=0$ for all $a\in{\Gamma(A)}$ if and only if ${\Gamma(S^{\leqslant n}(E))}$ is $\delta_a$-stable for all $n\in{\mathbb{N}}$ and $a\in{\Gamma(A)}$. Since ${\Gamma(\hat{S}^{>n}(E^\vee))}$ is the annihilator of ${\Gamma(S^{\leqslant n}(E))}$ according to Lemma \[Portugal\], ${\Gamma(S^{\leqslant n}(E))}$ is $\delta$-stable if and only if ${\Gamma(\hat{S}^{>n}(E^\vee))}$ is $\delta^*$-stable. Since $\delta^*_a$ is a derivation for all $a\in{\Gamma(A)}$, ${\Gamma(\hat{S}^{\geqslant n}(E))}$ is $\delta^*_a$-stable for all $n\in{\mathbb{N}}$ and $a\in{\Gamma(A)}$ if and only if $\delta^*_a(\varepsilon)\in{\Gamma(
\hat{S}^{\geqslant 1}(E{^{\vee}}))}$ for all $a\in{\Gamma(A)}$ and $\varepsilon\in{\Gamma(E{^{\vee}})}$.
(\[quattro\])$\Rightarrow$(\[tre\]) Given an infinitesimal $A$-action $\delta^*:{\Gamma(A)}\times{\Gamma(\hat{S}(E{^{\vee}}))}\to{\Gamma(\hat{S}(E{^{\vee}}))}$, set $$D(\alpha\otimes\sigma)=(d_A\alpha)\otimes\sigma
+\sum_j(\nu_j\wedge\alpha)\otimes(\delta^*_{v_j}\sigma)$$ for all $\alpha\in{\Gamma(\wedge A{^{\vee}})}$ and $\sigma\in{\Gamma(\hat{S}(E^\vee))}$. Here $v_1,\dots,v_l$ and $\nu_1,\dots,\nu_l$ are any pair of dual local frames of the vector bundles $A$ and $A{^{\vee}}$. The flatness of the infinitesimal action $\delta^*$ implies that $D^2=0$ while the other properties of $D$ follow immediately from those of $\delta^*$.
(\[tre\])$\Rightarrow$(\[quattro\]) Given $D$, the relation $$\delta^*_a(\sigma)=i_a D(1\otimes \sigma ), \quad\forall a\in {\Gamma(A)},
\sigma \in {\Gamma(\hat{S}(E^\vee))}$$ defines a map $\delta^*:{\Gamma(A)}\times{\Gamma(\hat{S}(E^\vee))}\to{\Gamma(\hat{S}(E^\vee))}$, which is indeed an infinitesimal action as $$\begin{split}
\delta^*_a (f\sigma) =& i_a D(1\otimes f\sigma) \\
=& i_a D\big\{(f\otimes 1)\cdot(1\otimes\sigma)\big\} \\
=& i_a \big\{ D(f\otimes 1)\cdot(1\otimes\sigma)
+ (f\otimes 1)\cdot D(1\otimes\sigma) \big\} \\
=& i_a \Big\{ \big((d_A f)\otimes 1\big)\cdot (1\otimes\sigma)
+f\cdot D(1\otimes\sigma) \Big\} \\
=& i_a \Big\{ (d_A f)\otimes\sigma + f\cdot D(1\otimes\sigma) \Big\} \\
=& \rho(a)f \cdot\sigma +f\cdot\delta^*_a(\sigma)
,\end{split}$$ for all $a\in{\Gamma(A)}$, $f\in C^\infty(M)$, and $\sigma\in{\Gamma(\hat{S}(E^\vee))}$.
(\[quattro\])$\Rightarrow$(\[cinque\])Given the derivation $D$ of ${\Gamma(\wedge A{^{\vee}}\otimes
\hat{S}(E{^{\vee}}))}$, set $$\label{Romania} \nabla^\natural_a \varepsilon = \varpi\big(i_a D(1\otimes\varepsilon)\big) ,\quad\forall a\in{\Gamma(A)}, \varepsilon\in{\Gamma(E{^{\vee}})},$$ where $\varpi$ denotes the canonical projection of $\hat{S}(E{^{\vee}})$ onto $S^1(E{^{\vee}})$.
For all $f\in C^\infty(M)$, we have $$\begin{split}
D(1\otimes f\varepsilon)
=& D\big((f\otimes 1)\cdot(1\otimes\varepsilon)\big) \\
=& D(f\otimes 1)\cdot(1\otimes\varepsilon)
+ (f\otimes 1)\cdot D(1\otimes\varepsilon) \\
=& \big((d_A f)\otimes 1\big)\cdot (1\otimes\varepsilon)
+f\cdot D(1\otimes\varepsilon) \\
=& (d_A f)\otimes\varepsilon + f\cdot D(1\otimes\varepsilon)
.\end{split}$$ It follows that $$i_a D(1\otimes f\varepsilon)=\rho(a)f\otimes\varepsilon
+f\cdot i_a D(1\otimes\varepsilon)$$ and, applying $\varpi$ to both sides, $$\nabla^\natural_a(f\varepsilon)=\rho(a)f\cdot \varepsilon
+f\cdot\nabla^\natural_a\varepsilon .$$ The map $\nabla^\natural:{\Gamma(A)}\times{\Gamma(E{^{\vee}})}
\to{\Gamma(E{^{\vee}})}$ is thus a connection. This connection is flat: since $D^2=0$ and $D$ preserves the filtration , we have $$d_A^{\nabla^\natural}(d_A^{\nabla^\natural}\varepsilon)
= \varpi\bigg(D\Big(\varpi\big(D
(1\otimes\varepsilon)\big)\Big)\bigg)
=\varpi\big(D^2(1\otimes\varepsilon)\big)
=0$$ for all $\varepsilon\in{\Gamma(E{^{\vee}})}$.
We use the same symbol $\nabla^\natural$ to denote the induced flat connection on the dual bundle $E$. Equation follows immediately from the identity $$\rho(a){\langle \varepsilon|e\rangle}=
{\langle \nabla^\natural_a\varepsilon|e\rangle}
+{\langle \varepsilon|\nabla^\natural_a e\rangle} ,$$ where $a\in{\Gamma(A)}$, $e\in{\Gamma(E)}$, and $\varepsilon\in{\Gamma(E{^{\vee}})}$, together with Equations and .
Since $D$ and $d_A^{\nabla^\natural}$ are derivations of the algebra ${\Gamma(\wedge A{^{\vee}}\otimes \hat{S}(E{^{\vee}}))}$ and $$D(\alpha\otimes 1)=(d_A\alpha)\otimes 1=
d_A^{\nabla^\natural}(\alpha\otimes 1) ,$$ we have $$\big(D-d_A^{\nabla^\natural}\big)(\alpha\otimes\sigma)
= (-1)^{k} (\alpha\otimes 1)
\cdot \big(D-d_A^{\nabla^\natural}\big)(1\otimes\sigma)$$ for all $\alpha\in{\Gamma(\wedge^k A{^{\vee}})}$ and $\sigma\in{\Gamma(\hat{S}(E{^{\vee}}))}$. Moreover, since $$\big(D-d_A^{\nabla^\natural}\big)(1\otimes\sigma)
\in{\Gamma(A{^{\vee}}\otimes \hat{S}^{\geqslant k+1}(E{^{\vee}}))}
\quad \forall \sigma\in{\Gamma(S^k(E{^{\vee}}))} ,$$ there exists $${\mathcal{R}}\in{\Gamma(A^\vee\otimes\hat{S}^{\geqslant 2}(E^\vee)\otimes E)}$$ such that $$\big(D-d_A^{\nabla^\natural}\big)(\alpha\otimes\sigma)
= \sum_k (\nu_k\wedge\alpha)\otimes
i_{v_k}{\mathcal{R}}(\sigma) .$$ Here $v_1,\dots,v_l$ and $\nu_1,\dots,\nu_l$ are any pair of dual local frames of the vector bundles $A$ and $A{^{\vee}}$ and $i_{v_k}{\mathcal{R}}$ ought to be seen as a derivation of the algebra ${\Gamma(\hat{S}(E{^{\vee}}))}$. Since $D^2=0$ and $(d_A^{\nabla^\natural})^2=0$, it follows that $d_A^{\nabla^\natural}{\mathcal{R}}+\frac{1}{2}{[{\mathcal{R}},{\mathcal{R}}]}=0$.
(\[cinque\])$\Rightarrow$(\[quattro\])Given a representation $\nabla^\natural$ of $A$ on $E$ and a solution ${\mathcal{R}}$ of the Maurer–Cartan equation, the derivation $D=d_A^{\nabla^\natural}+{\mathcal{R}}$ satisfies all requirements.
(\[cinque\])$\Rightarrow$(\[sei\])Let $\Pi:{\mathfrak{h}}\to{\mathfrak{a}}$ denote the natural projection of the graded Lie algebra $${\mathfrak{h}}=\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n A{^{\vee}}\otimes\hat{S}(E{^{\vee}})\otimes E)}$$ of vertical vector fields on $A[1]\oplus E$ (with Lie bracket given by Equation ) onto its abelian Lie subalgebra $${\mathfrak{a}}=\bigoplus_{n\geqslant 0}{\Gamma(\wedge^n A{^{\vee}}\otimes E)} .$$ The triple $({\mathfrak{h}},{\mathfrak{a}},\Pi)$ is a *V-algebra* in the sense of Cattaneo & Schätz [@Cattaneo-Schaetz Definition 2.3] as $$\Pi{[x,y]}=\Pi{[\Pi x,y]}+\Pi{[x,\Pi y]} ,\quad\forall x,y\in{\mathfrak{h}}.$$ It is simple to check that the derivation $d_A^{\nabla^\natural}+\operatorname{ad}_{{\mathcal{R}}}$ of degree $+1$ of the graded Lie algebra ${\mathfrak{h}}$ is an *adapted derivation* in the sense of Cattaneo & Schätz [@Cattaneo-Schaetz Definition 2.4].
According to Cattaneo & Schätz [@Cattaneo-Schaetz Theorem 2.5] and Voronov [@Voronov], the *higher derived brackets* $\lambda_k:S^k{\mathfrak{a}}\to{\mathfrak{a}}[1]$ defined by $\lambda_0=\Pi({\mathcal{R}})=0$ and, for $k\in{\mathbb{N}}$, $$\lambda_k(x_1,\cdots,x_k)=\Pi{[{[\cdots{[{[(d_A^{\nabla^\natural}+\operatorname{ad}_{{\mathcal{R}}}\big)x_1,x_2]},x_3]}\cdots,x_{k-1}]},x_k]}$$ make ${\mathfrak{a}}={\Gamma(\wedge A{^{\vee}}\otimes E)}$ an $L_\infty[1]$ algebra.
Straightforward computations show that $\lambda_1=d_A^\nabla$ and that each $\lambda_k$ with $k\geqslant 2$ is ${\Gamma(\wedge (A{^{\vee}}))}$-multilinear and satisfies $$\label{Switzerland}
\lambda_k(\xi_1\otimes e_1,\cdots,\xi_k\otimes e_k)
=(-1)^{{\left| \xi_1 \right|}+\cdots+{\left| \xi_k \right|}}
\xi_1\wedge\cdots\wedge\xi_k\wedge{\mathcal{R}}_k(e_1,\cdots,e_k)
,$$ for all $e_1,\dots,e_k\in{\Gamma(E)}$ and all homogeneous elements $\xi_1,\dots,\xi_k$ of ${\Gamma(\wedge A{^{\vee}})}$. Here ${\mathcal{R}}_k$ denotes the component of ${\mathcal{R}}$ in ${\Gamma(A{^{\vee}}\otimes S^k(E{^{\vee}})\otimes E)}$.
(\[sei\])$\Rightarrow$(\[cinque\])For $k\geqslant 2$, the multibracket $\lambda_k$ is ${\Gamma(\wedge A{^{\vee}})}$-multilinear and hence satisfies Equation for some ${\mathcal{R}}_k\in{\Gamma(A{^{\vee}}\otimes S^k(E{^{\vee}})\otimes E)}$. It follows from the generalized Jacobi identity satisfied by the sequence of multibrackets $(\lambda_k)_{k\in{\mathbb{N}}}$ that the fiberwise vertical vector field ${\mathcal{R}}:=\sum_{k\geqslant 2}{\mathcal{R}}_k\in{\mathfrak{h}}$ satisfies $$\Pi{[\cdots{[{[d_A^{\nabla^\natural}+\operatorname{ad}_{\mathcal{R}},d_A^{\nabla^\natural}+\operatorname{ad}_{\mathcal{R}}]}(e_1),e_2]}\cdots,e_n]} = 0
,\quad\forall e_1,\cdots,e_n\in{\Gamma(E)} .$$ Therefore, we have $$0={[d_A^{\nabla^\natural}+\operatorname{ad}_{\mathcal{R}},d_A^{\nabla^\natural}+\operatorname{ad}_{\mathcal{R}}]}=2 \operatorname{ad}_{d_A^{\nabla^\natural}{\mathcal{R}}+\frac{1}{2}{[{\mathcal{R}},{\mathcal{R}}]}} .\qedhere$$
According to (the proof of) Theorem \[Tasmania\], the homological vector field on a Kapranov dg-manifold $A[1]\oplus E$ is the sum $D=d_A^{\nabla^\natural}+{\mathcal{R}}$ of the Chevalley–Eilenberg differential induced by a representation $\nabla^\natural$ of the Lie algebroid $A$ on the vector bundle $E$ and a solution ${\mathcal{R}}=\sum_{k\geqslant 2}{\mathcal{R}}_k$, with ${\mathcal{R}}_k\in{\Gamma(A{^{\vee}}\otimes S^k(E{^{\vee}})\otimes E)}$, of the Maurer–Cartan equation $d_A^{\nabla^\natural}{\mathcal{R}}+{[{\mathcal{R}},{\mathcal{R}}]}=0$. Projecting the Maurer–Cartan equation on ${\Gamma(\wedge A{^{\vee}}\otimes S^2(E{^{\vee}})\otimes E)}$, we obtain $d_A^{\nabla^\natural}{\mathcal{R}}_2=0$. Hence ${\mathcal{R}}_2\in{\Gamma(A{^{\vee}}\otimes S^2(E{^{\vee}})\otimes E)}$ is a Chevalley–Eilenberg $1$-cocycle with respect to the representation of $A$ on $S^2(E{^{\vee}})\otimes E$ induced by $\nabla^\natural$. We have shown that the homological vector field of a Kapranov dg-manifold $A[1]\oplus E$ determines a cohomology class $[{\mathcal{R}}_2]$ in $H_{\operatorname{CE}}^1(A;S^2(E{^{\vee}})\otimes E)$.
\[Tagalog\] Suppose $D$ and $D'$ are two homological vector fields on the graded manifold $A[1]\oplus E$, which coincide along the submanifold $A[1]$ and endow $A[1]\oplus E$ with two distinct Kapranov dg-manifold structures. There exist natural bijections between the following classes of morphisms.
1. Automorphisms of the graded manifold $A[1]\oplus E$ mapping $D$ to $D'$ and fixing the dg-submanifold $A[1]$.
2. Automorphisms of the $R$-coalgebra ${\Gamma(S(E))}$ respecting the increasing filtration $${\Gamma(S^{\leqslant 0}(E))} \subset {\Gamma(S^{\leqslant 1}(E))} \subset
{\Gamma(S^{\leqslant 2}(E))} \subset {\Gamma(S^{\leqslant 3}(E))} \subset \cdots$$ and intertwining the infinitesimal $A$-actions by coderivations corresponding to $D$ and $D'$.
3. Automorphisms of the $R$-algebra ${\Gamma(\hat{S}(E{^{\vee}}))}$ respecting the decreasing filtration $$\cdots{\hookrightarrow}{\Gamma(\hat{S}^{\geqslant n+1}(E{^{\vee}}))}
{\hookrightarrow}{\Gamma(\hat{S}^{\geqslant n}(E{^{\vee}}))}{\hookrightarrow}\cdots ,$$ and intertwining the infinitesimal $A$-actions by derivations corresponding to $D$ and $D'$.
An automorphism $\Phi$ of the filtered $R$-algebra ${\Gamma(\hat{S}(E{^{\vee}}))}$ determines a sequence $(\phi_k)_{k\in{\mathbb{N}}}$ of bundle maps $\phi_k:E{^{\vee}}\to S^k(E{^{\vee}})$ such that $\phi_1$ is invertible: $$\Phi(\varepsilon)=\sum_{k=1}^\infty \phi_k(\varepsilon)
,\quad\forall{\varepsilon}\in{\Gamma(E{^{\vee}})} .$$ Abusing notations, we write $\Phi=(\phi_k)_{k\in{\mathbb{N}}}$.
Let $\Phi=(\phi_k)_{k\in{\mathbb{N}}}$ be an isomorphism of Kapranov dg-manifolds from $(A[1]\oplus E,d_A^{\nabla{^{\sharp}}}+\sum_{k\geqslant 2}{\mathcal{R}}{^{\sharp}}_k)$ to $(A[1]\oplus E,d_A^{\nabla{^{\flat}}}+\sum_{k\geqslant 2}{\mathcal{R}}{^{\flat}}_k)$ fixing the submanifold $A[1]$. Then $\phi_1: (E,\nabla{^{\sharp}})\to (E,\nabla{^{\flat}})$ is an isomorphism of $A$-modules and $$\phi_1^*: H_{\operatorname{CE},\nabla{^{\flat}}}^1(A;S^2(E{^{\vee}})\otimes E)\to
H_{\operatorname{CE},\nabla{^{\sharp}}}^1(A;S^2(E{^{\vee}})\otimes E)$$ maps $[{\mathcal{R}}{^{\flat}}_2]$ to $[{\mathcal{R}}{^{\sharp}}_2]$.
It follows from Equation that $\phi_1$ is an isomorphism of $A$-modules. Therefore, we may assume without loss of generality that $\phi_1=\operatorname{id}$ and $\nabla{^{\flat}}=\nabla{^{\sharp}}$ and it suffices to prove that $[{\mathcal{R}}{^{\sharp}}_2]=[{\mathcal{R}}{^{\flat}}_2]$ in $H_{\operatorname{CE}}^1(A;S^2(E{^{\vee}})\otimes E)$. For all $\varepsilon\in{\Gamma(E{^{\vee}})}$, we have $$(1\otimes\Phi)\Big(\big(d_A^\nabla+{\mathcal{R}}{^{\sharp}}_2+\sum_{k\geqslant 3}{\mathcal{R}}{^{\sharp}}_k\big)(\varepsilon)\Big)
=(d^{\nabla}_A+{\mathcal{R}}{^{\flat}}_2+\sum_{k\geqslant 3}{\mathcal{R}}{^{\flat}}_k)\big(\Phi(\varepsilon)\big)$$ in ${\Gamma(A{^{\vee}}\otimes\hat{S}(E^\vee))}$. Projecting onto ${\Gamma(A{^{\vee}}\otimes S^2(E^\vee))}$, we obtain $$\phi_2(d_A^\nabla(\varepsilon))+{\mathcal{R}}{^{\sharp}}_2(\varepsilon) = d_{A}^{\nabla}(\phi_2(\varepsilon))+{\mathcal{R}}{^{\flat}}_2(\varepsilon)$$ and ${\mathcal{R}}{^{\sharp}}_2-{\mathcal{R}}{^{\flat}}_2=d_A^\nabla\phi_2$.
\[Cantonese\] If a Kapranov dg-manifold $A[1]\oplus E$ is linearizable, then $[{\mathcal{R}}_2]$ vanishes in $H_{\operatorname{CE}}^1(A;S^2(E{^{\vee}})\otimes E)$.
Examples
--------
In this section, we describe a simple construction of Kapranov dg-manifolds.
Let $A\to M$ be a Lie algebroid and let $\pi:X\to M$ be a surjective submersion admitting a global section $\sigma:M\to X$. Each of the three spaces ${\Gamma(A)}$, ${\mathfrak{X}}(X)$, and ${\mathfrak{X}}(M)$ is simultaneously a left module over $C^\infty(M)$ and a Lie algebra over ${\mathbb{K}}$. Both the anchor map $\rho:{\Gamma(A)}\to{\mathfrak{X}}(M)$ and the differential of the projection $\pi_*:{\mathfrak{X}}(X)\to{\mathfrak{X}}(M)$ are simultaneously morphisms of left modules over $C^\infty(M)$ and morphisms of Lie algebras over ${\mathbb{K}}$.
An infinitesimal action of the Lie algebroid $A\to M$ on the surjective submersion $\pi:X\to M$ is a map $h:{\Gamma(A)}\to{\mathfrak{X}}(X)$, which is a morphism of left modules over $C^\infty(M)$, is a morphism of Lie algebras over ${\mathbb{K}}$, and makes the diagram $$\begin{tikzcd}[row sep=tiny]
& {\mathfrak{X}}(X) \arrow{dd}{\pi_*} \\
{\Gamma(A)} \arrow{ru}{h} \arrow{rd}[swap]{\rho} & \\
& {\mathfrak{X}}(M)
\end{tikzcd}$$ commute. If $h(a)$ is tangent to $\sigma(M)$ for all $a\in{\Gamma(A)}$, we say that the infinitesimal action $h$ preserves the section $\sigma$.
Suppose $h:{\Gamma(A)}\to{\mathfrak{X}}(E)$ is an infinitesimal action of a Lie algebroid $A\to M$ on a vector bundle $\pi:E\to M$ preserving the zero section. Elements of ${\Gamma(\hat{S}(E{^{\vee}}))}$ may be construed as $\pi$-fiberwise Taylor expansions of functions on $E$ along the zero section. Therefore, every derivation of $C^\infty(E)$ determines a derivation of the algebra ${\Gamma(\hat{S}(E{^{\vee}}))}$ and we obtain a morphism of Lie algebras $$\begin{tikzcd}
{\Gamma(A)} \arrow{r}{h} \arrow[bend right]{rr}[swap]{\delta^*}
&{\mathfrak{X}}(E) \arrow{r} & \operatorname{Der}\big({\Gamma(\hat{S}(E{^{\vee}}))}\big)
.\end{tikzcd}$$ Actually, it suffices that $A$ acts infinitesimally on a neighborhood of the zero section in $E$ in order for $\delta^*$ to be defined. From $\pi_*\circ h=\rho$, it follows that $$h(a)\big[\pi^*f\cdot g\big] = \pi^*(\rho(a)f)\cdot g+\pi^*f\cdot h(a)\big[g\big]$$ for all $a\in{\Gamma(A)}$, $f\in C^\infty(M)$, and $g\in C^\infty(E)$ and, consequently, $$\delta^*_a(f\cdot\sigma)=\rho(a)f\cdot\sigma+f\cdot\delta^*_a\sigma$$ for all $a\in{\Gamma(A)}$, $f\in C^\infty(M)$, and $\sigma\in{\Gamma(\hat{S}(E{^{\vee}}))}$. Thus $\delta^*$ is an infinitesimal action of $A$ on ${\Gamma(\hat{S}(E{^{\vee}}))}$ by derivations. Moreover, for all $a\in{\Gamma(A)}$ and $\varepsilon\in{\Gamma(E{^{\vee}})}$, we have $\delta^*_a\varepsilon \in{\Gamma(\hat{S}^{\geqslant 1}(E{^{\vee}}))}$ since $h(a)$ is tangent to the zero section of $\pi:E\to M$. Therefore, $A[1]\oplus E$ is a Kapranov dg-manifold according to Theorem \[Tasmania\]. Finally, if, for each $a\in{\Gamma(A)}$, the flow of the vector field $h(a)$ preserves the linear structure of the fibers of $\pi:E\to M$, then $E$ is an $A$-module and the Kapranov dg-manifold structure on $A[1]\oplus E$ is linear.
\[Patagonia\] Let $\pi:X\to M$ be a surjective submersion admitting a global section $\sigma:M\to X$ and let $N_\sigma\to M$ denote the normal bundle to the submanifold $\sigma(M)$ inside $X$. Suppose $h:{\Gamma(A)}\to{\mathfrak{X}}(X)$ is an infinitesimal action of a Lie algebroid $A\to M$ on $\pi:X\to M$ preserving the section $\sigma$ and suppose $\psi:N_\sigma\to X$ is a diffeomorphism from a neighborhood of the zero section of $N_\sigma\to M$ to a tubular neighborhood of the submanifold $\sigma(M)$ inside $X$ such that $$\begin{tikzcd}
N_\sigma \arrow{r}{\psi} \arrow{d} & X \arrow{d}{\pi} \\
M \arrow{r}[swap]{\operatorname{id}} & M
\end{tikzcd}$$ commutes. Then the morphism of Lie algebras $$\begin{tikzcd}
{\Gamma(A)} \arrow{r}{h} \arrow[bend right]{rrr}[swap]{\delta^*}
& {\mathfrak{X}}(X) \arrow{r}{\psi^{-1}_*} & {\mathfrak{X}}(N_\sigma) \arrow{r}
& \operatorname{Der}\big({\Gamma(\hat{S}(N_\sigma{^{\vee}}))}\big)
\end{tikzcd}$$ is an infinitesimal action of $A$ on ${\Gamma(\hat{S}(N_\sigma{^{\vee}}))}$ by derivations which respect the filtration $$\cdots \subset \hat{S}^{\geqslant k+1}(N_\sigma{^{\vee}}) \subset \hat{S}^{\geqslant k}(N_\sigma{^{\vee}}) \subset \hat{S}^{\geqslant k-1}(N_\sigma{^{\vee}}) \subset \cdots .$$ Therefore, $A[1]\oplus N_\sigma$ is a Kapranov dg-manifold according to Theorem \[Tasmania\]. It follows from Lemma \[Tagalog\] that the Kapranov dg-manifold structures on $A[1]\oplus N_\sigma$ resulting from the choice of any two distinct identifications $\psi$ and $\psi'$ but the same infinitesimal action $h$ are canonically isomorphic.
Kapranov dg-manifolds stemming from Lie pairs
=============================================
PBW isomorphism, Atiyah class, and Kapranov dg-manifold {#Paris}
-------------------------------------------------------
In this section, given a Lie algebroid pair $(L,A)$, we construct a Kapranov dg-manifold on $A[1]\oplus L/A$ canonical up to isomorphism. To get a rough idea of the principle of the construction, let us consider for a moment a real Lie algebroid pair $(L,A)$, i.e. let us work over the field ${\mathbb{R}}$. In this case, Lie’s fundamental theorems assert the existence of a pair $({\mathscr{L}},{\mathscr{A}})$ of local Lie groupoids, which the Lie functor transforms into $L$ and $A$ respectively. The Lie groupoid ${\mathscr{A}}$ acts from the right on the groupoid ${\mathscr{L}}$ and the resulting homogeneous space ${\mathscr{L}}/{\mathscr{A}}$ obviously contains the unit space $M$ of $({\mathscr{L}},{\mathscr{A}})$ as a distinguished embedded submanifold [@LWX:CMP]. Now the Lie groupoid ${\mathscr{A}}$ acts from the left on ${\mathscr{L}}/{\mathscr{A}}$. Therefore we obtain an infinitesimal $A$-action on ${\mathscr{L}}/{\mathscr{A}}$, which stabilizes $M$. Furthermore, the normal bundle $N_M$ to $M$ inside ${\mathscr{L}}/{\mathscr{A}}$ is naturally isomorphic to $L/A$ and can thus be indentified to a tubular neighborhood of $M$ in ${\mathscr{L}}/{\mathscr{A}}$ using the exponential map developed in Section \[Urumqi\]. Therefore, we may apply the construction of Example \[Patagonia\] and endow $A[1]\oplus L/A$ with a Kapranov dg-manifold structure. This is essentially the geometry underlying our construction. For a Lie algebroid pair $(L,A)$ over an arbitrary field ${\mathbb{K}}$ of characteristic zero, the Lie groupoid picture is not valid anymore. Nevertheless, the construction can still be performed working on algebraic counterparts and thinking of them as ‘formal geometric objects.’
Consider the short exact sequence of vector bundles $$\label{Sydney}
\begin{tikzcd} 0 \arrow{r} & A \arrow{r}{i} & L \arrow{r}{q} & L/A \arrow{r} & 0 \end{tikzcd}$$ and choose an $L$-connection $\nabla$ on $L/A$ extending the Bott $A$-connection on $L/A$ (see Example \[Istambul\]). Its *torsion* is the bundle map ${T^\nabla}:\wedge^2 L\to L/A$ defined by $$\label{Tibet} {T^\nabla}(l_1,l_2)=\nabla_{l_1}q(l_2)-\nabla_{l_2}q(l_1)-q\big({[l_1,l_2]}\big), \quad\forall l_1,l_2\in{\Gamma(L)}
.$$ Since ${T^\nabla}(a,l)=0$ if $a\in{\Gamma(A)}$ and $l\in{\Gamma(L)}$, there exists a unique bundle map $\beta:\wedge^2(L/A)\to L/A$ making the diagram $$\begin{tikzcd} \wedge^2 L \arrow{d}[swap]{q} \arrow{r}{{T^\nabla}} &
L/A \\ \wedge^2 (L/A) \arrow{ru}[swap]{\beta} & \end{tikzcd}$$ commute.
\[Berlin\] Given an $L$-connection $\nabla$ on $L/A$, setting $$\nabla'_{l_1}q(l_2)=\nabla_{l_1}q(l_2)-\tfrac{1}{2}{T^\nabla}(l_1,l_2) ,\quad\forall l_1,l_2\in{\Gamma(L)}$$ defines a torsion-free $L$-connection $\nabla'$ on $L/A$.
The connection $\nabla$ extends by derivation to an $L$-connection on $S(L/A)$, which we also denote by $\nabla$ by abuse of notation. Its *curvature* is the bundle map ${R^\nabla}:\wedge^2 L\to\operatorname{Der}\big(S(L/A)\big)$ defined by $${R^\nabla}(l_1,l_2)={[\nabla_{l_1},\nabla_{l_2}]}-\nabla_{{[l_1,l_2]}} ,\quad\forall l_1,l_2\in{\Gamma(L)} .$$ The sections of $\operatorname{Der}\big(S(L/A)\big)$ are the derivations of the $R$-algebra ${\Gamma(S(L/A))}$.
Now choose a splitting $j:L/A\to L$ of the short exact sequence and consider the bundle map ${\mathcal{Z}}^{\nabla}:A\otimes S^2(L/A)\to L/A$ defined by $${\mathcal{Z}}^{\nabla}(a;b_1\odot b_2)=\tfrac{1}{2}\big\{{R^\nabla}\big(a,j(b_1)\big)b_2+{R^\nabla}\big(a,j(b_2)\big)b_1\big\}$$ for all $a\in{\Gamma(A)}$ and $b_1,b_2\in{\Gamma(L/A)}$. In particular, we have $${\mathcal{Z}}^{\nabla}(a;b^2)={R^\nabla}\big(a,j(b)\big)b, \quad \forall a\in{\Gamma(A)},b\in{\Gamma(L/A)} .$$ Since the Bott $A$-connection on $L/A$ is flat, ${\mathcal{Z}}^{\nabla}$ does not actually depend on the chosen splitting.
1. The element ${\mathcal{Z}}^{\nabla}\in{\Gamma(A{^{\vee}}\otimes S^2(L/A){^{\vee}}\otimes L/A)}$ is a Chevalley–Eilenberg 1-cocycle with respect to the representation of $A$ on $S^2(L/A){^{\vee}}\otimes L/A$ induced by the Bott connection.
2. Given two $L$-connections $\nabla{^{\sharp}}$ and $\nabla{^{\flat}}$ extending the Bott representation of $A$ on $L/A$, the difference ${\mathcal{Z}}^{\nabla{^{\sharp}}}-{\mathcal{Z}}^{\nabla{^{\flat}}}$ is a Chevalley–Eilenberg 1-coboundary with respect to the representation of $A$ on $S^2(L/A){^{\vee}}\otimes L/A$ induced by the Bott connection.
The cohomology class $[{\mathcal{Z}}^{\nabla}]\in H_{\operatorname{CE}}^1\big(A;S^2(L/A){^{\vee}}\otimes L/A\big)$, which is independent of the chosen connection $\nabla$, is called the *Atiyah class* of the Lie pair $(L,A)$ and is denoted $\alpha_{L/A}$.
The definition of the Atiyah class given above agrees with the definition of [@CSX] even though the Atiyah cocycle ${\mathcal{Z}}^\nabla$ defined here is slightly different from (but cohomologous to) the Atiyah cocycle defined in [@CSX]. Indeed, a straightforward computation (involving the Jacobi identity in the Lie algebra ${\Gamma(L)}$) yields $$\big(\nabla_a\beta\big)(b_1,b_2)={R^\nabla}\big(a,j(b_1)\big)b_2- {R^\nabla}\big(a,j(b_2)\big)b_1 ,$$ which implies that $${\mathcal{Z}}^{\nabla}(a;b_1\odot b_2)={R^\nabla}\big(a,j(b_1)\big)b_2-\tfrac{1}{2}\big(\nabla_a\beta\big)(b_1,b_2)$$ and $${\mathcal{Z}}^{\nabla'}(a;b_1\odot b_2)=R^{\nabla'}\big(a,j(b_1)\big)b_2 ,$$ where $\nabla'$ stands for the torsion-free connection obtained from $\nabla$ as in Lemma \[Berlin\].
Given a splitting $$\begin{tikzcd} 0 \arrow{r} & A \arrow{r}{i} & L \arrow[bend left]{l}{p}
\arrow{r}{q} & {L/A} \arrow[bend left]{l}{j} \arrow{r} & 0 \end{tikzcd}$$ of the short exact sequence , i.e. a pair of bundle maps $j:L/A\to L$ and $p:L\to A$ satisfying $p\circ i=\operatorname{id}_A$, $q\circ j=\operatorname{id}_{L/A}$, and $i\circ p+j\circ q=\operatorname{id}_L$, we can define a bundle map $\alpha:\wedge^2(L/A)\to A$ by setting $$\alpha(b_1,b_2)=p\big({[j(b_1),j(b_2)]}\big), \quad\forall b_1,b_2\in{\Gamma(L/A)} .$$
\[Ohio\] Let $(L,A)$ be a Lie pair. Given a splitting $j:L/A\to L$ of the short exact sequence of vector bundles $0\to A\to L\to L/A\to 0$ and an $L$-connection $\nabla$ on $L/A$ extending the Bott representation of $A$ on $L/A$, there exists a Kapranov dg-manifold structure on $A[1]\oplus L/A$ whose homological vector field $D\in{\mathfrak{X}}(A[1]\oplus L/A)$ is a derivation $${\Gamma(\wedge{^{\bullet}}A{^{\vee}}\otimes\hat{S}((L/A){^{\vee}}))}
{\xrightarrow{{D}}} {\Gamma(\wedge^{\bullet+1} A{^{\vee}}\otimes\hat{S}((L/A){^{\vee}}))}$$ of degree +1 of the algebra of functions on $A[1]\oplus L/A$ of the form $${D}=d_{A}^{\nabla^{\operatorname{Bott}}}+\sum_{k=2}^{\infty}{\mathcal{R}}_k ,$$ where
1. $d_{A}^{\nabla^{\operatorname{Bott}}}$ denotes the Chevalley–Eilenberg differential associated to the Bott representation of the Lie algebroid $A$ on the vector bundle $L/A$;
2. ${\mathcal{R}}_2\in{\Gamma(A{^{\vee}}\otimes S^2(L/A){^{\vee}}\otimes L/A)}$ is the Atiyah 1-cocycle ${\mathcal{Z}}^{\nabla}$ associated to the connection $\nabla$;
3. the terms ${\mathcal{R}}_k\in{\Gamma(A{^{\vee}}\otimes S^k(L/A){^{\vee}}\otimes L/A)}$ with $k\geqslant 3$ are algebraic functions of $j$, ${\mathcal{R}}_2$, $\alpha$, $\beta$, ${R^\nabla}$, and their higher order covariant derivatives.
Moreover, the various Kapranov dg-manifold structures on $A[1]\oplus L/A$ resulting from all possible choices of splitting and connection are canonically isomorphic.
Extending the Bott representation of $A$ on $L/A$ to the symmetric algebra $S(L/A)$ by derivation, we obtain an infinitesimal $A$-action (by coderivations) $\nabla^{\operatorname{Bott}}$ on the graded $R$-coalgebra ${\Gamma(S(L/A))}$, which respects the graduation.
The $R$-coalgebra $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ admits an infinitesimal $A$-action as well.
\[Nebraska\] Multiplication from the left by elements of ${\Gamma(A)}$ defines an infinitesimal $A$-action (by coderivations) on the $R$-coalgebra $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$, which preserves the filtration .
It is clear that the multiplication from the left by elements $a\in{\Gamma(A)}$ defines an infinitesmal $A$-action on the left $R$-module $\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ by coderivations. This infinitesimal action preserves the filtration because, for all $a\in{\Gamma(A)}$ and $l_1,l_2,\cdots,l_n\in{\Gamma(L)}$, we have $$a l_1 \cdots l_n =
\sum_{i=1}^{n} \big(l_1 \cdots l_{i-1} {[a,l_i]} l_{i+1} \cdots l_n \big) + l_1 \cdots l_n a ,$$ where every term of the right hand side belongs to $\mathcal{U}^{\leqslant n}(L)$ except for the last which belongs to ${\mathcal{U}(L)}{\Gamma(A)}$.
\[Montana\] Let $(L,A)$ be a Lie pair. The following assertions are equivalent.
1. \[een\] The Kapranov dg-manifold stemming from the Lie pair $(L,A)$ (see Theorem \[Ohio\]) is linearizable.
2. \[twee\] The Atiyah class $\alpha_{L/A}\in H_{\operatorname{CE}}^1(A;S^2(L/A){^{\vee}}\otimes L/A)$ of the Lie pair $(L,A)$ vanishes.
3. \[drie\] For every splitting $j:L/A\to L$, there exists an extension of the Bott $A$-connection on $L/A$ to an $L$-connection $\nabla$ on $L/A$ such that the associated Poincaré–Birkhoff–Witt map $\operatorname{pbw}^{\nabla,j}$ intertwines the infinitesimal $A$-actions on ${\Gamma(S(L/A))}$ and $\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$.
In the case of Lie algebroids over ${\mathbb{R}}$, the following two assertions may be added to the list above.
1. \[vier\] For every splitting $j:L/A\to L$, there exists an extension of the Bott $A$-connection on $L/A$ to an $L$-connection $\nabla$ on $L/A$ such that the fiberwise $\infty$-jet of the associated exponential map $\exp^{\nabla,j}$ along the zero section of $L/A\to M$ intertwines the infinitesimal actions of $A$ on $L/A$ and ${\mathscr{L}}/{\mathscr{A}}$.
2. \[vijf\] There exists a fiber-preserving diffeomorphism $\phi$ from a neighborhood of the zero section in $L/A$ to a neigborhood of the identity section of ${\mathscr{L}}/{\mathscr{A}}$, which fixes the submanifold $M$, and whose fiberwise $\infty$-jet along $M$ intertwines the infinitesimal actions of $A$ on $L/A$ and ${\mathscr{L}}/{\mathscr{A}}$.
Proofs of Theorems \[Ohio\] and \[Montana\]
-------------------------------------------
Every choice of a splitting $j:L/A\to L$ of the short exact sequence and an $L$-connection $\nabla$ on $L/A$ extending the Bott representation of $A$ on $L/A$ determines a ${\mathbb{K}}$-bilinear map $$\Delta:{\Gamma(L/A)}\times{\Gamma(L)}\to{\Gamma(L)}$$ defined by $$\Delta_b l=j(\nabla_l b)-{[l,j(b)]},\quad\forall b\in{\Gamma(L/A)},l\in{\Gamma(L)} .$$
\[Louisiana\] The map $\Delta$ is $R$-linear in its first argument and satisfies $$\Delta_b\big(j(c)\big)=\alpha(b,c)+j\big\{ \nabla_{j(b)} c
- \beta(b,c) \big\},\quad\forall b,c\in{\Gamma(L/A)}
.$$
Substituting $j(b)$ and $j(c)$ for $l_1$ and $l_2$ in Equation , we obtain $$\beta(b,c)=\nabla_{j(b)}c-\nabla_{j(c)}b-q\big({[j(b),j(c)]}\big)$$ and, applying $j$ to both sides, $$j\big(\beta(b,c)\big) =
j\big(\nabla_{j(b)}c\big)-j\big(\nabla_{j(c)}b\big)
-{[j(b),j(c)]}+\alpha(b,c) .$$ The result follows from the definition of $\Delta$.
Likewise, every choice of a splitting $j:L/A\to L$ of the short exact sequence of vector bundles and an $L$-connection $\nabla$ on $L/A$ extending the Bott representation of $A$ on $L/A$ determines a Poincaré–Birkhoff–Witt map $$\operatorname{pbw}^{\nabla,j}:{\Gamma(S(L/A))}\to\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}} ,$$ which is an isomorphism of filtered $R$-coalgebras according to Theorem \[Nairobi\]. In what follows, we will use the simplified symbol $\operatorname{pbw}$ to denote the Poincaré–Birkhoff–Witt isomorphism $\operatorname{pbw}^{\nabla,j}$.
Being a quotient of the universal enveloping algebra ${\mathcal{U}(L)}$ by a left ideal, the $R$-coalgebra $\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ is naturally a left ${\mathcal{U}(L)}$-module. Hence $\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ is endowed with a canonical infinitesimal $L$-action (by coderivations). Pulling back this infinitesimal action through $\operatorname{pbw}$, we obtain an infinitesimal $L$-action $\nabla^\lightning$ on ${\Gamma(S(L/A))}$: $$\label{Vermont} \nabla^\lightning_l(s)=\operatorname{pbw}^{-1}\big(l\cdot\operatorname{pbw}(s)\big) ,$$ for all $l\in{\Gamma(L)}$ and $s\in{\Gamma(S(L/A))}$.
At this stage, it is useful to introduce the bundle map $$\Theta: L\otimes S(L/A)\to S(L/A)$$ defined by $$\label{California}
\nabla^\lightning_l s=q(l)\odot s+\nabla_l s+\Theta(l;s), \quad\forall l\in{\Gamma(L)},s\in{\Gamma(S(L/A))}
.$$
It follows immediately from Equations and that $$\label{Oregon}
\Theta(l;1)=0 ,\quad\forall l\in{\Gamma(L)}$$ since $\operatorname{pbw}(1)=1$ and $\operatorname{pbw}(b)=j(b)$ for all $b\in{\Gamma(L/A)}$.
\[Utah\] For all $l\in{\Gamma(L)}$, the map $s\mapsto\Theta(l;s)$ is a coderivation of the $R$-coalgebra ${\Gamma(S(L/A))}$ which preserves the filtration $$\label{Texas}
\cdots {\hookrightarrow}S^{\leqslant n-1}(L/A) {\hookrightarrow}S^{\leqslant n}(L/A) {\hookrightarrow}S^{\leqslant n+1}(L/A) {\hookrightarrow}\cdots
.$$
The verification that $\nabla^\lightning_l$, $\nabla_l$ and the map $s\mapsto q(l)\odot s$ are coderivations of ${\Gamma(S(L/A))}$ for all $l\in{\Gamma(L)}$ is straightforward. Hence $s\mapsto\Theta(l,s)$ is a coderivation of the $R$-coalgebra ${\Gamma(S(L/A))}$ as well for all $l\in{\Gamma(L)}$. It follows from Equation and Theorem \[shark\] that the coderivation $s\mapsto\Theta(l,s)$ preserves the filtration .
\[Mozambique\] For all $n\in{\mathbb{N}}$ and all $b_0,b_1,\dots,b_n\in{\Gamma(L/A)}$, we have $$\sum_{k=0}^n \Theta\big(j(b_k);b_0\odot\cdots\odot
\widehat{b_k}\odot\cdot\odot b_n\big) =0 .$$
Set $b^{\{k\}}=b_0\odot\cdots\odot
\widehat{b_k}\odot\cdot\odot b_n$ and rewrite Equation as $$(n+1) \operatorname{pbw}(b_0\odot\cdots\odot b_n)
= \sum_{k=0}^n \Big\{j(b_k)\cdot\operatorname{pbw}(b^{\{k\}})
-\operatorname{pbw}\big(\nabla_{j(b_k)}(b^{\{k\}})
\big)\Big\} .$$ Applying $\operatorname{pbw}^{-1}$ to both sides, we obtain $$(n+1)\ b_0\odot\cdots\odot b_n =
\sum_{k=0}^n \Big\{
\nabla^\lightning_{j(b_k)}(b^{\{k\}}) - \nabla_{j(b_k)}(b^{\{k\}})
\Big\} ,$$ which is equivalent to $$\sum_{k=0}^n \Big\{
\nabla^\lightning_{j(b_k)}(b^{\{k\}}) - q\big(j(b_k)\big)\odot b^{\{k\}} - \nabla_{j(b_k)}(b^{\{k\}})
\Big\} =0 .$$ The result follows from Equation .
Taking $b_0=b_1=\dots=b_n=b$ in Proposition \[Mozambique\], we obtain $$\label{Togo}
\Theta\big(j(b);b^n\big)=0
.$$
Taking $b_0=\nabla_l b$ and $b_1=\dots=b_n=b$ in Proposition \[Mozambique\], we obtain $$\label{Angola}
-\Theta\big(j(b);\nabla_l(b^n)\big)
=\Theta\big(j(\nabla_l b);b^n\big)
.$$
Taking $n=m+1$, $b_0=c$, and $b_1=\dots=b_n=b$ in Proposition \[Mozambique\], we obtain $$\label{Brazil}
-\Theta\big(j(b);c\odot b^{m}\big)
=\tfrac{1}{m+1}\Theta\big(j(c);b^{m+1}\big)
.$$
Since $\nabla^\lightning$ is flat and $\nabla^\lightning_l s=q(l)\odot s+\nabla_l s+\Theta(l;s)$, we have $$\begin{split}
0 =&
\nabla^\lightning_l\nabla^\lightning_z s -\nabla^\lightning_z\nabla^\lightning_l s-\nabla^\lightning_{{[l,z]}}s \\
=& T^\nabla(l,z)\odot s +R^\nabla(l,z) s \\
& +q(l)\odot\Theta(z;s)-q(z)\odot\Theta(l;s)
+\nabla_l\big(\Theta(z;s)\big)-\nabla_z\big(\Theta(l;s)\big) \\
& +\Theta\big(l;q(z)\odot s\big)-\Theta\big(z;q(l)\odot s\big)
+\Theta(l;\nabla_z s)-\Theta(z;\nabla_l s) \\
& +\Theta\big(l;\Theta(z;s)\big)-\Theta\big(z;\Theta(l;s)\big)
-\Theta({[l,z]};s)
.\end{split}$$ Subsituting $j(b)$ for $z$ and $b^n$ for $s$ and making use of Equations , , and and Lemma \[Louisiana\], we obtain $$\begin{gathered}
\label{pig}
\Theta(l;b^{n+1})+\tfrac{1}{n+1}\Theta\big(j\circ q(l);b^{n+1}\big) \\
= b\odot\Theta\big(l;b^n\big)
+ \Big\{ \nabla_{j(b)}\big(\Theta(l;b^n)\big)
-\Theta(\Delta_b l;b^n)
-\Theta\big(l;\nabla_{j(b)}(b^n)\big) \Big\} \\
+\Theta\big(j(b);\Theta(l;b^n)\big)
-\beta\big(q(l),b)\odot b^n-R^\nabla\big(l,j(b)\big)(b^n)
.\end{gathered}$$
Consider the bundle maps $${\boldsymbol{R}}:A\otimes S(L/A) \to S(L/A) \quad\text{and}\quad {\boldsymbol{H}}:L/A\otimes S(L/A) \to S(L/A)$$ defined by $$\begin{gathered}
{\boldsymbol{R}}(a;s)=\Theta(i(a);s),\quad\forall a\in{\Gamma(A)},s\in{\Gamma(S(L/A))} ; \\
{\boldsymbol{H}}(b;s)=\Theta(j(b);s),\quad\forall b\in{\Gamma(L/A)},s\in{\Gamma(S(L/A))}
.\end{gathered}$$
It follows from Lemma \[Utah\] that, for all $a\in{\Gamma(A)}$ and $c\in{\Gamma(L/A)}$, the maps $s\mapsto{\boldsymbol{R}}(a;s)$ and $s\mapsto{\boldsymbol{H}}(c;s)$ are $R$-linear coderivations of ${\Gamma(S(L/A))}$ which preserve the filtration .
\[cow\] For all $l\in{\Gamma(L)}$ and $s\in{\Gamma(S(L/A))}$, we have $$\label{donkey} \Theta(l;s)={\boldsymbol{R}}\big(p(l);s\big)+{\boldsymbol{H}}\big(q(l);s\big) .$$ For all $a\in{\Gamma(A)}$, $b\in{\Gamma(L/A)}$, and $n\in{\mathbb{N}}$, we have $$\label{monkey}\begin{aligned}
{\boldsymbol{R}}(a;b^{n+1})=& \ n {\mathcal{Z}}^{\nabla}(a;b^2)\odot b^{n-1}
+b\odot{\boldsymbol{R}}(a;b^n)+{\boldsymbol{H}}\big(b;{\boldsymbol{R}}(a;b^n)\big) \\
& +\Big\{\nabla_{j(b)}\big({\boldsymbol{R}}(a;b^n)\big)-{\boldsymbol{R}}(\Delta_b a;b^n)-{\boldsymbol{R}}\big(a;\nabla_{j(b)}(b^n)\big)\Big\}
.\end{aligned}$$ For all $b,c\in{\Gamma(L/A)}$ and $n\in{\mathbb{N}}$, we have $$\label{snake}\begin{aligned}
\tfrac{n+2}{n+1}{\boldsymbol{H}}(c;b^{n+1})=& \
b\odot{\boldsymbol{H}}(c;b^n)+{\boldsymbol{H}}\big(b;{\boldsymbol{H}}(c;b^n)\big) \\
& +\Big\{\nabla_{j(b)}\big({\boldsymbol{H}}(c;b^n)\big)-{\boldsymbol{H}}(\nabla_{j(b)} c;b^n)-{\boldsymbol{H}}\big(c;\nabla_{j(b)}(b^n)\big)\Big\} \\
& -{\boldsymbol{R}}\big(\alpha(b,c);b^n\big)+{\boldsymbol{H}}\big(\beta(b,c);b^n\big) \\
& +\beta(b,c)\odot b^n+{R^\nabla}\big(j(b),j(c)\big)(b^n)
.\end{aligned}$$
Equation is a trivial consequence of $i\circ p+j\circ q=\operatorname{id}_L$. Taking $l=i(a)$ in Equation yields Equation since $$-{R^\nabla}\big(a,j(b)\big)(b^n)=-n b^{n-1}\odot{R^\nabla}\big(a,j(b)\big)b
=nb^{n-1}\odot{\mathcal{Z}}^{\nabla}(a;b^2) .$$ Finally, to establish Equation , take $l=j(c)$ in and use Lemma \[Louisiana\] to eliminate $\Delta$ from the r.h.s.
\[rabbit\] Since the coderivations ${\boldsymbol{R}}$ and ${\boldsymbol{H}}$ preserve the filtration of $S(L/A)$, Proposition \[cow\] and Equation \[Oregon\] allow us to compute all expressions of the form $\Theta(l;b^n)$ by induction on $n$. Indeed, we have devised an algebraic algorithm for computing $\Theta$ from the connection $\nabla$, its torsion $\beta$, its curvature ${R^\nabla}$, the splitting $j$, and the associated bundle map $\alpha$.
We note that, if $j(L/A)$ is a Lie subalgebroid of $L$ (i.e. $\alpha=0$), ${\boldsymbol{H}}$ may be computed inductively using Equation independently of ${\boldsymbol{R}}$.
We leave the proofs of the next three corollaries to the reader.
\[crab\] If $\alpha=0$, $\beta=0$, and ${R^\nabla}\big(j(c),j(b)\big)=0$ for all $b,c\in{\Gamma(L/A)}$, then ${\boldsymbol{H}}=0$ and $$\begin{gathered}
{\boldsymbol{R}}(a,b^{n+1})= n {\mathcal{Z}}^{\nabla}(a;b^2)\odot b^{n-1} +b\odot{\boldsymbol{R}}(a;b^n) \\
+ \Big\{ \nabla_{j(b)}\big({\boldsymbol{R}}(a;b^n)\big)-{\boldsymbol{R}}(\Delta_b a;b^n)-{\boldsymbol{R}}\big(a;\nabla_{j(b)}(b^n)\big) \Big\}
,\end{gathered}$$ for all $a\in{\Gamma(A)}$, $b\in{\Gamma(L/A)}$, and $n\in{\mathbb{N}}$.
For all $a\in{\Gamma(A)}$ and $b,c\in{\Gamma(L/A)}$, we have $$\begin{gathered}
{\boldsymbol{R}}(a;b)=0 \\
{\boldsymbol{H}}(c;b)=\tfrac{1}{2}\beta(b,c) \\
{\boldsymbol{R}}(a;b^2)=-{R^\nabla}\big(a,j(b)\big) b=-{\mathcal{Z}}^{\nabla}(a;b^2) \\
{\boldsymbol{H}}(c;b^2)=b\odot\beta(b,c)+\tfrac{1}{6}\beta\big(b,\beta(b,c)\big)
+\tfrac{1}{3}\big(\nabla_{j(b)}\beta\big)(b,c)+\tfrac{2}{3}{R^\nabla}\big(j(b),j(c)\big)b
.\end{gathered}$$
\[horse\] The following assertions are equivalent:
1. ${\mathcal{Z}}^\nabla=0$;
2. ${\boldsymbol{R}}=0$;
3. $a\cdot\operatorname{pbw}(s)=\operatorname{pbw}(\nabla_a s)$ for all $a\in{\Gamma(A)}$ and $s\in{\Gamma(S(L/A))}$;
4. $\operatorname{pbw}:{\Gamma(S(L/A))}\to\frac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$ is a morphism of left ${\mathcal{U}(A)}$-modules.
We now return to the central issue of Theorem \[Ohio\]: how to make the graded manifold $A[1]\oplus L/A$ a Kapranov dg-manifold.
\[dolphin\]
1. \[Egypt\] Every isomorphism of filtered $R$-coalgebras $$\psi:{\Gamma(S(L/A))}\to\tfrac{{\mathcal{U}(L)}}{{\mathcal{U}(L)}{\Gamma(A)}}$$ gives rise to a homological vector field $D_\psi$ on $A[1]\oplus L/A$ making it a Kapranov dg-manifold with the Chevalley–Eilenberg differential $d_A$ as homological vector field on the dg-submanifold $A[1]$.
2. \[Armenia\] Given any two such isomorphisms of filtered $R$-coalgebras $\phi$ and $\psi$, there exists a canonical isomorphism of Kapranov dg-manifolds between $(A[1]\oplus L/A,D_{\phi})$ and $(A[1]\oplus L/A,D_{\psi})$ fixing the dg-submanifold $(A[1],d_A)$.
(\[Egypt\])Since $\psi$ respects the filtrations, it follows from Lemma \[Nebraska\] that the map $\delta:{\Gamma(A)}\times{\Gamma(S(L/A))}\to{\Gamma(S(L/A))}$ defined by $$\delta_a(s)=\psi^{-1}(a\cdot\psi(s)), \quad\forall a\in{\Gamma(A)},s\in{\Gamma(S(L/A))}$$ is an infinitesimal $A$-action by coderivations on ${\Gamma(S(L/A))}$, which preserves the filtration. Hence, we have $\delta_a(1)=0$ for all $a\in{\Gamma(A)}$ according to Theorem \[shark\]. The conclusion follows from Theorem \[Tasmania\]. (\[Armenia\]) For every $a\in{\Gamma(A)}$, the automorphism $\phi^{-1}\circ\psi$ of the filtered $R$-coalgebra ${\Gamma(S(L/A))}$ intertwines the coderivations $\phi^{-1}\big(a\cdot\phi(-)\big)$ and $\psi^{-1}\big(a\cdot\psi(-)\big)$. Therefore, by Lemma \[Tagalog\], there exists a diffeomorphism of the graded manifold $A[1]\oplus L/A$ that fixes the submanifold $A[1]$ and maps the homological vector field $D_\phi$ onto $D_\psi$.
Forearmed with Proposition \[dolphin\], we can now prove Theorem \[Ohio\].
Taking $\psi=\operatorname{pbw}^{\nabla,j}$ in Proposition \[dolphin\], we obtain the Kapranov dg-manifold $(A[1]\oplus L/A,D_{\operatorname{pbw}^{\nabla,j}})$ associated to the infinitesimal $A$-action by coderivations $$\delta_a(s)=\operatorname{pbw}^{-1}\big(a\cdot\operatorname{pbw}(s)\big)$$ on ${\Gamma(S(L/A))}$. It follows from Equation that $$\nabla^\natural_a b = \operatorname{pr}_1(\delta_a b)
= \operatorname{pr}_1\big(q(a)\odot b+\nabla_a b+\Theta(a;b)\big)
= \operatorname{pr}_1\big(\nabla^{\operatorname{Bott}}_a b+{\boldsymbol{R}}(a;b)\big)=\nabla^{\operatorname{Bott}}_a b .$$ It follows from Equation that $$\begin{split} {\langle i_a{\mathcal{R}}(\varepsilon)|s\rangle}
=& {\langle i_a(D_{\operatorname{pbw}^{\nabla,j}}-d^{\nabla^{\operatorname{Bott}}}_A)(\varepsilon)|s\rangle} \\
=& {\langle \delta^*_a\varepsilon-\nabla^{\operatorname{Bott}}_a\varepsilon|s\rangle} \\
=& -{\langle \varepsilon|\delta_a s-\nabla^{\operatorname{Bott}} s\rangle} \\
=& -{\langle \varepsilon|q(a)\odot s+\Theta(a;s)\rangle} \\
=& -{\langle \varepsilon|{\boldsymbol{R}}(a;s)\rangle} ,\end{split}$$ for all $a\in{\Gamma(A)}$, $\varepsilon\in{\Gamma((L/A){^{\vee}})}$, and $s\in{\Gamma(S(L/A))}$. Therefore, we have $$D_{\operatorname{pbw}^{\nabla,j}} = d_A^{\nabla^{\operatorname{Bott}}} + \sum_{k\geqslant 2} {\mathcal{R}}_k$$ with $${\mathcal{R}}_2(a;b^2)=-\operatorname{pr}_1\big({\boldsymbol{R}}(a;b^2)\big)={\mathcal{Z}}^{\nabla}(a;b^2)$$ and $$\label{whale} {\mathcal{R}}_k(a;b^k)=-\operatorname{pr}_1\big({\boldsymbol{R}}(a;b^k)\big) ,\quad\forall k\geqslant 3 ,$$ for all $a\in{\Gamma(A)}$ and $b\in{\Gamma(L/A)}$. Remark \[rabbit\] suffices to explain why each ${\mathcal{R}}_k$ is an algebraic function of ${\mathcal{R}}_2$, $\alpha$, $\beta$, $R^\nabla$, and their higher order covariant derivatives.
Finally, given any other choice of connection $\nabla'$ and splitting $j'$, Proposition \[dolphin\] warrants the existence of a canonical isomorphism of Kapranov dg-manifolds between $(A[1]\oplus L/A,D_{\operatorname{pbw}^{\nabla,j}})$ and $(A[1]\oplus L/A,D_{\operatorname{pbw}^{\nabla',j'}})$ fixing the dg-submanifold $(A[1],d_A)$.
It follows from Corollary \[Cantonese\] and Theorem \[Ohio\] that (\[een\]) implies (\[twee\]). If the Atiyah class $\alpha_{L/A}$ vanishes, there exists an $L$-connection $\nabla$ on $L/A$ extending the Bott representation such that the corresponding 1-cocycle ${\mathcal{Z}}^{\nabla}$ vanishes. So the equivalence between (\[twee\]) and (\[drie\]) essentially follows from Corollary \[horse\]. The equivalence between (\[drie\]) and (\[vier\]) follows from Theorem \[Madagascar\]. It is obvious that (\[vier\]) implies (\[vijf\]). Finally, the fact that (\[vijf\]) implies (\[een\]) follows from Example \[Patagonia\].
\[zebra\] Given a Lie pair $(L,A)$, suppose there exists a Lie subalgebroid $B$ of $L$ such that $L=A\oplus B$ and a torsion-free flat $B$-connection $\nabla$ on $B$. Then there exists a Kapranov dg-manifold structure on $A[1]\oplus B$ with homological vector field $D=d_A^{\nabla^{\operatorname{Bott}}}+\sum_{k\geqslant 2}{\mathcal{R}}_k$ where ${\mathcal{R}}_2={\mathcal{Z}}^{\nabla^{\operatorname{Bott}}+\nabla}$ and ${\mathcal{R}}_{n+1}=d_B^{\nabla}{\mathcal{R}}_n$ for $n\geqslant 2$.
The obvious identification of $L/A$ with $B$ is a splitting of the short exact sequence $0 \to A \to L \to L/A \to 0$ and the sum $\nabla^{\operatorname{Bott}}+\nabla$ is an $L$-connection on $L/A$ extending the Bott representation of $A$ on $L/A$. Therefore, we may apply Theorem \[Ohio\] to obtain a Kapranov dg-manifold structure on $A[1]\oplus L/A$. The conclusion follows from Corollary \[crab\] and Equation .
Example from complex geometry
-----------------------------
Every complex manifold $X$ determines a Lie pair $(L=T_X\otimes{\mathbb{C}},A=T{^{0,1}}_X)$ for which $L/A\cong T{^{1,0}}_X$, the Atiyah class $\alpha_{L/A}$ is the classical Atiyah class of the holomorphic tangent bundle $T_X$, and $d_A^{\nabla^{\operatorname{Bott}}}=\overline{\partial}$. Therefore, as a consequence of Theorems \[Ohio\] and \[Montana\], we obtain
Let $X$ be a complex manifold.
1. Every $T{^{1,0}}_X$-connection $\nabla{^{1,0}}$ on $T{^{1,0}}_X$ determines a Kapranov dg-manifold structure on $T{^{0,1}}_X[1]\oplus T{^{1,0}}_X$ with homological vector field $$D=\overline{\partial}+\sum_{k\geqslant 2}{\mathcal{R}}_k ,$$ where ${\mathcal{R}}_2={\mathcal{Z}}^{\nabla^{\operatorname{Bott}}+\nabla{^{1,0}}}$ is the Atiyah cocyle associated to the $T_X\otimes {\mathbb{C}}$-connection $\nabla^{\operatorname{Bott}}+\nabla{^{1,0}}$ on $T{^{1,0}}_X$ and all $${\mathcal{R}}_k \in \Omega^{0, 1}\big(S^k{(T{^{1,0}}_X)}{^{\vee}}\otimes {(T{^{1,0}}_X)}{^{\vee}}\big)$$ with $k\geqslant 3$ are algebraic functions of ${\mathcal{R}}_2$, the curvature of $\nabla{^{1,0}}$, and their higher covariant derivatives.
2. The Dolbeault complex ${{\Omega}}^{0,\bullet}(T{^{1,0}}_X)$ is thus an $L_\infty[1]$ algebra. For $n\geqslant 2$, the $n$-th multibracket $\lambda_n$ is the composition of the wedge product $${{\Omega}}^{0,j_1}(T{^{1,0}}_X)\otimes\cdots\otimes{{\Omega}}^{0,j_n}(T{^{1,0}}_X)\to{{\Omega}}^{0,j_1+\cdots+j_n}(\otimes^n T{^{1,0}}_X)$$ with the map $${{\Omega}}^{0,j_1+\cdots+j_n}(\otimes^n T{^{1,0}}_X)\to{{\Omega}}^{0,j_1+\cdots+j_n+1}(T{^{1,0}}_X)$$ induced by ${\mathcal{R}}_n\in{{\Omega}}{^{0,1}}\big(\operatorname{Hom}(\otimes^n T{^{1,0}}_X,T{^{1,0}}_X)\big)$, while $\lambda_1$ is the Dolbeault operator $\overline{\partial}:\Omega^{0,j}(T{^{1,0}}_X)\to\Omega^{0,j+1}(T{^{1,0}}_X)$.
3. The Kapranov dg-manifold structures on $T{^{0,1}}_X[1]\oplus T{^{1,0}}_X$ resulting from different choices of connections $\nabla{^{1,0}}$ are all canonically isomorphic. Hence the $L_\infty[1]$ algebra structure on the Dolbeault complex ${{\Omega}}^{0,\bullet}(T{^{1,0}}_X)$ resulting from different choices of connections $\nabla{^{1,0}}$ are all canonically isomorphic.
4. This Kapranov dg-manifold $T{^{0,1}}_X[1]\oplus T{^{1,0}}_X$ is linearizable if and only if the Atiyah class of the holomorphic tangent bundle $T_X$ vanishes.
When $X$ is a Kähler manifold, the Levi-Civita connection $\nabla^{\operatorname{LC}}$ induces a $T{^{1,0}}_X$-connection $\nabla{^{1,0}}$ on $T{^{1,0}}_X$ as follows. First, extend the Levi-Civita connection ${\mathbb{C}}$-linearly to a $T_X\otimes{\mathbb{C}}$-connection $\nabla^{\mathbb{C}}$ on $T_X\otimes{\mathbb{C}}$. Since $X$ is Kähler, $\nabla^{\operatorname{LC}}J=0$ and $\nabla^{\mathbb{C}}$ restricts to a $T_X\otimes{\mathbb{C}}$-connection on $T{^{1,0}}_X$. It is easy to check that the induced $T{^{0,1}}_X$-connection on $T{^{1,0}}_X$ is the canonical infinitesimal $T{^{0,1}}_X$-action on $T{^{1,0}}_X$ — a section of $T{^{1,0}}_X$ is $T{^{0,1}}_X$-horizontal if and only if it is holomorphic — while the induced $T{^{1,0}}_X$-connection $\nabla{^{1,0}}$ on $T{^{1,0}}_X$ is flat and torsion free. Thus according to Proposition \[zebra\], we have a simpler formula for ${\mathcal{R}}_n$: $${\mathcal{R}}_{n+1}=d^{\nabla^{1,0}}{\mathcal{R}}_n
,\quad\forall n\geqslant 2 .$$ Thus we recover a classical theorem of Kapranov [@Kapranov].
Appendix: coderivations of the symmetric coalgebra
==================================================
Given a module $V$ over a commutative ring $R$ with identity $1$, deconcatenation (see Equation ) defines a comultiplication $\Delta$ on $S(V)$. For every $v\in V$ and every non-negative integer $n$, we have $$\Delta(v^n)=\sum_{k=0}^{n}\binom{n}{k} v^k\otimes v^{n-k} ,$$ where $v^k$ denotes the symmetric product of $k$ copies of $v$ while $v^0$ denotes the unit element $1$ of the ring $R$. Thus $\big(S(V),\Delta\big)$ is a coalgebra.
The following result is immediate.
\[squirrel\] $\ker\big(\Delta-(1\otimes\operatorname{id}+\operatorname{id}\otimes 1)\big)=S^1(V)$
\[chicken\] If $\delta\in\operatorname{End}\big(S(V)\big)$ is a coderivation of the symmetric coalgebra $\big(S(V),\Delta\big)$, then $\delta(1)\in S^1(V)$.
We have $\delta(1)=\varepsilon+v+H$ for some $\varepsilon\in R$, $v\in V$, and $H\in S^{\geqslant 2}(V)$. Since $\delta$ is a coderivation, we have $$\begin{aligned}
\Delta\circ\delta(1)&=(\delta\otimes\operatorname{id}+\operatorname{id}\otimes\delta)\circ\Delta(1) \\
\Delta(\varepsilon+v+H)&=(\delta\otimes\operatorname{id}+\operatorname{id}\otimes\delta)(1\otimes 1). \end{aligned}$$ Since $\Delta (v)=v\otimes 1+1 \otimes v$ and $\Delta (1)=1 \otimes 1$, it follows that $\varepsilon\cdot 1\otimes 1+\Delta(H)=\varepsilon\cdot 1\otimes 1+
H\otimes 1+\varepsilon\cdot 1\otimes 1+1\otimes H$, which implies that $$\Delta(H)-(1\otimes H+H\otimes 1)=\varepsilon\cdot 1\otimes 1 .$$ In this last equation, the right hand side belongs to $S^0(V)\otimes S^0(V)$ while the left hand side belongs to $S^{\geqslant 1}(V)\otimes S^{\geqslant 1}(V)$. Therefore, we obtain the system of equations $$\left\{
\begin{array}{l}
\Delta(H)-(1\otimes H+H\otimes 1)=0 \\
\varepsilon\cdot 1\otimes 1 =0
\end{array}
\right. .$$ By Lemma \[squirrel\], the only solution is $\varepsilon=0$, $H=0$. Hence $\delta(1)=v\in S^1(V)$.
We are now ready to state the main result.
\[shark\] A coderivation $\delta$ of the symmetric coalgebra $\big(S(V),\Delta\big)$ preserves the increasing filtration $$S^{\leqslant 0}(V) \subset S^{\leqslant 1}(V) \subset S^{\leqslant 2}(V) \subset S^{\leqslant 3}(V) \subset \cdots$$ if and only if $\delta(1)=0$.
Suppose $\delta\big(S^{\leqslant n}(V)\big)\subset S^{\leqslant n}(V)$ for all non-negative integers $n$. Then $\delta(1)\in S^0(V)$ as $1\in S^0(V)$. Since $\delta$ is a coderivation, we must have $\delta(1)\in S^1(V)$ by Lemma \[chicken\]. Therefore $\delta(1)=0$.
Conversely, assume that $\delta(1)=0$. Then, obviously, $\delta\big(S^0(V)\big)=\{0\}\subset S^0(V)$. Given $v\in V$, we have $\delta(v)=\varepsilon+w+H$ for some $\varepsilon\in R$, $w\in V$, and $H\in S^{\geqslant 2}(V)$. Since $\delta$ is a coderivation, we have $\Delta\circ\delta(v)=(\delta\otimes\operatorname{id}+\operatorname{id}\otimes\delta)\circ\Delta(v)$, which implies that $\varepsilon\cdot 1\otimes 1+\Delta(H)=1\otimes\varepsilon+1\otimes H+\varepsilon\otimes 1+H\otimes 1$. Therefore $$\Delta(H)-(1\otimes H+H\otimes 1)=\varepsilon\cdot 1\otimes 1 .$$ In this last equation, the right hand side belongs to $S^0(V)\otimes S^0(V)$ while the left hand side belongs to $S^{\geqslant 1}(V)\otimes S^{\geqslant 1}(V)$. Therefore, we obtain the system of equations $$\left\{
\begin{array}{l}
\Delta(H)-(1\otimes H+H\otimes 1)=0 \\
\varepsilon\cdot 1\otimes 1 =0
\end{array}
\right. .$$ By Lemma \[squirrel\], the only solution is $\varepsilon=0$, $H=0$. Hence $\delta(v)=w\in V$, which shows that $\delta\big(S^1(V)\big)\subset S^1(V)$. For $n\geqslant 2$, we argue by induction. Suppose $\delta(v^k)\in S^{\leqslant k}(V)$ for $0\leqslant k\leqslant n-1$. Given any $v\in V$, we have $\delta(v^n)=x+y$ for some $x\in S^{\leqslant n}(V)$ and $y\in S^{>n}(V)$. Since $\delta$ is a coderivation, we have $$\begin{aligned}
\Delta\circ\delta(v^n)=&(\delta\otimes\operatorname{id}+\operatorname{id}\otimes\delta)\circ\Delta(v^n) \\
\Delta(x+y)=&(\delta\otimes\operatorname{id}+\operatorname{id}\otimes\delta)\big(1\otimes v^n+\sum_{k=1}^{n-1}\binom{n}{k} v^k\otimes v^{n-k}+v^n\otimes 1 \big) \\\end{aligned}$$ which implies that $$\begin{gathered}
\Delta(y)-(1\otimes y+y\otimes 1)=1\otimes x+x\otimes 1-\Delta(x) \\ +\sum_{k=1}^{n-1}\binom{n}{k}\big(\delta(v^k)\otimes v^{n-k}+v^k\otimes\delta(v^{n-k})\big) .\end{gathered}$$ In this last equation, while the left hand side belongs to $\bigoplus_{p+q>n}\big(S^p(V)\otimes S^q(V)\big)$, the right hand side belongs to $\bigoplus_{p+q\leqslant n}\big(S^p(V)\otimes S^q(V)\big)$ by the induction hypothesis. Therefore, $y\in S^{>n}(V)\cap\ker\big(\Delta-(1\otimes\operatorname{id}+\operatorname{id}\otimes 1)\big)$ and it follows from Lemma \[squirrel\] that $y=0$. Hence $\delta(v^n)=x\in S^{\leqslant n}(V)$.
[^1]: In the literature, dg-manifolds are sometimes called $Q$-manifolds [@Kontsevich_formality; @MR1432574]. Non-positive grading is assumed for dg-manifolds in derived geometry [@MR1839580; @MR2496057].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'This paper presents the first combined measurement of the double-differential muon neutrino and antineutrino charged-current cross sections with no pions in the final state on hydrocarbon at the off-axis near detector of the T2K experiment. The data analyzed in this work comprise 5.8$\times$10$^{20}$ and 6.3$\times$10$^{20}$ protons on target in neutrino and antineutrino mode respectively, at a beam energy peak of 0.6 GeV. Using the two measured cross sections, the sum, difference and asymmetry were calculated with the aim of better understanding the nuclear effects involved in such interactions. The extracted measurements have been compared with the prediction from different Monte Carlo generators and theoretical models showing that the difference between the two cross sections have interesting sensitivity to nuclear effects.'
author:
- 'K.Abe'
- 'R.Akutsu'
- 'A.Ali'
- 'C.Alt'
- 'C.Andreopoulos'
- 'L.Anthony'
- 'M.Antonova'
- 'S.Aoki'
- 'A.Ariga'
- 'T.Arihara'
- 'Y.Asada'
- 'Y.Ashida'
- 'E.T.Atkin'
- 'Y.Awataguchi'
- 'S.Ban'
- 'M.Barbi'
- 'G.J.Barker'
- 'G.Barr'
- 'D.Barrow'
- 'C.Barry'
- 'M.Batkiewicz-Kwasniak'
- 'A.Beloshapkin'
- 'F.Bench'
- 'V.Berardi'
- 'L.Berns'
- 'S.Bhadra'
- 'S.Bienstock'
- 'A.Blondel'
- 'S.Bolognesi'
- 'B.Bourguille'
- 'S.B.Boyd'
- 'D.Brailsford'
- 'A.Bravar'
- 'D.Bravo Berguño'
- 'C.Bronner'
- 'A.Bubak'
- 'M.Buizza Avanzini'
- 'J.Calcutt'
- 'T.Campbell'
- 'S.Cao'
- 'S.L.Cartwright'
- 'M.G.Catanesi'
- 'A.Cervera'
- 'A.Chappell'
- 'C.Checchia'
- 'D.Cherdack'
- 'N.Chikuma'
- 'G.Christodoulou'
- 'M.Cicerchia'
- 'J.Coleman'
- 'G.Collazuol'
- 'L.Cook'
- 'D.Coplowe'
- 'A.Cudd'
- 'A.Dabrowska'
- 'G.De Rosa'
- 'T.Dealtry'
- 'P.F.Denner'
- 'S.R.Dennis'
- 'C.Densham'
- 'F.Di Lodovico'
- 'N.Dokania'
- 'S.Dolan'
- 'T.A.Doyle'
- 'O.Drapier'
- 'J.Dumarchez'
- 'P.Dunne'
- 'A.Eguchi'
- 'L.Eklund'
- 'S.Emery-Schrenk'
- 'A.Ereditato'
- 'P.Fernandez'
- 'T.Feusels'
- 'A.J.Finch'
- 'G.A.Fiorentini'
- 'G.Fiorillo'
- 'C.Francois'
- 'M.Friend'
- 'Y.Fujii'
- 'R.Fujita'
- 'D.Fukuda'
- 'R.Fukuda'
- 'Y.Fukuda'
- 'K.Fusshoeller'
- 'C.Giganti'
- 'T.Golan'
- 'M.Gonin'
- 'A.Gorin'
- 'M.Guigue'
- 'D.R.Hadley'
- 'J.T.Haigh'
- 'P.Hamacher-Baumann'
- 'M.Hartz'
- 'T.Hasegawa'
- 'S.Hassani'
- 'N.C.Hastings'
- 'T.Hayashino'
- 'Y.Hayato'
- 'A.Hiramoto'
- 'M.Hogan'
- 'J.Holeczek'
- 'N.T.Hong Van'
- 'F.Iacob'
- 'A.K.Ichikawa'
- 'M.Ikeda'
- 'T.Ishida'
- 'T.Ishii'
- 'M.Ishitsuka'
- 'K.Iwamoto'
- 'A.Izmaylov'
- 'M.Jakkapu'
- 'B.Jamieson'
- 'S.J.Jenkins'
- 'C.Jesús-Valls'
- 'M.Jiang'
- 'S.Johnson'
- 'P.Jonsson'
- 'C.K.Jung'
- 'X.Junjie'
- 'M.Kabirnezhad'
- 'A.C.Kaboth'
- 'T.Kajita'
- 'H.Kakuno'
- 'J.Kameda'
- 'D.Karlen'
- 'K.Kasetti'
- 'Y.Kataoka'
- 'T.Katori'
- 'Y.Kato'
- 'E.Kearns'
- 'M.Khabibullin'
- 'A.Khotjantsev'
- 'T.Kikawa'
- 'H.Kikutani'
- 'H.Kim'
- 'S.King'
- 'J.Kisiel'
- 'A.Knight'
- 'A.Knox'
- 'T.Kobayashi'
- 'L.Koch'
- 'T.Koga'
- 'A.Konaka'
- 'L.L.Kormos'
- 'Y.Koshio'
- 'A.Kostin'
- 'K.Kowalik'
- 'H.Kubo'
- 'Y.Kudenko'
- 'N.Kukita'
- 'S.Kuribayashi'
- 'R.Kurjata'
- 'T.Kutter'
- 'M.Kuze'
- 'L.Labarga'
- 'J.Lagoda'
- 'M.Lamoureux'
- 'M.Laveder'
- 'M.Lawe'
- 'M.Licciardi'
- 'T.Lindner'
- 'R.P.Litchfield'
- 'S.L.Liu'
- 'X.Li'
- 'A.Longhin'
- 'L.Ludovici'
- 'X.Lu'
- 'T.Lux'
- 'L.N.Machado'
- 'L.Magaletti'
- 'K.Mahn'
- 'M.Malek'
- 'S.Manly'
- 'L.Maret'
- 'A.D.Marino'
- 'L.Marti-Magro'
- 'J.F.Martin'
- 'T.Maruyama'
- 'T.Matsubara'
- 'K.Matsushita'
- 'V.Matveev'
- 'K.Mavrokoridis'
- 'E.Mazzucato'
- 'M.McCarthy'
- 'N.McCauley'
- 'J.McElwee'
- 'K.S.McFarland'
- 'C.McGrew'
- 'A.Mefodiev'
- 'C.Metelko'
- 'M.Mezzetto'
- 'A.Minamino'
- 'O.Mineev'
- 'S.Mine'
- 'M.Miura'
- 'L.Molina Bueno'
- 'S.Moriyama'
- 'J.Morrison'
- 'Th.A.Mueller'
- 'L.Munteanu'
- 'S.Murphy'
- 'Y.Nagai'
- 'T.Nakadaira'
- 'M.Nakahata'
- 'Y.Nakajima'
- 'A.Nakamura'
- 'K.G.Nakamura'
- 'K.Nakamura'
- 'S.Nakayama'
- 'T.Nakaya'
- 'K.Nakayoshi'
- 'C.Nantais'
- 'C.E.R.Naseby'
- 'T.V.Ngoc'
- 'K.Niewczas'
- 'K.Nishikawa'
- 'Y.Nishimura'
- 'T.S.Nonnenmacher'
- 'F.Nova'
- 'P.Novella'
- 'J.Nowak'
- 'J.C.Nugent'
- 'H.M.O’Keeffe'
- 'L.O’Sullivan'
- 'T.Odagawa'
- 'K.Okumura'
- 'T.Okusawa'
- 'S.M.Oser'
- 'R.A.Owen'
- 'Y.Oyama'
- 'V.Palladino'
- 'J.L.Palomino'
- 'V.Paolone'
- 'W.C.Parker'
- 'S.Parsa'
- 'J.Pasternak'
- 'P.Paudyal'
- 'M.Pavin'
- 'D.Payne'
- 'G.C.Penn'
- 'L.Pickering'
- 'C.Pidcott'
- 'G.Pintaudi'
- 'E.S.Pinzon Guerra'
- 'C.Pistillo'
- 'B.Popov'
- 'K.Porwit'
- 'M.Posiadala-Zezula'
- 'A.Pritchard'
- 'B.Quilain'
- 'T.Radermacher'
- 'E.Radicioni'
- 'B.Radics'
- 'P.N.Ratoff'
- 'E.Reinherz-Aronis'
- 'C.Riccio'
- 'E.Rondio'
- 'S.Roth'
- 'A.Rubbia'
- 'A.C.Ruggeri'
- 'C.Ruggles'
- 'A.Rychter'
- 'K.Sakashita'
- 'F.Sánchez'
- 'G.Santucci'
- 'C.M.Schloesser'
- 'K.Scholberg'
- 'J.Schwehr'
- 'M.Scott'
- 'Y.Seiya'
- 'T.Sekiguchi'
- 'H.Sekiya'
- 'D.Sgalaberna'
- 'R.Shah'
- 'A.Shaikhiev'
- 'F.Shaker'
- 'A.Shaykina'
- 'M.Shiozawa'
- 'W.Shorrock'
- 'A.Shvartsman'
- 'A.Smirnov'
- 'M.Smy'
- 'J.T.Sobczyk'
- 'H.Sobel'
- 'F.J.P.Soler'
- 'Y.Sonoda'
- 'J.Steinmann'
- 'S.Suvorov'
- 'A.Suzuki'
- 'S.Y.Suzuki'
- 'Y.Suzuki'
- 'A.A.Sztuc'
- 'M.Tada'
- 'M.Tajima'
- 'A.Takeda'
- 'Y.Takeuchi'
- 'H.K.Tanaka'
- 'H.A.Tanaka'
- 'S.Tanaka'
- 'L.F.Thompson'
- 'W.Toki'
- 'C.Touramanis'
- 'T.Towstego'
- 'K.M.Tsui'
- 'T.Tsukamoto'
- 'M.Tzanov'
- 'Y.Uchida'
- 'M.Vagins'
- 'S.Valder'
- 'Z.Vallari'
- 'D.Vargas'
- 'G.Vasseur'
- 'C.Vilela'
- 'W.G.S.Vinning'
- 'T.Vladisavljevic'
- 'V.V.Volkov'
- 'T.Wachala'
- 'J.Walker'
- 'J.G.Walsh'
- 'Y.Wang'
- 'D.Wark'
- 'M.O.Wascko'
- 'A.Weber'
- 'R.Wendell'
- 'M.J.Wilking'
- 'C.Wilkinson'
- 'J.R.Wilson'
- 'R.J.Wilson'
- 'K.Wood'
- 'C.Wret'
- 'Y.Yamada'
- 'K.Yamamoto'
- 'C.Yanagisawa'
- 'G.Yang'
- 'T.Yano'
- 'K.Yasutome'
- 'S.Yen'
- 'N.Yershov'
- 'M.Yokoyama'
- 'T.Yoshida'
- 'M.Yu'
- 'A.Zalewska'
- 'J.Zalipska'
- 'K.Zaremba'
- 'G.Zarnecki'
- 'M.Ziembicki'
- 'E.D.Zimmerman'
- 'M.Zito'
- 'S.Zsoldos'
- 'A.Zykova'
bibliography:
- 'bibliography.bib'
title: 'First combined measurement of the muon neutrino and antineutrino charged-current cross section without pions in the final state at T2K'
---
[^1]
[^2]
[^3]
[^4]
[^5]
[^6]
[^7]
[^8]
[^9]
[^10]
[^11]
[^12]
[^13]
[^14]
[^15]
[^16]
[^17]
[^18]
[^19]
[^20]
[^21]
[^22]
[^23]
[^24]
[^25]
[^26]
[^27]
[^28]
[^29]
[^30]
[^31]
[^32]
[^33]
[^34]
[^35]
[^36]
Introduction {#sec:Intro}
============
Current and future long-baseline neutrino oscillation experiments have as primary goals the measurements of the Charge-Parity (CP) violating phase ([$\delta_\text{CP}$]{}), the neutrino mass ordering and the octant determination of the mixing angle $\theta_{23}$ [@Abe:2011ks; @Adamson:2016tbqf; @Abe:2014oxa; @Abi:2018dnh]. To this end, the associated systematic error must be minimized. At present, the limited knowledge of (anti)neutrino-nucleus interactions dominates the uncertainties [@Abe:2018wpn; @Adamson:2017zcg]. The main obstacles behind a better understanding of such interactions are a result of limited modeling of the nuclear dynamics and the difficulties in measuring its effect on the cross section. Despite theoretical and experimental efforts in investigating the (anti)neutrino-nucleus cross section during the last decade, a comprehensive picture has not yet emerged [@Katori:2016yel; @Alvarez-Ruso:2014bla].
Measured values of the muon neutrino and antineutrino charged-current quasi elastic scattering (CCQE) cross sections at K2K [@Gran:2006jn], MiniBooNE [@AguilarArevalo:2010zc; @Aguilar-Arevalo:2013dva], MINOS [@Adamson:2014pgc] and SciBooNE [@AlcarazAunion:2009ku], and more recently by T2K [@Abe:2014iza; @Abe:2016tmq; @Abe:2017rfw; @Abe:2018pwo] and MINERVA [@Fiorentini:2013ezn; @Fields:2013zhk; @Walton:2014esl; @Rodrigues:2015hik; @Betancourt:2017uso; @Patrick:2018gvi; @Gran:2018fxa; @Lu:2018stk; @Ruterbories:2018gub] were found to be higher than predictions obtained using the Relativistic Fermi Gas (RFG) nuclear model. The results favored a higher value of the nucleon axial mass (M$_A^{QE}$) than those previously measured in bubble chamber experiments using deuterium as targets and pion electroproduction data [@Kuzmin:2017bzt; @Bodek:2007ym; @Bernard:2001rs]. Furthermore, the CCQE muon neutrino and antineutrino cross sections measured by the NOMAD Collaboration at energies above 3 GeV are in agreement with a value of M$_A^{QE}$ around 1 GeV/c$^2$ [@Lyubushkin:2008pe]. This discrepancy highlighted the need for a more detailed description of the (anti)neutrino-nucleus scattering in the few-GeV energy region.
In a muon neutrino CCQE interaction a negatively charged muon and proton are produced via *W* exchange with a neutron, while in an antineutrino interaction of the same type a positively charged muon and neutron are produced via *W* exchange with a proton:
$$\begin{aligned}
\centering
&{\ensuremath{\nu_\mu}\xspace}& + n \rightarrow \mu^- + p \nonumber\\
&{\ensuremath{\bar{\nu}_\mu}\xspace}& + p \rightarrow \mu^+ + n.\end{aligned}$$
As modern long-baseline neutrino experiments use relatively heavy nuclei as targets, nuclear dynamics plays an important role in the interpretation of the (anti)neutrino oscillations. Several theoretical models have proposed that these effects may explain the observed anomalies between data and Monte Carlo (MC) simulations [@Martini:2009uj; @Nieves:2011yp; @Benhar:2015ula; @Ankowski:2012ei; @Butkevich:2009cp; @Bodek:2016abf; @Leitner:2008ue; @Maieron:2003df; @Meucci:2003cv; @Lovato:2013cua; @Pandey:2014tza; @Amaro:2010sd]. If nuclear effects are considered, the particles produced in the (anti)neutrino-nucleus interaction can interact with other nucleons before exiting the nucleus. These, so-called Final State Interactions (FSI), can alter the type, number and kinematics of particles that exit the nucleus after such interactions. For example, in a muon neutrino resonant pion production interaction a pion, a proton and a muon are produced. The pion could be reabsorbed by the nucleus, with the result that only the muon and the proton are observed to exit the nucleus. This would be indistinguishable from a CCQE interaction. Anyway, the observed discrepancies between data and models cannot be explained by FSI alone. Martini *et al.* [@Martini:2009uj] indicates that further contributions to CCQE-like processes arise from two (or more) interacting nucleons, referred to as 2p2h excitations or multi-nucleon knockout. Such interactions eject low-energy nucleons (200-500 MeV), which cannot be easily detected. Multi-nucleon knockout is expected to be less significant for antineutrinos relative to neutrinos; in particular it has a different role in the vector-axial interference term which differs by a sign for the neutrino and antineutrino cross section [@Martini:2010ex]. Studying differences between CCQE-like cross sections for neutrino and antineutrino interactions could provide information about the role of multi-nucleon knockout in (anti)neutrino-nucleus interactions. The sum and the difference of the neutrino and antineutrino CCQE-like cross sections could yield this information. In the sum, the axial-vector interference term is eliminated whereas the difference isolates this term [@Bernard:2001rs]. In Ref. [@Ericson:2015cva], the predicted sum and difference of the neutrino and antineutrino CCQE-like cross-sections are compared with the equivalent values computed using the CCQE-like double-differential cross sections obtained by the MiniBooNE experiment. The analysis found that additional nuclear effects, other than FSI, would be needed to explain the discrepancy between the observed and predicted values of the sum and difference. However, the analysis in Ref. [@Ericson:2015cva] was limited as the neutrino and antineutrino beams peaked at different energies and the two cross sections were measured independently, implying that correlations between the two data sets were not taken into account.
A more rigorous analysis can be performed at the T2K near detector complex. Data has been taken with neutrino and antineutrino beams, both of which peak at the same energy. Combining the two data sets can exploit the correlation between them leading to a more precise cross-section measurement. This paper reports the first combined measurement of the double differential neutrino and antineutrino charged current cross sections on hydrocarbon without pions in the final state. This CCQE-like cross section will include contributions from CCQE events as well as events in which a pion was produced and then reabsorbed by the nucleus and multi-nucleon knockout events. The neutrino and antineutrino cross sections were used to compute the sum, difference and asymmetry. The neutrino-antineutrino cross-section asymmetry, which is the ratio between the difference and the sum, is a crucial quantity to control in order to avoid biases in the search for CP violation in neutrino oscillation. All these quantities have been compared with predictions made using several MC generators and models, which are discussed in this paper.
The paper is organized as follows. The T2K experiment is described in \[sec:T2Kexp\]. The data and MC samples used in this analysis are reported in \[sec:datamc\]. \[sec:anaStrategy\] describes the analysis procedure, including the event selection and the cross-section extraction method. Finally the results and their interpretation are discussed in \[sec:resultscomp\], followed by conclusions in \[sec:conclusions\].
The T2K experiment {#sec:T2Kexp}
==================
The Tokai-to-Kamioka (T2K) experiment [@Abe:2011ks] is a long-baseline experiment that studies neutrino oscillations in an accelerator-produced $\nu_{\mu}$ ($\bar{\nu}_{\mu}$) beam. The neutrino beam, produced by the J-PARC facility, utilizes a 30GeV proton beam. A proton spill consisting of 8 bunches with 580ns spacing is produced every 2.48s. At a beam power of 430kW, this spill and repetition rate correspond to $2.25\times 10^{14}$ protons on target (p.o.t) per spill. Secondary hadrons, mainly pions and kaons, are produced when the proton beam interacts with a graphite target. Three magnetic horns are used to perform focusing and charge selection of the pions and kaons. The polarity of the magnetic horns can be changed to select positively (forward horn current) or negatively (reverse horn current) charged pions and kaons to produce a beam that is predominantly made of $\nu_{\mu}$ in the forward horn current case or $\bar{\nu}_{\mu}$ for the reverse horn current. The selected hadrons decay in a 96m long decay volume, to produce a (anti)neutrino beam whose direction is parallel to that of the initial proton beam.
Both the neutrino and antineutrino beam consist of a mixture of $\nu_{\mu}, \bar{\nu}_{\mu}, \nu_e$ and $\bar{\nu}_{e}$. The compositions of the neutrino and antineutrino beams are shown in \[beam\]. In the neutrino beam mode, the “right-sign” $\nu_{\mu}$ (and $\nu_e$) flux is around 15% higher around the flux peak when compared with the right-sign $\bar{\nu}_{\mu}$ (and $\bar{\nu}_e$) flux in the antineutrino mode. The background antineutrino flux is also lower in the neutrino mode compared with the neutrino flux in the antineutrino mode, especially at high energy. These differences can be attributed to the higher production multiplicities of positively, rather than negatively, charged parent particles.
The Super-Kamiokande far detector is located $2.5^{\circ}$ off the beam axis, at a distance of 295 km from the production point. The near detector complex, located 280m downstream from the production target, contains two sets of detectors: INGRID and ND280. INGRID [@Abe:2011xv] is on-axis and monitors the flux and direction of the neutrino beam. The ND280 detector is positioned $2.5^{\circ}$ off-axis and is used to study the unoscillated beam. At an off-axis angle of $2.5^{\circ}$, the energy spectrum of the beam is narrowed and centered around 600MeV, which corresponds to the oscillation maximum for a baseline of 295km. In addition, this narrow energy spectrum suppresses the intrinsic $\nu_e$ ($\bar{\nu}_e$) and non-quasi-elastic interactions, leading to lower intrinsic backgrounds to the $\nu_e$ ($\bar{\nu}_e$) appearance search at the far detector.
This work has been performed using the off-axis near detector, ND280. \[nd280\] shows a schematic of such detector. The ND280 detector is formed from five sub-detectors; an upstream $\pi^0$ detector (P$\emptyset$D) [@Assylbekov:2011sh], two Fine-Grained Detectors (FGDs) [@Amaudruz:2012agx], three Time Projection Chambers (TPCs) [@Abgrall:2010hi], Electromagnetic Calorimeter (ECal) [@Allan:2013ofa] and a Side Muon Range Detector (SMRD) [@Aoki:2012mf]. The P$\emptyset$D, FGDs, TPCs and ECal are contained within a magnet that provides a 0.2 T field, whilst the SMRD is embedded in the magnet.
The measurements reported in this paper used the FGDs, TPCs, ECal and SMRD to select charged-current $\nu_{\mu}$ and $\bar{\nu}_{\mu}$ interactions. The most upstream FGD (FGD1) is formed from layers constructed from polystyrene scintillator bars. The scintillator layers are perpendicular to the beam’s direction and alternating layers are orientated orthogonal to each other. The FGD is composed of 86.1% carbon, 7.4% hydrogen, 3.7% oxygen, 1.7% titanium, 1% silicon and 0.1% nitrogen by mass. The active region of FGD1 consists of scintillator layers only, whereas the downstream FGD (FGD2) has alternating layers of scintillator and water. The drift gas mixture used in the TPCs is Ar:CF$_4$:*i*C$_4$H$_{10}$ (95:3:2). The TPCs (TPC1 the most upstream, TPC2 the central and TPC3 the most downstream) provide excellent particle identification and accurate measurement of momentum. Together the TPCs and FGDs form the tracker region of ND280. The ECal surround the tracker and consists of 13 modules made up of plastic scintillator bars alternating with lead sheets. SMRD consists of 440 modules of plastic scintillator counters.
Data and Monte Carlo samples {#sec:datamc}
============================
The studies reported in this paper use $5.80\times10^{20}$ p.o.t forward horn-current ($\nu$-mode) data and $6.27\times 10^{20}$ p.o.t of reverse horn-current ($\bar{\nu}$-mode) data broken into run periods shown in \[tab:runs-pot\].
-------- ----------------------- ------------------- ------------------------
Run Dates $\nu$-mode p.o.t $\bar{\nu}$-mode p.o.t
Period ($\times10^{20}$) ($\times10^{20}$)
Run 2 Nov. 2010 - Mar. 2011 0.79 –
Run 3 Mar. 2012 - Jun. 2012 1.58 –
Run 4 Oct. 2012 - May 2013 3.42 –
Run 5 Jun. 2014 – 0.43
Run 6 Nov. 2014 - Apr. 2015 – 3.40
Run 7 Feb. 2016 - May 2016 – 2.44
Total Nov. 2010 - May 2016 5.80 6.27
-------- ----------------------- ------------------- ------------------------
: T2K neutrino and antineutrino mode runs and their associated p.o.t, filtered for spills where all ND280 detectors were flagged with good data quality.[]{data-label="tab:runs-pot"}
The MC simulation used for this analysis consist of a sample corresponding to ten times the data p.o.t. It is performed generating (anti)neutrino interactions according with the flux predicted at ND280. The simulation of the $\nu_{\mu}$ and $\bar{\nu}_{\mu}$ fluxes reaching the near detector are described in detail in Ref. [@Abe:2012av]. The neutrino and antineutrino interactions in the ND280 sub-detectors, as well as events inside the magnet yoke and in the rock surrounding the ND280 pit, were simulated using the <span style="font-variant:small-caps;">Neut</span> MC generator version `5.3.2` [@Hayato:2002sd]. The CCQE neutrino-nucleon cross section is simulated according to the Llewellyn-Smith formalism [@LlewellynSmith:1971uhs] with a dipole axial form factor and BBBA05 vector form factors [@Bradford:2006yz]. The nuclear model uses a spectral function (SF), developed in Ref. [@Benhar:1994hw] with an axial mass $M_A^{QE} = 1.21$ GeV/c$^2$ based on the K2K measurement of the [$\nu_\mu$]{}CCQE cross section [@Gran:2006jn]. It utilizes the multi-nucleon interaction model (2p2h) from Nieves *et al.* [@Nieves:2011pp] to simulate interactions with nucleon pairs. The model for resonant pion production (RES) is based on the Rein-Sehgal model [@Berger:2007rq] with updated nucleon form factors [@Graczyk:2007bc] and an invariant hadronic mass W $\leq$ 2 GeV. The DIS interaction is calculated for W $>$ 1.3 GeV, using GRV98 parton distribution functions [@Gluck:1998xa] with Bodek-Yang corrections [@Bodek:2003wc]. Single pion production through DIS is suppressed for W $\leq$ 2 GeV to avoid double counting with RES and it uses a custom hadronization model. For values of the invariant hadronic mass W $>$ 2 GeV, <span style="font-variant:small-caps;">Pythia/JetSet</span> [@Sjostrand:1993yb] is used for hadronization. FSI, i.e. interactions of the hadrons produced by neutrino interactions with the other nucleons before leaving the nuclear environment, are simulated using a semiclassical intranuclear cascade model [@Bertini:1972vz; @Oset:1986sy].
The propagation of the final state particles through the ND280 sub-detectors is simulated using the package <span style="font-variant:small-caps;">Geant4</span> version `4.9.4` [@Agostinelli:2002hh] as detailed in Ref. [@Abe:2011ks] employing the following physics lists: `QGSP_BERT` for the hadronic physics, `emstandard_opt3` for the electromagnetic physics and `G4DecayPhysics` for the particle decays.
Analysis strategy {#sec:anaStrategy}
=================
A joint measurement of neutrino and antineutrino cross-sections, fully accounting for correlations in the systematic uncertainties, has been performed. Given the relatively large background of neutrino interactions in the antineutrino sample, such a joint analysis is mandatory for a robust antineutrino cross-section measurement. Indeed, since the neutrino and antineutrino cross-sections are largely driven by the same underlying physics, it would be inconsistent to assume to know the former while measuring the latter. A further advantage of a joint measurement, is that it exploits the full, high-statistics, neutrino sample minimizing the correlated detector and flux systematic uncertainties and thus resulting in a more precise antineutrino measurement. Finally, a joint analysis enables interesting measurements, as explained in \[sec:Intro\].
An unregularized binned likelihood fit with control sample to constrain the background is performed as in Ref. [@Abe:2016tmq; @Abe:2018pwo; @Abe:2018uhf]. This analysis method guarantees a negligible dependence on the signal model used in the simulation for the correction of detector effects, provided that a too coarse binning is not used. A simultaneous fit is applied to the antineutrino sample and the neutrino sample, the former being further sub-divided in different signal and background samples depending on the direction of the outgoing muon, while the latter depending on the kinematics of the outgoing muon and proton. The number of selected events in each bin of reconstructed kinematics ($j$) for each signal and background sample ($s$) is computed as
$$\begin{aligned}
N^s_j = \sum_i^\text{true bins}\Bigg[ & c{_i}{^{{\ensuremath{\nu_\mu}\xspace}}} \left(N^\text{MC {\ensuremath{\nu_\mu}\xspace}{\ensuremath{\text{CC-}0\pi}\xspace}}_i \prod_x w(x)_{i}^{{\ensuremath{\nu_\mu}\xspace}{\ensuremath{\text{CC-}0\pi}\xspace}} \right) + \nonumber \\
& c{_i}{^{{\ensuremath{\bar{\nu}_\mu}\xspace}}} \left(N^\text{MC {\ensuremath{\bar{\nu}_\mu}\xspace}{\ensuremath{\text{CC-}0\pi}\xspace}}_i \prod_x w(x)_{i}^{{\ensuremath{\bar{\nu}_\mu}\xspace}{\ensuremath{\text{CC-}0\pi}\xspace}} \right) + \nonumber \\
& \sum_k^\text{bkg reactions} N^\text{MC bkg $k$}_{i} \prod_b w(b)_{i}^{k} \Bigg]\times \nonumber \\
& t_{ij}^\text{det} d_j \sum_n^{\text{E}_{{\ensuremath{\nu_\mu}\xspace}}\:_\text{or}\:_{{\ensuremath{\bar{\nu}_\mu}\xspace}}} w_n^i f_n
\label{eq:Nreco}\end{aligned}$$
where $N_i^\text{MC}$ is the true number of events in MC with the superscript indicating which interaction type they correspond to. The index $i$ runs over the bins of the “true" muon kinematics prior to detector smearing effects, $k$ runs over the possible background reactions and $n$ runs over the neutrino or antineutrino energy bins. Both $c{_i}{^{{\ensuremath{\nu_\mu}\xspace}}}$ and $c{_i}{^{{\ensuremath{\bar{\nu}_\mu}\xspace}}}$ are the parameters of interest which adjust the [$\nu_\mu$]{}and [$\bar{\nu}_\mu$]{}[$\text{CC-}0\pi$]{}number of events in MC, in order to match the observed number of events in data. The transfer matrix $t_{ij}^\text{det}$, relates the true ($i$) and reconstructed ($j$) muon kinematics bins and $d_j$ represents the nuisance parameters in the fit describing the detector systematics which are constrained by a prior covariance matrix. The flux parameters $f_n$ and weights $w^i_n$, describe the neutrino energy distribution for each bin of $p^{true}_\mu,\cos\theta^{true}_\mu$. The $f_n$ are nuisance parameters in the fit constrained by a prior covariance matrix. The product $\prod_{x,b}$ runs over the systematics related to the theoretical modeling of the interaction channels contributing to the signal ($x$) or the background ($b$). Each $w(x)_{i}^{{\ensuremath{\nu_\mu}\xspace}{\ensuremath{\text{CC-}0\pi}\xspace}}$ and $w(b)^{k}_{i}$ term is a weighting function describing how the generated muon kinematics change (in bins $i$ for each signal and background process) as a function of the value of a particular theoretical parameter. All the parameters $x$ and $b$ are nuisance parameters in the fit and are constrained by a prior covariance matrix. Signal modeling parameters $x$ are not fitted to avoid model dependence but they must be included to account for their effect on the uncertainty of the efficiency corrections. The parameters of interest $c{_i}{^{{\ensuremath{\nu_\mu}\xspace}}}$ reweights the neutrino signal in neutrino mode and neutrino background in the antineutrino mode, whereas $c{_i}{^{{\ensuremath{\bar{\nu}_\mu}\xspace}}}$ reweights the antineutrino background in neutrino mode and antineutrino signal in antineutrino mode. The nuisance parameters may be different in each sample and their correlations between samples are fully taken into account.
The best fit parameters are those that minimize the following log-likelihood:
$$\begin{aligned}
\chi^2 & = \chi^2_\text{stat,$\nu$} + \chi^2_\text{stat,$\bar{\nu}$} + \chi^2_\text{syst} \nonumber\\
& = \sum_j^\text{reco bins} 2\left(N^{{\ensuremath{\nu_\mu}\xspace}}_j-N_j^\text{{\ensuremath{\nu_\mu}\xspace}obs}+N_j^\text{{\ensuremath{\nu_\mu}\xspace}obs} \ln\frac{N_j^\text{{\ensuremath{\nu_\mu}\xspace}obs}}{N^{{\ensuremath{\nu_\mu}\xspace}}_j}\right) \nonumber\\
& + \sum_j^\text{reco bins} 2\left(N^{{\ensuremath{\bar{\nu}_\mu}\xspace}}_j-N_j^\text{{\ensuremath{\bar{\nu}_\mu}\xspace}obs}+N_j^\text{{\ensuremath{\bar{\nu}_\mu}\xspace}obs} \ln\frac{N_j^\text{{\ensuremath{\bar{\nu}_\mu}\xspace}obs}}{N^{{\ensuremath{\bar{\nu}_\mu}\xspace}}_j}\right) \nonumber\\
& + \;\;\;\sum_p\left(\vec{p}-\vec{p}_\text{prior}\right)\left(V^\text{syst}_\text{cov}\right)^{-1}\left(\vec{p}-\vec{p}_\text{prior}\right)
\label{eq:chi2} \end{aligned}$$
where $N_j^{\nu_\mu}$ ($N_j^{\bar{\nu}_\mu}$) is the expected total number of events in the neutrino (antineutrino) sample and $N_j^\text{{\ensuremath{\nu_\mu}\xspace}obs}$ ($N_j^\text{{\ensuremath{\bar{\nu}_\mu}\xspace}obs}$) is the observed number of events. $\chi^2_\text{syst}$ is a penalty term for the systematics, where $\vec{p}$ are the parameters that describe the effect of nuisance parameters, $\vec{p}_\text{prior}$ are the prior values of these systematic parameters and $V^\text{syst}_\text{cov}$ is their covariance matrix which describes the confidence in the nominal parameter values, as well as, correlations between them.
To minimize the dependence of the results on the signal model used in the simulation, two-dimensional differential cross-sections are extracted as a function of muon momentum and angle. Those are kinematic quantities directly observable in the detector and they represent all the relevant variables to characterize the detector acceptance and efficiency. The signal is defined by the absence of pions in the final state, avoiding model-dependent corrections for pion re-absorption in the nucleus.
The flux-integrated cross-sections are evaluated per nucleon and for each bin $i$ of detector unsmeared muon momentum and angle: $$\begin{aligned}
\label{eq:numuxsec}
\frac{\text{d}\sigma_{{\ensuremath{\nu_\mu}\xspace}}}{\text{d}p_\mu \text{d}\cos\theta_\mu}&=& \frac{N^{{\ensuremath{\nu_\mu}\xspace}\text{{\ensuremath{\text{CC-}0\pi}\xspace}}}_i}{\epsilon^{{\ensuremath{\nu_\mu}\xspace}}_i \Phi^{{\ensuremath{\nu_\mu}\xspace}} N^\text{FV}_\text{nucleons}} \times \frac{1}{\Delta p_\mu \Delta \cos\theta_\mu}\\
\frac{\text{d}\sigma_{{\ensuremath{\bar{\nu}_\mu}\xspace}}}{\text{d}p_\mu \text{d}\cos\theta_\mu}&=& \frac{N^{{\ensuremath{\bar{\nu}_\mu}\xspace}\text{{\ensuremath{\text{CC-}0\pi}\xspace}}}_i}{\epsilon^{{\ensuremath{\bar{\nu}_\mu}\xspace}}_i \Phi^{{\ensuremath{\bar{\nu}_\mu}\xspace}} N^\text{FV}_\text{nucleons}} \times \frac{1}{\Delta p_\mu \Delta \cos\theta_\mu}
\end{aligned}$$ where $N^{{\ensuremath{\nu_\mu}\xspace}\text{{\ensuremath{\text{CC-}0\pi}\xspace}}}_i$ and $N^{{\ensuremath{\bar{\nu}_\mu}\xspace}\text{{\ensuremath{\text{CC-}0\pi}\xspace}}}_i$ are the number of neutrino and antineutrino [$\text{CC-}0\pi$]{}events respectively evaluated by the fit, $\epsilon^{{\ensuremath{\nu_\mu}\xspace}}_i$ and $\epsilon^{{\ensuremath{\bar{\nu}_\mu}\xspace}}_i$ are the efficiency evaluated from MC, $N^\text{FV}_\text{nucleons}$ is the number of target nucleons in the fiducial volume, $\Phi^{{\ensuremath{\nu_\mu}\xspace}}$ and $\Phi^{{\ensuremath{\bar{\nu}_\mu}\xspace}}$ are the integrated fluxes for neutrino and antineutrino, $\Delta p_\mu$ and $\Delta \cos\theta_\mu$ are the bin widths of the muon momentum and cosine of the muon scattering angle w.r.t. the bean direction. The number of nucleons, computed using the areal density of the different elements composing the fiducial volume [@Amaudruz:2012agx], is equal to 5.9$\times$10$^{29}$ and it is used to extract both cross sections. The cross sections are normalized in all bins of muon kinematics with the same integrated flux to avoid a model-dependent mapping of such bins into energy intervals of the incoming neutrino.
The binning of the true muon kinematics has been optimized to reduce the bin-by-bin fluctuation derived by the extrapolation of an unsmeared quantity, as the cross-section is, and also to ensure that the systematic uncertainty are smaller than the statistical uncertainty. If the binning is too coarse, the results do not give much information about the shape of the cross section, while on the other hand if the binning is too fine, some bins could be empty causing problems with the minimization algorithm. The best binning lies in between these extreme cases and requires that the bin width is always greater than the resolution of the muon kinematics. A MC sample simulated using the version of <span style="font-variant:small-caps;">Neut</span> described in \[sec:datamc\] has been used as the prior of the fitting algorithm. This choice does not introduce model dependencies as extensively demonstrated in previous analyses [@Abe:2018pwo; @Abe:2018uhf]. The stability of the results has been confirmed by using alternative models in the fitting framework. To this end, a set of mock data samples has been created by modifying the amount of 2p2h interactions, the nuclear or the background model, and the input MC. Through them, it has been verified that the extracted cross section is always in agreement, within the uncertainties, with the mock data set predicted cross section and also produces a small $\chi^2$ computed considering the final cross section covariance matrix.
Event selections {#sec:selections}
----------------
The event selections developed for this analysis aim to select [$\nu_\mu$]{}and [$\bar{\nu}_\mu$]{}[$\text{CC-}0\pi$]{}interactions in the FGD1 and to provide appropriate control samples to constrain the main background sources.
In previous analyses, the selection criteria were optimized to select forward going muons (with respect to the beam direction) originating from FGDs [@Abe:2016tmq; @Abe:2013jth; @Abe:2016aoo; @Abe:2019arf; @Abe:2014iza]. For this analysis, the phase space of the muon kinematics was enlarged, including also high-angle and backward-going tracks. The acceptance has been increased using all the ND280 sub-detectors and the time of flight (ToF) of the particles between different sub-detectors which gives information about the direction of the track, i.e. if it is forward or backward with respect to the beam direction, following the same strategy described in Ref. [@Abe:2018uhf].
In addition to the common goal of enlarging the acceptance, the event selections have several common features:
- The selection criteria have been optimized by employing a MC sample simulated using the version of <span style="font-variant:small-caps;">Neut</span> described in Sec. \[sec:datamc\];
- Particles that enter the TPCs or are fully contained in FGD1 are identified through the TPC or FGD particle identification (PID), based on d$E$/d$x$ measurements;
- ECal PID is performed if there is an associated ECal segment, which reduces the shower-like contamination (mostly $\pi^0$);
- The ratio between the track length and the electromagnetic energy associated with the track is used to reduce the proton contamination;
- Particles stopping in the SMRD are identified as muons, since most likely this is the only particle that will reach this detector.
Each selection applies a set of cuts which have been optimized to give the best signal efficiency and purity. Two requirements are common to both selections:
- Events must occur within the time window of one of the eight beam bunches of the spill structure of the beam and when all ND280 sub-detectors are functioning correctly;
- The interaction vertex, defined as the starting position of the muon candidate, must be inside the FGD1 fiducial volume (FV). Compared with the previous analyses where both a true and a reconstructed vertex in the first two scintillator layers were rejected [@Abe:2016tmq; @Abe:2013jth], in this analysis the full span has been taken as the FV. Depending on the direction of the muon, the events with a reconstructed vertex in the first (forward-going muon) or the last (backward-going muon) layer have been rejected.
In the following sections, the selection strategy is discussed in detail.
### [$\nu_\mu$]{}CC event selection {#sec:numuevtsel}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
The selection described in this section is an improved version of the one used in Analysis I in Ref. [@Abe:2016tmq], and similar to that detailed in Ref. [@Abe:2018pwo] where it has been extensively described.
The target for [$\nu_\mu$]{}interactions is FGD1. This is used also as a tracker with TPC1, TPC2, ECal and SMRD. After the first requirements on the data quality and the position of the vertex are fulfilled, the selection requires tracks with a TPC segment with good reconstruction quality. For such tracks, the negatively charged one with the highest momentum, and compatible with the muon hypothesis according to the TPC PID is identified as a muon candidate. Tracks fully contained in the FGD and compatible with the energy loss by a muon have also been selected.
Protons are selected by looking for a track which starts in the FGD1 FV. The track should be identified as a positively charged in a TPC, and passes both the TPC track quality cut and PID criteria. Alternatively, if the track stops within the FGD it is identified as a proton if the track is consistent with the FGD proton hypothesis. To ensure the cross section is fully inclusive in terms of numbers of protons, events without a reconstructed proton are also included. Proton selection helps in further enlarging the phase space to high-angle and backward muons, as shown in Analysis I of Ref. [@Abe:2016tmq]. The selected events are divided into five signal samples:
Sample I
: characterized by events with only a muon candidate in one of the TPCs (TPC2 if the muon is going forward and TPC1 if it is going backward),
Sample II
: a muon candidate in one of the TPCs and one proton candidate in TPC2,
Sample III
: a muon candidate in one of the TPCs and a proton candidate in FGD1,
Sample IV
: a muon candidate in FGD1 and one proton in TPC2;
Sample V
: only a muon candidate in FGD1 that reaches the ECal or SMRD.
Events with a muon candidate in FGD or TPC and more than one proton in the final state, with the leading proton in TPC, have been selected as well. As these events only accounts for 0.8%, they have been added to the signal samples II-IV, accordingly with the muon candidate position (track in FGD only or in TPC). \[fig:numusignalsamples\] summarizes the signal samples described above.
{width="0.49\linewidth"} {width="0.49\linewidth"}
{width="0.49\linewidth"} {width="0.49\linewidth"}
The kinematics of the muon candidate in each sample for the [$\text{CC-}0\pi$]{}signal and the various backgrounds are shown in \[fig:eventsDistributionsNuMu\] where the MC is broken down by true topologies. The selection is dominated by events with one reconstructed muon and no other tracks. The signal samples where the muon is reconstructed in the TPC have very similar momentum distributions, although events with a reconstructed proton tend to have muons at slightly larger angles. The sample with the muon in the FGD and the proton in the TPC have muons with much smaller momenta and larger angles.
The [$\nu_\mu$]{}[$\text{CC-}0\pi$]{}cross section is extracted by adding together the contributions from all the samples, but it is important to keep the events with and without protons and with the muon in different sub-detectors separated in the analysis because they are affected by different systematics and backgrounds.
The main background arises from charged-current events with one true positively charged pion ([$\text{CC-}1\pi^+$]{}), or any number of true pions ([$\text{CC-Other}$]{}) which are misidentified or not reconstructed, neutral current interactions ([$\text{NC}$]{}) and interactions that occurred outside the FV ([$\text{out FV}$]{}) but were reconstructed inside constitute a smaller background. Two control samples were selected to constrain charged current event rates with single-pion and multiple-pion production: the [$\text{CC-}1\pi^+$]{}is made up of events with exactly two tracks, one negatively charged muon and one positively charged pion, and the [$\text{CC-Other}$]{}, made of events with more than one pion in the final state. Pions have been identified in different ways according to their charge. A $\pi^+$ can be identified by looking at the curvature of the track in the TPC and by requiring that the energy loss in this detector is consistent with a pion. $\pi^-$ are only identified by looking at the curvature of the tracks while $\pi^0$ are identified by looking for tracks in the TPC with charge depositions consistent with an electron from a $\gamma$ conversion. The kinematic distributions of the control samples are shown in \[fig:eventsDistributionsBkgNuMu\]. The data-MC discrepancy in the kinematic distributions of the [$\text{CC-}1\pi^+$]{}control sample was already observed in previous analysis [@Abe:2016tmq; @Abe:2018pwo] and is corrected by the likelihood fit as shown in \[sec:resultscomp\].
### [$\bar{\nu}_\mu$]{}CC event selection
After the first two common requirements described at the beginning of the Sec. \[sec:selections\] are fulfilled, the events are divided in four samples depending on the length of the muon candidate track in the TPCs and its direction, following the same strategy described in a recent T2K publication [@Abe:2018uhf]:
- If the muon candidate travels forward w.r.t. the beam direction and the associated track has more than 18 hits in TPC2 then the event belongs to the forward (FWD) sample;
- If it travels backward and the associated track has more than 18 hits in TPC1 then the event belongs to the backward (BWD) sample;
- If the muon candidate travels forward but the track has less than 19 hits in TPC2 then the event belongs to the high angle forward (HAFWD) sample;
- if it travels backward and the associated track has less than 19 hits in TPC1 then the event belongs to the high angle backward (HABWD) sample.
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.38\linewidth"} {width="0.38\linewidth"}
{width="0.49\linewidth"} {width="0.49\linewidth"}
{width="0.49\linewidth"} {width="0.49\linewidth"}
For each sample, different sets of cuts have been developed to reduce the background as much as possible without decreasing the efficiency.
One of the main backgrounds is caused by interactions that occur outside the FV ([$\text{out FV}$]{}) but are incorrectly reconstructed as starting inside the FV. This can be due to a failure of the reconstruction algorithms or a scattering of the particle which can lead to two unmatched track segments, one of which may start in the FV. The ratio of the momentum of the muon candidate to the other unmatched track and also the minimal distance between the tracks are used to reduce this background. The ratio should be lower than 1 if the two segments belong to the same track. Different cut values have been chosen for the event falling in the FWD, HAFWD and HABWD samples. These cuts are not applied in the selection of the BWD sample since signal events could be rejected. Another misreconstruction pathology can break a single track into two segments, with a reconstructed vertex inside the FV and a forward-going track into the downstream TPC. This often happens near the downstream edge of the FGD and the second track is considered as the muon candidate. Therefore, events for which the start position of the track associated to the muon candidate is in one of the last two layers of FGD1 are rejected.
The muon candidate is identified as the highest momentum track that is consistent with the muon PID. If the muon candidate enters a TPC, the track must pass the TPC muon PID. If the track does not enter a TPC, the ECal portion of the reconstructed object must be consistent with the ECal muon PID. In the case where the muon candidate enters a TPC, the charge of the track will be included in the selection. For particles entering ECal the information on the charge is not available, therefore this sample of events presents a high contamination of negatively charged muons that is constrained by measuring at the same time the [$\nu_\mu$]{}cross section.
In the selection of the FWD sample two additional cuts have been applied to reduce the pion and proton contamination: if the muon candidate stops in FGD2 and has a momentum greater than 280 MeV/c, the candidate is most likely a pion or a proton and the event is rejected; if it reaches the ECal it must have an ECal PID compatible with the muon hypothesis.
The described cuts select samples of muon antineutrino CC events with muons in every direction. Every sample is then split in three sub-samples according to the event pion multiplicity: events without a reconstructed pion, with one negatively charged pion, or with more than one pion in the final state. The TPC pion selection is similar to the one described previously in \[sec:numuevtsel\]. If the pion-candidate track is contained in the FGD1, pions are identified in two ways: by requiring a charge deposition consistent with a pion, or using the delayed energy deposition in the FGD due to a decay electron coming from $\pi \rightarrow \mu$ decay. In the latter case, the pion is tagged as positively charged since the $\pi^-$ are more likely to be absorbed.
The kinematics of the muon candidate for the signal sample are shown in \[fig:eventsDistributionsAntiNuMu\] where CC events without pions in the final state have been divided in four samples depending on the direction of the muon: [$\text{CC-}0\pi$]{}FWD, [$\text{CC-}0\pi$]{}BWD, [$\text{CC-}0\pi$]{}HAFWD and [$\text{CC-}0\pi$]{}HABWD. In the [$\bar{\nu}_\mu$]{}sample, the [$\nu_\mu$]{}contamination is larger than the [$\nu_\mu$]{}contamination in the [$\bar{\nu}_\mu$]{}sample. Moreover, positively charged pions (and, to a smaller extent, protons) produced in [$\nu_\mu$]{}interactions can be misidentified as muons constituting an irreducible background. In the high-angle selection, the charge is not reconstructed, therefore negatively charged muons are also selected (\[sec:anaStrategy\]). The statistics of the BWD sample is limited as the antineutrino cross section is suppressed for backwards going muons. \[fig:antinumusignalsamples\] summarizes the [$\bar{\nu}_\mu$]{}signal samples.
The background mostly arises from events with one true negatively charged pion ([$\text{CC-}1\pi^-$]{}), any number of true pions ([$\text{CC-Other}$]{}) and [$\text{out FV}$]{}, that contributes more to the BWD and HABWD samples. The [$\text{CC-}1\pi^-$]{}and [$\text{CC-Other}$]{}samples identified through the pion tagging are employed to constrain such backgrounds. For the [$\text{out FV}$]{}background there is not a dedicated control sample. The majority of them are [$\nu_\mu$]{}CC interactions that are constrained by the existing control samples. An uncertainty on the prediction of the [$\text{out FV}$]{}interactions is taken into account as reported in \[sec:uncertainties\]. The kinematic distributions of the control samples are shown in \[fig:eventsDistributionsBkgAntiNuMu\]. As shown in the legend, the purity is lower than the [$\nu_\mu$]{}selection, at 48% for the [$\text{CC-}1\pi^-$]{}sample and 24% for the [$\text{CC-Other}$]{}sample. Indeed, positively charged pions generated in [$\nu_\mu$]{}interactions are mis-identified as positively charged muons decreasing the purity. This difference with the [$\nu_\mu$]{}selection is caused by the higher [$\nu_\mu$]{}contamination of the antineutrino beam compared with the smaller [$\bar{\nu}_\mu$]{}component in the neutrino beam. The data-MC discrepancy observed for the [$\text{CC-}1\pi^-$]{}control sample is mainly due to an overestimation of the antineutrino coherent pion production cross-section as is implemented in NEUT version `5.3.2` [@Higuera:2014azj]. Also in this case, the discrepancy is corrected by the likelihood fit (see \[sec:resultscomp\]).
Sources of uncertainties {#sec:uncertainties}
------------------------
The uncertainties can be split into the following categories: statistical uncertainty, flux uncertainty, detector systematic uncertainties, uncertainty on the modeling of signal and background processes.
**Statistical Uncertainty.** To compute the data statistical uncertainty, the nominal MC was normalized to the number of protons on target in the data. Then 1,000 toy samples were generated where the MC was varied in each reconstructed bin according to a Poisson distribution and the fit was performed on each toy. The statistical error is taken as the width of the variation induced on the cross sections distribution in each true bin.
**Flux Uncertainty.** The evaluation of the uncertainties on the flux prediction are described in detail in Ref. [@Abe:2012av]. It is around 10% at the energy peak and is dominated by the hadron production model and is evaluated using data published by the NA61/SHINE experiment using a thin Carbon target [@Abgrall:2011ae; @Abgrall:2011ts; @Abgrall:2015hmv]. The flux covariance matrix was used to generate many toy MC sets. The flux bins include separate bins for the “right-sign" and “wrong-sign" components of the flux in both neutrino-mode and antineutrino-mode. The fit includes 32 nuisance parameters for the fluxes which are constrained by the fit, reducing the flux uncertainties by around 60%. In previous analyses the flux was not constrained by the fit since this procedure could introduce a model dependencies [@Abe:2016tmq]. For this reason dedicated mock data studies have been performed as described in \[sec:anaStrategy\].
**Detector Systematic Uncertainties.** Detector uncertainties can be grouped into three categories depending on the way they are propagated: efficiency-like, observable variation and normalization systematics. The systematics belonging to the first group have been propagated by applying a weight that depends on one or more observables, the second by adjusting the observables and reapplying the selection, the last by applying a single weight applied to all events. Dedicated data and MC samples have been used to quantify the detector uncertainties in the modeling of FGD and TPC responses, of neutrino interactions outside of the FGD1 FV, pion and proton secondary interactions. The differences between data and MC observed in control samples have been applied as correction factors to the nominal MC to take into account the observed discrepancies, while the error on these factors has been taken as detector systematic uncertainty. The [$\nu_\mu$]{}and [$\bar{\nu}_\mu$]{}selections are affected by the same detector uncertainties since they employ similar features of the sub-detectors. The dominant systematics are due to the uncertainties on the amount of background from interactions occurring outside of the fiducial volume, the modeling of the pion secondary interactions and the TPC PID. The detector systematics have been stored in a covariance matrix corresponding to the uncertainties on the total number of reconstructed events in each bin and in each signal and control regions. The systematics on the cross sections have been propagated by repeating the fit over many toy MC data sets where the detector parameters have been varied according with their covariance matrix but have been kept fixed in the fit. This choice has been driven by the necessity to ensure the stability and convergence of the fit by reducing the number of nuisance parameters. The uncertainty associated with the number of targets has been computed separately. A covariance matrix has been formed from the uncertainties on the areal densities of all the elements present in FGD1 and, the uncertainty calculated from toy experiments sampling such covariance matrix.
Parameter Prior Error
------------------------------------------------- ------- -------
$M_A^{QE}$ (GeV/c$^2$) 1.2 0.3
$p_F^C$ (MeV/c) 217 30
$E_B^C$ (MeV) 25 9
2p2h $\nu$ 1 1
2p2h [$\bar{\nu}$]{} 1 1
$C_A^5$ (GeV/c$^2$) 1.01 0.12
$M_A^{Res}$ (GeV/c$^2$) 0.95 0.15
I$_{1/2}$ 1.3 0.2
DIS Multiple pion 0.0 0.4
CC Coherent on C 1.0 1.0
CC-1$\pi$ $E_\nu < $ 2.5 GeV 1.0 0.5
CC-1$\pi$ $E_{\bar\nu} < $ 2.5 GeV 1.0 1.0
CC-1$\pi$ $E_\nu >$ 2.5 GeV 1.0 0.5
CC-1$\pi$ $E_{\bar\nu} >$ 2.5 GeV 1.0 1.0
CC Multile $\pi$ 1.0 0.5
CC-DIS $\nu$ 1.0 0.035
CC-DIS [$\bar{\nu}$]{} 1.0 0.065
NC Coherent 1.0 0.3
NC Other 1.0 0.3
Pion production 0.0 0.5
Pion absorption 0.0 0.41
Pion inelastic int. for $p_{\pi}\:<$ 500 MeV/c 0.0 0.41
Pion inelastic int. for $p_{\pi}\:>$ 400 MeV/c 0.0 0.34
Pion charge exchange for $p_{\pi}\:<$ 500 MeV/c 0.0 0.57
Pion charge exchange for $p_{\pi}\:>$ 400 MeV/c 0.0 0.28
: Prior values and errors of the cross section model parameters used in this analysis.
\[tbl:xsecparam\]
{width="1.\linewidth"}
{width="1.\linewidth"}
**Modeling of Signal and Background.** The signal efficiency and number of background events in each bin are affected by uncertainties in our cross-section model. \[tbl:xsecparam\] summarizes the cross-section parameters used for this analysis along with their prior value and error. The parameters include shape variations of the CCQE cross section ($M_{A}^{QE}$, Fermi momentum $p_{F}^{C}$, binding energy $E_{B}^{C}$) and the normalization of the amount of 2p2h interactions in neutrino and antineutrino. Fermi momentum and binding energy variations are modeled using RFG. The signal modeling parameters only affect the efficiency and are not constrained by the fit to avoid model dependencies. Other parameters control the shape and normalization of the background processes: the axial mass $M_{A}^{Res}$ and the form factor $C_{A}^{5}$ control the shape of the RES cross section; I$_{1/2}$ the normalization of non-resonant pion production; *CC-1$\pi$* the normalization of such background in different (anti)neutrino energy ranges; *DIS Multiple pion*, *CC-DIS $\nu$* and *CC-DIS [$\bar{\nu}$]{}* the normalization of the DIS; *CC Coherent on C* the normalization of such process; *NC Coherent* and *NC Other* the normalization of NC interactions. Pions that are produced in neutrino interactions can be affected by FSI as they leave the nuclear medium, changing their kinematics, charge and multiplicity. Dedicated systematic parameters have been included in the fit to describe the pion production, absorption, charge exchange and quasi-elastic scattering of the exiting pions. Again these modify not only the selected number of events, but the selection efficiency as well. Similarly protons are also subjected to FSI: the uncertainty is evaluated by comparing two different <span style="font-variant:small-caps;">NuWro</span> [@Golan:2012wx] MC simulations[^37], with and without FSI. The difference in the efficiency as a function of the muon kinematics between the two simulations has been taken as the uncertainty due to the proton FSI. To be conservative, it has been added in quadrature to the other efficiency uncertainties. Since in [$\bar{\nu}_\mu$]{}interactions proton FSI has a negligible impact, this uncertainty has been added to [$\nu_\mu$]{}cross section only. The cross section parameters have been propagated by throwing from Gaussian distributions that have as mean and sigma the prior and error values reported in \[tbl:xsecparam\].
All of the systematic errors in each bin are summarized in \[fig:nuerror,fig:antinuerror\]. In most bins, the dominant uncertainty is due to the statistical error on data. The systematic uncertainties are typically dominated by the detector systematics. The modeling errors are generally subdominant and smaller than 10% and closer to 1% in regions of high purity. Uncertainties related to the flux, modeling of the background and pion FSI have been propagated together since they are anti-correlated. The errors on the sum, difference and asymmetry has been computed numerically from toy experiments sampling the covariance matrix that includes the uncertainty and correlations between the two cross sections.
Results and comparisons with models {#sec:resultscomp}
===================================
The distribution of reconstructed events in bins used to evaluate the cross sections and in the background control samples is shown in \[fig:numucc0pirecopostfitvsdata,fig:antinumucc0pirecopostfitvsdata\]. The data are compared to MC predictions before and after the fit. The fit is able to reproduce the observed distributions in data by varying the parameters of interest to describe the signal cross section and the nuisance parameters describing the systematic uncertainties, as explained in \[sec:anaStrategy\]. As \[fig:numucontrolsamples,fig:antinumucontrolsamples\] show, the large discrepancy in the pre-fit MC prediction in the CC1$\pi^\pm$ control region is well corrected by the fit by varying the nuisance parameters describing the pion production cross-section listed in \[tbl:xsecparam\].
In the following, the measured cross sections, and their combinations, are compared to different (anti)neutrino-interaction models using the framework <span style="font-variant:small-caps;">Nuisance</span> [@Stowell:2016jfr] and the agreement is quantified by the $\chi^2$ statistic. Since the [$\nu_\mu$]{}and the [$\bar{\nu}_\mu$]{}cross sections are extracted simultaneously, a global $\chi^2$ computed using the full covariance matrix, i.e. the one containing the correlation between the two cross sections, is reported. It should be noted that, especially for the neutrino and antineutrino cross sections, as well as their sum, the overall normalization uncertainty (fully correlated between bins) constitutes a relatively large fraction of the uncertainty. In particular it contributes to 48% and 35% of the total systematic uncertainty for neutrino and antineutrino cross sections respectively and to 49% for the sum, while for the difference it decreases to 19% and for the asymmetry to 5%. Therefore the $\chi^2$ statistics may suffer from Peelle’s Pertinent Puzzle [@ppp] and may not be a reliable estimation of the data-MC agreement. This issue does not affect the shape-only $\chi^2$ which is reported as well.
The models considered for the data-MC comparisons are as follows:
- <span style="font-variant:small-caps;">Neut</span> (version `5.4.1`) Local Fermi Gas (LFG) assuming an axial mass $M_A^{QE}=1.03$ GeV/c$^2$, and corrections from the Random Phase Approximation (RPA) approach with and without 2p2h. A 1p1h and 2p2h model is used in this case from Ref. [@Nieves:2011pp];
- <span style="font-variant:small-caps;">NuWro</span> (version `18.02.1`) LFG [@Golan:2012wx] assuming an axial mass $M_A^{QE}=1.03$ GeV/c$^2$ with 2p2h and RPA corrections also from Ref. [@Nieves:2011pp];
- <span style="font-variant:small-caps;">Genie</span> (version `3.00.04`, configuration `G18_10b_000_00`) LFG assuming an axial mass $M_A^{QE}=0.99$ GeV/c$^2$ with 2p2h and RPA corrections from Ref. [@Nieves:2011pp];
- <span style="font-variant:small-caps;">NuWro</span> Spectral Function (SF), as developed in Ref. [@Benhar:1994hw], using the same 2p2h model as <span style="font-variant:small-caps;">Neut</span>;
- <span style="font-variant:small-caps;">GiBUU 2019</span> LFG in a coordinate- and momentum-dependent nuclear potential, as described in Ref. [@Gallmeister:2016dnq], using a 2p2h model based on Ref. [@OConnell:1972edu] and further tuned in Ref. [@Dolan:2018sbb], which uses the T2K measurements of final-state muon and proton kinematics and correlations in charged-current pionless interactions discussed in Ref. [@Abe:2018pwo];
- <span style="font-variant:small-caps;">SuSav2</span> is a complete implementation of the SuSAv2 model [@Gonzalez-Jimenez:2014eqa; @Megias:2014qva; @Megias:2016lke; @Megias:2016fjk] in GENIE, as described in [@Dolan:2019bxf], where 1p1h is based on the Relativistic Mean Field approach [@Caballero:2005sj] and 2p2h is based on the calculation from Ref. [@Simo:2016ikv]. The pion production and FSI models are the same as in <span style="font-variant:small-caps;">Genie</span> version `3.00.04`, configuration `G18_10b_000_00`;
- Martini *et al.* is the model described in Ref. [@Martini:2009uj]. It employs a LFG 1p1h model and RPA corrections including contribution from 2p2h.
The contribution of pion production, subsequently reabsorbed by FSI, is included in all the generators but not in the model by Martini *et al.*. It accounts for about 10%(5%) of the neutrino (antineutrino) measured cross section and, in order to properly compare this model with others, a prediction of this component obtained using <span style="font-variant:small-caps;">Neut</span> version `5.4.1` has been added on top of the Martini *et al.* prediction. This model is also missing antineutrino interactions on hydrogen which have been added to the antineutrino [$\text{CC-}0\pi$]{}cross section using the same strategy described above.
The comparisons with the models described above are shown in . The full and shape-only $\chi^2$ are reported in the legends (shape-only $\chi^2$ is reported in parenthesis) and are summarized in \[tab:chi2\]. In \[tab:redchi2\] the reduced $\chi^2$ is reported as well.
In order to evaluate the sensitivity to the 2p2h process, the measured cross sections, and their combinations, are compared in to <span style="font-variant:small-caps;">Neut</span> LFG with and without 2p2h. The full and shape-only $\chi^2$ show that the sensitivity is limited. Some conclusions can be drawn looking at each angular bin. In the intermediate and high-angle region both neutrino and antineutrino data tend to prefer the presence of 2p2h, as already shown in the previous T2K neutrino analysis [@Abe:2016tmq]. The $\chi^2$ in each angular bin has been computed, further confirming the preference for the presence of 2p2h in the intermediate and high-angle region. The effect is particularly evident in the sum of the neutrino and antineutrino cross sections, where the statistical uncertainty is smaller. For instance, in the angular bin 0.6 $<\cos\theta_\mu<$ 0.7 the reduced $\chi^2$ is 0.8 and 2.4 with and without 2p2h respectively. On the other hand, a clear overestimation of the cross section is visible in the forward region below 1 GeV, both for neutrinos and antineutrinos. This may point to incorrect 1p1h predictions, notably in the region of small energy transfer to the nucleus, where the treatment of various nuclear effects, like binding energy, is crucial. This issue is further discussed below, in the comparison to different 1p1h models. As expected the neutrino-antineutrino cross section difference emphasizes the 2p2h cross section, due to the change of sign of the axial-vector component. The statistical and systematic uncertainties, which are dominated by the flux, are still too large for a measurement of this component. Future foreseen reduction of such uncertainties, with more ND280 data and relying on NA61/SHINE T2K replica target data for flux tuning [@Berns:2018tap], will improve the sensitivity to the axial-vector 2p2h component. In some bins the difference is negative since antineutrinos can interact with the hydrogen of the hydrocarbon molecule, leading to a cross-section for antineutrino higher than for neutrino. The neutrino-antineutrino cross-section asymmetry shows a very small 2p2h dependence. The fractional change of the asymmetry with and without 2p2h is very small, except in the low momentum region where, at forward angle, it may reach 50%. The sensitivity to such observable is drastically limited by the statistical uncertainty. Despite most of the systematic uncertainties cancel out a residual detector systematic not correlated between neutrino and antineutrino dominate the systematic error.
A more sophisticated assessment of the 2p2h sensitivity is shown in , where the results are compared to different 2p2h models. The 2p2h model in <span style="font-variant:small-caps;">Neut</span> and the 2p2h model by Martini *et al.* [@Martini:2009uj] are both implemented on top of a similar 1p1h LFG model while the <span style="font-variant:small-caps;">SuSav2</span> model includes different 1p1h [@Caballero:2005sj] and 2p2h [@Simo:2016ikv] predictions. For the comparison with the model from Martini et al. the number of degrees of freedom (ndof) have been reduced to 96 for the cross sections and to 48 for their combinations in terms of sum, difference and asymmetry (w.r.t. 116 and 58 respectively) because the model predicts the cross section only for muon momentum lower than 3 GeV/c. Thus 10 high-momentum bins have been removed from the covariance matrix to compute the full $\chi^2$. Similarly, a complete shape-only covariance matrix has been obtained and those 10 bins have been removed afterwards to compute the shape-only $\chi^2$. An extended implementation of this model would be crucial for a better comparison with other models. None of the model is able to well describe the measured neutrino and antineutrino cross-sections in the entire phase space. As previously mentioned, the disagreement with cross-section measurements can be interpreted both in terms of 1p1h or 2p2h mismodeling. On the other hand, the various 2p2h models have quite different predictions for the axial-vector component, making the measurement of the neutrino-antineutrino cross-section difference a powerful probe to test the physics implemented in the different 2p2h predictions. To further investigate the dependence of the results on the 1p1h model, the measured cross-sections, and their combinations, are compared to different LFG implementations in . The <span style="font-variant:small-caps;">Neut</span>, <span style="font-variant:small-caps;">NuWro</span> and <span style="font-variant:small-caps;">Genie</span> LFG implementations differ mainly in the treatment of the nucleon binding energy. None of the generators is able to describe the measured neutrino and antineutrino cross-sections in the entire phase space. Among the different combinations the cross-sections difference show the lowest full $\chi^2$ in the comparison with <span style="font-variant:small-caps;">Genie</span>. The measured cross-sections, and their combinations, are also compared to a SF model in . The SF cross-section shows a different angular dependence than the LFG one: smaller for the backward and high-angle region and larger in the forward region. Interestingly, while SF is a more sophisticated model, the full $\chi^2$ is the largest (see \[tab:chi2\]). This may be due to an incomplete implementation of SF or to the merging with a 2p2h simulation modeled using RFG as nuclear model. The difference between LFG and SF tends to cancel in the neutrino-antineutrino cross-section difference and asymmetry. A more complete implementation of an SF model (including a 2p2h contribution) is likely required to investigate this further.
The integrated cross sections per nucleon and their combinations are reported in \[tab:integrated\] and compared with the model described above. It is striking that the models which exhibit best agreement in shape and in normalization are different, calling for further measurements with smaller systematic uncertainties and further model development. In summary, even if some conclusion can be drawn looking at the comparisons in some angular bins, none of the models is able to simultaneously describe [$\nu_\mu$]{}and [$\bar{\nu}_\mu$]{}[$\text{CC-}0\pi$]{}cross sections in all the phase space. Among the different combination, the difference between neutrino and antineutrino cross sections shows interesting sensitivity to different 2p2h models, which is limited by large uncertainties.
The poor (anti)neutrino-nucleus interaction modeling highlighted by this analysis is a limiting factor for the future neutrino oscillation experiments that have as primary goal the measurement of the CP violation, calling for a deeper understanding of the underlying processes involved in (anti)neutrino-nucleus interactions and for new cross-section analyses with larger statistics and improved systematic uncertainties.
----------------------------------------------------------------- --------------- --------------------- --------------- --------------------- --------------- --------------------- --------------- ---------------------
Generator/model
Full $\chi^2$ Shape-only $\chi^2$ Full $\chi^2$ Shape-only $\chi^2$ Full $\chi^2$ Shape-only $\chi^2$ Full $\chi^2$ Shape-only $\chi^2$
<span style="font-variant:small-caps;">Genie</span> LFG w/ 2p2h 333.1 444.7 101.3 141.3 76.2 102.0 143.6 134.4
<span style="font-variant:small-caps;">Neut</span> LFG w/ 2p2h 366.7 459.1 123.4 175.7 79.5 113.8 150.5 147.8
<span style="font-variant:small-caps;">Neut</span> LFG w/o 2p2h 236.7 388.7 82.5 126.5 87.6 154.8 160.0 169.4
<span style="font-variant:small-caps;">NuWro</span> LFG w/ 2p2h 408.9 481.5 122.2 158.1 87.0 121.6 162.9 142.4
<span style="font-variant:small-caps;">NuWro</span> SF w/ 2p2h 650.0 838.8 233.5 358.1 97.6 149.7 170.6 185.0
<span style="font-variant:small-caps;">GiBUU</span> 488.2 474.3 133.5 136.3 120.1 140.1 157.7 148.0
Martini *et al.* 368.6 573.4 142.0 227.4 119.6 289.8 93.9 131.2
<span style="font-variant:small-caps;">SuSAv2</span> 565.9 563.1 170.6 186.8 119.2 137.9 152.6 146.3
----------------------------------------------------------------- --------------- --------------------- --------------- --------------------- --------------- --------------------- --------------- ---------------------
----------------------------------------------------------------- --------------- --------------------- --------------- --------------------- --------------- --------------------- --------------- ---------------------
Generator/model
Full $\chi^2$ Shape-only $\chi^2$ Full $\chi^2$ Shape-only $\chi^2$ Full $\chi^2$ Shape-only $\chi^2$ Full $\chi^2$ Shape-only $\chi^2$
<span style="font-variant:small-caps;">Genie</span> LFG w/ 2p2h 2.9 3.8 1.7 2.4 1.3 1.8 2.5 2.3
<span style="font-variant:small-caps;">Neut</span> LFG w/ 2p2h 3.2 4.0 2.1 3.0 1.4 2.0 2.6 2.5
<span style="font-variant:small-caps;">Neut</span> LFG w/o 2p2h 2.0 3.3 1.4 2.2 1.5 2.7 2.7 2.9
<span style="font-variant:small-caps;">NuWro</span> LFG w/ 2p2h 3.5 4.1 2.1 2.7 1.5 2.1 2.8 2.4
<span style="font-variant:small-caps;">NuWro</span> SF w/ 2p2h 5.6 7.2 4.0 6.2 1.7 2.6 2.9 3.2
<span style="font-variant:small-caps;">GiBUU</span> 4.2 4.1 2.3 2.3 2.1 2.4 2.7 2.5
Martini *et al.* 3.8 6.0 3.0 4.7 2.5 6.0 2.0 2.7
<span style="font-variant:small-caps;">SuSAv2</span> 4.9 4.8 2.9 3.2 2.0 2.4 2.6 2.5
----------------------------------------------------------------- --------------- --------------------- --------------- --------------------- --------------- --------------------- --------------- ---------------------
----------------------------------------------------------------- --------------------------------- --------------------------------------- -------------------------- ------------------------------- --------------------------
[$\nu_\mu$]{}$\times$10$^{-39}$ [$\bar{\nu}_\mu$]{}$\times$10$^{-39}$ Sum $\times$10$^{-39}$ Difference $\times$10$^{-39}$ Asymmetry
cm$^2$/nucleon cm$^2$/nucleon cm$^2$/nucleons cm$^2$/nucleon
4.35 $\pm$ 0.06(*stat.*) 1.30 $\pm$ 0.04(*stat.*) 5.65 $\pm$ 0.07(*stat.*) 3.05 $\pm$ 0.07(*stat.*) 0.54 $\pm$ 0.01(*stat.*)
$\pm$ 0.30(*syst.*) $\pm$ 0.10(*syst.*) $\pm$ 0.30(*syst.*) $\pm$ 0.20(*syst.*) $\pm$ 0.02(*syst.*)
<span style="font-variant:small-caps;">Genie</span> LFG w/ 2p2h 3.76 1.14 4.90 2.62 0.53
<span style="font-variant:small-caps;">Neut</span> LFG w/ 2p2h 3.74 1.21 4.95 2.53 0.51
<span style="font-variant:small-caps;">Neut</span> LFG w/o 2p2h 3.20 1.03 4.23 2.17 0.51
<span style="font-variant:small-caps;">NuWro</span> LFG w/ 2p2h 3.91 1.28 5.19 2.63 0.51
<span style="font-variant:small-caps;">NuWro</span> SF w/ 2p2h 3.68 1.25 4.93 2.43 0.49
<span style="font-variant:small-caps;">GiBUU</span> 4.33 1.34 5.67 2.99 0.53
Martini 4.50 1.16 5.67 3.34 0.59
<span style="font-variant:small-caps;">SuSAv2</span> 4.35 1.35 5.70 3.00 0.53
----------------------------------------------------------------- --------------------------------- --------------------------------------- -------------------------- ------------------------------- --------------------------
Conclusions {#sec:conclusions}
===========
The T2K experiment has measured the first combined double-differential $\nu_{\mu}$ and $\bar{\nu}_{\mu}$ cross sections with no pions in the final state in the full phase space using $5.8\times 10^{20}$ POT of neutrino data and $6.2\times10^{20}$ POT of antineutrino data. The inclusion of ToF, in the selection of backward-going and high-angle tracks, enabled the exploration of the full phase space with better efficiency over previously reported T2K measurements of neutrino cross sections [@Abe:2016tmq]. The sum, difference and asymmetry of neutrino and antineutrino cross sections were measured, including full treatment of the correlations between the neutrino and antineutrino samples. Such observables have been compared with different models to shed light on the nuclear effects involved in the (anti)neutrino-nucleus interactions. Although none of the models considered in this work are able to describe the full phase space of the neutrino and antineutrino [$\text{CC-}0\pi$]{}cross section, it is difficult to determine the source of the problem. A precise understanding of this mis-modelling may be of critical importance for the next generation of neutrino oscillation experiments. Further investigation would benefit from smaller uncertainties and a mitigation of some of the approximations built into generator implementations of the models.
This analysis opens the road to joint cross-section measurements putting together different samples to minimize systematic uncertainties and to account properly for correlated systematics, enabling more complete and precise tuning of neutrino-nucleus interactions. A promising observable, measured here for the first time, is the difference between neutrino and antineutrino cross sections which shows interesting sensitivity to different 2p2h models, that can be further explored with more statistics and improved systematics uncertainties.
The data release for the results presented in this analysis is posted at the link in Ref. [@datarelease]. It contains the [$\nu_\mu$]{}and [$\bar{\nu}_\mu$]{}double-differential cross sections central values, their combinations and associated covariance matrices.
Acknowledgements
================
We thank the J-PARC staff for superb accelerator performance. We thank the CERN NA61/SHINE Collaboration for providing valuable particle production data. We acknowledge the support of MEXT, Japan; NSERC (Grant No. SAPPJ-2014-00031), NRC and CFI, Canada; CEA and CNRS/IN2P3, France; DFG, Germany; INFN, Italy; National Science Centre (NCN) and Ministry of Science and Higher Education, Poland; RSF (Grant \#19-12-00325) and Ministry of Science and Higher Education, Russia; MICINN and ERDF funds, Spain; SNSF and SERI, Switzerland; STFC, UK; and DOE, USA. We also thank CERN for the UA1/NOMAD magnet, DESY for the HERA-B magnet mover system, NII for SINET4, the WestGrid and SciNet consortia in Compute Canada, and GridPP in the United Kingdom. In addition, participation of individual researchers and institutions has been further supported by funds from ERC (FP7), “la Caixa" Foundation (ID 100010434, fellowship code LCF/BQ/IN17/11620050), the European Union’s Horizon 2020 Research and Innovation Programme under the Marie Sklodowska-Curie grant agreements no. 713673 and no. 754496, and H2020 Grants No. RISE-RISE-GA822070-JENNIFER2 2020 and RISE-GA872549-SK2HK; JSPS, Japan; Royal Society, UK; French ANR Grant No. ANR-19-CE31-0001; and the DOE Early Career program, USA, RFBR, project number 20-32-70196.
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
[NuMuCC1piPmuVsData]{} (42,22) [{width="0.20\linewidth"}]{}
{width="0.40\linewidth"}
{width="0.40\linewidth"} {width="0.40\linewidth"}
[AntiNuMuCC1piPmuVsData]{} (42,22) [{width="0.20\linewidth"}]{}
{width="0.40\linewidth"}
{width="0.40\linewidth"} {width="0.40\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
{width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"} {width="0.36\linewidth"}
[^1]: also at INFN-Laboratori Nazionali di Legnaro
[^2]: also at J-PARC, Tokai, Japan
[^3]: also at J-PARC, Tokai, Japan
[^4]: also at J-PARC, Tokai, Japan
[^5]: also at J-PARC, Tokai, Japan
[^6]: also at J-PARC, Tokai, Japan
[^7]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^8]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^9]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^10]: also at J-PARC, Tokai, Japan
[^11]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^12]: also at National Research Nuclear University “MEPhI” and Moscow Institute of Physics and Technology, Moscow, Russia
[^13]: also at J-PARC, Tokai, Japan
[^14]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^15]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^16]: also at J-PARC, Tokai, Japan
[^17]: also at J-PARC, Tokai, Japan
[^18]: also at J-PARC, Tokai, Japan
[^19]: also at the Graduate University of Science and Technology, Vietnam Academy of Science and Technology
[^20]: deceased
[^21]: also at J-PARC, Tokai, Japan
[^22]: also at JINR, Dubna, Russia
[^23]: also at J-PARC, Tokai, Japan
[^24]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^25]: also at Nambu Yoichiro Institute of Theoretical and Experimental Physics (NITEP)
[^26]: also at J-PARC, Tokai, Japan
[^27]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^28]: also at J-PARC, Tokai, Japan
[^29]: also at J-PARC, Tokai, Japan
[^30]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^31]: also at J-PARC, Tokai, Japan
[^32]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^33]: deceased
[^34]: also at Nambu Yoichiro Institute of Theoretical and Experimental Physics (NITEP)
[^35]: also at BMCC/CUNY, Science Department, New York, New York, U.S.A.
[^36]: affiliated member at Kavli IPMU (WPI), the University of Tokyo, Japan
[^37]: The version used is `11q`. The models implemented in this version are the same as present in version `18.02.1` described in \[sec:resultscomp\].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'M. Hugbart[^1], J. A. Retter, F. Gerbier, A. Varon, S. Richard, J. H. Thywissen, D. Clement, P. Bouyer, A. Aspect'
date: 'Received: date / Revised version: date'
subtitle: '\'
title: 'Coherence length of an elongated condensate: a study by matter-wave interferometry'
---
Introduction
============
The high atomic phase-space density provided by a Bose-Einstein condensate (BEC) [@BEC] has driven interest in guiding atoms in a manner analogous to guiding laser light through single-mode optical fibres. The most advanced technology to date is based on *atom chips*, where the fields which trap and guide the atoms are created by microfabricated structures [@ReiHan99; @Cor99; @Dek00; @Smi01; @Zim02; @Hin03]. Many groups around the world have already succeeded in preparing BEC on an atom chip [@Han01; @Ott01; @Gus02; @Sch03; @Vul04; @Est04; @Val04] and one can envisage using this technology to create integrated atom interferometers [@Hin01; @HanA01; @And02; @Wang04]. In this context, a precise characterization of the phase coherence properties of the condensate is crucial [@Det01; @Shv02; @Ric03].
Trapping quantum gases on atom chips naturally involves highly elongated, guide-like traps. Changing the dimensionality of the system from three-dimensional (3D) towards one-dimensional (1D) has a profound effect on the phase coherence of the condensate. In 3D condensates of modest aspect ratio, experimental results show that phase coherence extends across the whole cloud [@Hag99; @Ste99], even at finite temperature [@Blo00]. However, in the 1D regime, thermal excitations of the low energy axial modes lead to phase fluctuations which degrade the phase coherence [@Pet00; @Cas00]. Condensates exhibiting such phase fluctuations are known as [*quasi-condensates*]{}.
In the intermediate regime of elongated 3D condensates with a high aspect ratio, a behaviour similar to the 1D case is observed [@Pet01]: below a characteristic temperature $T_{\phi}$ determined by the atom number and the trapping frequencies, the condensate is nearly phase coherent, but above $T_{\phi}$ the population of the axial modes is high and phase fluctuations may be pronounced. For weakly elongated condensates, $T_\phi$ can be higher than the transition temperature $T_{\rm{c}}$, so that the condensate is nearly phase coherent at all temperatures. In contrast, atom chips can easily produce traps with high aspect ratios ($\sim\!1000$) for which $T_\phi$ can be much smaller than $T_{\rm{c}}$. Phase fluctuations are therefore likely to impose limits on the performance of atom chip devices and need to be well understood.
In an elongated condensate, the wavelength of the low energy axial modes is longer than the radial size of the condensate, so these excitations have a 1D character. However, the wavelength of these excitations can be much shorter than the axial length of the condensate, reducing the phase coherence length in the axial direction. An important feature of quasi-condensates is that density fluctuations remain suppressed in the trap, due to the mean field energy, even in the presence of large phase fluctuations [@Pet01; @Ric03]. Therefore, the 3D quasi-condensate has the usual parabolic profile in the Thomas-Fermi limit.
Phase-fluctuating condensates in elongated traps were first observed by the conversion of phase fluctuations into density fluctuations after a sufficiently long free expansion [@Det01] and by a condensate-focussing technique [@Shv02]. Quantitative measurements of the phase coherence length have since been obtained from the momentum distribution [@Ric03] and the second-order correlation function [@Hel03]. The results of each of these experiments showed good agreement with theory [@Pet01] in the strongly phase-fluctuating regime (at temperatures $T\gg T_{\phi}$). However, neither experiment explored the cross-over region ($T\sim T_\phi$) between phase-coherent and strongly phase-fluctuating condensates.
In this article we describe a new experiment using a matter-wave interferometer and Fourier-space analysis to measure the spatial correlation function, thereby extending our measurements into the cross-over region $(T\sim T_{\phi})$. Our results agree with the predicted shapes of the correlation function: for $T\gg T_{\phi}$, we find exponential-like correlation functions as predicted for significant phase fluctuations, whereas at $T\sim T_\phi$ we find a gaussian-like shape, as expected when the phase profile is nearly flat and the correlation function decay is dominated by the density profile. The coherence length as a function of $T/T_{\phi}$ follows the trend predicted by theory, showing that the coherence length increases smoothly as the temperature falls and that there is no sharp transition at $T_{\phi}$. This highlights the fact that phase fluctuations occur at all finite temperatures, even if these effects are too small to be resolved experimentally for more spherical traps. However, whereas our previous measurements based on momentum spectroscopy [@Ric03], realized for high $T/T_\phi$, were in full agrement with the theory, two observations remain unexplained in the interferometric method. First, as in a previous experiment [@Hag99], our experimental measurements of the coherence length are shifted from the theoretical prediction, by about $20\%$ for $T/T_\phi=0$, even after taking the limitations of our imaging system into account. Second, our interferometer produces unexplained supplementary fringes outside of the region where the condensates overlap, and we note that similar unexplained fringes appear in other published data [@Sim00]. These supplementary fringes do not seem to be compatible with interference of the thermal cloud observed in [@KetterleThermal].
Measurement of the coherence length by atom interferometry
==========================================================
A natural method to study the coherence length along the long axis $z$ of a condensate, or a quasi-condensate, is to use atom interferometry. With atomic beam-splitters, one produces two daughter copies of the initial condensate with a separation $s$, and observes the interference pattern appearing in the atomic density: $$\begin{aligned}
n(\Vec{r}) &\propto& |\Psi_0(\vec{r}-\frac {s}{2}\vec{e}_z)
+e^{i\phi_{\rm{rel}}}\Psi_0(\vec{r}+\frac {s}{2}\vec{e}_z)
|^2 \nonumber\\
&\propto& |\Psi_0(\vec{r}-\frac {s}{2}\vec{e}_z)|^2
+|\Psi_0(\vec{r}+\frac
{s}{2}\vec{e}_z) |^2 \nonumber\\
&+& 2{\rm Re}[e^{i\phi_{\rm{rel}}}\Psi_0^{\ast}(\vec{r}-\frac
{s}{2}\vec{e}_z)\Psi_0(\vec{r}+\frac {s}{2}\vec{e}_z)],
\label{equ:interfometer}\end{aligned}$$ where $\vec{e}_z$ is the axial unit vector, $\Psi_0(\vec{r})$ is the wavefunction describing the initial condensate and $\phi_{\rm{rel}}$ a relative phase shift produced by the interferometer and the free fall of the condensate.
Let us first consider the behaviour of a fully phase coherent condensate. During free expansion, it acquires a phase distribution proportional to $z^2$ [@Cas96]. The phase difference $\phi_{\rm{rel}}$ between two displaced copies of the condensate is therefore proportional to $zs$, giving rise to an interference pattern of straight fringes, uniformly spaced along the longitudinal direction, the spatial frequency of the fringes being proportional to the separation $s$. The fringe contrast integrated over the entire condensate gives the first-order correlation function at $s$: $$\begin{aligned}
\label{equ:C1}
C^{(1)}(s) = \int {\mathrm{d}}^3 \vec{r}\,\Psi_0^{\ast}(\vec{r}-\frac
{s}{2}\vec{e}_z)\Psi_0(\vec{r}+\frac {s}{2}\vec{e}_z).\end{aligned}$$ Repeating this contrast measurement for different separations $s$, one can study the decay of $C^{(1)}(s)$ with increasing $s$ [@Hag99]. For a fully phase coherent condensate, the first-order correlation function decay reflects only the width of the density profile $n(\vec{r})$ [@Zam00].
In the case of a quasi-condensate, 1D thermal excitations cause the phase to fluctuate along the longitudinal axis, both spatially and temporally. In our experiment, these fluctuations are small compared to the parabolic phase developed during free expansion. Therefore when we image the overlapping condensates after free expansion, we still observe straight fringes, but they are no longer strictly periodic. Small local phase shifts add a “jitter” to the fringe spacing, which in Fourier space has the effect of broadening the peak at the spatial frequency of the fringes and thereby reducing its height. As $s$ is increased, the fringes are perturbed more strongly, because the condensate phase becomes less correlated at larger separations. The contrast therefore decreases faster with $s$ than in the fully phase-coherent case. The greater the amplitude of the phase fluctuations, the faster the contrast decreases with $s$. Therefore by measuring the width of the correlation function at different temperatures, we extract the temperature dependence of the coherence length. Further information is obtained from the shape of the correlation function [@Ger03].
In the presence of phase fluctuations, each realization of the experiment gives a different interference pattern, even with fixed experimental conditions. Expression (\[equ:C1\]) must therefore be generalized to: $$\begin{aligned}
\label{equ:Cmean}
C^{(1)}(s) = \int {\mathrm{d}}^3 \vec{r}\,\langle
\Psi_0^{\ast}(\vec{r}-\frac {s}{2}\vec{e}_z)\Psi_0(\vec{r}+\frac
{s}{2}\vec{e}_z)\rangle,\end{aligned}$$ where the brackets $\langle\rangle$ denote a statistical average of the random process describing the random phase. In practice, one must repeat the experiment at a given separation $s$ and average the contrast measurements over many quasi-condensates to obtain $C^{(1)}(s)$.
This principle was used by Michelson in his famous astronomical interferometer, whose goal was to measure the spatial coherence of the light field arriving from a star, in order to deduce the diameter of the star [@Michelson]. However, Michelson’s method is plagued by the existence of a randomly fluctuating relative phase between the two inputs of the interferometer, and various methods insensitive to the relative phase fluctuations had to be developed [@Labeyrie]. A similar problem appears when one tries to determine the coherence length of a condensate or a quasi-condensate with interferometry. In addition to the controlled relative phase between the two components interfering in (\[equ:interfometer\]), there is an uncontrollable relative phase due to experimental problems such as a residual phase shift between the lasers creating the two copies or a random velocity kick imparted to the sample. In order to overcome this problem [@Bragg], D. Hellweg *et al.* [@Hel03] have used an analysis analogous to the Hanbury Brown and Twiss method [@HBT], since it is based on the measurement of the second-order correlation function which is insensitive to global phase shifts. In contrast, our method is in line with the initial method of Michelson who could visually evaluate the contrast of the randomly moving fringes he was observing. The decrease of this contrast as a function of the telescopes’ separation gave a direct measurement of the coherence length. Similarly, we directly evaluate the contrast of the fringes by taking the modulus of the Fourier transform of the fringe pattern. The decrease of that contrast as a function of the separation $s$ yields the coherence length of the quasi-condensate.
Experiment
==========
Creation of elongated Bose-Einstein condensates
-----------------------------------------------
In our experimental setup [@Des99], a Zeeman-slowed atomic beam of $^{87}$Rb is trapped in a MOT, and after optical pumping into the $5S_{1/2}|F=1,m_F=-1\rangle$ state is transferred to a magnetic Ioffe-Pritchard trap created by an iron-core electromagnet. Our design allows us to lower the magnetic field at the bottom of the trap to a few Gauss and thus to obtain very tight radial confinement [@Bou00]. Using this trap, we are able to create condensates very close to the 1D Thomas-Fermi regime [@Str98], as was demonstrated in [@QGLD; @RicT03; @GerT03]. In the present experiment, we produce condensates further into the 3D regime so that we can explore the cross-over regime ($T\sim T_{\phi}$). We use two different trap configurations: in the first, the final radial and axial trap frequencies are respectively $\omega_\bot=2\pi \times 395\,$Hz and $\omega_z=2\pi \times 8.67\,$Hz, giving an aspect ratio of 45; for the second trap configuration, the final frequencies are $\omega_\bot=2\pi \times 655\,$Hz and $\omega_z=2\pi \times
6.55\,$Hz, with aspect ratio 100. In this way, we obtain needle-shaped condensates containing around $3\times 10^5$atoms, with a typical half-length $L\,\simeq\,85\,\mu$m in the first trap and $L\,\simeq\,120\,\mu$m in the second. We control the final number of atoms by holding the condensate for a variable time, typically a few seconds, in the presence of an rf shied. The absolute number of atoms is calibrated from a measurement of the critical temperature, taking into account the effects of interactions [@Ger04L]. For condensates with small condensate fractions (less than 60%), the temperature is obtained by fitting a gaussian distribution to the thermal wings of the cloud. The temperature is then extrapolated from the final frequency of the rf ramp to lower temperatures for which the thermal fraction is indiscernible [@Ger04; @GerT03].
Interferometry set-up and timing
--------------------------------
As shown in Figure \[fig:interfero\], we implement the interferometer using a sequence of two $\frac{\pi}{2}$-Bragg pulses, which act as matter-wave beamsplitters. The set-up consists of two laser beams counter-propagating along the longitudinal trap axis, each of intensity $\sim\!2\,$mWcm$^{-2}$, red-detuned by $\Delta=6.6\,$GHz from the $^{87}$Rb D2 line at $\lambda=780.02$nm. Two acousto-optic modulators driven by frequency synthesizers produce a small relative detuning $\delta$, tuned to the two-photon Bragg resonance $\delta=2\hbar k_{\rm{L}}^2/m$, with $m$ the atomic mass and $k_{\rm L}=2\pi/\lambda$. The momentum width along $z$ of the expanding condensate corresponds to a frequency width of 200Hz. Therefore we use Bragg pulses of $\sim\!100\,\mu$s, short enough such that the corresponding $1.6$kHz frequency width (full width at half maximum) is sufficient to couple the entire condensate. The thermal cloud surrounding the condensate has a momentum distribution with a frequency width ranging from 12kHz to 60 kHz, much larger than that of the condensate. Thus, only a small fraction of the thermal cloud is coupled by the Bragg pulses [@Ger04]. By controlling the Bragg pulse length, we realize a $\frac{\pi}{2}$-pulse which splits the condensate into a coherent superposition of two wavepackets with velocities differing by $2
v_{\rm{R}} = 2\hbar k_{\rm L}/m = 11.72$ mms$^{-1}$, where $v_{\rm{R}}$ is the 2-photon recoil velocity. The interferometer sequence is illustrated in Figure \[fig:interfero\]. The condensate is held in the trap for at least 2s at the end of the final rf evaporative-cooling ramp, to allow residual oscillations to be damped [@Ric03; @Shv02]. After switching off the trap, the condensate is allowed to expand freely for 2ms before the first $\frac{\pi}{2}$-pulse is applied. During this expansion the condensate density reduces by two orders of magnitude, so collisions between the diffracted wavepacket and the original condensate become negligible. During a free-evolution time $2$ms$< T_s <10$ms, the two wavepackets separate to a distance $s\,=\,2 v_{\rm R} T_s$. The second $\frac{\pi}{2}$-pulse completes the interferometer, and we observe interference in each of the two output ports, which differ in momentum by $p\,=\,2\hbar k_{\rm L}$. The condensate is imaged by absorption perpendicular to the long axis $z$ after a 29ms total time-of-flight [@TOF].
For a given set of experimental conditions (condensate atom number and temperature), the experimental correlation function is acquired by taking a sequence of interference images with different condensate separations $s$ ranging from $0.2L$ to $1.2L$, varied by changing $T_s$. For smaller separations, we do not observe enough fringes to obtain a reliable measurement of the contrast. At the maximum value of $s$, the contrast has reduced such that the fringes are no longer discernible above the noise. Typical images for $0.3 L \leq s \leq 1.1 L$ are shown in Figure \[fig:franges\]. For each value of $s$, typically 5 images are taken, so that a statistical average can be performed. The fringe contrast is then measured, giving the correlation function. Correlation functions have been obtained at various temperatures $T$ between 100 and $230\,$nK and for condensate atom numbers $N_0$ between $0.5\times 10^5$ and $2.5\times 10^5$. These conditions correspond to $0.8< T/T_{\phi}<8$, where $T_\phi=15
\hbar^2 N_0/16 m k_B T L^2$ [@Pet01].
Analysis of interferograms
==========================
Interferogram
-------------
As shown in [@Pet00], a quasi-condensate is well-described by a fluctuating complex field $\Psi(\rho,z)= \sqrt{n(\rho,z)}
\textrm{e}^{i\Phi(\rho,z)}$, with fixed density distribution $n(\rho,z)$ and fluctuating phase $\Phi(\rho,z)$. In the following, $\Psi(\rho,z)$ represents the wavefunction of the condensate after the free-fall expansion. The first Bragg pulse is applied after 2ms of free expansion, at which time the density has reduced such that interactions between the atoms are negligible. We therefore assume that the different copies of the condensate propagate independently. The phase distribution can be expressed as $\Phi(\rho,z)=\alpha z^2 + \beta \rho ^2 +
\phi_{\rm{th}}(z)$, where $\phi_{\rm{th}}(z)$ represents the thermal phase fluctuations and the quadratic terms represent the parabolic phase developed during expansion.
We now consider the interference pattern produced at one of the output ports of the interferometer. For a separation $s$, we obtain the atomic density distribution: $$\begin{aligned}
n_{\rm{out}}(\rho,z) &=& \frac{1}{4} |\Psi(\rho,z-s/2)
+\Psi(\rho,z+s/2)
|^2 \nonumber\\
&=& \frac{n_+}{4}+ \frac{n_-}{4}+ \frac{ \sqrt{n_+ n_-}}{2}\cos
[\Delta \Phi(z) + \phi_{\rm{g}}], \label{equ:density}\end{aligned}$$ with $n_\pm =n(\rho,z\pm s/2)$ and $\Delta
\Phi(z)=\Phi(z+s/2)-\Phi(z-s/2)$. The global phase shift $\phi_{\rm{g}}$ is due to random, uncontrolled phase shifts between the two Bragg pulses. The phase difference between the two copies is $\Delta \Phi(z) =\alpha z s + \phi_{\rm{th}}(z+s/2)
- \phi_{\rm{th}}(z-s/2)$. The density of the condensate is small when the first Bragg pulse is applied, so we neglect a small relative velocity due to repulsion between the two copies. At the second output port, the two condensate copies have an additional $\pi$ relative phase shift due to the Bragg pulses, thereby producing a complementary fringe pattern. In our data analysis, the images of each output port are treated separately.
Since the global phase shift $\phi_{\rm{g}}$ fluctuates from shot to shot, we cannot average over different images at the same separation $s$. Instead, we take the contrast of each image individually and then average the contrast.
Analysis in Fourier Space
-------------------------
\[sec:2\] The atoms are imaged by absorption along the vertical $y$-axis, perpendicular to the long axis of the trap. The image we obtain, rescaled to units of 2D atomic density, is the integrated density: $$\begin{aligned}
\label{equ:image}
I(x,z) = {\rm const} \int {\mathrm{d}}y\,n_{\rm{out}}.\end{aligned}$$ We take the 2D Fourier transform of this image and extract its profile along the zero radial frequency $k_x=0$ axis: $$\begin{aligned}
\tilde{I}[0,k_z] = {\rm const} \int {\mathrm{d}}^3 r\,n_{\rm{out}}e^{ik_z
z}.\end{aligned}$$ Typical images and their 2D Fourier transform profiles are shown in Figure \[fig:franges\] for different separations $s$. The contrast of the fringe pattern is given by the ratio $2\tilde{I}[0,k_0(s)]/\tilde{I}[0,0]$, where $k_0(s) \simeq\alpha
s$ is the dominant spatial frequency of the fringe pattern. The profiles of Figure \[fig:franges\] show clearly the increasing spatial frequency and decreasing contrast as a function of $s$.
To extract the correlation function, we take the complex amplitude of the Fourier peak at the spatial frequency $k_0$ of the fringes: $$\begin{aligned}
\label{equ:Itilde}
\tilde{I}[0,k_0(s)] = e^{i\phi_{\rm{g}}}\int {\mathrm{d}}^3
r\,\sqrt{n_+n_-}e^{i\Delta\Phi_{\rm{th}}}.\end{aligned}$$ In the absence of the global phase shifts $\phi_{\rm{g}}$, the correlation function $C^{(1)}(s)$ would be obtained by taking the statistical average of equation (\[equ:Itilde\]): $$\begin{aligned}
\label{C1(s)}
C^{(1)}(s)=\langle\tilde{I}[0,k_0(s)]\rangle = \int {\mathrm{d}}^3
r\,\langle\sqrt{n_+n_-}e^{i\Delta\Phi_{\rm{th}}}\rangle\end{aligned}$$ which is identical to relation (\[equ:Cmean\]). However, $\phi_{\rm{g}}$ changes from shot to shot and prevents us from averaging Fourier transforms directly in this way [@meanI]. To eliminate this random phase, we can take the absolute value of the Fourier transform before averaging. Thus we obtain an effective correlation function: $$\begin{aligned}
\label{equ:C(s)}
C^{\rm{eff}}(s)=\langle\left|\tilde{I}[0,k_0]\right|\rangle =
\langle \left| \int {\mathrm{d}}^3
r\,\sqrt{n_+n_-}e^{i\Delta\Phi_{\rm{th}}}\right| \rangle.\end{aligned}$$ Note that although it has similar behaviour, our effective correlation function is expected to be quantitatively different from $C^{(1)}(s)$. Taking the absolute value of the transform reduces the cancelling effect between the random thermal phase shifts in the statistical average, so that the effective correlation function $C^{\rm{eff}}(s)$ decays more slowly with $s$ than $C^{(1)}(s)$, as calculated in the next section.
Simulation
----------
The 1st order correlation function $C^{(1)}(s)$ can be calculated analytically [@Ger03], using the theory of [@Pet01] to account for the phase fluctuations. However, it is not possible to obtain an analytic expression for the effective correlation function $C^{\rm{eff}}(s)$ (equation (\[equ:C(s)\])) which we measure. Therefore, to calculate $C^{\rm{eff}}(s)$, we first simulate the quasi-condensate phase fluctuations following the theory in [@Pet01]. The phase operator is given by: $$\begin{aligned}
\hat\phi_{\rm{th}}(\mathbf r) = [4n(\mathbf r)]^{-1/2}\sum_j
f^+_j(\mathbf r)\hat a_j+h.c.,\end{aligned}$$ where $\hat a_j$ is the annihilation operator of the excitation with quantum number $j$. The solution of the Bogoliubov-de Gennes equations for “low energy” excitations (with energies $\epsilon_\nu < \hbar \omega_{\bot}$) gives the wavefunctions of these modes: $$\begin{aligned}
f^+_j(\mathbf r) = \sqrt{\frac{(j+2)(2j+3)gn(\mathbf
r)}{4\pi(j+1)R^2L\epsilon_j}}P^{(1,1)}_j(z/L),\end{aligned}$$ where $P^{(1,1)}_j$ are Jacobi polynomials, $g=4\pi\hbar^2 a/m$ is the interaction strength, $a$ is the scattering length, $R$ and $L$ are the size of the condensate in the trap and $\epsilon_j=\hbar\omega_z\sqrt{j(j+3)/4}$ is the energy of mode $j$ [@Str98]. We assume the Thomas-Fermi approximation for the density $n(\mathbf{r})$, taking $R$ and $L$ from fits to images after expansion. To simulate numerically the phase fluctuations, we replace the operators $\hat a_j$ and $\hat a^\dagger_j$ by complex Gaussian random variables $\alpha_j$ and $\alpha^\ast_j$. These variables have a mean value of zero and the correlation $\langle \alpha_j \alpha^\ast_{j'} \rangle = \delta_{jj'}N_j$ where $N_j = k_B T/(\hbar \omega_z \sqrt{j(j+3)/4})$ is the occupation number for the quasiparticle mode $j$ at a given temperature $T$. We assume that the phase fluctuations do not evolve on the time scale of the expansion [@Ger03]. We have verified this by studying images of condensates after the same time-of-flight, but without the interferometer pulses. In this case, we observe a smooth density profile, with no extra features appearing in the Fourier transform.
For a given $T/T_\phi$, $20$ condensates are generated at each value of $s$, with $s$ ranging from $0.2 L$ to $2L$. Each condensate is integrated over $y$ as in equation (\[equ:image\]). These simulated absorption images are analysed in exactly the same way as the real experimental images. The absolute values of the Fourier transforms of the images are averaged, and the contrast extracted as in equation (\[equ:C(s)\]). The points in Figure \[fig:sim\_points\] show the typical contrast extracted from our simulations for $T/T_\phi
= 1$. We found that these points are very well fitted by the product of a gaussian and an exponential, for all $T/T_\phi$. We use these fits as our theoretical effective correlation functions $C^{\rm{eff}}(s)$.
The effective correlation function was simulated for $0\leq
T/T_\phi \leq 20$. Figure \[fig:simulation\]a presents results for different $T/T_\phi$. At $T=0$, $C^{\rm{eff}}(s)$ coincides with $C^{(1)}(s)$ [@Ger03]. This function is simply the integrated overlap function between the two condensates, and is approximately a gaussian function of the separation $s$ [@Zam00]. As $T/T_{\phi}$ increases, the width of the function decreases and its form gradually becomes exponential. In Figure \[fig:simulation\]b we plot the $1/e$ widths of the simulated $C^{\rm{eff}}(s)$ functions as a function of $T/T_\phi$. For comparison we show also the width of $C^{(1)}(s)$ [@Ger03] which decreases much faster with $T/T_{\phi}$.
Experimental results
====================
Figure \[fig:contraste\] shows two examples of effective correlation curves measured using our interferometer and analysed as described above. The points shown in Figure \[fig:contraste\]a were obtained using a magnetic trap with an aspect ratio of 45, and with an atom number and temperature corresponding to $T/T_\phi=1.35$, that is for small-amplitude phase fluctuations. The points in Figure \[fig:contraste\]b were obtained using a trapping aspect ratio of 100, and with $T/T_\phi=4.86$. The contrast is plotted as function of $s/L$, obtained from a fit to a truncated parabola. Each point corresponds to an average over 5 condensates. The difference in the range of $s/L$ explored is due to different expansion dynamics after release from the two different traps. In the more tightly confined trap (Figure \[fig:contraste\]b), the axial expansion is much slower, and thus the fringe spacing decreases more slowly with $s$. At the smallest values of $s$, it is therefore impossible to measure the contrast reliably since we do not observe enough fringes. We can extract information about the phase fluctuations from both the shape and the width of these effective correlation functions.
Shape of the effective correlation functions
--------------------------------------------
First, we observe qualitatively that the shape of the effective correlation functions $C^{\rm{eff}}(s)$ changes as $T/T_{\phi}$ increases. For small $T/T_{\phi}$, as in Figure \[fig:contraste\]a, the curves are clearly gaussian, as shown by the fit in the figure. As we increase $T/T_\phi$, the profiles become rapidly exponential. We see in Figure \[fig:contraste\]b that at $T/T_\phi=4.86$, the curve is already better fitted by an exponential than a gaussian. At intermediate values of $T/T_\phi$, we can use the product of a gaussian and an exponential to fit a smooth curve through the data. The contribution of the exponential increases rapidly in importance at finite $T$, in agreement with the simulation, reflecting the increasing amplitude of the phase fluctuations with $T/T_\phi$.
Comparison of coherence length ${L_{\rm{C}}}$
---------------------------------------------
In order to extract quantitative information from the effective correlation functions, we define a coherence length $L_{\rm{C}}$, equal to the $1/e$ width of the effective correlation curve $C^{\rm{eff}}(s)$. We then use this parameter to compare the widths of the measured and simulated effective correlation functions. A smooth curve is fitted through the data (using the product of a gaussian and an exponential) and the $1/e$ width extracted. Although the thermal cloud plays no role in the interference pattern we observe, it appears behind the condensates in the $p = 0$ output of the interferometer and thus reduces the measured contrast. Independent measurements of the thermal fraction (between 60% and 80% for this experiment) allow us to renormalize the experimental effective correlation functions to take account of this effect. When fitting the renormalized curves, we then fix the value at $s/L=0$ to unity.
In Figure \[fig:Graphe\_TTphi\], we plot $L_{\rm{C}}/L$ as a function of $T/T_{\phi}$. Importantly, we see that the coherence length varies smoothly as a function of $T/T_\phi$, even when the temperature is close to $T_\phi$. This is what we should expect since $T_{\phi}$ is simply a characteristic temperature, defined as that at which the mean square fluctuations of the phase difference between two points separated by a distance $L$ is equal to 1. Therefore it should be borne in mind that even condensates at temperatures below $T_\phi$ are not necessarily fully coherent.
The inset of Figure \[fig:Graphe\_TTphi\] compares the measured coherence widths with the results of the simulation (Figure \[fig:simulation\]b). Despite the offset between the two curves, the trend of the data follows very well that of the simulation. In Figure \[fig:Graphe\_TTphi\] the experimental data is fitted by a curve $A+B\exp\left(-b T /T\phi\right)$, giving $b=0.35$ and $B=0.32$. This is in reasonable agreement with the simulation, for which the fit yields $b=0.32$ and $B=0.39$. In the following discussion we consider possible explanations for the observed reduction in contrast.
Discussion
----------
As shown in the inset of Figure \[fig:Graphe\_TTphi\], the measured coherence length is offset by about $20\%$ from the results of the simulation at $T/T_{\phi}=0$. In order to eliminate various possible causes of this discrepancy, and to understand better the limitations of our experiment, we have performed several tests.
Since the accuracy of our experiment relies on the comparison of fringe contrast at different spatial frequencies, it is important to take great care in setting up and characterising the imaging system. Therefore we measured the modulation transfer function (MTF) (see [*e.g.*]{} [@Goodman]) of our complete imaging system *in situ*, using a USAF1940 resolution target engraved with 3-bar square wave patterns of spatial frequency 4–200 lines/mm, covering the range of spatial frequencies observed in our interference experiment. By Fourier transforming images of different regions of the target, we were able to compare the magnitude of the different Fourier components in the image with those of the target pattern, thereby obtaining the MTF of our system. We found that the MTF is approximately linear, falling from 1 at zero spatial frequency to 0 at $118\,$lines/mm. This resolution limit at $8.5\,\mu$m is in agreement with earlier characterisations of the system [@RicT03]. The shape of the MTF is due almost entirely to the CCD camera [@PixelFly], and is surprisingly significant at fringe spacings much greater than the effective pixel size of $2.5\,\mu$m. All contrast measurements were corrected by this MTF. In fact, for the data obtained using the second trapping aspect ratio, the axial expansion (and thus the phase difference developed) was sufficiently small that the maximum observed fringe spatial frequency was $16\,$lines/mm and therefore the MTF correction had a negligible effect on the experimental effective correlation functions. In the first set of data, where the maximum fringe frequency was $38\,$lines/mm (a fringe spacing of $25\,\mu$m), the correction was more significant, changing the width of the curves by typically 10%, though still leaving a 20% discrepancy with the simulation.
We also considered the error introduced by a small focussing error. The imaging system is focussed onto the condensate to within $\pm 0.2\,$mm by minimising the imaged size of a small condensate as a function of the objective lens position. However, by varying the time-of-flight used, we may have introduced an error of up to $\pm 1\,$mm. Such an error could also be introduced by small changes to the residual magnetic fields, which lead to changes in the release velocity of the condensate when the magnetic trap is switched off. We measured effective correlation curves for different foci of the imaging system, but found that the width of the effective correlation curve changed by less than the existing spread in the points.
Other possible sources of error, such as the alignment of the imaging beam with respect to the condensates’ fringes and correct background subtraction have also been eliminated.
There remains in our experiment an unexplained phenomenon regarding the distribution of the fringe pattern. We expect to see fringes only in the region where the two condensates overlap. However, it can be seen in the images of Figure \[fig:franges\] that the fringes extend to the edges of each condensate. Moreover, these “extra” fringes have the same spatial frequency and phase as the central fringes. Although it is possible that a small fraction of the thermal cloud is coupled by the Bragg beams, interference fringes produced in this way [@KetterleThermal] would have a much smaller fringe spacing, less than $6\,\mu$m. More importantly, the contributions from different parts of the original thermal cloud, whose width is $\sim\!100\,\mu$m, would sum incoherently to wash out the fringe pattern. It is more likely that these extra fringes arise as a result of interactions during the application of the Bragg pulses, but better modelling is still needed before evaluating whether their presence should increase or decrease the overall measured contrast.
Conclusion {#sec:1}
==========
We have demonstrated a new type of matter-wave interferometry using Bragg beam-splitter pulses and Fourier space analysis. Our results show that the expected shape of the correlation functions changes from a gaussian-like shape to an exponential-like shape when the amplitude of phase fluctuations is increased. The coherence length of elongated condensates varies smoothly at temperatures close to $T_\phi$, as predicted by theory. This highlights the fact that the characteristic phase temperature $T_\phi$ does not indicate a transition to full phase coherence, but rather that condensates exhibit phase fluctuations at all finite temperatures, albeit of small amplitude. This may place constraints on the trapping geometries which can be used for creating measurement devices based on the phase coherence of condensates.
Acknowledgements
================
We would like to acknowledge support from IXSEA-OCEANO (M.H.), from the BEC2000+ programme and the Marie Curie Fellowships programme of the European Union (J.R.), from IXCORE (F.G.) and from the Chateaubriand Program, CNRS, and NSERC (J.H.T.). This work was supported by Délégation Générale de l’Armement, the European Union (grants IST-2001-38863 and MRTN-CT-2003-505032) and INTAS (Contract No. 211-855).
D.S. Petrov, G.V. Shlyapnikov and J.T.M. Walraven, Phys. Rev. Lett. [**87**]{}, 050404 (2001)
D. Hellweg, L. Cacciapuoti, M. Kottke, T. Shulte, K. Sengstock, W. Ertmer and J.J. Arlt, Phys. Rev. Lett. [**[91]{}**]{}, 010406 (2003)
J.E. Simsarian, J. Denschlag, M. Edwards, C.W. Clark, L. Deng, E.W. Hagley, K. Helmerson, S.L. Rolston and W.D. Phillips, Phys. Rev. Lett. [**[85]{}**]{}, 2040 (2000)
Y. Castin and R. Dum, Phys. Rev. Lett. [**[77]{}**]{}, 5315-5319 (1996)
F. Zambelli, L. Pitaevskii, D. M. Stamper-Kurn and S. Stringari, Phys. Rev. A [**[61]{}**]{}, 063608 (2000)
F. Gerbier, J.H. Thywissen, S. Richard, M. Hugbart, P. Bouyer and A. Aspect, Phys. Rev. A [**[67]{}**]{}, 051602 (2003)
A.A. Michelson, Astroph. J. [**[51]{}**]{}, 257 (1920)
A. Labeyrie, in *Progress in Optics XIV*, edited by E. Wolf (North-Holland, 1976) p. 47-87
Note that the Bragg momentum spectroscopy method does not suffer from this problem.
R. Hanbury Brown and R.Q. Twiss, Nature (London) [**[177]{}**]{}, 27 (1956); [**[178]{}**]{}, 1046 (1956)
B. Desruelle [*[et al]{}*]{}, Phys. Rev. A [**[60]{}**]{}, R1759 (1999)
P. Bouyer, A. Aspect, M. Lécrivain, B. Desruelle and V. Boyer, patent 00 02704 (2000)
S. Stringari, Phys. Rev. A [**[58]{}**]{}, 2385 (1998)
P. Bouyer, J.H. Thywissen, F. Gerbier, M. Hugbart, S. Richard, J. Retter and A. Aspect, J. Phys. IV France [**[116]{}**]{}, 219 (2004)
S. Richard, Ph.D. thesis, University of Paris XI, 2003
F. Gerbier, Ann. Phys. Fr. 29, (2004)
F. Gerbier, J.H. Thywissen, S. Richard, M. Hugbart, P. Bouyer, and A. Aspect, Phys. Rev. A [**[70]{}**]{}, 013607 (2004)
Our cell allows maximum 25ms free-fall. We increase the time-of-flight by applying, after the Bragg pulses, a magnetic field gradient opposed to the gravity field.
This would be equivalent to averaging the original images: the fringe contrast would be averaged out to the non-interfering density profile.
J.W. Goodman [*Introduction to Fourier Optics*]{} 2nd edition, McGraw-Hill, 1996.
[^1]: e-mail: [email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Consider the problem of recovering an unknown signal from undersampled measurements, given the knowledge that the signal has a sparse representation in a specified dictionary $\D$. This problem is now understood to be well-posed and efficiently solvable under suitable assumptions on the measurements and dictionary, if the number of measurements scales roughly with the sparsity level. One sufficient condition for such is the $\D$-restricted isometry property ($\D$-RIP), which asks that the sampling matrix approximately preserve the norm of all signals which are sufficiently sparse in $\D$. While many classes of random matrices are known to satisfy such conditions, such matrices are not representative of the structural constraints imposed by practical sensing systems. We close this gap in the theory by demonstrating that one can subsample a fixed orthogonal matrix in such a way that the $\D$-RIP will hold, provided this basis is sufficiently incoherent with the sparsifying dictionary $\D$. We also extend this analysis to allow for weighted sparse expansions. Consequently, we arrive at compressive sensing recovery guarantees for structured measurements and redundant dictionaries, opening the door to a wide array of practical applications.'
author:
- 'Felix Krahmer, Deanna Needell, and Rachel Ward'
bibliography:
- 'bib.bib'
title: Compressive Sensing with Redundant Dictionaries and Structured Measurements
---
Introduction
============
Compressive Sensing
-------------------
The compressive sensing paradigm, as first introduced by Cand[è]{}s and Tao [@RefWorks:577] and Donoho [@RefWorks:70], is based on using available degrees of freedom in a sensing mechanism to tune the measurement system so as to allow for efficient recovery of a particular type of signal or image of interest from as few measurements as possible. A model assumption that allows for such signal recovery is *sparsity*: the signal can be well-approximated by just a few elements of a given representation system.
Often, a near-optimal strategy is to choose the measurements completely at random, for example following a Gaussian distribution [@RefWorks:285]. Typically, however, some additional structure is imposed by the application at hand, and randomness can only be infused in the remaining degrees of freedom. For example, magnetic resonance imaging (MRI) is known to be well modeled by inner products with Fourier basis vectors. This structure cannot be changed, and the only aspect that can be decided at random is which Fourier basis vectors to select.
An important difference between completely random measurement systems and structured random measurement systems is in the aspect of universality. Gaussian measurement systems, among many other systems with minimal imposed structure, are oblivious to the basis in which the underlying signal is sparse, and achieve equal reconstruction quality for all different orthonormal basis representations. For structured measurement systems, this is, in general, no longer the case. When the measurements are uniformly subsampled from an orthonormal basis such as the Fourier basis, one requires, for example, that the measurement basis and the sparsity basis are *incoherent*; the inner products between vectors from the two bases are small. If the two bases are not incoherent, more refined concepts of incoherence are required [@KW14; @adcock2013breaking]. An important example is that of the Fourier measurement basis and a wavelet sparsity basis. Since both contain the constant vector, they are maximally coherent.
Motivation
----------
All of the above mentioned works, however, exclusively cover the case of sparsity in *orthonormal* basis representations. On the other hand, there are a vast number of applications in which sparsity is expressed not in terms of a basis but in terms of a redundant, often highly overcomplete, dictionary. Specifically, if $\f \in \C^{n}$ is the signal of interest to be recovered, then one expresses $\f = \D\x$, where $\D\in\C^{n\times N}$ is an overcomplete dictionary and $\x\in\C^N$ is a sparse (or nearly sparse) coefficient vector. Redundancy is widespread in practice, either because no sparsifying orthonormal basis exists for the signal class of interest, or because the redundancy itself is useful and allows for a significantly larger, richer class of signals which can be sparsely represented in the resulting dictionary. For example, it has been well documented that overcompleteness is the key to a drastic reduction in artifacts and recovery error in the denoising framework [@RefWorks:342; @RefWorks:344].
In the compressive sensing framework, results using overcomplete dictionaries are motivated by the broad array of tight (and often Parseval) frames appearing in practical applications. For example, if one assumes sparsity with respect to the Discrete Fourier Transform (DFT), this is implicitly assuming that the signal is well-represented by frequencies lying along the lattice of the DFT. To allow for more flexibility in this rigid assumption, one instead may employ the oversampled DFT frame, containing frequencies on a much finer grid, or even over intervals of varying widths. Gabor frames are used in imaging as well as radar and sonar applications, which are often highly redundant [@RefWorks:224]. Many of the most widely-used frames in imaging applications such as undecimated wavelet frames [@RefWorks:345; @RefWorks:342], curvelets [@RefWorks:273], shearlets [@labate2005sparse; @easley2008sparse], framelets [@cai2008framelet], and many others, are overcomplete with highly correlated columns.
Related Work
------------
Due to the abundance of relevant applications, a number of works have studied compressive sensing for overcomplete frames. The first work on this topic aimed to recover the coefficient vector $\x$ directly, and thus required strong incoherence assumptions on the dictionary $\D$ [@RefWorks:583]. More recently, it was noted that if one instead aims to recover $\f$ rather than $\x$, recovery guarantees can be obtained under weaker assumptions. Namely, one only needs that the measurement matrix $\A$ respects the norms of signals which are sparse in the dictionary $\D$. To quantify this, Candès et al. [@RefWorks:60] define the $\D$-restricted isometry property ($\D$-RIP in short, see Definition \[DRIP-orig\] below). For measurement matrices that have this property, a number of algorithms have been shown to guarantee recovery under certain assumptions. Optimization approaches such as $\ell_1$-analysis [@RefWorks:340; @RefWorks:60; @RefWorks:607; @RefWorks:581; @liu2012compressed; @rauhut2013analysis; @giryes2014sampling] and greedy approaches [@Paper5; @RefWorks:607; @giryes2013greedy; @peleg2013performance; @giryes2014near] have been studied.
This paper establishes the $\D$-RIP for structured random measurements formed by subsampling orthonormal bases, allowing for these types of recovery results to be utilized in more realistic settings. To date, most random matrix constructions known to yield $\D$-RIP matrices with high probability are random matrices with a certain concentration property. As shown in [@RefWorks:60], such a property implies that for arbitrary dictionary $\D$, one obtains the $\D$-RIP with high probability. This can be interpreted as a dictionary version of the universality property discussed above. It has now been shown that several classes of random matrices satisfy this property as well as subsampled structured matrices, after applying random column signs [@Dasgupta; @RefWorks:230]. The matrices in both of these cases are motivated by application scenarios, but typically in applications they appear without the randomized column signs. In many cases one is not able to apply column signs in practice; for example in cases where the measurements are fixed such as in MRI, one has no choice but to use unsigned Fourier samples and cannot pre-process the data to incorporate column signs. Without these signs however, such measurement ensembles will not work for arbitrary dictionaries in general. This is closely related to the underlying RIP matrix constructions not being universal. For example, it is clear that randomly subsampled Fourier measurements will fail for the oversampled Fourier dictionary (for reasons similar to the basis case). In this work, we address this issue, deriving recovery guarantees that take into account the dictionary. Similar to the basis case, our analysis will be coherence based. A similar approach has also been taken by Poon in [@P15] (completed after the first version of our paper), for an infinite dimensional version of the problem.
Contribution
------------
Our main result shows that a wide class of orthogonal matrices having uniformly bounded entries can be subsampled to obtain a matrix that has the $\D$-RIP and hence yields recovery guarantees for sparse recovery in the setting of redundant dictionaries. As indicated above, our technical estimates below will imply such guarantees for various algorithms. As an example we focus on the method of $\ell_1$-analysis, for which the first $\D$-RIP based guarantees were available [@RefWorks:60]. Our technical estimates will also provide more general guarantees for weighted $\ell_1$-analysis minimization (a weighted version of , see in Section \[sec:main\] for details) in case one has prior knowledge of the underlying sparsity support.
Recall that the method of $\ell_1$-analysis consists of estimating a signal $\f$ from noisy measurements $\y = \A\f+{\mathrm{e}}$ by solving the convex minimization problem $$\f^{\sharp}= \operatorname*{arg min}_{\tilde \f\in\C^n} \|\D^*\tilde\f\|_1 \text{\quad such that \quad} \|\A \tilde\f-\y\|_2\leq \varepsilon \tag{$P_1$}, \label{eq:l1ana}$$ where $\varepsilon$ is the noise level, that is, $\|{\mathrm{e}}\|_2\leq \varepsilon$. The $\ell_1$-analysis method (like alternative approaches) assumes that for the signal $\f=\D\x$, not only is the underlying (synthesis) coefficient sequence $\x$ sparse (typically unknown and hard to obtain), but also the analysis coefficients $\D^*\f$ are nearly sparse, i.e., dominated by a few large entries. We refer the reader to Theorem \[thm:l1anarec\] below for the precise formulation of the resulting recovery guarantees (as derived in [@RefWorks:60]). The assumption has been observed empirically for many dictionaries used in practice such as the Gabor frame, undecimated wavelets, curvelets, etc. (see, e.g., [@RefWorks:60] for a detailed description of such frames) and is also key for a number of thresholding approaches to signal denoising [@RefWorks:273; @shearlet07; @easley2008sparse].
A related and slightly stronger signal model, in which the analysis vector $\D^*\f$ is sparse or nearly sparse, has been considered independently from coefficient sparsity (e.g., [@RefWorks:581]), and is commonly called the co-sparsity model.
The results in this paper need a similar, but somewhat weaker assumption to hold for all signals corresponding to sparse synthesis coefficients $\x$. Namely, one needs to control the as we now introduce.
\[unrec-orig\] For a dictionary $\D\in\C^{n\times N}$ and a sparsity level $s$, we define the ** as $$\label{eq:unrec}
\unrec_{s, \D} =\unrec {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} \frac{{\lVert {\D^*\D\z} \rVert}_{1}}{\sqrt{s}}.$$
In the following we will mainly consider dictionaries, which form Parseval frames, that is, $\D^*$ is an isometry. Then the term is appropriate because this quantity can be viewed as the factor by which sparsity is preserved under the gram matrix map $\D^{*}\D$, compared to the case where $\D$ is an orthonormal basis and $\D^{*} \D = {\mathbf{I}}_n$ and $\eta \equiv 1$ by Cauchy-Schwarz. For a general family of such frames parameterized by the redundancy $N/n$, $\unrec$ will increase with the redundancy; families of dictionaries where such growth is relatively slow will be of interest. For example, new results on the construction of unit-norm tight frames give a constructive method to generate *spectral tetris* frames [@CFMWZ11], whose gram matrices have at most $2\lceil N/n\rceil+6$ non-zeros per row, guaranteeing that $\unrec$ is proportional only to the redundancy factor $N/n$. In fact, one can show that these are sparsest frames possible [@CHKK11], suggesting that families of tight frames with depending linearly on the redundancy factor should be essentially optimal for $\ell_1$-analysis reconstruction. We pose as a problem for subsequent work to obtain such estimates for dictionaries of practical interest, discussing a few examples in Section \[sec:main\]. First, to illustrate that our results go strictly beyond existing theory, we show that harmonic frames (frames constructed by removing the high frequency rows of the DFT – see ) with small redundancy indeed have a bounded . To our knowledge, this important case is not covered by existing theories. In addition, we also bound the for redundant Haar wavelet frames. We expect, however, that it will be difficult to efficiently compute the of an arbitrary dictionary. For bounded s, we prove the following theorem.
\[thm:intro\] Fix a sparsity level $s < N$. Let $\D\in\C^{n\times N}$ be a Parseval frame – i. e., $\D\D^*$ is the identity – with columns $\{\d_1, \ldots, \d_N\}$, and let $\B=\{\b_1,\dots, \b_n\}$ be an orthonormal basis of $\C^n$ which is [*incoherent*]{} to $\D$ in the sense that $$\label{eq:incoh}
\sup_{i \in [n]} \sup_{ j\in [N]} | \langle{ \b_i, \d_j \rangle} | \leq K n^{-1/2}$$ for some constant $K \geq 1$.
Consider the (unweighted) $\unrec = \unrec_{s,\D}$ as in Definition \[unrec-orig\]. Construct $\widetilde{\B}\in\C^{m\times n}$ by sampling row vectors from $\B$ i.i.d. uniformly at random. Provided $$\label{mboundloc}
m \geq C s K^2 \unrec^2\log^3(s\unrec^2) \log(N), \quad$$ then with probability $1-N^{-\log^3 (2s)}$, $\sqrt{\frac{n}{m}}\widetilde \B$ exhibits uniform recovery guarantees for $\ell_1$-analysis. That is, for every signal $\f$, the solution $\f^{\sharp}$ of the minimization problem with $\y=\sqrt{\frac{n}{m}}\widetilde\B\f+{\mathrm{e}}$ for noise ${\mathrm{e}}$ with $\|{\mathrm{e}}\|_2\leq \varepsilon$ satisfies $$\label{eq:intrec}
\|\f^{\sharp}-\f\|_2 \leq C_1 \varepsilon + C_2 \frac{\|\D^*\f - (\D^*\f)_s \|_1}{\sqrt{s}}.$$ Here, $C, C_1$, and $C_2$ are absolute constants independent of the dimensions and the signal $\f$.
Above and throughout, $(\u)_s$ denotes the best $s$-sparse approximation to a signal $\u$, that is, $(\u)_{s} {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\emph{arg}\min_{{\lVert {\z} \rVert}_{0} \leq s}{{\lVert {\u-\z} \rVert}_2}$. [**Remarks.**]{}\
[**1.**]{} Our stability result for $\ell_1$ analysis guarantees recovery for the particular $\f = \D\z$ only up to the scaled $\ell_1$ norm of the tail of $\D^*\D\z$. In fact, the quantity $\mathcal{E}^{*}_{\f} = \frac{\|\D^*\f - (\D^*\f)_s\|_1}{\sqrt{s}}$, referred to as the *unrecoverable energy* of the signal $\f$ in $\D$ [@RefWorks:60], is closely related to the $\eta_{s}$: $$\eta_{s} = \sup_{\f = \D \z: \hspace{.5mm} \| \f \|_2 = 1, \| \z \|_0 \leq s} \frac{\|\D^*\f \|_1}{\sqrt{s}} \leq \sup_{\f = \D \z: \hspace{.5mm} \| \f \|_2 = 1, \| \z \|_0 \leq s} \mathcal{E}^{*}_{f} + 1$$ [**2.**]{} As mentioned, the proof of this theorem will proceed via the $\D$-RIP, so our analysis yields similar guarantees for other recovery methods as well. [**3.**]{} It is interesting to remark on the role of incoherence in this setting. While prior work in compressed sensing has required the measurement matrix $\widetilde\B$ itself be incoherent, we now ask instead for incoherence between the basis from which measurements are selected and the dictionary $\D$, rather than incoherence within $\D$. Of course, note that the coherence of the dictionary $\D$ itself impacts the localization factor $\unrec_{s, \D}$. We will see later in Theorem \[thm:main\] that the incoherence requirement between the basis and dictionary can be weakened even further by using a weighted approach.
[**4.**]{} The assumption that $\D$ is a Parseval frame is not necessary but made for convenience of presentation. Several results have been shown for frames which are not tight, e.g., [@liu2012compressed; @rauhut2013analysis; @giryes2014sampling]. Indeed, any frame $\D \in\C^{n\times N}$ with linearly independent rows is such that $$\tilde{\D} := (\D \D^{*})^{-1/2} \D$$ is a tight frame. As observed in the recent paper [@Foucart15], in this case one can use for sampling the adapted matrix $\tilde{\A} = (\tilde{\B} (\D \D^{*})^{-1/2}),$ as measuring the signal $\f = \D \z$ through $\y = \tilde{\A} \f + {\mathrm{e}}$ is equivalent to measuring the signal $\tilde{\f} = \tilde{\D} \z$ through $\y = \tilde{\B} \tilde{\f} + {\mathrm{e}}$. Working through this extension poses an interesting direction for future work.
Organization
------------
The rest of the paper is organized as follows. We introduce some notation and review some technical background in Section \[sec:notate\] before presenting a technical version of our main result in Section \[sec:main\], which demonstrates that bases which are incoherent with a dictionary $\D$ can be randomly subsampled to obtain $\D$-RIP matrices. In that section, we also discuss some relevant applications and the implications of our results, including the regime where the sampling basis is coherent to the dictionary. The proof of our main theorem is presented in the final section.
Notation and technical background {#sec:notate}
=================================
Throughout the paper, we write $C$, $C'$, $C_1$, $C_2,\ldots$ to denote absolute constants; their values can change between different occurrences. We write ${\mathbf{I}}_n$ to denote the $n\times n$ identity matrix and we denote by $[n]$ the set $\{1, 2, \ldots, n\}$. For a vector $\u$, we denote the $j$th index by $\u[j]$ or $\u(j)$ or $\u_j,$ depending on which notation is the most clear in any given context; its restriction to only those entries indexed by a subset $\Lambda$ is donated by $u_\Lambda$. These notations should not be confused with the notation $(\u)_s$, which we use to denote the best $s$-sparse approximation to $\u$, that is, $(\u)_{s} {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\text{arg}\inf_{{\lVert {\z} \rVert}_{0} \leq s}{{\lVert {\u-\z} \rVert}_2}$. Similarly for a matrix $\A$, $\A_j$ is the $j$th column and $\A_\Lambda$ is the submatrices consisting of the columns with indices in $\Lambda$.
The unweighted case with incoherence
------------------------------------
Recall that a dictionary $\D\in\C^{n\times N}$ is a [*Parseval frame*]{} if $\D\D^*={\mathbf{I}}_n$ and that a vector $\x$ is [*$s$-sparse*]{} if $\|\x\|_0 := |\operatorname*{supp}(\x)| \leq s$. Then the restricted isometry property with respect to the dictionary $\D$ ($\D$-RIP) is defined as follows.
\[DRIP-orig\] Fix a dictionary $\D\in\C^{n\times N}$ and matrix $\A\in\C^{m\times n}$. The matrix $\A$ satisfies the $\D$-RIP with parameters $\delta$ and $s$ if $$\label{DRIPbound-orig}
(1-\delta){{{\lVert {\D\x} \rVert}_2}^2} \leq {{{\lVert {\A\D\x} \rVert}_2}^2} \leq (1+\delta){{{\lVert {\D\x} \rVert}_2}^2}$$ for all $s$-sparse vectors $\x\in\C^N$.
Note that when $\D$ is the identity, this definition reduces to the standard definition of the restricted isometry property [@RefWorks:48]. Under such an assumption on the measurement matrix, the following results bound the reconstruction error for the $\ell_1$-analysis method.
\[thm:l1anarec\] Let $\D$ be a tight frame, $\varepsilon>0$, and consider a matrix $\A$ that has the $\D$-RIP with parameters $2s$ and $\delta<0.08$. Then for every signal $\f\in\C^n$, the reconstruction $\f^{\sharp}$ obtained from noisy measurements $\y=\A\f+{\mathrm{e}}$, $\|{\mathrm{e}}\|_2\leq \varepsilon$, via the $\ell_1$-analysis problem above satisfies
$$\label{L1bound}
{{\lVert {\f-\f^{\sharp}} \rVert}_2} \leq \varepsilon + \frac{{\lVert {\D^*\f - (\D^*\f)_s} \rVert}_{1}}{\sqrt{s}}.$$
Thus as indicated in the introduction, one obtains good recovery for signals $\f$ whose analysis coefficients $\D^*\f$ have a suitable decay.
Both the RIP and the $\D$-RIP are closely related to the Johnson-Lindenstrauss lemma [@RefWorks:327; @RefWorks:326; @RefWorks:86]. Recall the classical variant of the lemma states that for any $\varepsilon\in (0, 0.5)$ and points $\x_1, \ldots, \x_d \in \mathbb{R}^n$, that there exists a Lipschitz function $f : \mathbb{R}^n \rightarrow \mathbb{R}^m$ for some $m = O(\varepsilon^{-2}\log d)$ such that $$\label{JL}
(1-\varepsilon) \|\x_i - \x_j\|_2^2 \leq \|f(\x_i) - f(\x_j)\|_2^2 \leq (1+\varepsilon)\|\x_i-\x_j\|_2^2$$ for all $i, j \in \{1, \ldots, d\}$. Recent improvements have been made to this statement which in particular show that the map $f$ can be taken as a random linear mapping which satisfies with high probability (see, e.g., [@RefWorks:90] for an account of such improvements). Indeed, any matrix $\A\in\C^{m\times n}$ which for a fixed[^1] vector $\z\in\C^n$ satisfies $$\P\left((1-\delta){{{\lVert {\z} \rVert}_2}^2} \leq {{{\lVert {\A\z} \rVert}_2}^2} \leq (1+\delta){{{\lVert {\z} \rVert}_2}^2}\right) \leq Ce^{-cm}$$ will satisfy the $\D$-RIP with high probability as long as $m$ is at least on the order of $s\log(n/s)$ [@RefWorks:60]. From this, any matrix satisfying the Johnson-Lindenstrauss lemma will also satisfy the $\D$-RIP (see [@RefWorks:86] for the proof of the RIP for such matrices, which directly carries over to the $\D$-RIP). Random matrices known to have this property include matrices with independent subgaussian entries (such as Gaussian or Bernoulli matrices), see for example [@Dasgupta]. Moreover, it is shown in [@RefWorks:230] that any matrix that satisfies the classical RIP will satisfy the Johnson-Lindenstrauss lemma and thus the $\D$-RIP with high probability after randomizing the signs of the columns. The latter construction allows for structured random matrices with fast multiplication properties such as randomly subsampled Fourier matrices (in combination with the results from [@RefWorks:285]) and matrices representing subsampled random convolutions (in combination with the results from [@rauhut2012restricted; @KMR12]); in both cases, however, again with randomized column signs. While this gives an abundance of such matrices, as mentioned above, it is not always practical or possible to apply random column signs in the sampling procedure.
An important general set-up for structured random sensing matrices known to satisfy the regular RIP is the framework of *bounded orthonormal systems*, which includes as a special case the subsampled discrete Fourier transform measurements (without column signs randomized). Such measurements are the only natural measurements possible in many physical systems where compressive sensing is of interest, such as in MRI, radar, and astronomy [@RefWorks:82; @RefWorks:79; @RefWorks:289; @bobin2008compressed], as well as in applications to polynomial interpolation [@rw11; @burq2012weighted; @rauhut2011sparse] and uncertainty quantification [@hampton2015compressive]. In the following, we recall this set-up (in the discrete setting), see [@ra09-1 Sec.4] for a detailed account of bounded orthonormal systems and examples.
\[BONS\] Consider a probability measure $\nu$ on the discrete set $[n]$ and a system $\{ \r_j \in \C^n$, $j \in [n] \}$ that is orthonormal with respect to $\nu$ in the sense that $$\sum_{i} \r_k(i) \r_j(i) \nu_i = \delta_{j,k}$$ where $\delta_{j,k}=\begin{cases}
1 & \text{ if } j = k\\
0 & \text{ else.}
\end{cases}$ is the Kronecker delta function. Suppose further that the system is uniformly bounded: there exists a constant $K \geq 1$ such that $$\label{local}
\sup_{i \in [n]} \sup_{j \in [n]} | \r_j(i) | \leq K.$$ Then the matrix $\A \in \C^{n \times n}$ whose rows are $\r_j$ is called a *bounded orthonormal system* matrix.
Drawing $m$ indices $i_1, i_2, \dots, i_m$ independently from the orthogonalization measure $\nu$, the sampling matrix $\widetilde{\A} \in \C^{m \times n}$ whose rows are indexed by the (re-normalized) sampled vectors $\sqrt{\frac{1}{m}} \r_j(i_k) \in \C^n$ will have the restricted isometry property with high probability (precisely, with probability exceeding $1-n^{-C \log^3(s)}$) provided the number of measurements satisfies $$\label{BOS:RIP:cond}
m \geq CK^2 s \log^2(s) \log(n).$$ This result was first shown in the case where $\nu$ is the uniform measure by Rudelson and Vershynin [@RefWorks:285], for a slightly worse dependence of $m$ on the order of $s\log^2 s \log(s\log n) \log(n)$. These results were subsequently extended to the general bounded orthonormal system set-up by Rauhut [@ra09-1], and the dependence of $m$ was slightly improved to $s\log^3 s\log n$ in [@rauhut2013interpolation].
An important special case where these results can be applied is that of [*incoherence*]{} between the measurement and sampling bases. Here the [*coherence*]{} between two sets of vectors $\Phi=\{\phi_i\}$ and $\Psi=\{\psi_j\}$ is given by $\mu=\sup_{i,j} |\langle \phi_i, \psi_j\rangle|$. Two orthonormal bases $\Phi$ and $\Psi$ of $\C^n$ are called incoherent if $\mu \leq K n^{-1/2}$. In this case, the renormalized system $\tilde\Phi=\{\sqrt{n} \phi_i \Psi^*\}$ is an orthonormal system with respect to the uniform measure, which is bounded by $K$. Then the above results imply that signals which are sparse in basis $\Psi$ can be reconstructed from inner products with a uniformly subsampled subset of basis $\Phi$. These incoherence-based guarantees are a standard criterion to ensure signal recovery, such results had first been observed in [@RefWorks:379].
Generalization to a weighted setup and to local coherence
---------------------------------------------------------
Recently, the criterion of incoherence has been generalized to the case where only *most* of the inner products between sparsity basis vectors and measurement basis vectors are bounded. If some vectors in the measurement basis yield larger inner products, one can adjust the sampling measure and work with a preconditioned measurement matrix [@KW14], and if some vectors in the sparsity basis yield larger inner products, one can incorporate weights into the recovery schemes and work with a weighted sparsity model [@rauhut2013interpolation]. Our results can accommodate both these modifications. In the remainder of this section, we will recall some background on these frameworks from [@KW14; @rauhut2013interpolation] and formulate our definitions for the more general setup.
Consider a set of positive weights $\omega = (\omega_j)_{j \in [N]}$. Associate to these weights the norms $$\label{weighted_norm}
\| \x \|_{\omega,p} := \left( \sum_j | x_j |^p \omega_j^{2-p} \right)^{1/p}, \quad 0 < p \leq 2,$$ along with the weighted “$\ell_0$-norm", or weighted sparsity: $\| \x \|_{\omega,0} := \sum_{j: | x_j | > 0} \omega_j^2$, which equivalently is defined as the limit $\| \x \|_{\omega,0} = \lim_{p \rightarrow 0} \| \x \|_{\omega,p}^p$. We say that a vector $\x$ is [**weighted $s$-sparse**]{} if $\| \x \|_{\omega,0} := \sum_{j: | x_j | > 0} \omega_j^2 \leq s$. In line with this definition, the weighted size of a finite set $\Lambda \subset \NN$, is given by $\omega(\Lambda):= \sum_{j\in\Lambda} \omega_j^2$; thus a vector is weighted $s$-sparse if its support has weighted size at most $s$. When $\omega_j \geq 1$ for all $j$, the weighted sparsity of a vector is at least as large as its unweighted sparsity, so that the class of weighted $s$-sparse vectors is a strict subset of the $s$-sparse vectors. We make this assumption in the remainder. Note in particular the special cases $\| \x \|_{\omega,1} = \sum_j | x_j | \omega_j$ and $\| \x \|_{\omega,2} = \| \x \|_{2} = \sqrt{\sum_j | x_j |^2};$ by Cauchy-Schwarz, it follows that $ \| \x \|_{\omega,1} \leq \sqrt{s} \| \x \|_2$ if $\x$ is weighted $s$-sparse. Indeed, we can extend the notions of and $\D$-RIP to the weighted sparsity setting. It should be clear from context which definition we refer to in the remainder.
\[eq:Wunrec\] For a dictionary $\D\in\C^{n\times N}$, weights $\omega_1, \omega_2, \dots, \omega_N \geq 1$, and a sparsity level $s$, we define the (weighted) ** as $$\label{unrec-full}
\unrec_{\omega, s, \D} =\unrec {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{\omega,0}\leq s} \frac{{\lVert {\D^*\D\z} \rVert}_{\omega,1}}{\sqrt{s}},$$
\[DRIP\] Fix a dictionary $\D\in\C^{n\times N}$, weights $\omega_1, \omega_2, \dots, \omega_N \geq 1$, and matrix $\A\in\C^{m\times n}$. The matrix $\A$ satisfies the $\D$-$\omega$RIP with parameters $\delta$ and $s$ if $$\label{DRIPbound}
(1-\delta){{{\lVert {\D\x} \rVert}_2}^2} \leq {{{\lVert {\A\D\x} \rVert}_2}^2} \leq (1+\delta){{{\lVert {\D\x} \rVert}_2}^2}$$ for all [**weighted $s$-sparse**]{} vectors $\x\in\C^N$.
When $\D$ is an orthonormal matrix, this definition reduces to the definition $\omega$RIP from [@rauhut2013interpolation]. More generally, weights allow the flexibility to incorporate prior information about the support set and allow for weaker assumptions on the dictionary $\D$. In particular, a larger weight assigned to a dictionary element will allow for this element to have larger inner products with the measurement basis vectors. In this regard, a basis version of our result is the following variant of a theorem from [@rauhut2013interpolation]:
\[weights:RW\] Fix a probability measure $\nu$ on $[n]$, sparsity level $s < n$, and constant $0 < \delta < 1$. Let $\D\in\C^{n \times n}$ be an orthonormal matrix. Let $\A$ be an orthonormal system with respect to $\nu$ as in Definition \[BONS\] with rows denoted by $\r_i$, and consider weights $\omega_1, \omega_2, \dots, \omega_n \geq 1$ such that $\max_i | \langle{ \r_i, \d_j \rangle} | \leq \omega_j$. Construct an $m\times n$ submatrix $\widetilde{\A}$ of $\A$ by sampling rows of $\A$ according to the measure $\nu$ and renormalizing each row by $\sqrt{\frac{1}{m}}$. Provided $$\label{mbound}
m \geq C\delta^{-2} s \log^3(s) \log(n),$$ then with probability $1-n^{c\log^3 s}$, the submatrix $\widetilde{\A}$ satisfies the $\omega$RIP with parameters $s$ and $\delta$.
An adjusted sampling density for the case when certain vectors in the measurement basis yield larger inner products is obtained via the [**local coherence**]{} as defined in the following.
The *local coherence* of a set $\Phi=\{\varphi_i\}_{i=1 }^k \subset \C^n$ with respect to another set $\Psi=\{\psi_j\}_{j=1 }^\ell \subseteq \C^n$ is the function $\mu^{\text{loc}}(\Phi, \Psi) \in \R^k$ defined coordinate-wise by
[ $$\mu^{\text{loc}}_i(\Phi, \Psi) = \mu^{{\text loc}}_i = \sup\limits_{1\leq j\leq \ell } |\langle \varphi_i, \psi_j\rangle|, \quad \quad i = 1,2,\dots, k.$$ ]{}
When $\Phi$ and $\Psi$ are orthonormal bases, and a subset of $\Phi$ is used for sampling while $\Psi$ is the basis in which sparsity is assumed, the main point is that renormalizing the different vectors $\varphi_i$in the sampling basis by respective factors $\frac{1}{\mu^{{\text loc}}_i}$ does not affect orthogonality of the rows of $\Phi^{*} \Psi$. In this way, the larger inner products can be reduced in size. To compensate for this renormalization and retain the orthonormal system property in the sense of Definition \[BONS\], one can then adjust the sampling measure. The resulting preconditioned measurement system with variable sampling density will then yield better bounds $K$ in Equation . This yields the following estimate for the restricted isometry property.
\[thm:wBOS\] Let $\Phi=\{\varphi_j\}_{j=1}^n$ and $\Psi =\{\psi_k\}_{k=1}^n$ be orthonormal bases of $\C^n$. Assume the local coherence of $\Phi$ with respect to $\Psi$ is pointwise bounded by the function $\kappa/\sqrt{n} \in \C^n$, that is $ \sup\limits_{1\leq k\leq n} |\langle \varphi_j, \psi_k\rangle| \leq \kappa_j$. Suppose $$\label{wBOS:RIP:cond}
m \geq C\delta^{-2} \|\kappa \|_2^2 s \log^3(s) \log(n),$$ and choose $m$ (possibly not distinct) indices $j\in [n]$ i.i.d. from the probability measure $\nu$ on $[n]$ given by $$\nu(j) = \frac{\kappa^2_j}{\|\kappa \|_2^2 }.$$ Call this collection of selected indices $\Omega$ (which may possibly be a multiset). Consider the matrix $\A \in \C^{m \times n}$ with entries $$\label{def:wPhi:matrix}
A_{j,k} = \langle \varphi_j, \psi_k\rangle, \quad j \in \Omega, k \in [n],$$ and consider the diagonal matrix $\W = \operatorname{diag}(\w) \in \C^{n \times n}$ with $w_{j} = \| \kappa \|_2 / \kappa_j$. Then with probability at least $1-n^{-c \log^3(s)},$ the preconditioned matrix $\frac{1}{\sqrt{m}} \W \A$ has the restricted isometry property with parameters $\delta$ and $s$.
[**Remark.**]{} In case $\Phi$ and $\Psi$ are incoherent, or if $\kappa_j = K n^{-1/2}$ uniformly for all $j$, then local coherence sampling as above reduces to the previous results for incoherent systems: in this case, the associated probability measure $\nu$ is uniform, and the preconditioned matrix reduces to $\frac{1}{\sqrt{m}} \W \A = \sqrt{ \frac{n}{m} } \A$. Our main result on $\D$-RIP for redundant systems and structured measurements (Theorem \[thm:main\]) implies a strategy for extending local coherence sampling theory to dictionaries, of which Theorem \[thm:intro\] is a special case. Indeed, if $\Psi = \{ \psi_j \}$, more generally, is a Parseval frame and $\Phi = \{\varphi_i \}$ is an orthonormal matrix, then renormalizing the different vectors $\varphi_i$ in the sampling basis by respective factors $\frac{1}{\mu^{{\text loc}}_i}$ still does not affect orthogonality of the rows of $\Phi^{*} \Psi$. We have the following corollary of our main result; we will establish a more general result, Theorem \[thm:main\] in the next section, from which both Theorem \[thm:intro\] and Corollary \[thm:variable\] follow.
\[thm:variable\] Fix a sparsity level $s < N$, and constant $0 < \delta < 1$. Let $\D \in\C^{n\times N}$ be a Parseval frame with columns $\{\d_1, \ldots, \d_N\}$, let $\B$ with rows $\{\b_1,\dots, \b_n\}$ be an orthonormal basis of $\C^n$, and assume the local coherence of $\D$ with respect to $\B$ is pointwise bounded by the function $\kappa \in \C^n$, that is $ \sup\limits_{1\leq j \leq N} |\langle \d_j, \b_k\rangle| \leq \kappa_k$.
Consider the probability measure $\nu$ on $[n]$ given by $\nu(k) = \frac{\kappa^2_k}{\|\kappa \|_2^2 }$ along with the diagonal matrix $\W = \operatorname{diag}(\w) \in \C^{n \times n}$ with $w_{k} = \| \kappa \|_2 / \kappa_k$. Note that the renormalized system $\Phi = \frac{1}{\sqrt{m}} \W \B$ is an orthonormal system with respect to $\nu$, bounded by $\| \kappa \|^2$. Construct $\widetilde{\B}\in\C^{m\times n}$ by sampling vectors from $\B$ i.i.d. from the measure $\nu$, and consider the $\unrec = \unrec_{\D,s}$ of the frame $\D$ as in Definition \[unrec-orig\],
As long as $$m \geq C\delta^{-2} \unrec^2 \|\kappa \|_2^2 s \log^3(s \unrec^2) \log(N)$$ then with probability $1-N^{-\log^3(2s)}$, the following holds for every signal $\f$: the solution $\f^{\sharp}$ of the weighted $\ell_1$-analysis problem $$\f^{\sharp}= \operatorname*{arg min}_{\tilde \f\in\C^n} \|\D^*\tilde\f\|_1 \text{\quad such that \quad} \| \frac{1}{\sqrt{m}} \W (\widetilde{\B} \tilde\f-\y) \|_2\leq \varepsilon \tag{$P_{1,w}$}, \label{eq:l1anaW}$$ with $\y=\B \f+{\mathrm{e}}$ for noise ${\mathrm{e}}$ with weighted error $\| \frac{1}{\sqrt{m}} \W {\mathrm{e}}\|_2 \leq \varepsilon$ satisfies $$\|\f^{\sharp}-\f\|_2 \leq C_1 \varepsilon + C_2 \frac{\|\D^*\f - (\D^*\f)_s\|_1}{\sqrt{s}}.$$ Here $C, C_1$, and $C_2$ are absolute constants independent of the dimensions and the signal $\f$.
The resulting bound involves a weighted noise model, which can, in the worst case, introduce an additional factor of $\sqrt{\frac{n}{m}}\max_i \sqrt{\nu(i)}$. We believe however that this noise model is just an artifact of the proof technique, and that the stability results in Corollary \[thm:variable\] should hold for the standard noise model using the standard $\ell_1$-analysis problem . In the important case where $\D$ is a wavelet frame and $\B$ is the orthonormal discrete Fourier matrix, we believe that *total variation minimization*, like $\ell_1$-analysis, should give stable and robust error guarantees with the standard measurement noise model. Indeed, such results were recently obtained for variable density sampling in the case of orthonormal wavelet sparsity basis [@poon14], improving on previous bounds for total variation minimization [@KW14; @needell2013stable; @needell2013near; @adcock2013breaking]. Generalizing such results to the dictionary setting is indeed an interesting direction for future research.
Our main result and its applications {#sec:main}
====================================
As mentioned in the previous section, the case of incoherence between the sampling basis and the sparsity dictionary, as in Theorem \[thm:intro\], is a special case of a bounded orthonormal system. The following more technical formulation of our main result in the framework of such systems covers both weighted sparsity models and local coherence and as we will see, implies Theorem \[thm:intro\]. The proof of Theorem \[thm:main\] is presented in Section \[sec:proof\].
\[thm:main\] Fix a probability measure $\nu$ on $[N]$, sparsity level $s < N$, and constant $0 < \delta < 1$. Let $\D\in\C^{n\times N}$ be a Parseval frame. Let $\A$ be an orthonormal systems matrix with respect to $\nu$ with rows $\r_i$ as in Definition \[BONS\], and consider weights $\omega_1, \omega_2, \dots, \omega_N \geq 1$ such that $$\label{eq:wts}
\max_i | \langle{ \r_i, \d_j \rangle} | \leq \omega_j.$$ Define the unrecoverable energy $\unrec = \unrec_{\D,s}$ as in Definition \[unrec-full\]. Construct an $m\times n$ submatrix $\widetilde{\A}$ of $\A$ by sampling rows of $\A$ according to the measure $\nu$. Then as long as $$\begin{aligned}
\label{mbound3}
m &\geq& C\delta^{-2} s \unrec^2\log^3(s \unrec^2) \log(N), \quad \text{and} \nonumber \\
m &\geq& C\delta^{-2} s \unrec^2 \log(1/\gamma)\end{aligned}$$ then with probability $1-\gamma$, the normalized submatrix $\sqrt{\tfrac{1}{m}} \widetilde{\A}$ satisfies the $\D$-$\omega$RIP with parameters $s$ and $\delta$.
Proof of Theorem \[thm:intro\] assuming Theorem \[thm:main\]: {#proof-of-theoremthmintro-assuming-theoremthmmain .unnumbered}
-------------------------------------------------------------
Let $\B$, $\D$ and $s$ be given as in Theorem \[thm:intro\]. We will apply Theorem \[thm:main\] with $\omega_j = 1$ for all $j$, with $\nu$ the uniform measure on $[n]$, sparsity level $2s$, $\gamma = N^{-\log^3 (2s)}$, $\delta = 0.08$, and matrix $\A=\sqrt{n}\B$. We first note that since $\B$ is an orthonormal basis, that $\sqrt{n}\B$ is an orthonormal systems matrix with respect to the uniform measure $\nu$ as in Definition \[BONS\]. In addition, implies for matrix $\A=\sqrt{n}\B$ with $K=\omega_j = 1$. Furthermore, setting $\gamma = N^{-\log^3(2s)}$, implies both inequalities of (adjusting constants appropriately). Thus the assumptions of Theorem \[thm:main\] are in force. Theorem \[thm:main\] then guarantees that with probability $1-\gamma = 1 - N^{-\log^3 (2s)}$, the uniformly subsampled matrix $\sqrt{\frac{1}{m}}(\sqrt{n}\widetilde{\B})$ satisfies the $\D$-$\omega$RIP with parameters $2s$ and $\delta$ and weights $\omega_j = 1$. A simple calculation and the definition of the $\D$-$\omega$RIP (Definition \[DRIP\]) shows that this implies the $\D$-RIP with parameters $2s$ and $\delta$. By Theorem \[thm:l1anarec\], holds and this completes the proof. $\square$
Proof of Corollary \[thm:variable\] assuming Theorem \[thm:main\] {#proof-of-corollary-thmvariable-assuming-theorem-thmmain .unnumbered}
-----------------------------------------------------------------
Corollary \[thm:variable\] is implied by Theorem \[thm:main\] because the preconditioned matrix $\Phi = \frac{1}{\sqrt{m}} \W \B$ formed by normalizing the $i$th row of $\B$ by $ \| \kappa \|_2 / \kappa_j$ constitutes an orthonormal system with respect to the probability measure $\nu(i)$ and uniform weights $\omega_j = \| \kappa \|_2$. Then the preconditioned sampled matrix $\widetilde{\Phi}$ satisfies the $\D$-RIP for sparsity $2s$ and parameter $\delta$ according to Theorem \[thm:main\] with probability $1-\gamma = 1 - N^{-\log^3 (2s)}$. Applying Theorem \[thm:l1anarec\] to the preconditioned sampling matrix $\widetilde{\Phi}$ and Parseval frame $\D$ produces the results in Corollary \[thm:variable\]. $\square$
Example: Harmonic frame with $L$ more vectors than dimensions
-------------------------------------------------------------
It remains to find examples of a dictionary with bounded and an associated measurement system for which incoherence condition holds. Our main example is that of sampling a signal that is sparse in an oversampled Fourier system, a so-called [*harmonic frame*]{} [@vale2004tight]; the measurement system is the standard basis. Indeed, one can see by direct calculation that the standard basis is incoherent in the sense of to any set of Fourier vectors, even of non-integer frequencies. We will now show that if the number of frame vectors exceeds the dimension only by a constant, such a dictionary will also have bounded . This setup is a simple example, but our results apply, and it is not covered by previous theory.
More precisely, we fix $L\in\NN$ and consider $N=n+L$ vectors in dimension $n$. We assume that $L$ is such that $Ls\leq \tfrac{N}{4}$. Then the harmonic frame is defined via its frame matrix $\D=(d_{jk})$, which results from the $N\times N$ discrete Fourier transform matrix $\F=(f_{jk})$ by deleting the last $L$ rows. That is, we have $$\label{harmonic}
d_{jk} = \tfrac{1}{\sqrt{n+L}}\exp(\tfrac{2\pi i j k}{n+L})$$ for $j=1\ldots n$ and $k=1\ldots N$. The corresponding Gram matrix satisfies $(\D^*\D)_{ii}=\tfrac{n}{n+L}$ and, for $j\neq k$, by orthogonality of $\F$, $$|(\D^*\D)_{jk}|= \left|(\F^*\F)_{jk} - \sum_{\ell=n+1}^N \bar f_{\ell j} f_{\ell k} \right| = \left| \sum_{\ell=n+1}^N \tfrac{1}{n+L} \exp(\tfrac{-2\pi i \ell j}{n+L}) \exp(\tfrac{2\pi i \ell k}{n+L}) \right| \leq \tfrac{L}{n+L}$$ As a consequence, we have for $\z$ $s$-sparse and $j\notin \operatorname*{supp}\z$ $$|(\D^*\D \z)[j]| = \left|\sum_{k\in\operatorname*{supp}\z} (\D^*\D)_{jk} \z[k] \right|\leq \tfrac{L}{n+L} \|\z\|_1$$ where we write $\z[k]$ to denote the $k$th index of the vector $\z.$ Similarly, for $j\in \operatorname*{supp}\z$, we obtain $$|(\D^*\D \z)[j] - \z[j]| = \left|\sum_{k\in\operatorname*{supp}\z, k\neq j} (\D^*\D)_{jk} \z[k] \right|\leq \tfrac{L}{n+L} \|\z\|_1.$$ So we obtain, using that $\D^*$ is an isometry, $$\begin{aligned}
\|\D\z\|^2_2 & =\|\D^*\D\z\|^2_2\geq \sum _{k\in \operatorname*{supp}\z}\left( (\D^*\D \z)[k] \right)^2
\geq \sum _{k\in \operatorname*{supp}\z}(|\z[k]|-|(\D^*\D \z)[k] - \z[k]|)^2 \\
&\geq \sum _{k\in \operatorname*{supp}\z}|\z[k] |^2 - 2 \tfrac{L}{n+L} \|\z\|_1 |\z[k] |\\
&= \|\z\|_2^2-\tfrac{2L}{n+L} \|\z\|_1^2 \geq (1-\tfrac{2Ls}{N})\|\z\|_2^2\geq \tfrac{1}{2} \|\z\|_2^2.\end{aligned}$$ That is, for $\z$ with $\|\D\z\|_2=1$, one has $\|\z\|_2\leq \sqrt{2}$. Consequently, $$\begin{aligned}
\unrec&=\sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} \frac{{\lVert {\D^*\D\z} \rVert}_{1}}{\sqrt{s}} = \sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} \frac{{\lVert {(\D^*\D\z)|_{(\operatorname*{supp}\z)}} \rVert}_1 + {\lVert {(\D^*\D\z)|_{(\operatorname*{supp}\z)^c}} \rVert}_{1}}{\sqrt{s}}\\
&\leq \sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} {\lVert {\D^*\D\z} \rVert}_2 + \frac{{\lVert {(\D^*\D\z)|_{(\operatorname*{supp}\z)^c}} \rVert}_{1}}{\sqrt{s}} = 1 + \sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s}\tfrac{1}{\sqrt{s}} \sum_{j\in(\operatorname*{supp}\z)^c} |(\D^*\D \z)[j]| \\
&\leq 1 + \sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s}\tfrac{1}{\sqrt{s}} (N-s)\tfrac{L\|\z\|_1}{N} \leq 1 + \tfrac{(N-s)L\|\z\|_2}{N} \leq 1 + L \sqrt{2}.$$ $\square$ [**Remark.** ]{} We note here that the dependence on $L$ seems quite pessimistic. Indeed, one often instead considers the redundancy of the dictionary, namely $r=N/n$. In this case, even for redundancy $r=2$ we would have $L=n$ and require a full set of measurements according to this argument. However, we conjecture that rather than a dependence on $L$ one can reduce the dependence to one on the redundancy $r$. Intuitively, due to the restriction of Parseval normalization on the frame $\D$, there is a tradeoff between redundancy and coherence; an increase in redundancy should be in some way balanced by a decrease in coherence. This conjecture is supported by numerical experiments and by theoretical lower bounds for robust recovery in the co-sparse model [@GPV14], but we leave a further investigation to future work. This example also plays an important role in the so-called *off the grid* compressed sensing setting [@tang2012compressive; @stoica2012sparse]. In this framework, the signal frequencies are not assumed to lie on a lattice but instead can assume any values in a continuous interval. When the signal parameters do not lie on a lattice, the signal may not be truly sparse in the discrete dictionary, and refining the grid to incorporate finer frequencies may lead to numerical instability. In addition, typical compressed results are difficult to apply under discretization, making off the grid approaches advantageous. The example we discuss above gives a possible compromise to nearly on the grid recovery from linear measurements.
Another line of research that relates to this example is that of superresolution (see, e.g., [@reading2; @reading1] and many followup works). These works study the recovery of frequency sparse signals from equispaced samples. No assumptions regarding an underlying grid are made, but rather one assumes a separation distance between the active frequencies. In this sense, the nearly-on-the-grid example just discussed satisfies their assumption as the corresponding separation distance is close to the grid spacing. However, the nature of the resulting guarantees is somewhat different. For example, in these works, every deviation from an exactly sparse signal must be treated as noise, whereas in our result above we have an additional term capturing the compressibility of a signal. We think that an in-depth comparison of the two approaches is an interesting topic for future work.
Example: Fourier measurements and Haar frames of redundancy 2 {#sec:variable}
-------------------------------------------------------------
In this subsection, we present a second example of a sampling setup that satisfies the assumptions of incoherence and of Theorem \[thm:main\]. In contrast to the previous example, one needs to precondition and adjust the sampling density to satisfy these assumptions according to Corollary \[thm:variable\], which allows us to understand the setup of the Fourier measurement basis and a 1D Haar wavelet frame with redundancy 2, as introduced in the following. Let $n = 2^p$. Recall that the univariate discrete Haar wavelet basis of $\C^{2^p}$ consists of $h^0 = 2^{-p/2} (1, 1, \dots, 1), h=h_{0,0}=2^{-p/2} (1,1,\dots, 1,-1,-1,\dots, -1)$ and the frame basis elements $h_{\ell,k}$ given component wise by $$\begin{aligned}
h_{\ell,k}[j] = h[2^{\ell} j -k]
&= \begin{cases}
2^{\frac{\ell-p}{2}} \quad&\text{for }\qquad \qquad \qquad k2^{p-\ell} \leq j< k2^{p-\ell} +2^{p-\ell-1}\\
-2^{\frac{\ell-p}{2}} \quad&\text{for }\quad\ k2^{p-\ell} +2^{p-\ell-1}\leq j < k2^{p-\ell} +2^{p-\ell}\\
0\quad &\text{else,}
\end{cases}\end{aligned}$$ for any $(\ell,k)\in\mathbb{Z}^2$ satisfying $0<\ell<p$ and $0\leq k<2^\ell$. The corresponding basis transformation matrix is denoted by $\H$.
One can now create a [*wavelet frame of redundancy 2*]{} by considering the union of this basis and a circular shift of it by one index. That is, one adds the vector $ \tilde{h}^{0} = h^{0} = 2^{-p/2}(1,1,\dots,1)$ and vectors of the form $\tilde{h}_{\ell,k}[j] = h_{\ell,k}[j+1]$ for all $(\ell,k)$. Here we identify $2^p + 1 = 1$. This is also an orthonormal basis – its basis transformation matrix will be denoted by $\tH$ in the following, and the matrix $\D \in \mathbb{C}^{2^p \times 2^{p+1}}$ with columns $$\begin{aligned}
\label{frameHaar}
\D(:,(\ell,2k-1)) &=& \frac{1}{\sqrt{2}} h_{\ell,k}, \nonumber \\
\D(:,(\ell,2k)) &=& \frac{1}{\sqrt{2}} \tilde{h}_{\ell,k}\end{aligned}$$ forms a Parseval frame with redundancy 2.
Corollary \[thm:variable\] applies to the example where sparsity is with respect to the redundant Haar frame and where sampling measurements are rows $\{ \r_k \}$ from the $n \times n$ orthonormal DFT matrix. Indeed, following Corollary 6.4 of [@KW14], we have the following coherence estimates: $$\max_{\ell, j} \{ | {\left\langle}\r_k, h_{\ell,j} {\right\rangle}|, | {\left\langle}\r_k, \tilde{h}_{\ell,j} {\right\rangle}| \} \leq \kappa_k := 3 \sqrt{2\pi} / \sqrt{k}.$$ Since $\| \kappa \|_2^2 = 18\pi \sum_{k=1}^n k^{-1} \leq 18 \pi \log_2(n)$ grows only mildly with $n$, the Fourier / wavelet frame example is a good fit for Corollary \[thm:variable\], provided the of the Haar frame is also small. We will show that the of the Haar frame is bounded by $\unrec \leq \sqrt{2\log_2(n)}$, leading to the following corollary.
\[fourier:haar\] Fix a sparsity level $s < N$, and constant $0 < \delta < 1$. Let $\D \in\C^{n\times N}$ ($N = 2n$) be the redundant Haar frame as defined in and let $\A \in \C^{n \times n}$ with rows $\{\r_1,\dots, \r_n\}$ be the orthonormal DFT matrix. Consider the diagonal matrix $\W = \operatorname{diag}(\w) \in \C^{n \times n}$ with $w_{k} = C' (\log_2(n))^{-1/2} k^{1/2}$, and construct $\widetilde{\A}\in\C^{m\times n}$ by sampling rows from $\A$ i.i.d. from the probability measure $\nu$ on $[n]$ with power-law decay $\nu(k) = C k^{-1}$.
As long as the number of measurements satisfies $m \geq C_1\delta^{-2} s \log^3(s \log(n)) \log^{3}(n),$ then with probability $1-n^{-\log^3 s}$, the following holds for every signal $\f$: the solution $\f^{\sharp}$ of the weighted $\ell_1$-analysis problem with $\y=\widetilde{\A} \f+{\mathrm{e}}$ for noise ${\mathrm{e}}$ with weighted error $\| \frac{1}{\sqrt{m}} \W {\mathrm{e}}\|_2 \leq \varepsilon$ satisfies $$\|\f^{\sharp}-\f\|_2 \leq C_2 \varepsilon + C_3 \frac{\|\D^*\f - (\D^*\f)_s\|_1}{\sqrt{s}} \nonumber$$ Here $C, C_1, C_2$, and $C_3$ are absolute constants independent of the dimensions and the signal $\f$.
To derive this result from Corollary \[thm:variable\], it suffices to prove that the redundant Haar frame as defined above has [ at most $\unrec \leq \sqrt{2\log_2(n)}$]{}. We will show that for each $s$-sparse $\z\in\R^N$, $\D^*\D\z$ is at most $3 s \log_2 n$-sparse. Here, $\D^{*} \D$ is the $N \times N$ Gramian matrix. For that, consider first either of the two (equal) frame elements $\tfrac{1}{\sqrt{2}}h^0$ and $\tfrac{1}{\sqrt{2}} \tilde{h}^{0}$. Each of these frame elements has non-zero inner product with exactly two frame elements: $\tfrac{1}{\sqrt{2}}h^0$ and $\tfrac{1}{\sqrt{2}}\tilde{h}^{0}$. So the corresponding columns of $\D^*\D$ have only two non-vanishing entries.
Consider then a non-constant frame element $\tfrac{1}{\sqrt{2}} h_{\ell,k}$. Because $\H$ is an orthonormal basis, it is orthogonal to all ${h}_{\ell', k'}$ with $(\ell',k') \neq (\ell,k)$, which is why $(\D^*\D)_{(k,2\ell), (k',2\ell')}=(\D^*\D)_{(k,2\ell-1), (k',2\ell'-1)}=0$ for $(\ell',k') \neq (\ell,k)$. So it remains to consider correlations with the $\tilde h_{\ell',k'}$. Again by orthogonality, one has for $(\ell',k')\neq (\ell,k)$ $$\langle \tilde{h}_{\ell',k'}, h_{\ell,k}\rangle = \delta_{(\ell',k'), (\ell,k)}+ \langle \tilde{h}_{\ell',k'}, h_{\ell,k}-\tilde{h}_{\ell,k}\rangle,$$ where $\delta_{(a,b),(c,d)}=\begin{cases}
1 & \text{ if }(a,b)=(c,d)\\
0 & \text{ else.}
\end{cases}
$ denotes the Kronecker delta. By definition, $h_{\ell,k}-\tilde{h}_{\ell,k}$ has only three non-zero entries. On the other hand, for each level $\ell'$, the supports of the $\tilde{h}_{\ell',k'}$ are disjoint for different values of $k'$. So for each $\ell'$, the supports of at most $3$ of the $\tilde{h}_{\ell',k'}$ can intersect the support of $h_{\ell,k}-\tilde{h}_{\ell,k}$. As for $\ell'=\ell$, one of these $\tilde{h}_{\ell',k'}$ must be $\tilde{h}_{\ell,k}$, we conclude that for each $\ell'$ at most $3$ of the $|\langle \tilde{h}_{\ell',k'}, h_{\ell,k}\rangle|$ can be nonzero. As there are $\log_2 n$ levels $\ell'$, this contributes at most $3\log_2 n$ nonzero entries in the column of $\D^*\D$ indexed by $(k,2\ell)$. Together with $(\D^*\D)_{(k,2\ell), (k,2\ell)}=1$ and noting that a similar analysis holds for columns indexed by $(k,2\ell-1)$, we obtain that each column of $\D^*\D$ has at most $3\log_2(n)+1$ non-zero entries. Now for each $s$-sparse $\z$, $\D^*\D\z$ is a linear combination of the $s$ columns corresponding to $\operatorname*{supp}\z$. Consequently, $\|\D^*\D\z-(\D^*\D\z)_s\|_0\leq 3s\log_2(n)$ and thus, by Cauchy-Schwarz and noting that for a Parseval frame $\D$, $\D \D^*$ is the identity, $$\begin{aligned}
\unrec &= \sup_{{{\lVert {\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} \frac{{\lVert {\D^*\D\z} \rVert}_1 }{\sqrt{s}} =\sup_{{{\lVert {\D^*\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} \frac{{\lVert {(\D^*\D\z)_{s}} \rVert}_1 + {\lVert {\D^*\D\z - (\D^*\D\z)_{s}} \rVert}_{1}}{\sqrt{s}}\\
&\leq 1 + \sup_{{{\lVert {\D^*\D\z} \rVert}_2}=1, {\lVert {\z} \rVert}_{0}\leq s} \frac{\sqrt{3s\log_2 n} {\lVert {\D^*\D\z - (\D^*\D\z)_{s}} \rVert}_{2}}{\sqrt{s}}\\
&\leq 1 + \sup_{{{\lVert {\D^*\D\z} \rVert}_2}=1} \sqrt{3\log_2 n} {\lVert {\D^*\D\z} \rVert}_{2} = 1 + \sqrt{3\log_2 n}.\end{aligned}$$
[**Remark.**]{} Note that this proof is closely related to the observation that each signal that is $s$-sparse in the Haar frame of redundancy $2$ is also $O(s\log n)$-sparse in the Haar wavelet basis. So a number of measurements comparable to the one derived here also allows for recovery of the wavelet basis coefficients. In addition, however, a more refined analysis suggests that the entries of $\D^*\D\z$ decay quickly - provided not too many approximate cancellations happen. We conjecture that the number of such approximate cancellations can be controlled using a more sophisticated analysis, but we leave this to follow-up work. We hence conjecture that the logarithmic factor can be removed, making the required number of measurements smaller than for a synthesis approach.
Proof of main result {#sec:proof}
====================
Our proof of Theorem \[thm:main\] extends the analysis in [@rauhut2013interpolation], which extends the analysis in [@cheraghchi2013restricted] to weighted sparsity and improves on the analysis in [@RefWorks:285], from orthonormal systems to redundant dictionaries.
Consider the set $$\Dsn {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\{\u : \u\in\C^n, \u = \D\z, {\lVert {\z} \rVert}_{\omega,0} \leq s, {{\lVert {\u} \rVert}_2} = 1\}.$$ Let $\Atilde$ be the $m\times n$ subsampled matrix of $\A$ as in the theorem. Then observe that the smallest value $\delta = \delta_s$ which satisfies the $\D$-$\omega$RIP bound for $\Atilde$ is precisely $$\delta_s = \sup_{\u\in\Dsn} \u^*(\Atilde^*\Atilde-{\mathbf{I}}_n)\u.$$ Since $\Atilde^*\Atilde - {\mathbf{I}}_n$ is a self-adjoint operator, we may instead define for any self-adjoint matrix $\B$ the operator $${{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \B
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\sup_{\u\in\Dsn} {\left\langle {\B\u}, \, {\u} \right\rangle}.$$ and equivalently write $$\delta_s = {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \Atilde^*\Atilde - {\mathbf{I}}_n
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s}.$$ Our goal is thus to bound this quantity. To that end, let $\rtilde_1, \rtilde_2, \ldots, \rtilde_m \in \C^{n}$ denote the $m$ randomly selected rows of $\A$ that make up $\Atilde$. It follows from Definition \[BONS\] that $\E \rtilde_i^{*} \rtilde_i = {\mathbf{I}}_n$ for each $i$, where the expectation is taken with respect to the sampling measure $\nu$. We thus have $$\label{delta}
\delta_s = {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \Atilde^*\Atilde - {\mathbf{I}}_n
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} = {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \frac{1}{m}\sum_{i=1}^m \rtilde_i^{*} \rtilde_i - {\mathbf{I}}_n
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} = \frac{1}{m}{{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s}.$$ We may bound the moments of this quantity using a symmetrization argument (see, e.g., Lemma 6.7 of [@ra09-1]): $$\begin{aligned}
\label{symm}
\E{{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} &\leq 2\left(\E{{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \epsilon_i\rtilde_i^{*} \rtilde_i
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} \right) \nonumber \\
&= 2 \E_{\rtilde} \E_{\epsilon} \sup_{\x \in \Dsn} | \langle \sum_{i=1}^m \epsilon_{i} \rtilde_i^{*} \rtilde_i \x, \x \rangle | \nonumber \\
&= 2 \E_{\rtilde} \E_{\epsilon} \sup_{\x \in \Dsn} \left| \sum_{i=1}^m \epsilon_{i} | \langle \rtilde_i, \x \rangle |^2 \right|\end{aligned}$$ where $\epsilon_i$ are independent symmetric Bernoulli $(\pm 1)$ random variables.
Conditional on $(\rtilde_i)$, we have a subGaussian process indexed by $\Dsn$. For a set $T$, a metric $d$, and a given $t > 0$, the *covering number* $\N(T, d, t)$ is defined as the smallest number of balls of radius $t$ centered at points of $T$ necessary to cover $T$ with respect to $d$. For fixed $\rtilde_i,$ we work with the (pseudo-)metric $$\label{pseudo}
d(\x, \z) = \left( \sum_{i=1}^m ( | \langle \rtilde_i, \x \rangle |^2 - | \langle \rtilde_i, \z \rangle |^2 )^2 \right)^{1/2}$$ Then [ Dudley’s inequality]{} [@dudley1967sizes] implies $$\label{dudley0}
\E_{\epsilon} \sup_{\x \in \Dsn} | \langle \sum_{i=1}^m \epsilon_{i} \rtilde_i^{*} \rtilde_i \x, \x \rangle | \leq 4 \sqrt{2} \int_{0}^{\infty} \sqrt{\log(\N(\Dsn, d, t))}\,dt.$$ Continuing as in [@rauhut2013interpolation], inspired by the approach of [@cheraghchi2013restricted], we estimate the metric $d$ using H[ö]{}lder’s inequality with exponents $p \geq 1$ and $q \geq 1$ satisfying $1/p + 1/q = 1$ to be specified later. Using also the reverse triangle inequality, we have for $\x, \z \in \Dsn$, $$\label{continue}
d(\x,\z) \leq 2 \sup_{\u \in \Dsn} \left( \sum_{i=1}^m | \langle \rtilde_i, \u \rangle |^{2p} \right)^{1/(2p)} \left( \sum_{i=1}^m | \langle \rtilde_i, \x - \z \rangle |^{2q} \right)^{1/(2q)}.$$ We will optimize over the values of $p,q$ later on. To further bound this quantity, we have $ |{\left\langle {\rtilde_i}, \, {\d_j} \right\rangle}| \leq \omega_j$ by assumption. Recall the in Definition \[eq:Wunrec\]. Then, fixing $\u\in\Dsn$ such that $\u=\D\z$ with ${\lVert {\z} \rVert}_{\omega,0} \leq s$, we have for any realization of $(\rtilde_i)$, $$\begin{aligned}
\label{old88}
|{\left\langle {\rtilde_i}, \, {\u} \right\rangle}| &= |{\left\langle {\D\D^*\rtilde_i}, \, {\u} \right\rangle}|\notag\\
&= |{\left\langle {\D^*\rtilde_i}, \, {\D^*\D\z} \right\rangle}|\notag\\
&\leq \sum_j |{\left\langle {\rtilde_i}, \, {\d_j} \right\rangle}|_j |(\D^{*} \D \z)_j | \notag \\
&\leq \sum_j \omega_j |(\D^{*} \D \z)_j | \notag \\
&= {\lVert {\D^*\D\z} \rVert}_{\omega,1} \notag \\
&\leq \sqrt{s} \unrec.
\end{aligned}$$ The first line uses that $\D$ is a Parseval frame, while the last line uses the definition of $\unrec$. The quantity $s \unrec^2$ will serve as a rescaled sparsity parameter throughout the remaining proof.
Continuing from , we may bound $$\begin{aligned}
\label{bound_gen}
\sup_{\u \in \Dsn} \left( \sum_{i=1}^m | \langle \rtilde_i, \u \rangle |^{2p} \right)^{1/(2p)} &=
\sup_{\u \in \Dsn} \left( \sum_{i=1}^m | \langle \rtilde_i, \u \rangle |^{2}| \langle \rtilde_i, \u \rangle |^{2p-2} \right)^{1/(2p)} \notag \\
&\leq (s \unrec^2)^{(p-1)/(2p)} \left( \sup_{\u \in \Dsn} \sum_{i=1}^m | \langle \rtilde_i, \u \rangle |^2 \right)^{1/(2p)}.\end{aligned}$$ We now introduce the (semi-)norm $$\| \u \|_{X,q} := \left( \sum_{i=1}^m | \langle \rtilde_i, \u \rangle |^{2q} \right)^{1/(2q)}.$$ Using basic properties of covering numbers and the bound in , we obtain $$\begin{aligned}
\label{dudley2}
\mathbb{E}_{\epsilon} \sup_{\u \in \Dsn} | \langle \sum_{i=1}^m \epsilon_{i} \rtilde_i^{*} \rtilde_i \u, \u \rangle | &\leq C_1 (s \unrec^2)^{(p-1)/(2p)} \left( \sup_{\u \in \Dsn} \sum_{i=1}^m | \langle \rtilde_i, \u \rangle |^2 \right)^{1/(2p)} \int_{0}^{\infty} \sqrt{\log(\N(\Dsn, \| \cdot \|_{X,q}, t))}\,dt\end{aligned}$$ where $C_1$ is an absolute constant.
We now estimate the key integral above in two different ways, the first bound being better for small values of $t$, the second bound better for larger values of $t$.
Small values of $t$:
: Following and and since $\|\u\|_2=1$ for all $\u\in\Dsn$, we have that for any $\u \in \Dsn$, $$\label{simple_bound}
\| \u \|_{X,q} \leq (s \unrec^2)^{1/2} m^{1/(2q)}.$$ Standard covering arguments show that for any seminorm ${\lVert {\cdot} \rVert}$, one has for the unit ball $B_{{\lVert {\cdot} \rVert}}\subset \C^s$ that $\N(B_{{\lVert {\cdot} \rVert}}, {\lVert {\cdot} \rVert}, t) \leq (1+1/t)^{2s}$. Thus for $\Lambda\subset\{1, \dots, N\}$ with $|\Lambda|=s'$, one can define the seminorm ${\lVert {\cdot} \rVert}_{\Lambda}$ on $\C^{s'}$ by $${\lVert {\y} \rVert}_{\Lambda} {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\left( \sum_{i=1}^m | {\left\langle {\D_\Lambda^*\rtilde_i^{*}}, \, {\y} \right\rangle} |^{2q} \right)^{1/(2q)},$$ and observe that for $\u$ with $\| \u \|_{X,q} \leq 1$ and $\u=\D_\Lambda\y$ for $\y\in\C^{s'}$ that one has the equivalence $ {\lVert {\y} \rVert}_{\Lambda} = {\lVert {\u} \rVert}_{X,q}$. Applying this equivalence for $\Lambda=\operatorname*{supp}(\z)$ and $\y=\z_\Lambda$, we have that $$\begin{aligned}
\N(\Dsn, \| \cdot \|_{X,q}, t) &\leq \sum_{\omega(\Lambda)\leq s} \N(B_{\|\cdot\|_\Lambda}, m^{1/(2q)} \| \cdot \|_2 \sqrt{(s \unrec^2)},t ) \\
&\leq {\binom{N}{s}}(1 + \sqrt{(s \unrec^2)} m^{1/(2q)} t^{-1})^{2s}\\
&\leq (eN/s)^{s}(1 + \sqrt{(s \unrec^2)} m^{1/(2q)}t^{-1})^{2s},
\end{aligned}$$ where we have applied [@RefWorks:45 Proposition C.3] in the final line.
Large values of $t$:
: To get a bound for larger values of $t$, we will utilize another embedding, and for that reason we define the auxiliary set $$\DsnOne {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\{\u : {\lVert {\D^*\u} \rVert}_{\omega,1} \leq 1, \u=\D\z \in \C^n, \| \z \|_{\omega,0} \leq s\} = \bigcup_{\Lambda: \omega(\Lambda) \leq s}\BoneS,$$ where $\BoneS {\overset{\mathrm{\scriptscriptstyle{def}}}{=}}\{\u : {\lVert {\D^*\u} \rVert}_{\omega,1} \leq 1, \u=\D\z \in \C^n, \operatorname*{supp}(\z) = \Lambda\}$. Then we have the embedding $\Dsn \subset \sqrt{(s \unrec^2)}\DsnOne$ since for any $\u\in\Dsn$, ${\lVert {\D^*\u} \rVert}_{\omega,1} \leq \sqrt{(s \unrec^2)}$ by definition of the localization factor $\unrec$.
We now use a variant of Maurey’s lemma [@maurey], precisely, the variant as stated in Lemma 5.3 in [@rauhut2013interpolation], in order to deduce a different covering number bound:
\[maurey\] For a normed space $X$, consider a finite set $U \subset X$ of cardinality $N$, and assume that for every $L \in \mathbb{N}$ and $(\u_1, \dots, \u_L) \in U^L$, $\mathbb{E}_{\epsilon} \| \sum_{j=1}^L \epsilon_j \u_j \|_X \leq A \sqrt{L}$, where $\epsilon$ denotes a Rademacher vector. Then for every $t > 0$, $$\log{\N\left(\emph{conv}(U), \| \cdot \|_X, t \right)} \leq c (A/t)^2 \log(N),$$ where $\text{conv}(U)$ is the convex hull of $U$.
As a corollary, we have the following.
\[lem:covering\] For every $t > 0$, $$\log(\N(\Dsn, \| \cdot \|_{X,q}, t) \leq C \left( 2 e^{-1/2} \sqrt{2q} m^{1/(2q)} \sqrt{2(s \unrec^2)} t^{-1} \right)^2 \log(N).$$
Consider Lemma \[maurey\] with norm $\| \cdot \|_{X,q}$, parameter $A = 2 e^{-1/2} \sqrt{2q} m^{1/(2q)}$, and $U = \{ \pm \omega_j^{-1} \D^{*}({\mathrm{e}}_j), \pm i \omega_j^{-1} \D^*({\mathrm{e}}_j), j \in [N] \}$. First observe that $$\Dsn \subset \sqrt{(s \unrec^2)}\DsnOne \subset \sqrt{2(s \unrec^2)} \text{conv}(U).$$ For a Bernoulli random vector $\epsilon = (\epsilon_1, \dots, \epsilon_L)$ and $\u_1, \dots, \u_L \in U$ we have $$\begin{aligned}
\mathbb{E}_{\epsilon} \| \sum_{j=1}^L \epsilon_j \u_j \|_{X,q} &\leq \left( \mathbb{E} \| \sum_{j=1}^L \epsilon_j \u_j \|_{X,q}^{2q} \right)^{1/(2q)} \notag \\
&= \left( \mathbb{E} \sum_{\ell=1}^m | \langle{\rtilde_{\ell}, \sum_{j=1}^L \epsilon_j \u_j \rangle} |^{2q} \right)^{1/(2q)}
\notag \\
&= \left( \sum_{\ell=1}^m \mathbb{E} | \langle{\rtilde_{\ell}, \sum_{j=1}^L \epsilon_j \u_j \rangle} |^{2q} \right)^{1/(2q)}
\notag \\
&\leq 2 e^{-1/2} \sqrt{2q} \left( \sum_{\ell=1}^m \| ( \langle{\rtilde_{\ell},\u_j \rangle} )_{j=1}^L \|_2^{2q} \right)^{1/(2q)},\end{aligned}$$ where we applied Khintchine’s inequality [@khintchine1923dyadische] in the last step. Since each $\u_j$ consists of a multiple of a single column of $\D$, we also have for each $j$ and $i$ that $|{\left\langle {\rtilde_i}, \, {\u_j} \right\rangle}| = |{\left\langle {\D^*\rtilde_i^{*}}, \, {\omega_j^{-1} {\mathrm{e}}_{j}} \right\rangle}|$ for some coordinate vector ${\mathrm{e}}_{j}$ consisting of all zeros with a $1$ in the $j$th position. Thus $|{\left\langle {\rtilde_i}, \, {\u_j} \right\rangle}| \leq 1$ for each $j$ and $i$, which means that ${{\lVert {({\left\langle {\rtilde_i}, \, {\u_j} \right\rangle})_{k=1}^L} \rVert}_2} \leq \sqrt{L}$ for any $L$ and $$\mathbb{E}_{\epsilon} \| \sum_{j=1}^L \epsilon_j \u_j \|_{X,q} \leq 2e^{-1/2}\sqrt{2q}m^{1/(2q)} \sqrt{L} = A \sqrt{L}.$$ The corollary then follows from Lemma \[maurey\].
We have thus obtained the two bounds for $t > 0$: $$\begin{aligned}
\sqrt{\log(\N(\Dsn, {{\lVert {\cdot} \rVert}_X}, t))} &\leq \sqrt{ s \log(eN/s) + 2s \log\left( 1 + 2\sqrt{(s \unrec^2)}m^{1/(2q)}t^{-1} \right)} \nonumber \\
\sqrt{\log(\N(\Dsn, {{\lVert {\cdot} \rVert}_X}, t))} &\leq C \sqrt{q} m^{1/(2q)} \sqrt{(s \unrec^2)} t^{-1} \sqrt{\log(N)}.
\end{aligned}$$ We may now bound the integral in . Without loss, we take the upper integration bound as $t_0 = \sqrt{(s \unrec^2)} m^{1/(2q)}$ because, for $t >t_0$, we have $\N(\Dsn, \| \cdot \|_{X,q}, t) = 1$ by . Splitting the integral in two parts and using our covering number bounds, we have for $\alpha \in (0, t_0)$, $$\begin{aligned}
\int_{0}^{\infty} &\sqrt{\log(\N(\Dsn, \| \cdot \|_{X,q}, t))}\,dt\\
&\leq \int_{0}^{\alpha} \sqrt{ s \log(eN/s) + 2s \log\left( 1 + 2\sqrt{(s \unrec^2)}m^{1/(2q)}t^{-1} \right)} dt \hspace{.5mm} + C_1 \sqrt{q} m^{1/(2q)} \sqrt{(s \unrec^2)} \int_{\alpha}^{t_0} t^{-1} dt \nonumber \\
&\leq \alpha \sqrt{s \log(eN/s)} + \sqrt{2s}\alpha \sqrt{ \log(e(1 + \sqrt{(s \unrec^2)} m^{1/(2q)})\alpha^{-1}} + C_2 \sqrt{q m^{1/q} (s \unrec^2) \ln (4N)} \log(\sqrt{(s \unrec^2)} m^{1/(2q)}/\alpha)
\end{aligned}$$ where in the last step we have utilized that for $a>0$, $\int_0^a \sqrt{\ln(1+t^{-1})}\,dt \leq a\sqrt{\ln(e(1+a^{-1}))}$ (see Lemma 10.3 of [@ra09-1]). [ Choosing $\alpha = m^{1/(2q)}$ ]{} yields $$\int_{0}^{\infty} \sqrt{\log(\N(\Dsn, \| \cdot \|_{X,q}, t))}\,dt \leq C_3 \sqrt{q (s \unrec^2) m^{1/q} \log(N) \log^2(s \unrec^2)}.$$ Combining this with , and ,we have $$\begin{aligned}
\E \delta_s &\leq \frac{C_3 (s \unrec^2)^{(p-1)/(2p)} \sqrt{q m^{1/q} (s \unrec^2) \log(N) \log^2(s \unrec^2)}}{m} \E \sup_{\x \in \Dsn} \left( \sum_{i=1}^m | \langle \rtilde_i, \x \rangle |^2 \right)^{1/(2p)} \nonumber \\
&\leq \frac{C_3 (s \unrec^2)^{1/2 + (p-1)/(2p)} \sqrt{q \log(N) \log^2(s \unrec^2)}}{m^{1 - 1/(2q)}m^{-1/(2p)}}
\E \left( \frac{1}{m} {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \rtilde_i^{*} \rtilde_i - {\mathbf{I}}_n
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} + {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert {\mathbf{I}}_n
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} \right)^{1/(2p)} \nonumber \\
&\leq \frac{C_3 (s \unrec^2)^{1/2 + (p-1)/(2p)} \sqrt{q \log(N) \log^2(s \unrec^2)}}{m^{1/2}} \sqrt{ \E \delta_s + 1}.
\end{aligned}$$ Above we applied H[ö]{}lder’s inequality and used that $1/q + 1/p = 1$ as well as $p \geq 1$. [ We now choose $p = 1 + 1/\log(s \unrec^2)$ and $q = 1 + \log(s \unrec^2)$ to give]{} $(s \unrec^2)^{(p-1)/(2p)} \leq (s \unrec^2)^{(p-1)/2} = (s \unrec^2)^{1/(2\log(s \unrec^2))} = e^{1/2}$ and $$\E \delta_s \leq C_4 \sqrt{2 \log(N) \log^2(s \unrec^2)/m} \sqrt{ \E \delta_s + 1}.$$ Squaring this inequality and completing the square finally shows that $$\label{prob}
\E \delta_s \leq C_5 \sqrt{ \frac{(s \unrec^2) \log(N) \log^3(s \unrec^2)}{m}},$$ provided the term under the square root is at most 1. Then $\E \delta_s \leq \delta/2$ for some $\delta \in (0,1)$ if $$\label{end_m}
m \geq C_6 \delta^{-2} (s \unrec^2) \log^3(s \unrec^2) \log(N).$$
It remains to show that $\delta_s$ does not deviate much from its expectation. For this *probability bound*, we may write $\delta_s$ as the supremum of an empirical process as in \[[@ra09-1], Theorem 6.25\] and apply the following Bernstein inequality for the supremum of an empirical process:
Let ${\mathcal{F}}$ be a countable set of functions $f : \C^n \rightarrow \R$. Let $Y_1, \dots, Y_m$ be independent copies of a random vector $Y$ on $\C^n$ such that $\E f(Y) = 0$ for all $f \in {\mathcal{F}}$, and assume $f(Y) \leq 1$ almost surely. Let $Z$ be the random variable $Z = \sup_{f \in {\mathcal{F}}} \sum_{\ell=1}^m f(Y_{\ell})$, and $\E Z$ its expectation. Let $\sigma^2 > 0$ such that $\mathbb{E} [ f(Y)^2 ] \leq \sigma^2$ for all $f \in {\mathcal{F}}$. Then, for all $t > 0$, $$\mathbb{P}(Z \geq \E Z + t) \leq \exp\left( - \frac{t^2}{2(m \sigma^2 + 2 \E Z) + 2t/3} \right)$$
We apply this theorem to provide a probability bound. Let $f_{z,w}(\rtilde) = \text{Re}(\langle (\rtilde_i^{*} \rtilde_i - {\mathbf{I}})z, w \rangle )$ so that $$m \delta_s = {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} = \sup_{(z,w) \in {\mathcal{D}}, } \sum_{i=1}^m f_{z,w}(\rtilde_{i}).$$ Since $\E \rtilde_i^{*} \rtilde_i = {\mathbf{I}}$ we have $\E f_{z,w}(\rtilde) = 0$. Moreover, $| f_{z,w}(\rtilde) | \leq \max_{z \in {\mathcal{D}}} |\langle \rtilde_i, z \rangle|^2 + 1 \leq s \eta^2 + 1$. For the variance term, we have $\E | f_{z,w}(\rtilde_i)|^2 = \E \| (\rtilde_i^{*} \rtilde_i - {\mathbf{I}}) z \|_2^2 \leq (s \eta^2 + 1)^2.$
Fix $\delta \in (0,1)$, and suppose the number of measurements $m$ satisfies so that $\E \delta_s \leq \delta/2$. Then it follows $$\begin{aligned}
\label{tail}
\P(\delta_s \geq \delta) &\leq& \P(\delta_s \geq \E \delta_s + \delta/9) \nonumber \\
&=& \P \left({{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} \geq \E {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} + \delta m/9 \right) \nonumber \\
&=& \P \left(\frac{1}{s \eta^2 + 1} {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} \geq \frac{1}{s \eta^2 + 1} \E {{{\left\vert\kern-0.25ex\left\vert\kern-0.25ex\left\vert \sum_{i=1}^m \left(\rtilde_i^{*} \rtilde_i - \E\rtilde_i^{*} \rtilde_i\right)
\right\vert\kern-0.25ex\right\vert\kern-0.25ex\right\vert}}_s} + \frac{\delta m/9}{s \eta^2 + 1} \right) \nonumber \\
&\leq& \exp \left(- \frac{(\frac{\delta m/9}{s \eta^2 + 1})^2}{2m(1 + \frac{\delta}{s \eta^2+1}) + \frac{2}{3}(\frac{\delta m/9}{s \eta^2 + 1})} \right) \nonumber \\
&\leq& \exp{\left( -\frac{\delta^2 m}{C_7 (s \unrec^2)}\right)}.
\end{aligned}$$ The last term is bounded by $\gamma \in (0,1)$ if $m \geq C \delta^{-2} (s \unrec^2) \log(1/\gamma)$. Together, we have $\delta_s \leq \delta$ with probability at least $1-\gamma$ if $$m \geq C_8 \delta^{-2} s \unrec^2 \max\left\{ \log^3(s\unrec^2) \log(N), \log(1/\gamma) \right\}.$$ This completes the proof.
Conclusion {#sec:disc}
==========
We have introduced a coherence-based analysis of compressive sensing when the signal to be recovered is approximately sparse in a redundant dictionary. Whereas previous theory only allowed for unstructured random sensing measurements, our coherence-based analysis extends to structured sensing measurements such as subsampled uniformly bounded bases, bringing the theory closer to the setting of practical applications. We also extend the theory of variable density sampling to the dictionary setting, permitting some coherence between sensing measurements and sparsity dictionary. We further extend the analysis to allow for weighted sparse expansions. Still, several open questions remain. While we provided two concrete examples of dictionaries satisfying the bounded condition required by our analysis – the oversampled DFT frame and redundant Haar wavelet frame – these bounds can almost certainly be extended to more general classes of dictionaries, and improved considerably in the case of the oversampled DFT frame. We have also left several open problems related to the full analysis for variable density sampling in this setting, including the removal of a weighted noise assumption in the $\ell_1$-analysis reconstruction method. Finally, we believe that the $\D$-RIP assumption used throughout our analysis can be relaxed, and that a RIPless analysis [@RefWorks:595] should be possible and permit non-uniform signal recovery bounds at a reduced number of measurements. It would also be useful to extend our results to measurement matrices constructed in a deterministic fashion for those applications in which randomness is not admissable; of course this is a challenge even in the classical setting [@RefWorks:36].
Acknowledgments {#acknowledgments .unnumbered}
===============
The authors thank the reviewers of this manuscript for their helpful suggestions which significantly improved the paper. The authors would also like to thank the Institute for Computational and Experimental Research in Mathematics (ICERM) for its hospitality during a stay where this work was initiated. In addition, Krahmer was supported by the German Science Foundation (DFG) in the context of the Emmy Noether Junior Research Group KR 4512/1-1 (RaSenQuaSI). Needell was partially supported by NSF CAREER $\#1348721$ and the Alfred P. Sloan Foundation. Ward was partially supported by an NSF CAREER award, DOD-Navy grant N00014-12-1-0743, and an AFOSR Young Investigator Program award.
[^1]: By “fixed” we mean to emphasize that this probability bound must occur for a *single* vector $\z$ rather than for all vectors $\z$ like one typically sees in the restricted isometry property.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
bibliography:
- 'bibliography.bib'
---
[Fermionic Markov Chains]{}\
M. Fannes and J. Wouters[^1]\
`[email protected]`\
`[email protected]`\
Instituut voor Theoretische Fysica\
K.U.Leuven, Belgium\
Introduction {#sec:intro}
============
Quantum channels describe the black box dynamics of small open quantum systems, i.e. a quantum system evolving in contact with an inaccessible environment. Technically, a channel is a completely positive map, which maps an input density matrix into an output density matrix. It corresponds to a one-shot random evolution of the system. A classical channel is a stochastic matrix.
In this article, we consider the construction of a quantum process associated to a channel, much like a stochastic matrix generates a Markov process. Introducing [ in a quantum system]{} multi-time correlations that are compatible with a given channel is, however, much more delicate than in the classical context. Such an amplification to a process is not generally possible and, if possible, the process is not unique. We focus on entropic properties of such processes, both for determining extensions with minimal entropy and obtaining a measure of randomness in the given channel.
The construction uses a generalization of matrix product states which was introduced under the name of finitely correlated states, see [@Accardi81; @Fannes89]. While matrix product states prove to be a very useful class of pure states on quantum spin chains [@Perez07; @Klumper91], suitable for studying ground state properties, one needs to go beyond such states in the context of channels [due to the mixing of pure states by generic channels]{}. In fact, even in the classical context, the class of processes we consider includes hidden Markov processes.
In the study of the information carrying capacity of classical channels with memory, the entropy density of hidden Markov processes also arises [@Wouters09]. The entropy density of finitely correlated states is expected to play a role in the information capacity of quantum channels with memory.
In this article, however, we consider a simpler problem and introduce a construction of free Fermionic Markov processes compatible with a free Fermionic channel. We show in particular that, [instead of through a direct calculation]{}, the entropy density can also be obtained as the asymptotic entropy production under the shift dynamics.
The structure of the article is as follows. In Section \[sec:class\] we review properties of classical Markov and hidden Markov processes. We pay particular attention to a method to calculate the entropy density, based on a De L’Hôpital-like property of strongly subadditive functions. In Section \[sec:quant\] we introduce the quantum version of hidden Markov processes. In Section \[sec:fermionic\] we consider a free Fermionic version of such processes. We then turn to the main result in Section \[sec:density\]. We show that the De L’Hôpital-like property, which connects averages of functions to their growth rates, can be extended to a much wider class than the strongly subadditive functions. As the density matrices of free Fermionic systems are basically Toeplitz matrices, this amounts to an extension Szegö’s theorem for averages of functions to rates.
The classical case {#sec:class}
==================
We first consider classical Markov and hidden Markov processes. Let $P$ be a stochastic matrix over a finite state space $\Omega = \{ 1, 2, \ldots, d \}$: the entries of the $d \times d$ matrix $P$ are non-negative and the row sums are equal to 1. The entry $P_{\omega_1 \omega_2}$ specifies the jump probability from state $\omega_1$ to state $\omega_2$ and hence $P$ defines a stochastic dynamics in discrete time. Generically $P$ has a non-degenerate eigenvalue 1 and so has its transpose $P^{\textsf T}$. The Perron-Frobenius theorem asserts that the absolute values of the eigenvalues are not larger than 1 and that the entries of the eigenvector of $P^{\textsf T}$ corresponding to the eigenvalue 1 can be chosen non-negative. A proper normalization provides us therefore with a probability vector $\mu$ over $\Omega$ such that $P^{\textsf T}\mu = \mu$. For a generic $P$, $\mu$ is faithful and we have exponentially fast convergence to the invariant measure: there exist $C \ge 0$ and $0 \le \gamma < 1$ such that $$\bigl\Vert ( P^{\textsf T})^n \nu - \mu \bigr\Vert_1 \le C \gamma^n,\enskip n \in {\mathbb{N}},\ \text{$\nu$ probability vector}.$$
Moreover, $P$ generates a natural stationary stochastic process by assigning to a path $(\omega_0, \omega_1, \ldots, \omega_n)$ the probability $$\bigl{\langle}\omega_0, \omega_1, \ldots, \omega_n \bigr{\rangle}_{[0,n]} = \mu_{\omega_0} P_{\omega_0\, \omega_1} P_{\omega_1\, \omega_2} \cdots P_{\omega_{n-1}\, \omega_n},\enskip \omega_j \in \Omega,\ n = 0, 1, \ldots
\label{mc}$$ This is a one step Markov process: the probability for reaching the state $\omega_n$ at time $n$ given the full history $(\omega_0, \omega_1, \ldots, \omega_{n-1})$ is the same as that for reaching $\omega_n$ starting at time $n-1$ at $\omega_{n-1}$.
The entropy production or rather the mean entropy of this Markov chain is a natural way to quantify the randomness of $P$ $${\mathsf{h}}(P) := \lim_{n \to \infty} \frac{1}{n}\, {\mathsf{H}}\bigl( {\langle}\ {\rangle}_{[0,n-1]} \bigr)
\label{mc:ent}$$ Here, ${\mathsf{H}}$ is the usual Shannon entropy of a probability vector. In fact, in the event of multiple stationary measures for $P$ the entropy also depends on the chosen initial measure.
For strongly subadditive function, such as the Shannon entropy, the following discrete version of De L’Hôpital’s rule can be proven[@Alicki01; @Kay01]: $$\lim_{n \to \infty} \frac{1}{n}\, {\mathsf{H}}\bigl( {\langle}\ {\rangle}_{[0,n-1]} \bigr) = \lim_{n \to \infty} \Bigl( {\mathsf{H}}({\langle}\ {\rangle}_{[0,n]}) - {\mathsf{H}}({\langle}\ {\rangle}_{[0,n-1]}) \Bigr).
\label{be}$$ Using this equality a simple calculation shows that the entropy density of the Markov process (\[mc\]) is given by $${\mathsf{h}}(P) = {\langle}{\mathsf{H}}_\text{trans} {\rangle}_\mu \,, \label{eq:markov_density}$$ where ${\mathsf{H}}_\text{trans} (\omega)$ is the entropy of the conditional probabilities related to the transition from $\omega$ to the next state: $${\mathsf{H}}_\text{trans} (\omega) = - \sum_\sigma P_{\omega \sigma} \log P_{\omega \sigma} \label{class_density} \,.$$
This construction is however difficult to carry over to quantum systems due to the prominent role of paths. The probabilistic nature of quantum mechanics and the non-uniqueness of the choice of basis for the Hilbert space of the system make paths an unnatural concept in quantum mechanics. A construction that is better suited for generalization is based on positive maps. We consider stochastic matrices with $d$ rows and $d^2$ columns. An observable on the discrete state space $\Omega$ can be seen as a vector $f \in {{\mathbb R}}^d$ and we use the notation $\bm 1$ for the constant function 1, i.e. every entry of $\bm 1$ is equal to 1.
\[con2\] Let $P$ be a stochastic $d \times d$ matrix with invariant measure $\mu$ and let $Q$ be a $d \times d^2$ stochastic matrix that satisfies the compatibility condition $$Q\, (f\otimes \bm 1) = Q\, (\bm 1 \otimes f) = P\, f,\enskip f \in {{\mathbb R}}^d.
\label{cmc}$$ Any such $Q$ defines a stationary measure on the half-chain $\times^{{\mathbb{N}}} \Omega$ with marginals $$\bigl{\langle}f_n \bigr{\rangle}= \mu\Bigl( Q \bigl( Q \otimes {\mathbb{1}}) \cdots \bigl( Q \otimes {\mathbb{1}}\otimes \cdots \otimes {\mathbb{1}}\bigr) \bigl( \bm 1 \otimes f_n \bigr) \Bigr),\enskip f_n \in \operatorname*{\mbox{\Large\mbox{$\otimes$}}}_0^{n-1} {{\mathbb R}}^d.
\label{cfcs}$$
$\square$
The set of stochastic matrices $Q$ obeying (\[cmc\]) is closed, convex and non-empty. E.g. the Markov chain (\[mc\]) is obtained by choosing $$Q(f \otimes g) = P(fg),\enskip f,g \in {{\mathbb R}}^d,
\label{marext}$$ where $fg$ is the entrywise product of $f$ and $g$. For a general $Q$, the measure (\[cfcs\]) is a stationary hidden Markov process.
The mean entropy of a hidden Markov process can be computed using a method due to Blackwell, see [@Fannes92] and [@Blackwell57]. This computation is based on the discrete version of De L’Hôpital’s rule for the mean entropy given in Equation \[be\]. The relation between the $n$ and $n-1$ site marginals of a hidden Markov process is given by a transfer matrix like relation, as seen from (\[cfcs\]). This allows to express the mean entropy as an average of entropies of probability vectors over $\Omega$ $$\lim_{n \to \infty} \frac{1}{n}\, {\mathsf{H}}\bigl( {\langle}\ {\rangle}_n \bigr) = \int \!\varphi(d\nu)\, {\mathsf{H}}^ \prime_\text{trans}(\nu).$$ In this formula, $\nu$ is varies over the set of probability vectors over $\Omega$ and $\varphi$ is a measure on this set of probability vectors. The measure $\varphi$ is the unique stationary measure of a dynamical system on probability vectors over $\Omega$ that is determined by $Q$.
This formula bears some similarity to Eq. \[eq:markov\_density\]. The measure $\varphi$ weighs the possible past paths, much as the measure $\mu$ in Eq. \[eq:markov\_density\] weighs the possible configuration at the previous time step. As the hidden Markov process also has correlations with previous time steps, this weighting term gets significantly more complicated. The entropy term ${\mathsf{H}}^ \prime_\text{trans}(\nu)$ is an entropy function related to the transition from one state to the next, much as the function ${\mathsf{H}}_\text{trans}(\nu)$ in Eq. \[class\_density\] is determined by the transition probabilities $P_{\omega \sigma}$.
Numerical experiments suggest that among the extensions that satisfy the compatibility condition (\[qmc\]), the Markov chain extension (\[marext\]) has the smallest entropy.
The quantum case {#sec:quant}
================
The natural quantum analogue of a stochastic matrix is a unity preserving completely positive (CP) map $\Gamma$ acting on the $d$-dimensional complex matrices ${\mathcal{M}}_d$. Such maps send a pure state into a mixed one and are therefore stochastic. Generically $\Gamma$ has a non-degenerate eigenvalue 1, the corresponding eigenvector of the transpose, mostly called the dual, is a faithful density matrix $\rho$ and exponentially fast return to equilibrium holds: for any initial density matrix $\sigma$ $$\bigl\Vert \sigma \circ \Gamma^n - \rho \bigr\Vert_1 \le C \gamma^n,\enskip n \in {\mathbb{N}},\ \text{$\sigma$ density matrix}.$$
Unlike the classical case, where there is a trivial connection between stochastic matrices and Markov processes, there is no straightforward extension to a process. A first reason is that a general density matrix admits many convex decompositions in pure states, a mixed quantum state is not uniquely linked to an ensemble albeit that there is a preferred decomposition, namely the spectral decomposition. A second reason is that the map $\Gamma$ not only mixes pure states but also rotates them which prohibits a description in terms of paths.
This makes quantifying the randomness of a CP $\Gamma$ not evident. At least two proposals can be found in the literature: the minimal output entropy [@King01] and the map entropy [@Zyczkowski_duality_2004]. Here we propose an approach in the spirit of Markov chains.
\[con1\] Let $\Gamma: {\mathcal{M}}_d \to {\mathcal{M}}_d$ be a unity preserving CP map with invariant state $\rho$ and let $\Lambda$ be a unity preserving CP map from ${\mathcal{M}}_d \otimes {\mathcal{M}}_d$ to ${\mathcal{M}}_d$ that satisfies the compatibility condition $$\Lambda(X \otimes {\mathbb{1}}) = \Lambda({\mathbb{1}}\otimes X) = \Gamma(X),\enskip X \in {\mathcal{M}}_d.
\label{qmc}$$ The quantum Markov chain defined by $\Lambda$ and $\rho$ is then the finitely correlated state $\omega$ [@Fannes92-1] on the semi-infinite quantum spin chain $\otimes^{{\mathbb{N}}_0} {\mathcal{M}}_d$ with marginals $\rho_n$ on the sites $[1, 2, \ldots, n]$ given by $$\begin{split}
\omega\bigl( X_n \bigr)
&= \operatorname*{Tr}\bigl( \rho_n X_n \bigr) \\
&= \operatorname*{Tr}\bigl( \rho\, \Lambda \circ (\Lambda \otimes \operatorname{id}) \circ \cdots \circ (\Lambda \otimes \operatorname{id}\otimes \cdots \otimes \operatorname{id}) ({\mathbb{1}}\otimes X_n) \bigr)
\end{split}
\label{fcs}$$ where $X_n \in \operatorname*{\mbox{\Large\mbox{$\otimes$}}}_1^n {\mathcal{M}}_d$.
$\square$
Note that this construction contains the class of classical hidden Markov processes.
We can now associate an entropy to a quantum Markov chain as in the classical case $${\mathsf{h}}(\Lambda,\rho) := {\mathsf{s}}(\omega) := \lim_{n \to \infty} \frac{1}{n}\, {\mathsf{S}}(\rho_n)
\label{qme}$$ where ${\mathsf{S}}$ is the usual von Neumann entropy. Generically, $\Gamma$ has a non-degenerate eigenvalue 1 so that $\rho$ is uniquely determined by $\Gamma$ and that there is no $\rho$ dependence in (\[qme\]). Clearly a number of issues have to be addressed: for which $\Gamma$ can one find a $\Lambda$ that satisfies (\[qmc\])? How does ${\mathsf{h}}(\Gamma, \Lambda)$ depend on $\Lambda$? Can one compute the mean entropy (\[qme\])?
A Fermionic model {#sec:fermionic}
=================
Quantum states are mostly indirectly given, typically as ground or equilibrium states for a given interaction and are hence difficult to work with as there is for example no explicit density matrix. Also, in general one has to deal with an enormous amount of parameters when the number of particles grows. As the number of components grows, typically the number of parameters grows exponentially. Free Fermionic states [@Shale64; @Powers70; @Balslev68] form an exception in both respects. These states describe systems of non-interacting fermions. They are given by an explicit recipe, reducing the calculation of higher order correlation to a simple combinatorial combination of second order correlations. Hence not only can they be calculated explicitly, they are also fully described by their second order correlation, resulting in a significant reduction in parameters.
In this section we first discuss some of the properties of free Fermionic states. We then introduce a Markov construction similar to the one given in Section \[sec:quant\].
Free states and maps {#s3:1}
--------------------
The algebra ${\mathfrak{A}}({\mathfrak{H}})$ generated by the canonical anticommutation relations (CAR) describes the observables of a system of Fermions with one-particle space ${\mathfrak{H}}$. It is generated by the identity and the creation and annihilation operators $a^*(\varphi)$ and $a(\varphi)$ that obey the relations $$\begin{aligned}
&\phantom{i}i)\enskip \varphi \mapsto a^*(\varphi) \text{ is ${{\mathbb C}}$-linear} \\
&ii)\enskip \{ a(\varphi) \,,\, a(\psi) \} = 0 \enskip\text{and}\enskip \{ a(\varphi) \,,\, a^*(\psi) \} = {\langle}\varphi \,,\, \psi {\rangle}{\mathbb{1}}.\end{aligned}$$
A useful set of states, called free, quasi-free, Gaussian, or determinantal, is determined by a simple combinatorial rule. Given a symbol $Q \in {\mathfrak{B}}({\mathfrak{H}})$, the state $\omega_Q$ vanishes on every monomial except for $$\omega_Q\bigl( a^*(\varphi_1) \cdots a^*(\varphi_n) a(\psi_n) \cdots a(\psi_1) \bigr) = \det\bigl( \bigl[{\langle}\psi_k \,,\, Q \varphi_\ell {\rangle}\bigr] \bigr).
\label{qfs}$$ Positivity holds if and only if $0 \le Q \le {\mathbb{1}}$. The set of symbols $${\mathcal{Q}}({\mathcal{H}}) = \{ Q \mid Q \text{ linear operator on } {\mathcal{H}} \text{ such that } 0 \le Q \le {\mathbb{1}}\}$$ is convex and weakly compact. Convexity at the level of symbols is very different from convexity at the level of the free states. Nevertheless it can be shown that a free state is pure, i.e. extreme in the full state space of ${\mathcal{A}}({\mathcal{H}})$, if and only if its symbol is an orthogonal projector, i.e. an extreme point of ${\mathcal{Q}}({\mathcal{H}})$.
Important quantities like the entropy of free states are expressible in terms of symbols, e.g. $${\mathsf{S}}(Q) = - \operatorname*{Tr}Q \log Q - \operatorname*{Tr}({\mathbb{1}}- Q) \log ({\mathbb{1}}- Q).
\label{qfent}$$
Let $P$ be an orthogonal projection on ${\mathcal{H}}$, then the restriction of the free state $\omega_Q$ on ${\mathcal{A}}({\mathcal{H}})$ is a free state on the sub-CAR algebra ${\mathcal{A}}(P{\mathcal{H}})$ with symbol $PQP$. Conversely, a pair of free states $\omega_{Q_i}$ on ${\mathcal{A}}({\mathcal{H}}_i)$, $i=1,2$ extends to a free state $\omega_{Q_1 \oplus Q_2}$ on ${\mathcal{A}}({\mathcal{H}}_1 \oplus {\mathcal{H}}_2)$ by putting $$\omega_{Q_1 \oplus Q_2} (X_1 X_2) = \omega_{Q_1}(X_1)\, \omega_{Q_2}(X_2),\enskip X_i \in {\mathcal{A}}({\mathcal{H}}_i).$$
Free, identity preserving, CP maps $\Lambda_{A,B}: {\mathfrak{A}}({\mathfrak{H}}) \to {\mathfrak{A}}({\mathfrak{K}})$ are determined by a pair of linear operators $A: {\mathfrak{H}} \to {\mathfrak{K}}$ and $B: {\mathfrak{H}} \to {\mathfrak{H}}$. For monomials of degree two $$\Lambda_{A,B} \bigl( a^*(\varphi)a(\psi) \bigr) = a^*(A\varphi) a(A\psi) + {\langle}\psi \,,\, B\varphi {\rangle}{\mathbb{1}}.
\label{qfm}$$ For more details, see [@Dierckx08]. Complete positivity holds if and only if $0 \le B \le {\mathbb{1}}- A^*A$. As for free states, we introduce the set $$\begin{split}
{\mathcal{CP}}({\mathcal{H}}, {\mathcal{K}}) = \bigl\{ (A,B) \,\bigm|\,
&A: {\mathcal{H}} \to {\mathcal{K}} \text{ and } B: {\mathcal{H}} \to {\mathcal{H}} \text{ linear} \\
&\text{operators such that } 0 \le B \le {\mathbb{1}}- A^*A \bigr\}.
\end{split}$$ We use ${\mathcal{CP}}({\mathcal{H}})$ for ${\mathcal{CP}}({\mathcal{H}}, {\mathcal{H}})$. The set of free, CP maps extends that of free states by putting $$Q \in {\mathcal{Q}}({\mathcal{H}}) \mapsto (0,Q) \in {\mathcal{CP}}({\mathcal{H}}, {\mathcal{K}}).$$ Another special distinguished class of maps are the free homomorphism from ${\mathcal{A}}({\mathcal{H}})$ to ${\mathcal{A}}({\mathcal{K}})$ $$\{ (V,0) \in {\mathcal{CP}}({\mathcal{H}}, {\mathcal{K}}) \mid V: {\mathcal{H}} \to {\mathcal{K}} \text{ isometric} \}.$$ The set ${\mathcal{CP}}({\mathcal{H}}, {\mathcal{K}})$ is also convex and weakly compact. Free CP maps transform free states into free states and one checks from (\[qfs\]) and (\[qfm\]) that $$\omega_Q \circ \Lambda_{A,B} = \omega_{A^*QA + B}.$$
The construction of the quantum Markov process consists of using a completely positive map to contract the observable and then applying a single-party state that is invariant under a completely positive map. We have the following lemma concerning the existence of such invariant states.
\[lem2\] Let $\Lambda_{A,B}$ be a completely positive free transformation of ${\mathcal{A}}({\mathcal{H}})$ as in (\[qfm\]) and assume that $\dim({\mathcal{H}}) < \infty$, then $\Lambda_{A,B}$ has a unique invariant state if and only if ${\lVertA\rVert} < 1$. Moreover, the unique invariant state is free with symbol $Q$ determined by $$Q = A^*Q\,A + B.
\label{lem2:1}$$
The condition ${\lVertA\rVert} < 1$ is equivalent to the non-existence of non-trivial solutions to the homogeneous equation $Q = A^*Q\, A$. It has to be satisfied to have uniqueness of the solution of the invariance condition (\[lem2:1\]) for symbols. Conversely, suppose that ${\lVertA\rVert} < 1$, then there exists by the fixed point theorem for contractions a unique $Q$ such that $$Q = A^*Q\,A + B.$$ This $Q$ satisfies $0 \le Q \le {\mathbb{1}}$ as we may obtain $Q$ by iterating the map $X \mapsto A^*X\,A + B$ with initial value 0. It is then easily checked that $$\lim_{n\to\infty} \Lambda_{A,B}^n = \omega_Q$$ which guarantees both the uniqueness of the invariant state and its free character.
Constructing a chain {#s3:2}
--------------------
We now have the necessary ingredients to introduce the free Fermionic counterpart of Construction \[con1\]. There are natural embeddings $$a(\varphi) \mapsto a(\varphi \oplus 0) \enskip\text{and}\enskip a(\psi) \mapsto a(0 \oplus \psi)
\label{car:emb}$$ of ${\mathfrak{A}}({\mathcal{H}})$ and ${\mathfrak{A}}({\mathcal{K}})$ into ${\mathfrak{A}}({\mathcal{H}} \oplus {\mathcal{K}})$. Both factors together generate the large algebra and they satisfy graded commutation relations as creation operators in different factors anticommute. We can transport the construction of the quantum Markov chain (\[qmc\],\[fcs\]) and its entropy (\[qme\]) to the free Fermionic setting. The spin chain algebra $\otimes^{{\mathbb{N}}} {\mathcal{M}}_d$ is replaced by a semi-infinite Fermionic chain ${\mathfrak{A}}(\operatorname*{\mbox{\Large\mbox{$\oplus$}}}^{{\mathbb{N}}} {\mathcal{H}})$ where ${\mathfrak{A}}({\mathcal{H}})$ is now the one site algebra.
The basic ingredient is a free CP transformation $\Lambda_{A,B}$ of ${\mathfrak{A}}({\mathcal{H}})$ and we look for free CP maps $\Lambda_{C,D}$ from ${\mathfrak{A}}({\mathcal{H}} \oplus {\mathcal{H}})$ to ${\mathfrak{A}}({\mathcal{H}})$ such that $$\label{eq:qf_compatibility}
\Lambda_{C,D} \circ \jmath_1 = \Lambda_{C,D} \circ \jmath_2 = \Lambda_{A,B}.$$ Here, $\jmath_1$ and $\jmath_2$ are the natural embeddings of ${\mathfrak{A}}({\mathcal{H}})$ into the first and second factor of ${\mathfrak{A}}({\mathcal{H}} \oplus {\mathcal{H}})$ $$\jmath_1(a(\varphi)) = a(\varphi \oplus 0) \enskip\text{and}\enskip \jmath_2(a(\varphi)) = a(0 \oplus \varphi).$$ Applying the compatibility condition (\[eq:qf\_compatibility\]) to monomials $a(\varphi)$ and $a^*(\varphi)a(\psi)$ we see that $$C = \begin{bmatrix} A & A \end{bmatrix} \enskip\text{and}\enskip D = \begin{bmatrix} B & X \\ X^* & B \end{bmatrix},
\label{qfcomp}$$ where $X$ is as of yet undetermined and allows for some freedom in the choice of $D$. Because of the structure of free CP maps, the compatibility conditions (\[qfcomp\]) are not only necessary but also sufficient and we can rephrase the whole construction on the level of symbols. Doing so, graded tensor products become direct sums.
\[con3\] Let $(A,B) \in {\mathcal{CP}}({\mathcal{H}})$ and let $Q \in {\mathcal{Q}}({\mathcal{H}})$ be such that $\omega_Q$ is invariant under $\Lambda_{A,B}$: $$Q = A^* Q A + B.
\label{inv}$$ Let $X: {\mathcal{H}} \to {\mathcal{H}}$ satisfy the compatibility condition $$(C,D) \in {\mathcal{CP}}({\mathcal{H}} \oplus {\mathcal{H}}, {\mathcal{H}}) \enskip\text{with}\enskip C \text{ and } D \text{ as in~(\ref{qfcomp})}.
\label{qfcomp2}$$ The free Markov chain defined by $X$ and $Q$ is the symbol $$Q_\infty = \underset{n\to\infty}{\text{w-lim}}\, P_n R_n P_n^*\enskip \text{on } \oplus^{{\mathbb{N}}} {\mathcal{H}}
\label{qinf}$$ where $$\begin{aligned}
&P_n: {\mathcal{H}} \oplus \Bigl( \oplus_{k=0}^{n-1} {\mathcal{H}} \Bigr) \to \Bigl( \oplus_{k=0}^{n-1} {\mathcal{H}} \Bigr): \varphi \oplus \psi_n \mapsto \psi_n \\
&R_0 = Q \enskip\text{and}\enskip R_{n+1} = \Bigl( C^* \oplus \left( \oplus^n {\mathbb{1}}\right) \Bigr) R_n \Bigl(C \oplus \left( \oplus^n {\mathbb{1}}\right) \Bigr) + \Bigl( D\oplus \left( \oplus^n 0 \right) \Bigr).
\label{rn}\end{aligned}$$
$\square$
There is some freedom in choosing the channel $\Lambda_{C,D}$. The operator $X$ has to be chosen such that the $\Lambda_{C,D}$ is completely positive, i.e. $0 \leq D \leq {\mathbb{1}}- C^* C$. One may wonder if and when this is possible. The question of existence of compatible channels is answered by the following lemma.
\[lem1\] The compatibility condition (\[qfcomp2\]) is satisfiable if and only if $$A^*A \le \min \bigl( \{\tfrac{1}{2}\, {\mathbb{1}}, {\mathbb{1}}- B \}\bigr).$$
We look for the necessary and sufficient conditions for the existence of a $X: {\mathcal{H}} \to {\mathcal{H}}$ such that $$\begin{bmatrix} B &X \\ X^* &B \end{bmatrix} \ge 0 \enskip\text{and}\enskip \begin{bmatrix} {\mathbb{1}}- A^*A - B &-A^*A - X \\ -A^*A - X^* &{\mathbb{1}}- A^*A - B \end{bmatrix} \ge 0.$$ Clearly $0 \le B \le {\mathbb{1}}- A^*A$ as $(A,B) \in {\mathcal{CP}}({\mathcal{H}})$. The remaining positivity conditions are then the existence of $S$ and $T$ with $$\begin{split}
&{\lVertS\rVert} \le 1,\enskip {\lVertT\rVert} \le 1, \enskip X = B^{\frac{1}{2}} S B^{\frac{1}{2}}, \enskip \text{and } \\
&A^*A + X = ({\mathbb{1}}- A^*A - B)^{\frac{1}{2}} T ({\mathbb{1}}- A^*A - B)^{\frac{1}{2}}.
\end{split}$$ Replacing $S$ and $T$ by their Hermitian parts, we may restrict to Hermitian $X$ and so we need $$[-B \,,\, B] \cap [-{\mathbb{1}}+ B \,,\, {\mathbb{1}}- 2A^*A - B ] \neq \emptyset$$ or, equivalently, that $$[{\mathbb{1}}\,,\, {\mathbb{1}}+ 2B] \cap [2B \,,\, 2 {\mathbb{1}}- 2A^*A ] \neq \emptyset.$$ But this is the case if and only if $$\max \bigl( \{ {\mathbb{1}}, 2B \}\bigr) \le 2{\mathbb{1}}- 2A^*A \enskip\text{or}\enskip A^*A \le \min \bigl( \{\tfrac{1}{2}\, {\mathbb{1}}, {\mathbb{1}}- B \}\bigr).$$
Let us look at this compatibility condition for a simple case.
\[ex:compat\] If ${\mathcal{H}} = {{\mathbb C}}$, we have $$D = \begin{bmatrix} b & x \\ \overline{x} & b \end{bmatrix} \text{ and } C = \begin{bmatrix} a & a \end{bmatrix} \,,$$ with $a,x \in {{\mathbb C}}$ and $b \in {{\mathbb R}}$. From the complete positivity of $\Lambda_{a,b}$, we know that $0 \leq b \leq 1-|a|^2$. The complete positivity conditions for $C$ and $D$ limit the choice for $x$. From $0 \leq D$ we see that $$|x| \leq b \,.$$ From $D \leq {\mathbb{1}}- C^* C $ on the other hand, we get that $$| x + |a|^2| \leq 1 - |a|^2 -b \,.$$ These two inequalities means that $x$ has to lie in the intersection of two circles in the complex plane, one centred at $0$ with radius $r_1=b$ and another one centred at $-|a|^2$ with radius $r_2=1-|a|^2 - b$. These two circles have an intersection when the distance between the centres is smaller than the sum of the radii. Hence, the channel is extendible if $$|a|^2 \leq \frac{1}{2}$$ which corresponds to the conditions in the lemma.
Given the constituents of the Markov construction, the channel $\Lambda_{C,D}$ and the invariant symbol $Q$, the symbol of the full process $Q_\infty$ can easily be determined
\[pro1\] The symbol $Q_\infty$ in (\[qinf\]) is an Hermitian block Toeplitz matrix with entries $$\bigl( Q_\infty \bigr)_{i\,i} = Q \enskip\text{and}\enskip \bigl( Q_\infty \bigr)_{i\,i+n} = (A^*)^n (Q - B + X).$$ Here $i=0,1,2,\ldots$ and $n=1,2,3,\ldots$
The proof consists in a straightforward computation of the consecutive $R_n$ in (\[rn\]) combined with the invariance equation (\[inv\]).
Fermionic Entropy density {#sec:density}
=========================
In this section we compute the entropy $h$ for the Fermionic Markov process constructed in Section \[sec:fermionic\]. We can associate an entropy to a Fermionic Markov chain using (\[qfent\]) $$\begin{aligned}
{\mathsf{h}}(X,Q)
&:= \lim_{n \to \infty} \frac{1}{n}\, {\mathsf{S}}(P_n R_n P_n^*) \\
&:= \lim_{n \to \infty} \frac{1}{n}\, \Bigl( - \operatorname*{Tr}P_n R_n P_n^* \log(P_n R_n P_n^*) \nonumber \\
&\phantom{:= \lim_{n \to \infty} \frac{1}{n}\, \Bigl(\ }- \operatorname*{Tr}({\mathbb{1}}- P_n R_n P_n^*) \log({\mathbb{1}}- P_n R_n P_n^*) \Bigr).
\label{qfme}\end{aligned}$$
A first method to compute this relies directly on the expression (\[qfent\]) for the entropy of a free state in terms of its symbol and on the structure of the symbols $Q_\infty$ in Proposition \[pro1\]. A second way is to rewrite the entropy as the asymptotic rate of disorder, as in the classical case, see Section \[sec:class\]. This last approach was used in [@Blackwell57; @Fannes92] to compute the entropy of a hidden Markov process. The first method uses the full local restrictions of the state while the second relies on the incremental structure of the local states given by a transfer matrix like construction, see (\[fcs\]) and (\[rn\]).
Direct approach
---------------
The first approach to calculating the entropy density uses an extension of Szegö’s theorem to block Toeplitz matrices $\hat T$. This theorem allows to calculate asymptotic densities of trace functions of Toeplitz matrices. A block Toeplitz matrix is a block matrix $\hat T$ where the blocks along diagonals are equal $${\hat T}_{i,j} = {\hat T}_{i+k,j+k} \,,$$ where ${\hat T}_{i,j}$ denotes a block elements. Using Szegö’s theorem, we can write densities $$\lim_{n \rightarrow \infty} \operatorname*{Tr}\frac{f(\hat{T}_n)}{n}$$ of a matrix function $f$ and the finite projections ${\hat T}_n = P_n {\hat T} P_n$ in terms of a generating function $T(\theta)$. The Fourier coefficients of this $T(\theta)$ are the elements on the diagonals of $\hat T$. We will now formulate this more precisely.
Let $T: [-\pi,\pi[ \to {\mathcal{M}}_d$ be an essentially bounded measurable matrix-valued function on the circle and denote its Fourier coefficients by $$\hat T(k) := \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, T(\theta)\, {\mathrm{e}}^{-ik\theta} \in {\mathcal{M}}_d.$$ A function $T$ is essentially bounded if there exist a constant $M$ such that $|T(\theta)| \leq M$ almost everywhere. The operator $$\hat T = \begin{pmatrix}
\hat T(0) &\hat T(1) &\hat T(2) &\ldots \\
\hat T(-1) &\hat T(0) &\hat T(1) &\ldots \\
\hat T(-2) &\hat T(-1) &\hat T(0) &\ldots \\
\vdots & \vdots & \vdots &\ddots \\
\end{pmatrix}$$ defined on $\ell^0({\mathbb{N}}) \otimes {{\mathbb C}}^d$ extends to a bounded linear transformation of $\ell^2({\mathbb{N}}) \otimes {{\mathbb C}}^d$. Operators of this type are block Toeplitz matrices and one has $${\lVert\hat T\rVert} = {\lVertT\rVert}_\infty = \underset{\theta}{\text{ess\,sup}}\, {\lVertT(\theta)\rVert} \,,$$ where the essential supremum of $T$ is the infimum of all constants $M$ that bound $|T(\theta)|$ almost everywhere.
The Toeplitz matrices we are interested in are symbols and hence self-adjoint. For such Toeplitz matrices, we have that ${\hat T}^* = \hat T$ if and only if the function $T$ takes values in the Hermitian matrices.
### Szegö’s theorem
An extension of Szegö’s theorem to block Toeplitz matrices characterizes the limiting spectrum of principal submatrices $P_n \hat T\, P_n$ in terms of the generating function $T$, see [@Miranda00]. Here $P_n$ projects on the first $n$ blocks in $\ell^2({\mathbb{N}}) \otimes {{\mathbb C}}^d$. We obtain here a more general characterization of such limiting submatrices.
Let us denote for a simply connected compact subset ${\mathcal{K}}$ of ${{\mathbb C}}$ by ${\mathcal{H}}({\mathcal{K}})$ the set of continuous functions $f: {\mathcal{K}} \to {{\mathbb C}}$ that are holomorphic in the interior $\overset{\circ}{{\mathcal{K}}}$ of ${\mathcal{K}}$. Mergelyan’s theorem [@Rudin87] asserts that the complex polynomials in the indeterminate $z$ are dense in ${\mathcal{H}}({\mathcal{K}})$: for any $f \in {\mathcal{H}}({\mathcal{K}})$ and $\epsilon > 0$ there exists a polynomial $p^\epsilon$ such that $$\max_{z \in {\mathcal{K}}} \bigl| f(z) - p^\epsilon(z) \bigr| \le \epsilon.$$ Finally, let us denote by ${\mathbb{E}}_n$ the conditional expectation from ${\mathcal{B}} \bigl( \ell^2({\mathbb{N}}) \bigr) \otimes {\mathcal{M}}_d \to {\mathcal{M}}_d$ which traces out the first $n$ blocks $${\mathbb{E}}_n(X) := \frac{1}{n}\, \sum_{j=0}^{n-1} X_{jj} \in {\mathcal{M}}_d.$$
We get the following generalization of Szegö’s theorem [@Grenander01].
\[thm:szego\] Let $\{T_1, T_2, \ldots, T_k \} \subset {\mathcal{L}}^\infty \bigl( [-\pi, \pi[, {\mathcal{M}}_d \bigr)$ be such that every $T_j(\theta)$ is $\theta$-a.e. diagonalizable, let $f_j \in {\mathcal{H}} \bigl( \{ z \in {{\mathbb C}}\mid {\lvertz\rvert} \le {\lVertT_j\rVert}_\infty \} \bigr)$ for $j = 1, 2, \ldots, k$ and let $A_j \in {\mathcal{M}}_d$, $j = 1, 2, \ldots, k+1$, then $$\begin{split}
&\lim_{n \to \infty} {\mathbb{E}}_n \Bigl( ({\mathbb{1}}\otimes A_1)\, f_1 \bigl( P_n \hat T_1 P_n \bigr)\, ({\mathbb{1}}\otimes A_2)\, \cdots f_k \bigl( P_n \hat T_k P_n \bigr)\, ({\mathbb{1}}\otimes A_{k+1}) \Bigr) \\
&\quad= \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, A_1\, f_1(T_1(\theta))\, A_2 \cdots f_k(T_k(\theta))\, A_{k+1}.
\end{split}
\label{thm:szego:1}$$
The proof relies on a continuity argument combined with a standard counting argument. First remark that given $\epsilon > 0$ every $f_j$ can be approximated by a suitable complex polynomial $p^\epsilon_j$ $$\max_{{\lvertz\rvert} \le {\lVertT_j\rVert}_\infty} \bigl| f_j(z) - p^\epsilon_j(z) \bigr| \le \epsilon.$$ Next, as ${\lVertT_j(\theta)\rVert} \le {\lVertT_j\rVert}_\infty$ a.e., we can use von Neumann’s inequality [@Halmos74] to get $$\begin{aligned}
&{\lVertf_j(T_j(\theta))\rVert} \le \max_{{\lvertz\rvert} \le {\lVertT_j\rVert}_\infty} {\lvertf_j(z)\rvert} \enskip\text{and} \\
&\bigl\Vert f_j(T_j(\theta)) - p^\epsilon_j(T_j(\theta)) \bigr\Vert = \bigl\Vert \bigl( f_j - p^\epsilon_j \bigr)(T_j(\theta)) \bigr\Vert
\nonumber \\
&\quad \le \max_{{\lvertz\rvert} \le {\lVertT_j(\theta)\rVert}} \bigl( f_j - p^\epsilon_j \bigr)(z) \le \max_{{\lvertz\rvert} \le {\lVertT_j\rVert}_\infty} \bigl( f_j - p^\epsilon_j \bigr)(z) \le \epsilon.\end{aligned}$$ These estimates allow to replace the $f_j$ in (\[thm:szego:1\]) by polynomials. It then remains to verify the statement for monomials, but this reduces to a standard counting argument.
In the case where there is only one function $f(X)=X^k$ and $A_j = {\mathbb{1}}$, the density limit can be worked out as follows: $$\begin{aligned}
\lim_{n \rightarrow \infty} \frac{1}{n} \operatorname*{Tr}(P_n \hat T P_n)^k &= \lim_{n \rightarrow \infty} \frac{1}{n} \sum_{i_1,\ldots, i_k =0}^n \hat T_{i_1, i_2} \hat T_{i_2, i_3} \ldots \hat T_{i_k, i_1} \\
&= \lim_{n \rightarrow \infty} \frac{1}{n} \sum_{i_1,\ldots, i_k =0}^n \hat T( i_2 - i_1 ) \hat T( i_3 - i_2 ) \ldots \hat T( i_k - i_1 )\end{aligned}$$ By substituting $v_1 = i_2 - i_1 \,, \ldots \,, v_{k-1} = i_k - i_{k-1}$, this sum becomes: $$\lim_{n \rightarrow \infty} \frac{1}{n} \sum_{v_1, \ldots, v_{k-1} = -n}^n \,\, \sum_{i_1 \in \mathcal{S}_n(v_1,\ldots, v_{k-1})} \hat T(v_1) \ldots \hat T(v_{k-1}) \hat T(-v_1 - \ldots - v_{k-1}) \,,$$ where $\mathcal{S}_n(v_1,\ldots, v_{k-1})$ is the set of indices $i$ such that $v_1+i, v_1+v_2+i, \ldots, v_1 + \ldots v_{k-1} + i \in [0,n]$. The number of elements in this set increases by exactly one when $n$ goes to $n+1$, so in the limit we get $$\lim_{n \rightarrow \infty} \sum_{v_1, \ldots, v_{k-1} = -\infty}^\infty \hat T(v_1) \ldots \hat T(v_{k-1}) \hat T(-v_1 - \ldots - v_{k-1}) \,.$$ This is exactly the zeroth Fourier coefficient of $T(\theta)^k$, so we get that the density equals $$\frac{1}{2 \pi} \int_{-pi}^\pi d \theta f(T(\theta)) \,.$$ The general case of the theorem can be worked out in a similar manner.
To deal with entropy we don’t need the full amalgamated extension of Theorem \[thm:szego\] of Szegö’s theorem but we may restrict ourselves for an Hermitian $T$ to the asymptotic eigenvalue distribution of the principal blocks $P_n \hat T P_n$. Taking the trace of (\[thm:szego:1\]) with a single $f$ and all $A_j = {\mathbb{1}}$ we recover the result [@Miranda00]. We denote by $\inf(T)$ and $\sup(T)$ the largest and smallest real numbers such that $$\inf(T) \le T \le \sup(T)\enskip \text{a.e.}$$ The increasingly ordered eigenvalues $\bigl( \tau_1(\theta), \tau_2(\theta), \ldots, \tau_d(\theta) \bigr)$ of $T(\theta)$ are measurable functions of $\theta$ that satisfy $$\inf(T) \le \tau_1(\theta) \le \cdots \le \tau_d(\theta) \le \sup(T).$$ The eigenvalue distribution of $P_n \hat T P_n$ is the atomic probability measure $$\delta_n = \frac{1}{nd}\, \sum_{\lambda \in \sigma(P_n\hat T P_n)} \delta_\lambda.$$
\[thm1\] With the assumptions of above $$\underset{n \to \infty}{{\mathrm{w^*\!\text{-}\,lim}}}\, \delta_n = \delta_\infty,$$ where $$\delta_\infty \bigl( ]-\infty, t] \bigr) = \frac{1}{d}\, \sum_{k=1}^d \frac{1}{2\pi}\, \int_{\tau_k(\theta) \le t} \!d\theta.
\label{limdis}$$
An equivalent way to express this result is saying that for any continuous complex function $f$ on $[\inf(T), \sup(T)]$ $$\lim_{n \to \infty} \frac{1}{nd}\, \operatorname*{Tr}f(P_n \hat T P_n) = \frac{1}{2\pi} \int_{-\pi}^\pi \!d\theta\, \frac{1}{d}\, \operatorname*{Tr}f(T(\theta)).
\label{szego2}$$ This version is in some sense more natural as it doesn’t involve the reordering of the eigenvalue functions $\tau_k$ used in the definition of the distribution function of the limiting eigenvalue distribution (\[limdis\]).
We can apply Theorem \[thm1\] to the computation of the entropy, replacing the Toeplitz operator $T$ by $Q_\infty$ in Proposition \[pro1\] and choosing in (\[szego2\]) $$f(\lambda) = -\lambda \log(\lambda) - (1-\lambda)\log(1-\lambda)\enskip \text{on } (0,1).$$ The generating function $T$ becomes $$\begin{aligned}
\theta \mapsto Q + &(Q -B +X) A {\mathrm{e}}^{\imath \theta}({\mathbb{1}}- A {\mathrm{e}}^{\imath \theta})^{-1} \\ &+ A^* {\mathrm{e}}^{-\imath \theta}({\mathbb{1}}- A^* {\mathrm{e}}^{-\imath \theta})^{-1} (Q -B +X^*).\end{aligned}$$
In case of a single particle space $\mathcal{H}=\mathbb{C}$ as in the Example \[ex:compat\], the entropy can be calculated from the scalar version of the above function $$\theta \mapsto q + (q -b +x) a {\mathrm{e}}^{\imath \theta}(1 - a {\mathrm{e}}^{\imath \theta})^{-1} \\ + \text{h.c.} \,$$ where $x$ lies within the two circles determining the compatibility condition, as explained in Example \[ex:compat\].
This scalar function is linear in $x$ and the function $f$ is concave in it’s argument. Hence the minimal entropy is obtained on the border of the compatibility region, much like in the case of the classical Markov process and compatible hidden Markov processes described at the end of Section \[sec:class\].
Entropy rate approach
---------------------
The second approach expresses the entropy as an asymptotic rate. Let $\omega$ be a translation invariant state on a quantum spin chain $\otimes^{\mathbb{Z}}{\mathcal{M}}_d$ and denote by $\rho_{(0,n-1)}$ its reduced density matrices, i.e.$$\omega(X) = \operatorname*{Tr}\bigl( \rho_{(0,n-1)} X \bigr)\enskip \text{for } X \in \otimes_{k=0}^{n-1} {\mathcal{M}}_d.$$ As we have seen before, subadditivity combined with translation invariance guarantee the existence of the mean entropy of $\omega$ for intervals $${\mathsf{s}}(\omega) = \lim_{n \to \infty} \frac{1}{n}\, {\mathsf{S}}(\rho_{(0,n-1)}).$$ Moreover, strong subadditivity in conjunction with translation invariance also guarantees that $$\begin{aligned}
&n \mapsto {\mathsf{S}}(\rho_{(0,n-1)})\enskip \text{is monotonically increasing and}
\label{ent:mon} \\
&{\mathsf{s}}(\omega) = \lim_{n \to \infty} \frac{1}{n}\, {\mathsf{S}}(\rho_{(0,n-1)}) = \lim_{n \to \infty} \Bigl( {\mathsf{S}}(\rho_{(0,n)}) - {\mathsf{S}}(\rho_{(0,n-1)}) \Bigr).
\label{ent:hop}\end{aligned}$$ Both properties (\[ent:mon\]) and (\[ent:hop\]) fail for general quantum states or for general finite local regions [@Kay01]. These results for quantum spin chains extend to Fermionic lattices using the natural embeddings (\[car:emb\]) and restricting to even states [@Araki03]. The equality of both limits in (\[ent:hop\]) can be seen as a discrete version of de l’Hôpital’s rule. Obviously, the existence of the limit of the differences is a much stronger requirement than that of the averages.
For free Fermionic states we can work at the level of symbols. E.g., strong subadditivity of entropy amounts to $${\mathsf{S}}(Q_{123}) + {\mathsf{S}}(Q_2) \le {\mathsf{S}}(Q_{12}) + {\mathsf{S}}(Q_{23})$$ where ${\mathsf{S}}$ is defined in (\[qfent\]) and where the symbols in the inequality are as follows $$Q_{123} = \begin{bmatrix} Q_1 &T &S \\ T^* &Q_2 &R \\ S^* &R^* &Q_3 \end{bmatrix},\enskip Q_{12} = \begin{bmatrix} Q_1 &T \\ T^* &Q_2 \end{bmatrix}, \enskip\text{and}\enskip Q_{23} = \begin{bmatrix} Q_2 &R \\ R^* &Q_3 \end{bmatrix}.$$ For more on functions that satisfy such strong subadditivity, see [@audenaert_strongly_2010].
Below, we extend the equality of the limit of differences with that of averages, as in (\[ent:hop\]), to a much wider class of functions than the strongly subadditive ones, like the entropy of a symbol (\[qfent\]). The argument relies on regularity of the functions and not on subadditivity or convexity which rarely hold. Szegö’s theorem follows as a consequence.
We first show that the theorem holds for polynomials.
\[lem3\] With the notation and assumptions on an Hermitian Toeplitz operator at the beginning of this section, for any polynomial $p$ $$\lim_{n \to \infty} \Bigl( \operatorname*{Tr}p(P_{n} \hat T P_{n}) - \operatorname*{Tr}p(P_{n-1} \hat T P_{n-1}) \Bigr) = \frac{1}{2\pi} \int_{-\pi}^\pi \!d\theta\, \operatorname*{Tr}p(T(\theta)).$$
It suffices to consider $p(\lambda) = \lambda^k$ for $k \in {\mathbb{N}}$. We have $$\begin{aligned}
&\lim_{n \to \infty} \operatorname*{Tr}(P_n \hat T P_n)^k - \operatorname*{Tr}(P_{n-1} \hat T P_{n-1})^k \\
&\quad= \lim_{n \to \infty} \sum_{i=1}^n \operatorname*{Tr}\bigl( (P_n \hat T P_n)^k \bigr)_{ii} - \sum_{i=1}^{n-1} \operatorname*{Tr}\bigl( (P_{n-1} \hat T P_{n-1})^k) \bigr)_{ii} \\
&\quad= \lim_{n \to \infty} \Bigl( \sum_{i_1,\ldots,i_k=1}^n - \sum_{i_1,\ldots,i_k=1}^{n-1} \Bigr) \\
&\qquad \qquad \qquad \operatorname*{Tr}\Bigl\{ (P_n \hat T P_n)_{i_1i_2} \cdots (P_n \hat T P_n)_{i_{k-1}i_k} (P_n \hat T P_n)_{i_ki_1} \Bigr\} \\
&\quad= \lim_{n \to \infty} \Bigl( \sum_{i_1,\ldots,i_k=1}^n - \sum_{i_1,\ldots,i_k=1}^{n-1} \Bigr) \hat T (i_2 - i_1) \ldots \hat T (i_1 - i_k) \,,\end{aligned}$$ where $(Q)_{ij}$ denotes the block at position $(i,j)$ inside of a block matrix $Q$.
By substituting $v_1 = i_2 - i_1 \,, \ldots \,, v_{k-1} = i_k - i_{k-1}$, this sum becomes: $$\begin{aligned}
&\lim_{n \rightarrow \infty} \sum_{v_1, \ldots, v_{k-1} = -n}^n \,\, \Bigl(\sum_{i_1 \in \mathcal{S}_n(v_1,\ldots, v_{k-1})} - \sum_{i_1 \in \mathcal{S}_{n-1}(v_1,\ldots, v_{k-1})} \Bigr) \\
& \qquad \qquad \qquad \qquad \qquad \qquad \hat T(v_1) \ldots \hat T(v_{k-1}) \hat T(-v_1 - \ldots - v_{k-1}) \,,\end{aligned}$$ where $\mathcal{S}_n(v_1,\ldots, v_{k-1})$ is the set of indices $i$ such that $v_1+i, v_1+v_2+i, \ldots, v_1 + \ldots v_{k-1} + i \in [0,n]$. For fixed $v_1, \ldots, v_{k-1}$, the number of elements in these sets increases by exactly one when $n$ goes to $n+1$. Hence, the difference of sums between brackets equals one and we arrive at the expression prescribed by the lemma.
We can now use this lemma and an approximation argument to prove the general case.
\[thm:lhopital\] With the notation and assumptions on an Hermitian Toeplitz operator at the beginning of this section, for any function $f$ that is absolutely continuous on the interval $[ \inf(T), \sup(T) ]$ $$\lim_{n \to \infty} \Bigl( \operatorname*{Tr}f(P_{n+1} \hat T P_{n+1}) - \operatorname*{Tr}f(P_n \hat T P_n) \Bigr) = \frac{1}{2\pi} \int_{-\pi}^\pi \!d\theta\, \operatorname*{Tr}f(T(\theta)).$$
By the continuity of the eigenvalues of a matrix and by the minimax principle [@Courant89] we can label the eigenvalues of $P_n \hat T P_n$ as $$\begin{aligned}
&\{\tau^n_{k\,j} \mid k = 1,2,\ldots, d,\ j = 1,2, \ldots, n\} \enskip \text{with}
\nonumber \\
&\inf(T) \le \tau^n_{1\,j} \le \tau^n_{2\,j} \le \cdots \le \tau^n_{d\,j} \le \sup(T) \enskip\text{and}\enskip \tau^{n+1}_{k\,j} \le \tau^n_{k\,j} \le \tau^{n+1}_{k\,j+1}.
\label{thm:lhopital:0}\end{aligned}$$ See [@Horn90] for a proof of this interlacing property.
Let $f: [\inf(T), \sup(T)] \to {{\mathbb C}}$ be absolutely continuous with integrable derivative $g$, then for any $\lambda,\tau \in [\inf(T), \sup(T)]$ $$f(\lambda) = f(\tau) + \int_\tau^\lambda \!dx\, g(x).$$ Therefore $$\begin{aligned}
\frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, f(\tau(\theta))
&= \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, \Bigl\{ f(\tau) + \int_\tau^{\tau(\theta)} \! dx\, g(x) \Bigr\} \\
&= f(\tau) + \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta \int_\tau^{\tau(\theta)} \!dx\, g(x) \\
&= f(\tau) + \int_{\inf(T)}^{\sup(T)} \!dx\, g(x)\, \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, \eta(\tau,x,\theta).
\label{thm:lhopital:1}\end{aligned}$$ Here, $\eta$ is defined as $$\eta(\tau,x,\theta) =
\begin{cases} 1 &\tau < x < \tau(\theta) \\-1 &\tau(\theta) < x < \tau \\0 &\text{otherwise} \end{cases}.$$ By (\[thm:lhopital:1\]) we rewrite the increment of traces of $f(P_n \hat T P_n)$ as $$\begin{aligned}
&\operatorname*{Tr}f(P_{n+1} \hat T P_{n+1}) - \operatorname*{Tr}f(P_n \hat T P_n) \\
&\quad= \sum_{k=1}^d \Bigl\{ \sum_{j=1}^{n+1} f(\tau^{n+1}_{k\,j}) - \sum_{j=1}^n f(\tau^n_{k\,j}) \Bigr\} \\
&\quad= \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, \sum_{k=1}^d f(\tau_k(\theta)) - \sum_{k=1}^d \int_{\inf(T)}^{\sup(T)} \!dx\, g(x)\, \frac{1}{2\pi} \int_{-\pi}^\pi \!d\theta
\nonumber\\
&\phantom{\quad= \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\ } \Bigl\{ \sum_{j=1}^{n+1} \eta(\tau^{n+1}_{k\,j}, x, \tau_k(\theta)) - \sum_{j=1}^n \eta(\tau^n_{k\,j}, x, \tau_k(\theta)) \Bigr\} \\
&\quad= \frac{1}{2\pi}\, \int_{-\pi}^\pi \!d\theta\, \operatorname*{Tr}f(\hat T(\theta)) - \sum_{k=1}^d \int_{\inf(T)}^{\sup(T)} \!dx\, g(x)\, \frac{1}{2\pi} \int_{-\pi}^\pi \!d\theta\, h^n_k(x,\theta), \end{aligned}$$ with $$h^n_k(x,\theta) = \sum_{j=1}^{n+1} \eta(\tau^{n+1}_{k\,j}, x, \tau_k(\theta)) - \sum_{j=1}^n \eta(\tau^n_{k\,j}, x, \tau_k(\theta)).$$ The functions $h^n_k$ are piecewise constant with values -1, 0 or 1 due to the interlacement (\[thm:lhopital:0\]) of the $\tau^n_{k\,j}$. As any integrable $g$ on $[\inf(T), \sup(T)]$ can be arbitrarily well approximated in ${\mathcal{L}}^1$-norm by polynomials, the theorem follows from Lemma \[lem3\].
Conclusion
==========
We have studied a free Fermionic version of quantum Markov processes. Due to the free Fermionic nature of the states we can characterize all possible Markov processes that one can construct. The density matrices of these states can be described by a Toeplitz matrix. By studying the behaviour of the eigenvalues of subsequent Toeplitz matrices, we have proved a new Szegö theorem that allows to calculate the asymptotic entropy rate. This is what corresponds in the free Fermionic case to the method proposed by Blackwell [@Blackwell57].
It would be interesting to look for other quantum Markov processes for which an explicit calculation of the entropy rate is possible. Processes with a high symmetry are obvious first choices. Hopefully, such a calculation can lead to a quantum version of the Blackwell dynamical system.
[^1]: Current affiliation: Meteorologisches Institut, University of Hamburg
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Shaoli Huang and Dacheng Tao, [^1]'
bibliography:
- 'egbib.bib'
title: 'Real Time Fine-Grained Categorization with Accuracy and Interpretability'
---
visual categorization (FGVC) refers to the task of indentifying objects from subordinate categories and is now an important subfield in object recognition. FGVC applications include, for example, recognizing species of birds [@welinder2010caltech; @wah2011caltech; @berg2014birdsnap], pets [@khosla2011novel; @parkhi2012cats], flowers [@nilsback2008automated; @angelova2013image], and cars [@stark2011fine; @maji2013fine]. Lay individuals tend to find it easy to quickly distinguish basic-level categories (e.g., cars or dogs), but identifying subordinate classes like “*Ringed-billed gull*” or “*California gull*” can be difficult, even for bird experts. Tools that aid in this regard would be of high practical value.
This task is made challenging due to the small inter-class variance caused by subtle differences between subordinaries and the large intra-class variance caused by negative factors such as differing pose, multiple views, and occlusions. However, impressive progress [@wah2011multiclass; @berg2014birdsnap; @vedaldi2014understanding; @krause2015fine; @xu2015augmenting] has been made over the last few years and fine-grained recognition techniques are now close to practical use in various applications such as for wildlife observation and in surveillance systems.
![Overview of the proposed approach. We propose to classify fine-grained categories by modeling the subtle difference from specific object parts. Beyond classification results, the proposed DPS-CNN architecture also offers human-understandable instructions on how to classify highly similar object categories explicitly.[]{data-label="fig:title"}](fig1){width="\linewidth"}
Whilst numerous attempts have been made to boost the classification accuracy of FGVC [@deng2013fine; @chai2013symbiotic; @branson2014bird; @lin2015bilinear; @wang2015multiple], an important aspect of the problem has yet to be addressed, namely the ability to generate a human-understandable “manual” on how to distinguish fine-grained categories in detail. For example, ecological protection volunteers would benefit from an algorithm that could not only accurately classify bird species but also provide brief instructions on how to distinguish very similar subspecies (a “*Ringed-billed*” and “*California gull*”, for instance, differ only in their beak pattern, see Figure \[fig:title\]), aided by some intuitive illustrative examples. Existing fine-grained recognition methods that aim to provide a visual field guide mostly follow a “part-based one-vs.-one features” (POOFs) [@berg2013poof; @berg2013you; @berg2014birdsnap] routine or employ human-in-the-loop methods [@kumar2012leafsnap; @branson2014ignorant; @van2015building]. However, since the amount of available data requiring interpretation is increasing drastically, a method that simultaneously implements and interprets FGVC using deep learning methods [@krizhevsky2012imagenet] is now both possible and advocated.
{width="1.\linewidth"}
It is widely acknowledged that the subtle differences between fine-grained categories mostly reside in the unique properties of object parts [@rosch1976basic; @berg2013poof; @chai2013symbiotic; @maji2014part; @zhang2014part; @zhang2014fused]. Therefore, a practical solution to interpreting classification results as human-understandable manuals is to discover classification criteria from object parts. Some existing fine-grained datasets provide detailed part annotations including part landmarks and attributes [@wah2011caltech; @maji2013fine]. However, they are usually associated with a large number of object parts, which incur a heavy computational burden for both part detection and classification. From this perspective, a method that follows an object part-aware strategy to provide interpretable prediction criteria at minimal computational effort but deals with large numbers of parts is desirable. In this scenario, independently training a large convolutional neural network (CNN) for each part and then combining them in a unified framework is impractical [@zhang2014part].
Here we address the fine-grained categorization problem not only in terms of accuracy and efficiency when performing subordinate-level object recognition but also with regard to the interpretable characteristics of the resulting model. We do this by learning a new part-based CNN for FGVC that models multiple object parts in a unified framework with high efficiency. Similar to previous fine-grained recognition approaches, the proposed method consists of a localization module to detect object parts (“where pathway”) and a classification module to classify fine-grained categories at the subordinate level (“what pathway”). In particular, our key point localization network structure is composed of a sub-network used in contemporary classification networks (AlexNet [@krizhevsky2012imagenet] and BN-GoogleNet [@ioffe2015batch]) and a 1x1 convolutional layer followed by a softmax layer to predict evidence of part locations. The inferred part locations are then fed into the classification network, in which a two-stream architecture is proposed to analyze images at both the object level (global information) and part level (local information). Multiple parts are then computed via a shared feature extraction route, separated directly on feature maps using a part cropping layer, concatenated, and then fed into a shallower network for object classification. Except for categorical predictions, our method also generates interpretable classification instructions based on object parts. Since the proposed deeper network architecture-based framework employs a sharing strategy that stacks the computation of multiple parts, we call it *Deeper Part-Stacked CNN* (DPS-CNN).
{width="0.8\linewidth"}
This paper makes the following contributions:
1. DPS-CNN is the first efficient framework that not only achieves state-of-the-art performance on Caltech-UCSD Birds-200-2011 but also allows interpretation;
2. We explore a new paradigm for key point localization, which has exceed state of the art performance on Birds-200-2011 dataset;
3. Our classification network follows a two-stream structure that captures both object level (global) and part level (local) information, in which a new share-and-divide strategy is presented to compute multiple object parts. As a result, the proposed architecture is very efficient with a capacity of $32$ frames/sec [^2] without sacrificing the fine-grained categorization accuracy. Also, we propose a new strategy called scale mean-max (SMM) for feature fusion learning.
This paper is not a direct extension of our previous work [@huang2016part] and several other state-of-the-art fine-grained classification models [@zhang2015fine; @zhangspda2016; @zhang2016picking; @lin2015bilinear] but a significant development regarding the following aspects: Different to [@zhang2015fine] who adapts FCN for part localization, we propose a new paradigm for key point localization that first samples a small number of representable pixels and then determine their labels via a convolutional layer followed by a softmax layer; We also propose a new network architecture and enrich the methodology used in [@huang2016part]; Further, we introduce a simple but effective part feature encoding (named Scale Average Max) method in contrast to Bilinear in [@lin2015bilinear], Spatially Weighted Fisher Vector in [@zhang2016picking], and Part-based Fully Connected in [@zhang2016picking].
The remainder of the paper is organized as follows. Related works are summarized in Section \[sec:relatedwork\], and the proposed architecture including the localization and classification networks is described in Section \[sec:dpscnn\]. Detailed performance studies and analysis are presented in Section \[sec:exp\], and in Section \[sec:conclusion\] we conclude and propose various applications of the proposed DPS-CNN architecture.
Related Work {#sec:relatedwork}
============
**Keypoint Localization**. Subordinate categories generally share a fixed number of semantic components defined as ’parts’ or ’key points’ but with subtle differences in these components. Intuitively, when distinguishing between two subordinate categories, the widely accepted approach is to align components containing these fine differences. Therefore, localizing parts or key points plays a crucial role in fine-grained recognition, as demonstrated in recent works [@berg2013poof; @zhang2014panda; @maji2014part; @zhang2014part; @gkioxari2015actions; @zhu2015deepm].
Seminal works in this area have relied on prior knowledge about the global shape [@milborrow2008locating; @cootes2001active; @matthews2004active; @saragih2009face]. For example, the active shape model (ASM) uses a mixture of Gaussian distributions to model the shape. Although these techniques provide an effective way to locate facial landmarks, they cannot usually handle a wide range of differences such as those seen in bird species recognition. The other group of methods [@branson2014bird; @liu2013bird; @liu2014part; @zhang2014part; @shih2015part; @lin2015deep; @zhangspda2016; @yu2016deep] trains a set of key point detectors to model local appearance and then uses a spatial model to capture their dependencies and has become more popular in recent years. Among them, the part localization method proposed in [@shih2015part; @lin2015deep; @zhangspda2016] is most similar to ours. In [@shih2015part], a convolutional sub-network is used to predict the bounding box coordinates without using a region candidate. Although its performance is acceptable because the network is learned by jointly optimizing the part regression, classification, and alignment, all parts of the model need to be trained separately. To tackle this problem, [@lin2015deep] and [@zhangspda2016] adopt the similar pipeline of Fast R-CNN[@girshick2015fast], in which part region candidates are generated to learn the part detector. In this work, we discard the common proposal-generating process and regard all receptive field centers [^3] of a certain intermediate layer as potential candidate key points. This strategy results in a highly efficient localization network, since we take advantage of the natural properties of CNNs to avoid the process of proposal generation.
{width="0.8\linewidth"}
Our work is also inspired by and inherited from fully convolutional networks (FCNs) [@long2015fully], which produces dense predictions with convolutional networks. However, our network structure is best regarded as a fast and effective approach to predict sparse pixels since we only need to determine the class labels of the centers of the receptive fields of interest. Thus, FCN is more suited to segmentation, while our framework is designed for sparse key point detection. As FCN aims to predict intermediate feature maps then upsample them to match the input image size for pixel-wise prediction. Recent works [@zhang2015fine; @wei2016convolutional] borrow this idea directly for key point localization. During training, both of these works resize the ground truths to the size of the output feature maps and then use them to supervise the network learning, while, during testing, the predicted feature maps are resized to match the input size to generate the final key point prediction. However, these methods cannot guarantee accurate position prediction due to the upsampling process.
**Fine-Grained Visual Categorization**.A number of methods have been developed to classify object categories at the subordinate level. The best performing methods have gained performance improvements by exploiting the following three aspects: more discriminative features (including deep CNNs) for better visual representation [@bo2010kernel; @sanchez2011fisher; @krizhevsky2012imagenet; @szegedy2014going; @simonyan2014very]; explicit alignment approaches to eliminate pose displacements [@branson2014bird; @gavves2015local]; and part-based methods to examine the impact of object parts [@berg2013poof; @zhang2014panda; @maji2014part; @zhang2014part; @gkioxari2015actions; @zhu2015deepm]. Another approach has been used to explore human-in-the-loop methods [@branson2010visual; @deng2013fine; @wah2014similarity] to identify the most discriminative regions for classifying fine-grained categories. Although such methods provide direct and important information about how humans perform fine-grained recognition, they are not scalable due to the need for human interactions during testing. Of these, part-based methods are thought to be most relevant to fine-grained recognition, in which the subtle differences between fine-grained categories mostly relate to the unique object part properties.
Some part-based methods [@berg2013poof; @zhang2014part] employ strong annotations including bounding boxes, part landmarks, or attributes from existing fine-grained recognition datasets [@wah2011caltech; @parkhi2012cats; @maji2013fine; @vedaldi2014understanding]. While strong supervision significantly boosts performance, the expensive human labelling process motivates the use of weakly-supervised fine-grained recognition without manually labeled part annotations, , discovering object parts in an unsupervised fashion [@simon2015neural; @krause2015fine; @lin2015bilinear]. Current state-of-the-art methods for fine-grained recognition include [@zhang2015fine] and [@lin2015bilinear], which both employ deep feature encoding method, while DPS-CNN is largely inherited from [@zhang2014part], who first detected the location of two object parts and then trained an individual CNN based on the unique properties of each part. Compared to part-based R-CNN, the proposed method is far more efficient for both detection and classification. As a result, we can use many more object parts than [@zhang2014part], while still maintaining speed during testing.
Lin [@lin2015bilinear], argued that manually defined parts were sub-optimal for object recognition and thus proposed a bilinear model consisting of two streams whose roles were interchangeable as detectors or features. Although this design exploited the data-driven approach to possibly improve classification performance, it also made the resulting model difficult to interpret. In contrast, our method attempts to balance the need for classification accuracy and model interpretability in fine-grained recognition systems.
Deeper Part-Stacked CNN {#sec:dpscnn}
=======================
A key motivation of our proposed method is to produce a fine-grained recognition system that not only considers recognition accuracy but also addresses efficiency and interpretability. To ensure that the resulting model is interpretable, we employ strong part-level annotations with the potential to provide human-understandable classification criteria. We also adapt several strategies such as sparse prediction instead of dense prediction to eliminate part proposal generation and to share computation for all part features. For the sake of classification accuracy, we learn a comprehensive representation by incorporating both global (object-level) and local (part-level) features. Based on these, in this section we present the model architecture of the proposed Deeper Part-Stacked CNN (DPS-CNN).
According to the common framework for fine-grained recognition, the proposed architecture is decomposed into a localization network (Section \[subsec:localization\]) and a classification network (Section \[subsec:classification\]). In our previous work [@huang2016part], we adopted CaffeNet [@jia2014caffe], a slightly modified version of the standard seven-layer AlexNet architecture [@krizhevsky2012imagenet], as the basic network structure. In this paper, we use a deeper but more powerful network (BN-GoogleNet) [@ioffe2015batch] as a substitute. A unique feature of our architecture is that the message transferring operation from the localization network to the classification network, which uses the detected part locations to perform part-based classification, is conducted directly on the *Inception-4a* output feature maps within the data forwarding process. This is a significant departure from the standard two-stage pipeline of part-based R-CNN, which consecutively localizes object parts and then trains part-specific CNNs on the detected regions. Based on this design, sharing schemes are performed to make the proposed DPS-CNN fairly efficient for both learning and inference. Figure \[fig:architecture\] illustrates the overall network architecture.
Localization Network {#subsec:localization}
--------------------
The first stage in our proposed architecture is a localization network that aims to detect the location of object parts. We employ the simplest form of part landmark annotation, where a 2D key point is annotated at the center of each object part. Assume that $M$ - the number of object parts labeled in the dataset is sufficiently large to offer a complete set of object parts in which fine-grained categories are usually different. A naive approach to predicting these key points is to directly apply FCN architecture [@long2015fully] for dense pixel-wise prediction. However, this method usually biases the learned predictor because, in this task and unlike semantic segmentation, the number of key point annotations is extremely small compared to the number of irrelevant pixels.
{width="\linewidth"}
Motivated by the recent progress in object detection [@ren2015faster] and semantic segmentation [@long2015fully], we propose to use the centers of receptive fields as key point candidates and use a fully convolutional network to perform sparse pixel prediction to locate the key points of object parts (see Figure \[fig:kp\](b)). In the field of object detection, box candidates expected to be likely objects are first extracted using proposal-generating methods such as selective search [@uijlings2013selective] and region proposal networks [@ren2015faster]. Then, CNN features are learned to represent these box candidates and finally used to determine their class label. We adapt this pipeline to key point localization but omit the candidate generation process and simply treat the centers of receptive fields corresponding to a certain layer as candidate points. As shown in Figure \[fig:kp\](a), the advantage of using this method is that each candidate point can be represented by a $1D$ cross-channel feature vector in the output feature maps. Also, in our candidate point evaluation experiments in Table \[tab:recall\], we find that given an input image of size $448$x$448$ and using the receptive fields of the *inception-4a* layer in BN-GoogleNet generates $28$x$28$ candidate points and $100$% recall at PCK@$0.1$.\
**Fully convolutional network.** An FCN is achieved by replacing the parameter-rich fully connected layers in standard CNN architectures constructed by convolutional layers with kernels of spatial size $1\times1$. Given an input RGB image, the output of an FCN is a feature map of reduced dimension compared to the input. The computation of each unit in the *feature map* only corresponds to pixels inside a region of fixed size in the input image, which is called its *feature map*. We prefer FCNs because of the following reasons: (1) feature maps generated by FCNs can be directly utilized as the part locating results in the classification network, as detailed in Section \[subsec:classification\]; (2) the results of multiple object parts can be obtained simultaneously; (3) FCNs are very efficient for both learning and inference.\
**Learning.** We model the part localization process as a multi-class classification problem on sparse output spatial positions. Specifically, suppose the output of the last FCN convolutional layer is of size $h\times w\times d$, where $h$ and $w$ are spatial dimensions and $d$ is the number of channels. We set $d=M+1$. Here, $M$ is the number of object parts and $1$ denotes an additional channel to model the background. To generate corresponding ground-truth labels in the form of feature maps, units indexed by $h \times w$ spatial positions are labeled with their nearest object part; units that are not close to any of the labeled parts (with an overlap with respect to a receptive field) are labeled as background. In this way, ground-truth part annotations are transformed into the form of corresponding feature maps, while in recent works that directly apply FCNs [@zhang2015fine; @wei2016convolutional], the supervision information is generated by directly resizing the part ground-truth image.
Another practical problem here is determining the model depth and the input image size for training the FCN. Generally, layers at later stages carry more discriminative power and, therefore, are more likely to generate good localization results; however, their receptive fields are also much larger than those of previous layers. For example, the receptive field of the *inception-4a* layer in BN-GoogleNet has a size of $107\times107$ compared to the $224\times224$ input image, which is too large to model an object part. We propose a simple trick to deal with this problem, namely upsampling the input images so that the fixed size receptive fields denoting object parts become relatively smaller compared to the whole object, while still using later stage layers to guarantee discriminative power. In the proposed architecture, the input image is upsampled to double the resolution and the *inception-4a* layer is adopted to guarantee discrimination.
The localization network is illustrated in Figure \[fig:fcn\]. The input images are warped and resized into a fixed size of $448\times448$. All layers from the beginning to the *inception-4a* layer are cut from the BN-GoogleNet architecture, so the output size of the *inception-4a* layer is $28\times28\times576$. Then, we further introduce an $1\times1$ convolutional layer with $M+1$ outputs termed *conv* for classification. By adopting a location-preserving softmax that normalizes predictions at each spatial location of the feature map, the final loss function is a sum of softmax loss at all $28\times28$ positions:
$$\label{eqn:fcnloss}
L = -\sum_{h=1}^{28}\sum_{w=1}^{28} \log\sigma(h,w,\hat{c}),$$
where $$\sigma(h,w,\hat{c}) = \frac{\exp(f_{conv}(h,w,\hat{c}))}{\sum_{c=0}^M \exp(f_{conv}(h,w,c))}.$$ Here, $\hat{c}\in[0,1,...,M]$ is the part label of the patch at location $(h,w)$, where the label $0$ denotes background. $f_{conv}(h,w,c)$ stands for the output of *conv* layer at spatial position $(h,w)$ and channel $c$.\
**Inference.** Inference starts from the output of the learned FCN, , $(M+1)$ part-specific heat maps of size $28\times28$, in which we introduce a Gaussian kernel $\mathcal{G}$ to remove isolated noise in the feature maps. The final output of the localization network are $M$ locations in the $28\times28$ *conv* feature map, each of which is computed as the location with the maximum response for one object part.
Meanwhile, considering that object parts may be missing in some images due to varied poses and occlusion, we set a threshold $\mu$ that if the maximum response of a part is below $\mu$, we simply discard this part’s channel in the classification network for this image. Let $g(h,w,c)=\sigma(h,w,c)*\mathcal{G}$, the inferred part locations are given as:
$$(h_c^*,w_c^*)=
\left\{
\begin{array}{ll}
\operatorname*{argmax}_{h,w} g(h,w,c) & \text{if } g(h_c^*,w_c^*,c)>\mu, \\
(-1,-1) & \text{otherwise.}
\end{array}
\right.$$
Classification network {#subsec:classification}
----------------------
The second stage of the proposed DPS-CNN is a classification network with the inferred part locations given as an input. As shown in Figure \[fig:architecture\], it follows a two-stream architecture with a *Part Stream* and a *Object Stream* to capture semantics from different angles. The outputs of both two streams are fed into a feature fusion layer followed by a fully connected layer and a softmax layer.\
**Part stream.** The part stream is the core of the proposed DPS-CNN architecture. To capture object part-dependent differences between fine-grained categories, one can train a set of part CNNs, each one of which conducts classification on a part separately, as proposed by Zhang [@zhang2014part]. Although such method works well for situations employing two object parts [@zhang2014part], we argue that this approach is not applicable when the number of object parts is much larger, as in our case, because of the high time and space complexities.
We introduce two strategies to improve part stream efficiency, the first being model parameter sharing. Specifically, model parameters of layers before the part crop layer and *inception-4e* are shared among all object parts and can be regarded as a generic part-level feature extractor. This strategy reduces the number of parameters in the proposed architecture and thus reduces the risk of overfitting. We also introduce a *part crop* layer as a computational sharing strategy. The layer ensures that the feature extraction procedure of all parts only requires one pass through the convolutional layers.
After performing the shared feature extraction procedure, the computation of each object part is then partitioned through a *part crop* layer to model part-specific classification cues. As shown in Figure \[fig:architecture\], the input for the part crop layer is a set of feature maps (the output of *inception-4a* layer in our architecture) and the predicted part locations from the previous localization network, which also reside in *inception-4a* feature maps. For each part, the part crop layer extracts a local neighborhood centered on the detected part location. Features outside the cropped region are simply discarded. In practice, we crop $l\times h$ neighborhood regions from the $28\times 28$ *inception-4a* feature maps. The cropped size of feature regions may have an impact on recognition performance, because larger crops will result in redundancy when extracting multiple part features, while smaller crops cannot guarantee rich information. For simplicity, we use $l=h=7$ in this paper to ensure that the resulting receptive field is large enough to cover the entire part.\
**Object stream.** The object stream captures object-level semantics for fine-grained recognition. It follows the general architecture of BN-GoogleNet, in which the input of the network is a $448\times 448$ RGB image and the output of *incenption-5b* layer are $14\times 14$ feature maps. Therefore, we use $14\times 14$ average pooling instead of $7 \times 7$ in original setting.
The design of the two-stream architecture in DPS-CNN is analogous to the famous Deformable Part-based Models [@felzenszwalb2010object], in which object-level features are captured through a root filter in a coarser scale, while detailed part-level information is modeled by several part filters at a finer scale. We find it critical to measure visual cues from multiple semantic levels in an object recognition algorithm.
We conduct the standard gradient descent to train the classification network. It should be noted, however, that the gradient of each element $\frac{\partial E}{\partial X_{i,j}}$ in *inception-4a* feature maps is calculated by the following equation: $$\frac{\partial E}{\partial X_{i,j}} = \sum_{c=1}^M \phi (\frac{\partial E}{\partial X^c_{i,j}}),$$ where $E$ is the loss function, $X^c_{i,j}$ is the feature maps cropped by part $c$ and $$\phi (\frac{\partial E}{\partial X^c_{i,j}})=
\begin{cases}
\frac{\partial E}{\partial X^c_{i,j}} &\mbox{$X_{i,j}$ corresponding to $X^c_{i,j}$}, \\
0 &\mbox{otherwise}.
\end{cases}$$
Specifically, the gradient of each cropped part feature map (in $7\times 7$ spatial resolution) is projected back to the original size of *inception-4a* ($28\times28$ feature maps) according to the respective part location and then summed. The computation of all other layers simply follows the standard gradient rules. Note that the proposed DPS-CNN is implemented as a two stage framework, after training the FCN, weights of the localization network are fixed when training the classification network.\
**Feature Fusion**
The commonest method [@lin2015deep; @zhang2014part] for combining all part-level and object-level features is to simply concatenate all these feature vectors as illustrated in Figure \[fig:fusion\](a). However, this approach may cause feature redundancy and also suffer from high-dimensionality when part numbers become large. To effectively utilize all part- and object-level features, we present three options for learning fusion features: scale sum (SS), scale max (SM), and scale mean-max (SMM), as illustrated in Figure \[fig:fusion\](a), Figure \[fig:fusion\](b), and Figure \[fig:fusion\](d), respectively. All three methods include the shared process of placing a scale layer on top of each branch. Nevertheless, as indicated by their names, the scale sum feature is the element-wise sum of all output branches, the scale max feature is generated by an element-wise maximum operation, while the scale average-max feature is the concatenation of element-wise mean and max features. In our previous work [@huang2016part] based on the standard CaffeNet architecture, each branch from the part stream and the object stream was connected with an independent *fc6* layer to encourage diversity features, and the final fusion feature was the sum of all the outputs of these *fc6* layers. As this fusion process requires $M+1$ times model parameters more than the original *fc6* layer in CaffeNet and consequently incurs a huge memory cost, a $1\times1$ convolutional layer is used for dimensionality reduction. Here we redesign this component for simplicity and to improve performance. First, a shared inception module is placed on top of the cropped part region to generate higher level features. Also, a scale layer follows each branch feature to encourage diversity between parts. Furthermore, the scale layer has fewer parameters than the fully connected layer and, therefore, reduces the risk of overfitting and decreases the model storage requirements.
[17.5cm]{}[cccccccccccccccccc]{}
$\alpha$ & Ba & Bk & Be & Br & Cr & Fh & Le & Ll & Lw & Na & Re & Rl & Rw & Ta & Th & Avg\
$0.05$ & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100 & 100\
$0.02$ & 90.8 & 89.8 & 90.8 & 90.4 & 90.9 & 91.4 & 90.4 & 90.4 & 90.0 & 90.7 & 90.3 & 89.9 & 90.3 & 90.5 & 90.3 & 90.5\
$0.01$ & 26.8 & 26.3 & 9.1 & 11.2 & 5.2 & 4.1 & 40.4 & 9.4 & 10.8 & 14.6 & 9.9 & 11.9 & 9.6 & 11.2 & 22.3 & 13.2\
Experiments {#sec:exp}
===========
In this section we present experimental results and a thorough analysis of the proposed method. Specifically, we evaluate the performance from four different aspects: localization accuracy, classification accuracy, inference efficiency, and model interpretation.
{width="\linewidth"}
Dataset and implementation details
----------------------------------
Experiments are conducted on the widely used fine-grained classification benchmark the Caltech-UCSD Birds dataset (CUB-200-2011) [@wah2011caltech]. The dataset contains $200$ bird categories with roughly $30$ training images per category. In the training phase we adopt strong supervision available in the dataset, we employ 2D key point part annotations of altogether $M=15$ object parts together with image-level labels and object bounding boxes.
The labeled parts[^4] imply places where people usually focus on when being asked to classify fine-grained categories; thus they provide valuable information for generating human-understandable systems.
The proposed Deeper Part-Stacked CNN architecture is implemented using the open-source package Caffe [@jia2014caffe]. Specifically, input images are warped to a fixed size of $512\times512$, randomly cropped into $448\times448$, and then fed into the localization network and the part stream in the classification network as input. We employ a pooling layer with kernel $7\times7$ to guarantee synchronization between the two streams in the classification network.
Candidate keypoints
-------------------
For the key point localization task, we follow the proposal-based object detection method pipeline; centers of receptive fields corresponding to a certain layer are first regarded as candidate points and then forwarded to a fully convolutional network for further classification. Similar to object detection using proposals, whether selected candidate points have a good coverage of pixels of interest in the test image plays a crucial role in key point localization, since missed key points cannot be recovered in subsequent classification. Thus, we first evaluate the candidate point sampling method. The evaluation is based on the PCK metric [@yang2013articulated], in which the error tolerance is normalized with respect to the input image size. For consistency with evaluation of key point localization, a ground truth point is recalled if there exists a candidate point matched in terms of the PCK metric. Table \[tab:recall\] shows the localization recall of candidate points selected by *inception-4a* with different $\alpha$ values $0.05$, $0.02$ and $0.01$. As expected, candidate points sampled by layer *inception-4a* have a great coverage of ground truth using PCK metric with $\alpha > 0.02$. However, the recall drop dramatically when using $\alpha = 0.01$. This mainly because of the large stride(16) in *inception-4a* layer, which results in the distance between two closest candidate points is 16 pixels, while setting a input size of $448$ with $\alpha = 0.01$ requires the candidate point should be close to the ground truth within $4.48$ pixels.
{width="1\linewidth"}
Localization Results
--------------------
Following [@long2014convnets], we consider a key point to be correctly predicted if the prediction lies within a Euclidean distance of $\alpha$ times the maximum of the input width and height compared to the ground truth. Localization results are reported on multiple values of $\alpha\in \{0.1, 0.05, 0.02\}$ in the analysis below. The value $\alpha$ in the PCK metric is introduced to measure the error tolerance in key point localization. To investigate the effect of the selected layer for key point localization, we perform experiments using the *inception-4a*,*inception-4b*,*inception-4c* and *inception-4d* layers as part detector layers. As shown in Table \[tab:loc\], a higher layer with a larger receptive field tends to achieve better localization performance than a lower layer with $\alpha=0.1$. This is mainly because the larger receptive fields are crucial for capturing spatial relationships between parts and improve performance (see Table \[tab:rf\]). However, in contrast, for $\alpha = 0.05$ or $0.02$, the performance decreases at deeper layers. One possible explanation is that although higher layers obtain better semantic information about the object, they lose more detailed spatial information. To evaluate the effectiveness of our key point localization approach, we also compare it with recent published works [@huang2016part; @zhang2015fine; @yu2016deep] providing PCK evaluation results on CUB-200-2011 along with experimental results using a more reasonable evaluation metric called average precision of key points (APK), which correctly penalizes both missed and false-positive detections [@yang2013articulated]. As can be seen from the Table \[tab:loc\], our method outperforms existing techniques with various $\alpha$ setting in terms of PCK. In addition, the most striking result is that our approach outperforms the compared methods with large margins when using small $\alpha$ value.
The part localization architecture adopted in DPS-CNN achieves a highest average *[email protected]* $88.5\%$ on the CUB-200-2011 test set for $15$ object parts. Specifically, the employed Gaussian smoothing kernel delivers $2\%$ improvements over methods that use standard convolutional layers in BN-GoogleNet.
[cc]{} Layer & Rec. Field\
Inception-4a & $107\times107$\
Inception-4b & $139\times139$\
Inception-4c & $171\times171$\
Inception-4d & $204\times204$\
Another interesting phenomenon of note is that parts residing near the birds’ heads tend to be located more accurately. It turns out that a bird’s head has a relatively stable structure with fewer deformations and a lower probability of occlusion. In contrast, parts that are highly deformable such as the wings and legs get lower PCK values. Figure \[fig:loc\] shows typical localization results using the proposed method.
[17.5cm]{}[cccccccccccccccccc|c]{}
$\alpha$ & Methods & Ba & Bk & Be & Br & Cr & Fh & Le & Ll & Lw & Na & Re & Rl & Rw & Ta & Th & Avg & APK\
& [@huang2016part] & 80.7 & 89.4 & 79.4 & 79.9 & 89.4 & 88.5 & 85.0 & 75.0 & 67.0 & 85.7 & 86.1 & **77.5** & 67.8 & 76.0 & 90.8 & 81.2 & 86.6\
& [@zhang2015fine] & 85.6 & 94.9 & 81.9 & 84.5 & 94.8 & **96.0** & **95.7** & 64.6 & 67.8 & 90.7 & **93.8** & 64.9 & 69.3 & 74.7 & 94.5 & 83.6 & -\
& [@yu2016deep] & **94.0** & 82.5 & **92.2** & 93.0 & 92.2 & 91.5 & 93.3 & 69.7 & 68.1 & 86.0 & **93.8** & 74.2 & 68.9 & 77.4 & 93.4 & 84.7 & -\
& Ours(4a) & 82.7 & 94.1 & 85.3 & 87.8 & 95.2 & 93.3 & 88.6 & 75.5 & 75.9 & 92.0 & 89.5 & 76.6 & 75.9 & 67.4 & 94.7 & 84.9 & 89.1\
& Ours(4b) & 87.4 & 93.6 & 87.4 & 88.9 & 95.2 & 93.7 & 88.3 & 73.3 & 77.6 & 93.4 & 88.9 & 76.3 & 79.0 & 70.5 & 94.5 & 85.9 & 88.9\
& Ours(4c) & 89.0 & **95.1** & 91.5 & 92.6 & **95.7** & 94.7 & 90.3 & **78.5** & **82.3** & 94.4 & 91.0 & 73.2 & 81.9 & 78.4 & 95.7 & 88.3 & 90.9\
& Ours(4d) & 89.0 & 95.0 & 92.2 & **93.2** & 95.2 & 94.2 & 90.5 & 73.2 & 81.5 & **94.4** & 91.6 & 75.5 & **82.3** & **83.2** & **95.8** & **88.5** & **91.2**\
& [@huang2016part] & 48.8 & 63.7 & 44.5 & 50.3 & 50.2 & 43.7 & 80.0 & 44.8 & 42.7 & 60.1 & 59.4 & 46.5 & 39.8 & 46.8 & 71.9 & 52.9 & 62.7\
& [@zhang2015fine] & 46.8 & 62.5 & 40.7 & 45.1 & 59.8 & 63.7 & 66.3 & 33.7 & 31.7 & 54.3 & 63.8 & 36.2 & 33.3 & 39.6 & 56.9 & 49.0 & -\
& [@yu2016deep] & 66.4 & 49.2 & 56.4 & 60.4 & 61.0 & 60.0 & 66.9 & 32.3 & 35.8 & 53.1 & 66.3 & 35.0 & 37.1 & 40.9 & 65.9 & 52.4 & -\
& Ours(4a) & **70.6** & **89.5** & **69.5** & **75.0** & **89.0** & **87.8** & **87.1** & **58.5** & **57.6** & **84.6** & **87.8** & **59.6** & **60.2** & **56.3** & **90.0** & **74.9** & **80.4**\
& Ours(4b) & 69.2 & 79.4 & 69.0 & 74.5 & 73.2 & 72.3 & 85.7 & 53.3 & 58.3 & 83.7 & 86.0 & 55.5 & 60.1 & 59.0 & 86.5 & 74.5 & 71.1\
& Ours(4c) & 62.3 & 57.1 & 67.6 & 72.2 & 49.1 & 47.0 & 84.6 & 49.7 & 57.6 & 79.3 & 84.9 & 44.1 & 56.9 & 63.7 & 82.6 & 63.0 & 67.9\
& Ours(4d) & 42.3 & 27.5 & 59.7 & 60.6 & 21.3 & 23.3 & 82.2 & 33.1 & 49.6 & 65.6 & 82.4 & 37.4 & 47.5 & 66.7 & 69.4 & 51.3 & 54.5\
& [@huang2016part] & 11.1 & 16.9 & 9.1 & 11.2 & 5.2 & 4.1 & 40.4 & 9.4 & 10.8 & 14.6 & 9.9 & 11.9 & 9.6 & 11.2 & 22.3 & 13.2 & 13.3\
& [@zhang2015fine] & 9.4 & 12.7 & 8.2 & 12.2 & 13.2 & 11.3 & 7.8 & 6.7 & 11.5 & 12.5 & 7.3 & 6.2 & 8.2 & 11.8 & 56.9 & 13.1 & -\
& [@yu2016deep] & 18.8 & 12.8 & 14.2 & 15.9 & 15.9 & 16.2 & 20.3 & 7.1 & 8.3 & 13.8 & 19.7 & 7.8 & 9.6 & 9.6 & 18.3 & 13.8 & -\
& Ours(4a) & **24.9** & **31.0** & **23.0** & **28.3** & **25.1** & **26.6** & **44.8** & **19.6** & **17.4** & **38.4** & **46.9** & **20.9** & **20.7** & **22.0** & **37.5** & **28.5** & **17.2**\
& Ours(4b) & 19.7 & 15.8 & 21.6 & 24.0 & 9.1 & 8.1 & 40.7 & 16.0 & 16.8 & 32.6 & 43.1 & 16.7 & 17.7 & 23.6 & 29.8 & 22.4 & 13.5\
& Ours(4c) & 12.5 & 5.9 & 17.9 & 17.9 & 2.6 & 3.0 & 41.4 & 12.0 & 15.0 & 22.2 & 41.4 & 8.9 & 14.9 & 24.0 & 23.1 & 17.5 & 11.8\
& Ours(4d) & 6.4 & 1.9 & 14.1 & 11.8 & 1.0 & 2.1 & 36.7 & 4.9 & 10.9 & 15.5 & 38.5 & 5.9 & 10.4 & 24.0 & 17.0 & 13.4 & 9.3\
[17.5cm]{}[cccccccccccccccccc]{}
Part & Ba & Bk & Be & Br & Cr & Fh & Le & Ll & Lw & Na & Re & Rl & Rw & Ta & Th\
Accuracy(%)&47.9 & 63.7 & 43.9 & 56.8 & 66.8 & 66.1 & 36.6 & 30.8 & 30.4 & 64.8 & 36.1 & 29.2 & 29.7 & 20.0 & 68.7\
Classification results
----------------------
We begin our classification analysis by studying the discriminative power of each object part. We select one object part each time as the input and discard the computation of all other parts. As shown in Table \[tab:single\], different parts produce significantly different classification results. The most discriminative part “*Throat*” achieves a quite impressive accuracy of $68.7\%$, while the lowest accuracy is $20.0\%$ for the part “*Tail*”. Therefore, to improve classification, it may be beneficial to find a rational combination or order of object parts instead of directly running the experiment on all parts altogether. More interestingly, when comparing the results between Table \[tab:loc\] and Table \[tab:single\] it can be seen that parts located more accurately such as *Throat*, *Nape*, *Forehead* and *Beak* tend to achieve better performance in the recognition task, while some parts like *Tail* and *Left Leg* with poor localization accuracy perform worse. This observation may support the hypothesis that a more discriminative part is easier to locate in the context of fine-grained categorization and vice versa.
To evaluate our framework’s overall performance, we first train a baseline model with accuracy $81.56\%$ using a BN-Inception architecture [@ioffe2015batch] with pre-training on ImageNet [@russakovsky2015imagenet]. By stacking certain part features and applying our proposed fusion method, our framework improves the performance to $85.12\%$. Also, to evaluate our proposed feature fusion method, we then train four DPS-CNN models with same experimental settings (maximum iteration and learning rate) but using different feature fusion methods. The results shown in Table \[tab:setting\] (Rows 2-5) demonstrate that SMM fusion achieves the best performance and outperforms the FC method by $1.69\%$.
To investigate which parts should be selected in our learning framework, we conduct the following experiments by employing two guiding principles: one concerns the feature discrimination and the other feature diversity. Here we consider parts with higher accuracy in Table \[tab:single\] are more discriminative, and combination of parts with distant location are more diverse. We firstly select top 6 parts with the highest accuracy from Table \[tab:single\] by only applying the discriminative principle, then choose 3,5,9 and 15 parts respectively by taking two principles into account. Experimental results are shown in Table \[tab:setting\] (Row 6-10), we observe that increasing part numbers generally bring slight improvement. However, all setting perform better than that with 6 most discriminative parts. This mainly because most of these parts are adjacent to each other so that it fails to produce diverse feature in our framework. Also, it should be noticed that using all parts feature does not guarantee the best performance, on the other hand, results in pool accuracy. This finding shows that the feature redundancy caused by appending exorbitant number of parts in learning, may degrade the accuracy, and suggests that an appropriate strategy for integrating multiple parts is critical.
[clc]{} Row & Setting & Acc(%)\
1 & Object Only(Baseline) & 81.56\
2 & 5-parts + FC & 81.86\
3 & 5-parts + SS & 83.06\
4 & 5-parts + SM & 83.41\
5 & 5-parts + SMM & **83.55**\
6 & 6-parts + SMM & 84.12\
7 & 3-parts + SMM & 84.29\
8 & 5-parts + SMM & 84.91\
**9** & **9 parts + SMM** & **85.12**\
10 & 15-parts + SMM & 84.45\
------------------------------------------------- ------ ------- ------ ------- -------------- -------- -----------
.
BBox Parts BBox Parts
Part-Stacked CNN [@huang2016part] AlexNet 20 76.62
Deep LAC [@lin2015deep] AlexNet - 80.26
Part R-CNN [@zhang2014part] AlexNet - 76.37
SPDA-CNN [@zhangspda2016] VGG16 - 84.55
SPDA-CNN [@zhangspda2016]+ensemble VGG16 - 85.14
Part R-CNN [@zhang2014part] without BBox AlexNet - 73.89
PoseNorm CNN [@branson2014bird] AlexNet - 75.70
Bilinear-CNN (M+D+BBox) [@lin2015bilinear] VGG16+VGGM 8 85.10
Bilinear-CNN (M+D) [@lin2015bilinear] VGG16+VGGM 8 84.10
Constellation-CNN [@simon2015neural] VGG19 - 84.10
Spatial Transformer CNN [@jaderberg2015spatial] Inception+BN - 84.10
Two-Level [@xiao2015application] VGG16 - 77.90
Co-Segmentation [@krause2015fine] VGG19 - 82.80
DPS-CNN with 9 parts Inception+BN **32** 85.12
**DPS-CNN ensemble with 4 models** Inception+BN 8 **86.56**
------------------------------------------------- ------ ------- ------ ------- -------------- -------- -----------
We also present the performance comparison between DPS-CNN and existing fine-grained recognition methods. As can be seen in Table \[tab:cls\], our approach using only keypoint annotation during training achieve 85.12% accuracy which is comparable with the state-of-the-art method [@lin2015bilinear] that achieves 85.10% using bounding box both in training and testing. Moreover, it is interpretable and faster - the entire forward pass of DPS-CNN runs at $32$ (NVIDIA TitanX), while B-CNN\[D,M\][@lin2015bilinear] runs at $8$ (NVIDIA K40)[^5]. In particular, our method is much faster than proposal based methods such as [@zhang2014part] and [@zhangspda2016] which require multiple network forward propagation for proposal evaluation, while part detection and feature extraction are accomplished efficiently by running one forward pass in our approach. In addition, we combine four models stemmed from integrating different parts(listed in Table \[tab:setting\] (Row 7-10)) to form an ensemble which leads to 86.56% accuracy on cub-200-2011.
To understand what features are learned in DPS-CNN, we use the aforementioned five-parts model and show its feature map visualization compared with that from BN-Inception model fine-tuning on cub-200-2011. Specifically, we pick the top six scoring feature maps of *Inception-4a* layer for visualization, where the score is the sum over each feature map. As shown in Figure \[fig:fm\], each example image from test set is followed by three rows of feature maps, from top row to bottom, which are selected from part stream, object stream and BN-inception base-line network respectively. Interestingly, by comparison, our part stream have learned feature maps that appear to be more intuitive than those learned by the other two methods. Specifically, it yields more focused and cleaner patterns which tend to be highly activated by the network. Moreover, we can observe that object stream and baseline network are more likely to activate filters with extremely high frequency details but at the expense of extra noise, while part stream tends to obtain a mixture of low and mid frequency information. The red dashed box in Figure \[fig:fm\] indicates a failure example, in which both our part stream and object stream fails to learn useful feature. This may be caused by our part localization network fails to locate *Crown* and *Left Leg* parts because the branch in this image looks similar to bird legs and another occluded bird also has an effect on locating the *Crown* part.
{width="1\linewidth"}
Model interpretation
--------------------
One of the most prominent features of DPS-CNN method is that it can produce human-understandable interpretation manuals for fine-grained recognition. Here we directly borrow the idea from [@huang2016part] for interpretation using the proposed method.
Different from [@berg2013poof] who directly conducted one-on-one classification on object parts, the interpretation process of the proposed method is conducted relatively indirectly. Since using each object part alone does not produce convincing classification results, we perform the interpretation analysis on a combination of bounding box supervision and each single object part. The analysis is performed in two ways: a “one-versus-rest” comparison to denote the most discriminative part to classify a subcategory from all other classes, and a “one-versus-one” comparison to obtain the classification criteria of a subcategory with its most similar classes.
- The *“one-versus-rest”* manual for an object category $k$. For every part $p$, we compute the summation of prediction scores of the category’s positive samples. The most discriminative part is then captured as the one with the largest accumulated score: $$p_k^* = \operatorname*{argmax}_p \sum_{i, y_i=k}S^{(p)}_{ip}.$$
- The *“one-versus-one”* manual obtained by computing as the part which results in the largest difference of prediction scores on two categories $k$ and $l$. We first take the respective two rows in the score matrix $S$, and re-normalize it using the binary classification criterion as $S'$. Afterwards, the most discriminative part is given as: $$p_{k\rightarrow l}^* = \operatorname*{argmax}_p (\sum_{i, y_i=k}S'^{(p)}_{ip}+\sum_{j, y_j=l}S'^{(p)}_{jp})$$
The model interpretation routine is demonstrated in Figure \[fig:interpret\]. When a test image is presented, the proposed method first conducts object classification using the DPS-CNN architecture. The predicted category is presented as a set of images in the dataset that are closest to the test image according to the feature vector of each part. Except for the classification results, the proposed method also presents classification criteria that distinguish the predicted category from its most similar neighboring classes based on object parts. Again we use part features but after part cropping to retrieve nearest neighbor part patches of the input test image. The procedure described above provides an intuitive visual guide for distinguishing fine-grained categories.
Conclusion {#sec:conclusion}
==========
In this paper, we propose a novel fine-grained recognition method called Deeper Part-Stacked CNN (DPS-CNN). The method exploits detailed part-level supervision, in which object parts are first located by a localization network and then by a two-stream classification network that explicitly captures object- and part-level information. We also present a new feature vector fusion strategy that effectively combines both part and object stream features. Experiments on CUB-200-2011 demonstrate the effectiveness and efficiency of our system. We also present human-understandable interpretations of the proposed method, which can be used as a visual field guide for studying fine-grained categorization.
DPS-CNN can be applied to fine-grained visual categorization with strong supervision and can be easily generalized to various applications including:
1. Discarding the requirement for strong supervision. Instead of introducing manually labeled part annotations to generate human-understandable visual guides, one can also exploit unsupervised part discovery methods [@krause2015fine] to define object parts automatically, which requires far less human labelling effort.
2. Attribute learning. The application of DPS-CNN is not restricted to FGVC. For instance, online shopping [@kiapour2015where] performance could benefit from clothing attribute analysis from local parts provided by DPS-CNN.
3. Context-based CNN. The role of local “parts” in DPS-CNN is interchangeable with global contexts, especially for objects that are small and have no obvious object parts such as volleyballs or tennis balls.
[^1]: Manuscript received
[^2]: For reference, a single CaffeNet runs at $82$ frames/sec under the same experimental setting.
[^3]: Here the receptive field means the area of the input image, to which a location in a higher layer feature map correspond.
[^4]: The $15$ object parts are back, beak, belly, breast, crown, forehead, left eye, left leg, left wing, nape, right eye, right leg, right wing, tail, and throat.
[^5]: note that the computational power of TitanX is around 1.5 times of that of K40).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We report an observation of a stable soliton-like structure on the surface of a ferrofluid, generated by a local perturbation in the hysteretic regime of the Rosensweig instability. Unlike other pattern-forming systems with localized 2D structures, magnetic fluids are characterized by energy conservation; hence their mechanism of soliton stabilization is different from the previously discussed gain/loss balance mechanism. The radioscopic measurements of the soliton’s surface profile suggest that locking on the underlying periodic structure is instrumental in its stabilization.'
author:
- 'Reinhard Richter$^1$, and I.V. Barashenkov$^2$'
title: 'Two-dimensional solitons on the surface of magnetic fluids'
---
To date, stable solitary waves have been experimentally observed in a variety of one-dimensional and quasi-one-dimensional physical systems. In 2D, dispersive nonlinear systems are prone to collapse instabilities and hence the 2D solitons turned out to be more elusive. (Here we use the term “soliton" in a broad physical sense, as a synonym of localized structure.) So far, the list of experimentally detectable 2D localized objects was confined mostly to vortices in superfluids, superconductors, and other media on one hand, and dissipative solitons in nonequilibrium systems on the other. While the stability of the former is due to their nontrivial topology, the latter come into being via the balance of strong dissipation and energy gain. Examples include current filaments in gas discharge systems [@filaments]; oscillons in fluids and granular materials [@oscillons]; breathing spots in chemical reactions [@chemical] and feedback and cavity solitons in optics [@optics]. Despite some encouraging theoretical insights, the question of whether 2D non topological solitons can arise in [*conservative*]{} systems has remained open.
In this Letter we report an experimental observation of a strongly localized, stable stationary soliton on the surface of magnetic fluid (MF) in a stationary magnetic field. MF is a dispersion of magnetic nanoparticles, and thus has a high relative permeability $\mu_{\rm r}$[@rosensweig1985]. This is a lossless system; a horizontal layer of MF in a vertically applied magnetic induction ${\boldmath B}$ is characterized by the energy density [@gailitis1977; @friedrichs2001]: $$\begin{aligned}
{\cal F}[h(x,y)]& = & \frac{\rho g}{2} h^2(x,y) -\int_0^{h} dz {\boldmath B}\frac{\mu_{\rm r}-1}{2}{\boldmath H_{\rm MF}(x,y,z)} \nonumber\\
& &+\sigma \sqrt{1+(\partial_x h(x,y) )^2+(\partial_y h(x,y) )^2}.
\label{functional}\end{aligned}$$ Here $\rho$ and $\sigma$ are the density and surface tension of the MF, $h(x,y)$ the local height of the liquid layer, and $\boldmath H_{\rm MF}(x,y,z)$ is the magnetic field in the presence of the MF. The three terms in Eq.(\[functional\]) represent the hydrostatic, magnetic and surface energy, respectively. As the surface profile deviates from the flat reference state, the first and last term grow whereas the magnetic energy decreases. For sufficiently large ${\boldmath B}$, this gives rise to the normal field, or Rosensweig, instability [@cowley1967; @rosensweig1985].
Our experimental setup is sketched in Fig.\[setup\]. A Teflon$^{\circledR}$ vessel with the radius $R=60\,\rm mm$ and depth of 3 mm [@reimann2003] is filled with MF up to the brim and placed on the common axis midway between two Helmholtz coils. An x-ray tube is mounted above the center of the vessel at a distance of 1606mm. The radiation transmitted through the fluid layer and the bottom of the vessel is recorded by an x-ray sensitive photodiode array detector (16 bit) connected to a computer. The full surface relief is then reconstructed from the calibrated radioscopic images. For details see [@richter2001]. The experiments were performed with the magnetic fluid EMG901, Lot F121901 AX from Ferrotec. Its material parameters have been measured to $\mu_{\rm r}=3.2$, $\rho=1.406 \rm\,g\,cm^{-3}$, and $\sigma=25 \pm 0.7 \rm\,mN/m$.
![Sketch of the experimental setup. See text for details. []{data-label="setup"}](fig1){width="75mm"}
Starting from a flat layer at $B=0 \rm mT$, we increase the external induction adiabatically in steps of 15 $\rm \mu T$, pausing for 30s after each increase. As shown in Fig.\[half\_profiles\](a), a deformation of the surface of the liquid appears first at the edge of the vessel. This is due to the magnetic field gradient induced by the discontinuous magnetization at the edge of the liquid layer. Increasing the induction further gives rise to a fully developed pattern of the Rosensweig instability, as shown in (b,c).
![Surface reliefs as reconstructed from the radioscopic images for a) $B=$ 8.922 mT, and b),c) 10.407 mT. Each color indicates a layer thickness of 1 mm.[]{data-label="half_profiles"}](fig2){width="75mm"}
![The amplitude of the pattern for $r<11 \rm mm$ versus the magnetic induction. The crosses (dots) mark the values for increasing (decreasing) induction respectively. The solid (dashed) lines display the least square fit to the roots , of the amplitude equation $ \epsilon A + \gamma (1+\epsilon) A^2 - g A^3=0$ of Ref. [@friedrichs2001] with $\gamma=0.281 $ and $g=0.062$. The full circles (squares) give the amplitude of the localized spike initiated at $B=8.91 \rm mT$ for increasing (decreasing) induction respectively. []{data-label="h_vs_B"}](fig3){width="86mm"}
We measured the top-to-bottom height $A$ of the stationary fluid pattern arising in the adiabatic increase and decrease of $B$. To avoid the edge-induced imperfections in the character of the bifurcation, we only consider spikes located within 11 mm from the center of the dish. Figure\[h\_vs\_B\] displays results obtained for 400 values of $B$. As $B$ is increased, a sudden transition to the upper branch occurs at $B_c=9.025 \rm mT$. For $B>B_c$, the entire surface is covered by a lattice of liquid spikes, which is hexagonal away from the boundary. Decreasing $B$, the order parameter $A$ remains on the upper branch all the way to $B^*=8.076 \rm mT$ where it drops to the flat reference level. Thus the diagram shows a subcritical bifurcation to hexagons. The solid and dashed lines display a fit to the roots of the corresponding amplitude equation of Ref.[@friedrichs2001].
![A single soliton surrounded by the unperturbed magnetic liquid. The magnetic induction generated by the Helmholtz coils amounts to $B=8.91 \rm mT$. The amplitude of the local pulse which produced the soliton was $B_+= 0.68 \rm mT$ at the bottom of the vessel. []{data-label="LS_1"}](fig4){width="75mm"}
To study the stability of the flat surface to local perturbations (in the hysteretic regime), a small air coil with the inner diameter of 8 mm was placed under the center of the vessel (see Fig.\[setup\]). This allows to increase, locally, the magnetic induction. A local pulse of $B_+=0.68 \mbox{\rm mT}$ added to the uniform field of $B=8.91 \rm mT$, produces a single stationary spike of fluid, surrounded by a circular dip, which does not disperse after $B_+$ has been turned off. Figure\[LS\_1\] presents a measured relief of this radially-symmetric state which will be referred to as the soliton. The soliton is a stable nondecaying structure; it remained intact for days. After its formation at the center of the dish, the soliton was often seen to float around (with $v \sim 0.1 \rm mm/s$), until reaching an equilibrium position somewhere near the edge of the dish. This behavior can be attributed to radial gradients of the magnetic field due to the discontinuous magnetization at the edge and the ring of spikes pinned along the perimeter of the dish.
We examined the range of stability of the soliton generated by a pulse with $B_+=0.68 \mbox{ \rm mT}$ added to the uniform induction $B=8.91 \rm mT$. Reducing $B$ adiabatically we measured the corresponding amplitude of the soliton (marked by full squares in Fig.\[h\_vs\_B\]). Similarly to the spikes in the hexagonal pattern, the height of the soliton decreases as $B$ is reduced. The soliton decays for $B<8.09 $mT, which is close to $B^*= 8.076 $mT, the lower stability boundary of the hexagonal pattern. As $B$ is increased, the amplitude of the soliton grows, as indicated in Fig.\[h\_vs\_B\] by full circles. At $B=9.055$mT, a sudden transition from the soliton to the fully developed Rosensweig pattern occurs. This value is somewhat larger than $B_c=9.025 \rm \,mT$; this is due to the fact that the birth of the soliton produces a slight drop in the flat layer thickness, which shifts $B_c$ - similarly to Fig.5 in Ref.[@friedrichs2001].
![ The filled squares mark the profile of one period of the hexagonal pattern, measured at $B= 9.07\rm mT$ in the center of the vessel; $r<8.8$mm. Azimuthally averaged height profiles of two different solitons, measured at the same induction are depicted by open symbols (one) and a dashed line (the other). []{data-label="height_profile"}](fig5){width="65mm"}
In order to illustrate the robustness of the soliton’s shape we show in Fig.\[height\_profile\] the azimuthally averaged profiles of two different solitons, produced in two separate experiments at $B=9.07 \mbox{\rm mT }$. The profiles are practically indistinguishable. Also plotted are two half-periods of the corresponding hexagonal lattice. In agreement with Fig.\[h\_vs\_B\], the soliton is about $1 \mbox{mm}$ taller than the spikes of the lattice. This may be attributed to the fact that the spikes emerge simultaneously, and thus have to share the liquid available. However, the width of the soliton is exactly equal to the period of the lattice. Therefore, there is a preferred wavelength in the system, defined by the lattice, to which the soliton locks. As we show now, this width locking is the central part of the soliton’s stabilization mechanism.
Consider the dispersion relation of a semi-infinite layer of inviscid MF [@cowley1967], $$\omega^2 = gk - \mu_0 \mu_r
\frac{(\mu_r - 1)^2}{\mu_r +1} \frac{1}{\rho}H^2k^2 + \frac{
\sigma}{\rho} k^3 .
\label{dispersion}$$ Here $\omega$ is the frequency, $k=|\vec k|$ the wavenumber, $H$ the strength of the external magnetic field, and $\mu_0$ the magnetic field constant. While the first and the third terms account for the gravity and capillary effects and are common for all fluids, the second term is specific just for the MF. As $H$ is increased above $H_c$, where $ H_c^2=\frac{2\,(\mu_r+1)}{\mu(\mu_r-1)^2 }\ (\rho g \sigma)^{1/2}$, a band of wavenumbers with $\omega^2<0$ appears around $k_c=(\rho g / \sigma)^{1/2}$ and the flat state loses its stability to the hexagonal pattern. Since the soliton should decay to the flat surface as ${\vec x}^2 \to \infty$, there can be no stable solitons for $H>H_c$.
Let now $H<H_c$. On the qualitative level, our system can be modelled by a conservative analogue of the Swift-Hohenberg equation: $${\ddot u} +(k_0^2+\nabla^2)^2 u+au =3bu^2 -2cu^3; \quad b,c>0.
\label{SHE}$$ Eq.(\[SHE\]) can be used as a model since (a) it has a nonmonotonic dispersion relation $\omega^2=a+k_0^4-2k_0^2k^2+k^4$, where, as in (2), the destabilizing $k^2$-term is opposed by a higher power of $k$; and (b) it has a symmetry-breaking hysteretic nonlinearity which was shown to provide a fairly accurate approximation of the amplitude of the hexagonal pattern [@friedrichs2001] (see Fig.\[h\_vs\_B\].) We have verified, numerically, that Eq.(\[SHE\]) does indeed have a stable stationary radially-symmetric soliton solution coexisting with stable hexagons in a broad parameter range. Its stability can be explained by a Derrick-type argument for the corresponding energy functional, $$E= \int [
{\dot u}^2 +V(u)-2k_0^2 (\nabla u)^2+(\nabla^2 u)^2
] d^2x.
\label{energy}$$ Here $V(u)=(a+k_0^4)u^2-2bu^3+cu^4$. A scaling perturbation $u({\vec x}) \to u(\kappa {\vec x})$ takes the stationary energy to $$E(\kappa^2)= \frac{1}{\kappa^2} \int V d^2x -
2k_0^2 \int (\nabla u)^2 d^2x
+ \kappa^2 \int (\nabla^2 u)^2 d^2x.$$ The first term (nonlinearity) opposes the dispersive broadening of the soliton (for which $\kappa \to 0$) while the last one prevents the nonlinear blow-up (for which $\kappa \to \infty$). In a similar way, the first and last term in (\[dispersion\]) make contributions to the energy which scale as $\kappa^{-1}$ and $\kappa$, respectively. The first term (along with the nonlinearity) opposes the spreading and the last one arrests the blow-up. Next, setting the derivative $(dE/d\kappa^2)_{\kappa=1}$ to zero, gives $\int V d^2x=\int (\nabla^2 u)^2 d^2x$. Using this relation, the second derivative, $ d^2 E/d (\kappa^2)^2$, is calculated to be $2\int (\nabla^2 u)^2 d^2x>0$, which means that the soliton renders the energy minimum. If Eq.(\[SHE\]) did not include the higher-derivative term, the energy would not have a nontrivial minimum. The introduction of a higher derivative (or, equivalently, the nonmonotonicity of the dispersion curve) sets a preferential wavelength in the system — to which the soliton locks and stabilizes. A similar stabilization mechanism was discussed before in the context of the wave front locking [@pomeau1986]; see also [@sakaguchi1997].
![Nine solitons at $B=8.91 \rm mT$.[]{data-label="LS_9"}](fig6){width="65mm"}
Applying, repeatedly, pulses of $B_+$ and allowing the newly born solitons to drift away from the site of the probe coil, we were able to generate two, three, and more solitons. Figure \[LS\_9\] presents an example of a 9-soliton configuration, with only one remaining at the center. In this way, it is possible to increase the surface energy of the liquid layer in steps. This is illustrated in Fig.\[E\_surface\] which also shows the surface energy of the Rosensweig pattern as a hysteretic function of $B$. Thus, one can reach the region between the two branches of this function which is not accessible for the standard Rosensweig instability.
![The surface energy of the liquid layer for increasing (open squares) and decreasing (circles) magnetic induction. The full circles mark the increase of $E_{\rm s}$ through the successive generation of solitons at $B=8.91 \mbox{\rm mT}$ (see also inset). To reduce the influence of the perimeter spikes, only the area $r<0.88\,R$ was covered, where $R$ is the radius of the vessel. []{data-label="E_surface"}](fig7){width="65mm"}
Could solitons serve as building blocks in the formation of periodic patterns? We have observed that, if additional care is taken to suppress the edge-induced inhomogeneity of the magnetization, solitons can form molecule-like clusters (Fig.\[molecules\].) This may seem to contradict the repulsive nature of the dipole-dipole interaction; however, there is a simple mechanism that can account for the binding. Indeed, each soliton is surrounded by concentric dips representing ring-like regions of depleted magnetic induction. The innermost, deepest, ring is clearly visible in Figs.\[LS\_1\],\[height\_profile\], and \[LS\_9\]; a higher-resolution measurement allows to discern another, shallower dip of larger radius. The dips create a potential relief which may capture the partner soliton(s). As the density of solitons grows, the multisoliton cluster evolves towards the hexagonal Rosensweig lattice. It still remains to be understood whether the cluster-lattice transition requires additional excitation energy.
![The solitons can form molecule-like clusters: (a) di-, (b) tri-, (c) tetra-, (d) penta-, (e) hepta-, and (f) oligomers. The height is indicated by switching between black and white after each mm. Here $B= 8.71\rm mT$; each panel covers the area of $(87\rm mm)^2$. []{data-label="molecules"}](fig8){width="86mm"}
In conclusion, we found stable 2D solitons on the surface of a ferrofluid in the hysteretic regime of the Rosensweig instability. These objects are easy to generate and control and they are easily set in motion; this opens ways for studying their binding and scattering. Due to the [*conservative nature*]{} of the ferrosolitons, and unlike the localized structures observed previously in dissipative systems, the balance of dissipation and energy gain plays no role in their stabilization. Instead, we suggest a stabilization mechanism which appeals to the locking of the soliton to the wavelength imposed by the nonmonotonic dispersion relation. This mechanism can also be at work in other conservative systems with preferred wavelengths, e.g. in electrostatics and elasticity [@taylor1965].
We thank I.Rehberg, W.Pesch and R.Friedrich for discussions and K.Staliunas for providing a code for Eq.(\[SHE\]). Support by NRF of South Africa, grant 205723, and Deutsche Forschungsgemeinschaft, grants HBFG051-201 and Ri 1054/1-4 is gratefully acknowledged.
[10]{}
Yu.A. Astrov and Yu.A. Logvin, Phys. Rev. Lett. [**79**]{}, 2983 (1997); I. Müller, E. Ammelt, and H.-G. Purwins, [*ibid.*]{}[**82**]{}, 3428 (1999)
P.B. Umbanhowar, F. Melo, and H.L. Swinney, Nature (London) [**382**]{}, 793 (1996); O. Lioubashevski, H. Arbell, and J. Fineberg, Phys. Rev. Lett. [**76**]{}, 3959 (1996); O. Lioubashevski [*et al*]{}, [*ibid.*]{} [**83**]{}, 3190 (1999); H. Arbell and J. Fineberg, [*ibid.*]{} [**85**]{}, 756 (2000); D. Astruc and S. Fauve, in: Fluid Mechanics and Its Applications, vol.62, pp.39-46 (Kluwer, 2001)
D. Haim [*et al*]{}, Phys. Rev. Lett. [**77**]{}, 190 (1996); V.K. Vanag and I.R. Epstein, [*ibid.*]{} [**92**]{}, 128301 (2004) V. B. Taranenko, K. Staliunas, and C. O. Weiss, Phys. Rev. A 56, 1582 (1997); B. Schäpers [*et al*]{}, Phys. Rev. Lett. [**85**]{}, 748 (2000); S. Barland [*et al*]{}, Nature 419, 699 (2002)
R. E. Rosensweig, [*Ferrohydrodynamics*]{} (Cambridge University Press, Cambridge, 1985).
A. Gailitis, J. Fluid Mech. [**82**]{}, 401 (1977).
R. Friedrichs and A. Engel, Phys. Rev. E [**64**]{}, 021406 (2001).
M. D. Cowley and R. E. Rosensweig, J. Fluid Mech. [**30**]{}, 671 (1967).
B. Reimann, R. Richter, I. Rehberg, and A. Lange, Phys. Rev. E [**68**]{}, 036220 (2003).
R. Richter and J. Bläsing, Rev. Sci. Instrum. [**72**]{}, 1729 (2001).
Y. Pomeau, Physica D [**23**]{}, 3 (1986).
H. Sakaguchi and H. R. Brand, Europhysics Lett. [**38**]{}, 341 (1997).
G. I. Taylor and A. D. McEwan, J. Fluid Mech. [**22**]{}, 1 (1965); G. J. Stroebel and W.H. Warner, J. Elast. [**3**]{}, 185 (1973).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: '3C 396 is a composite supernova remnant (SNR), consisting of a central pulsar wind nebula (PWN) and a bright shell in the west, which is known to be interacting with molecular clouds (MCs). We present a study of X-ray emission from the shell and the PWN of the SNR 3C 396 using archival [*Suzaku*]{} data. The spectrum of the SNR shell is clearly thermal, without a signature of a non-thermal component. The abundances of Al and Ca from the shell are slightly enhanced, which indicates the presence of metal-enriched supernova ejecta. The PWN spectra are well described by a power-law model with a photon index of $\sim$1.97 and a thermal component with an electron temperature of $\sim$0.93 keV. The analysis of about 11-years of [*Fermi*]{} data revealed an 18 sigma-detection of gamma-ray emission from the location overlapping with the position of 3C 396 / 4FGL J1903.8+0531. The spectrum of 3C 396 / 4FGL J1903.8+0531 is best-fitted with a log-parabola function with parameters of $\alpha$ = 2.66 and $\beta$ = 0.16 in the energy range of 0.2$-$300 GeV. The luminosity of 3C 396 / 4FGL J1903.8+0531 was found to be $>$10$^{35}$ erg s$^{-1}$ at 6.2 kpc, which rules out the inverse Compton emission model. Possible scenarios of gamma-ray emission are hadronic emission and bremsstrahlung processes, due to the fact that the SNR is expanding into dense MCs in the western and northern regions of the SNR.'
author:
- |
A. Sezer,$^{1}$[^1] T. Ergin$^{2}$[^2], N. Cesur$^{3}$[^3], S.J. Tanaka$^{4}$[^4],\
\
\
$^{1}$Department of Electrical-Electronics Engineering, Avrasya University, 61250, Trabzon, Turkey\
$^{2}$TUBITAK Space Technologies Research Institute, ODTU Campus, 06800, Ankara, Turkey\
$^{3}$Department of Astrophysics/IMAPP, Radboud University Nijmegen, 6500 GL Nijmegen, the Netherlands\
$^{4}$Department of Physics and Mathematics, Aoyama Gakuin University, 5-10-1 Fuchinobe, Sagamihara 252-5258, Japan\
$^{5}$Frontier Research Institute for Interdisciplinary Sciences, Tohoku University, Miyagi, 980-8578, Japan\
$^{6}$Department of Earth and Planetary Science, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan\
$^{7}$Institute of Laser Engineering, Osaka University, 2-6 Yamadaoka, Suita, Osaka 565-0871, Japan\
title: '[*Suzaku*]{} and [*Fermi*]{} view of the supernova remnant 3C 396'
---
\[firstpage\]
ISM: individual objects: 3C 396 (G39.2$-$0.3), 4FGL J1903.8+0531 $-$ ISM: supernova remnants $-$ X-rays: ISM $-$ gamma-rays: ISM.
Introduction
============
Composite supernova remnants (SNRs) are identified by the SNR shell and a pulsar wind nebula (PWN) (like MSH 11$-$62 and G327.1$-$1.1, see e.g. @Sl12 [@Te15]). Observations of composite SNRs provide information about the ejecta, evolution of the nebulae and structure of the surrounding circumstellar medium (CSM)/interstellar medium (ISM) (for review see @Sl17).
The gamma-ray emission in composite SNRs may originate from the interaction between the reverse shock and the PWN (e.g. @Sl12). The Large Area Telescope detector on board [*Fermi*]{} Gamma-Ray Space Telescope ([*Fermi*]{}-LAT) revealed gamma-rays emission from composite SNRs such as CTA 1 [@Ab08] and MSH 15$-$56 [@Te13].
3C 396 (also known as G39.2$-$0.3, HC24 or NRAO 593) is one of the known composite SNRs, which was first classified as a Crab-like SNR in the radio band [@Ca75; @Ca82]. Using [*ASCA*]{} data, @Ha99 presented the results of the X-ray study of 3C 396 and showed the unambiguous composite nature. They extracted the X-ray spectra from a circular region using a radius of about 4 arcmin that covers the entire SNR. Their analysis reveals a spectrum best described by a combination of thermal and non-thermal models. They reported that the non-thermal emission is from a central region, providing evidence for the presence of a PWN. The spectral index of the power-law component is $\Gamma$ $\sim$ 2.53 and the associated X-ray luminosity is $L_{\rm X}$ $\sim$ 2.28$\times10^{35} D_{10}^{2}$ erg s$^{-1}$ (0.2$-$4.0 keV), where $D_{10}$ is the distance to the SNR in units of 10 kpc. The thermal component arising from the interaction of the blast wave with the ISM has an electron temperature $kT_{\rm e}$ $\sim$ 0.62 keV. They found high absorption, $N_{\rm H}$ $\sim$ 4.65$\times10^{22}$ cm$^{-2}$. Their results implied a remnant age of $\sim$7000 yr in its Taylor-Sedov phase of evolution and an X-ray-emitting mass of $M_{\rm X}$ in the range of 40 and 300 $M_{\sun}$.
@Ol03 analysed the 100 ks [*Chandra*]{} observation of 3C 396 and detected a point source at the centre of the nebula ($\rmn{RA}(J2000)=19^{\rmn{h}} 04^{\rmn{m}} 04\fs7$, $\rmn{Dec.}~(J2000)=05\degr 27\arcmin 12\arcsec$). They also detected three extensions in the east and west regions. The spectrum of the entire central nebula was characterized by a single absorbed power-law model. The best-fitting model from the 0.5$-$7.5 keV spectrum yielded an absorption column density of $N_{\rm H}$ $\sim$ 5.3$\times10^{22}$ cm$^{-2}$ and a photon index of $\Gamma$ $\sim$ 1.5. They found no thermal emission features in their analysis. From the X-ray spectral analysis of the point source, they also derived a photon index of $\Gamma$ $\sim$ 1.2. Using the Very Small Array (VSA) telescope, the possible presence of anomalous microwave emission due to spinning dust in an SNR was first suggested by @Sc07, who reported an anomalously high emission at 33 GHz. @Le09 detected the \[Fe[ii]{}\] 1.64 $\mu$m and H$_2$ 2.12 $\mu$m filaments with the near-infrared (IR) \[Fe[ii]{}\] and H$_2$ Êline imaging and spectroscopic data of the SNR. These filaments are close to each other in the western rim of 3C 396.
@Su11 investigated the molecular environment and re-analysed [*Chandra*]{} ACIS observation of 3C 396. Using CO millimeter observations, they found that the western boundary of the remnant is perfectly confined by the western molecular wall at the local standard of rest (LSR) velocity of $V_{\rm LSR}$ $\sim$ $84$ km s$^{-1}$ and the multiwavelength properties are consistent with the presence of the $84$ km s$^{-1}$ molecular clouds (MCs). Along its western edge, the SNR has a broadened molecular emission, and also shows presence of OH maser at the same velocity. @Su11 also extracted [*Chandra*]{} X-ray spectra from five X-ray bright regions excluding the PWN and the point source. They used an absorbed non-equilibrium ionization (NEI) model, found Si and S for all regions, but Ar and Ca lines are prominent only in the northern and southern regions, where the absorption column density, $N_{\rm H}$, in the range of $\sim$ $(4.0-6.5)\times10^{22}$ cm$^{-2}$ and the electron temperature, $kT_{\rm e}$, in the range of 0.66$-$1.33 keV. They estimated a progenitor mass of 13$-$15 $M_{\sun}$ and the SNR age to be $\sim$3 kyr. In the study of @Ki16, the presence of the velocity-broadened $^{12}$CO $J=2-1$ emission centred around $+69$ km s$^{-1}$ to the north of 3C 396 was detected and it was shown to be extending to the west, where the broadened emission is centred around $+72$ km s$^{-1}$. Recently, @No18 searched for the iron K-shell line in some SNRs using the [*Suzaku*]{} data. They found that the flux ratio of Fe[i]{} K$\alpha$/Fe[xxv]{} He$\alpha$ in 3C 396 is consistent with the Galactic ridge X-ray emission (GRXE) within 1$\sigma$ errors.
From the H[i]{} absorption, @Ca75 derived a lower limit for the distance of $\sim$7.7 kpc to 3C 396. The X-ray H[i]{} column density and CO associations implied a distance of 6.2$-$8 kpc [@Ol03; @He09; @Su11]. @Su11 determined a distance of $\sim$6.2 kpc, with the LSR velocity of $\sim$84 km s$^{-1}$. We assume the distance to 3C 396 to be 6.2 kpc throughout the paper[^5].
3C 396 was searched in TeV and GeV gamma-ray bands by HEGRA and H.E.S.S. and by [*Fermi*]{}-LAT, respectively. In the 1st [*Fermi*]{}-LAT Supernova Remnant Catalog [@Ac16], the upper limits on the flux of this SNR were reported on Table 3, i.e. among the not-detected SNRs. A preliminary analysis of [*Fermi*]{}-LAT data by @Er16 revealed an excess of GeV gamma rays from 3C 396. Recently, the Fourth [*Fermi*]{}-LAT sources (4FGL; [@Fe19]) catalog contained a new gamma-ray source, 4FGL J1903.8+0531, which was detected at R.A.(J2000) = 19$^h$ 03$^m$ 54$^s\!$.0 and decl.(J2000) = $+$05$^{\circ}$ 31$'$ 17$''\!$.0, about 0$^{\circ}\!\!$.08 away from 3C 396, with a significance of 10$\sigma$. HEGRA and H.E.S.S. did not detect 3C 396 in TeV gamma-ray energies [@Bo11; @Ah01; @Ah05; @Ab18].
In this work, we analyse the high spectral resolution data from [*Suzaku*]{} to identify the nature and the spectral properties of the X-ray emission from 3C 396. We also analyse [*Fermi*]{}-LAT data to examine a gamma-ray emission in the GeV energy range. The paper is organized as follows. In Section \[obs\], we summarize X-ray and gamma-ray observations and data reduction. In Section \[analysis\], we describe how we estimated background components, explain our spectral analysis procedure and give results of the X-ray and gamma-ray analysis. In Section \[discussion\], we discuss the nature of X-ray and gamma-ray emission and the spectral properties of the SNR shell and the interior of the remnant. Finally, we draw our conclusions in Section \[conclusions\].
Observations and Data Reduction {#obs}
===============================
X-ray data
----------
3C 396 was observed with [*Suzaku*]{} X-ray Imaging Spectrometer (XIS; @Ko07) on 2014 April 26 (ObsID: 509038010; PI: T. Pannuti). The net exposure of the cleaned event data was $\sim$72 ks. We used the data obtained with front-side illuminated (FI) CCD chips (XIS0 and XIS3) and back-side illuminated (BI) chip (XIS1).
The X-ray data reduction and analysis were performed with [heasoft]{}[^6] version 6.20 and [xspec]{} version 12.9.1 [@Ar96] with atomic data base ([atomdb]{}) version 3.0.9[^7] [@Sm01; @Fo12]. We analysed cleaned event data that had already been preprocessed with the [*Suzaku*]{} team. To analyse the XIS data, we generated redistribution matrix files with ftool [xisrmfgen]{} and ancillary response files with [xissimarfgen]{} [@Is07]. The spectra were grouped to include at least 25 counts in each bin.
{width="52.00000%"}
Gamma-ray data reduction
------------------------
The gamma-ray observations were taken from 2008-08-04 to 2019-10-07. In this analysis, we made use of the analysis packages `fermitools`[^8] version `1.0.1` and `fermipy`[^9] version `0.17.4`. Using `gtselect` of `fermitools` we selected [*Fermi*]{}-LAT Pass 8 ‘source’ class and ‘front$+$back’ type events coming from zenith angles smaller than 90$^{\circ}$ and from a circular region of interest (ROI) with a radius of 20$^{\circ}$ centered at the SNR’s radio location[^10]. The [*Fermi*]{}-LAT instrument response function version *P8R3$_{-}$SOURCE$_{-}\!\!$V2* was used. For mapping the morphology and searching for new sources within the analysis region, events having energies in the range of 1$-$300 GeV were selected. To deduce the spectral parameters of the investigated sources, events with energies between 200 MeV and 300 GeV were chosen.
Analysis and Results {#analysis}
====================
{width="49.00000%"} {width="49.00000%"}
Analysis of [*Suzaku*]{} data
-----------------------------
### XIS image
Fig. \[Fig1\] displays an XIS3 image of 3C 396 in the 0.3$-$10.0 keV energy range, where the spectral analysis regions are: (i) The outermost circular region defining the SNR shell extraction region, excluding the PWN and the point source region; (ii) The eastern and western shells of the SNR; (iii) The $3.0$ arcmin-diameter circle defining the PWN region, excluding the innermost circle contaminated by emission from the point source.
Component Parameters Shell PWN
----------- ------------------------------------- ------------------------- ------------------------
TBABS $N_{\rm H}$ ($10^{22}$ cm$^{-2})$ $5.83_{-0.24}^{+0.27}$ $5.22_{-0.17}^{+0.24}$
VVNEI $kT_{\rm e}$ (keV) $1.12_{-0.03}^{+0.02}$ $0.93_{-0.09}^{+0.08}$
Mg (solar) $1.2_{-0.2}^{+0.2}$ (1)
Al (solar) $1.9_{-0.3}^{+0.4}$ (1)
Si (solar) $1.3_{-0.2}^{+0.1}$ $1.5_{-0.2}^{+0.2}$
S (solar) $1.1_{-0.2}^{+0.2}$ $1.4_{-0.3}^{+0.1}$
Ca (solar) $1.7_{-0.2}^{+0.4}$ (1)
$\tau$ ($10^{11}$ cm$^{-3}$ s) $2.3_{-0.5}^{+0.4}$ $3.6_{-0.4}^{+0.4}$
Norm$\dagger$ ($10^{-3}$ cm$^{-5}$) $38.12_{-2.49}^{+4.08}$ $4.79_{-0.32}^{+0.65}$
Power-law $\Gamma$ $-$ $1.97_{-0.38}^{+0.23}$
Norm$\ddagger$ ($10^{-5}$) $-$ $1.64_{-0.45}^{+0.28}$
Reduced $\chi^{2}$ (dof) 1.12 (828) 1.03 (379)
[**Notes.**]{} Errors are within a 90 per cent confidence level. Abundances are given relative to the solar values of @Wi00. (1) indicates that the elemental abundance was fixed at solar.\
$\dagger$ The normalization of the VVNEI, norm=$10^{-14}$ $\int n_{\rm e} n_{\rm H} dV$/($4\pi d^{2}$), where $d$ is the distance to the source (in cm), $n_{\rm e}$, $n_{\rm H}$ are the electron and hydrogen densities (in units of cm$^{-3}$), respectively, and $V$ is the emitting volume (in units of cm$^{3}$).\
$\ddagger$ The normalization of the power-law is in units of photons cm$^{-2}$ s$^{-1}$ keV$^{-1}$ at 1 keV.
Background estimation
---------------------
For 3C 396, @Ol03 and @Su11 extracted the background spectra by selecting regions from the field of view (FoV) of the [*Chandra*]{} observation in their analysis. We extracted the background spectra from the entire source-free region of the same chip area, excluding the calibration regions and the SNR region. The instrumental (non-X-ray) background (NXB) spectra of XIS were generated using [xisnxbgen]{} [@Ta08]. The CXB and GRXE are considered as astrophysical X-ray background components. The background spectrum was fitted with the following model:
{width="99.00000%"}
$${\rm Abs1} \times ({\rm power}\mbox{-}{\rm law}) + {\rm Abs2} \times ({\rm apec} + {\rm apec}),$$
where Abs1 and Abs2 represent the ISM absorption for CXB and GRXE, respectively. The apec is a collisional ionisation equilibrium (CIE) plasma model in the <span style="font-variant:small-caps;">xspec</span>. The CXB component parameters are fixed at those in @Ku02, where power-law fit with a photon index of ${\Gamma}=1.41$ and a surface brightness of 5.41$\times$10$^{-15}$ erg s$^{-1}$ cm$^{-2}$ arcmin$^{-2}$ in the 2.0$-$10.0 keV energy band were adopted. The absorption column density is assumed to be $N_{\rm H (CXB)}$=2 $\times$ $N_{\rm H (GRXE)}$ [@Uc13]. The GRXE component is represented with two CIE models (apec+apec) as a low-temperature plasma ($kT_{\rm e}$ $\sim$ 1 keV) and a high-temperature plasma ($kT_{\rm e}$ $\sim$ 7 keV) [@Uc13]. We generated a background spectrum file using the <span style="font-variant:small-caps;">fakeit</span> command in <span style="font-variant:small-caps;">xspec</span> and used it for each source spectrum as background.
### The SNR Shell
To examine the X-ray spectral properties of the SNR shell, we extracted spectra of the entire SNR from a circular region with a radius of 4.5 arcmin, shown by the outermost circle in Fig. \[Fig1\]. Emissions from the PWN and the point source were eliminated from the spectra.
After subtracting the background, the SNR has no significant X-ray emission above $\sim$6 keV, which is consistent with the result of @No18. We employed an absorbed single-component variable-abundance NEI plasma model (VVNEI in [xspec]{}), where the free parameters were the normalization, the absorption column density ($N_{\rm H}$), the electron temperature ($kT_{\rm e}$), the ionization parameter ($\tau$=$n_{\rm e}t$) and the abundances of Mg, Al, Si, S and Ca. Other metal abundances were fixed to the solar values [@Wi00]. We found an electron temperature of $kT_{\rm e}$=$1.12_{-0.03}^{+0.02}$ keV, an ionization time-scale of $n_{\rm e}t$=$2.3_{-0.5}^{+0.4}\times10^{11}$ cm$^{-3}$ s, an absorption column density of $N_{\rm H}$=$5.83_{-0.24}^{+0.27}$ $\times10^{22}$ cm$^{-2}$, with evidence for slightly enhanced abundances of Al and Ca. The [*Suzaku*]{} XIS spectra of the SNR shell in the 1.0$-$6.0 keV are illustrated in Fig. \[Fig2\] (left panel). The best-fitting parameters are given in Table 1.
To better understand the emission characteristics of the SNR shell surrounding the PWN, we also extracted XIS spectra from two selected regions (as shown in Fig. \[Fig1\]), which correspond to the X-ray, IR and radio - bright western and faint eastern shells. For the western shell, we obtained an absorption column density of $N_{\rm H}$=$6.03_{-0.34}^{+0.47}$$\times10^{22}$ cm$^{-2}$, which is higher than that for the eastern region.
### The PWN
We extracted the PWN spectra from an X-ray bright region with a 1.5 arcmin radius circle excluding the point source ($r$=0.16 arcmin). We fitted the spectra of the PWN with an absorbed VVNEI and power-law model. The spectra are well described by a power-law model with $N_{\rm H}$=$5.22_{-0.17}^{+0.24}$ $\times10^{22}$ cm$^{-2}$ and $\Gamma$=$1.97_{-0.38}^{+0.23}$ and a thermal component with an electron temperature of $0.93_{-0.09}^{+0.08}$ keV. The best-fitting spectral parameters are summarized in Table 1 and the XIS spectra of the PWN in the 1.0$-$10.0 keV energy band are given in Fig. \[Fig2\] (right panel).
Gamma-ray analysis and results
------------------------------
The background model of the analysis region consists of diffuse background sources, diffuse Galactic emission (GE) and the isotropic component (ISO), and all the extended and point sources from 4FGL Catalog [@Fe19] located within a 10$^{\circ}$ $\times$ 10$^{\circ}$ region centred on the ROI centre. We freed all parameters of GE (*gll$_{-}$iem$_{-}$v7.fits*) and ISO (*iso$_{-}$P8R3$_{-}$SOURCE$_{-}\!\!$V2$_{-}\!$v1.txt*). The normalization parameters of all sources within 3$^{\circ}$ are set free. In addition, we freed all sources that have test statistic (TS[^11]) values greater than 400 and fixed all sources with TS values smaller than 400.
### Position and Source Morphology
In this analysis, we used the gamma rays in the energy range of 1$-$300 GeV. We removed 4FGL J1903.8+0531 from the gamma-ray background model and the resulting distribution of the gamma-ray emission presents a point-like source morphology as shown on the TS map given in Fig. \[Fig3\] left panel. The best-fit position of the excess gamma-ray emission was found to be R.A.(J2000) = 285$^{\circ}\!\!$.957 $\pm$ 0$^{\circ}\!\!$.027 and decl.(J2000) = 5$^{\circ}\!\!$.475 $\pm$ 0$^{\circ}\!\!$.021 using the `localize` method of the `fermipy` analysis package. This position is 0$^{\circ}\!\!$.08 away from the radio location of 3C 396 and only 0$^{\circ}\!\!$.05 away from 4FGL J1903.8+0531, where the error ellipse of the best-fit position and the positional error of 4FGL J1903.8+0531 are overlapping. Therefore, herewith we name the excess of gamma rays found in this analysis as 4FGL J1903.8+0531 and/or 3C 396. If we insert 4FGL J1903.8+0531 back into the gamma-ray background model using the best-fit position found in this analysis, we obtain the TS map given in Fig. \[Fig3\] right panel.
### Gamma-ray Spectrum
The spectral measurement was performed in the 200 MeV$-$300 GeV energy range, where the spectrum of 4FGL J1903.8+0531 was fit to the Log-parabola (LP) spectral model below: $$\mbox{dN/dE} = \mbox{N}_0 ~(\mbox{E}/\mbox{E}_b)^{{\rm-(\alpha + \beta \log(E/E_{b}))}},$$ where E$_b$ is a scale parameter. ${\rm \beta}$ and ${\rm \alpha}$ are spectral indices of the LP spectral model. N$_0$ is the normalization parameter.
The TS value of 4FGL J1903.8+0531 in this energy range was found to be higher (TS = 338) than what was reported in the 4FGL catalog [@Fe19]. The gamma-ray spectral data points and their statistical error bars are shown in Fig. \[Fig4\]. The fitted LP-type spectral model is drawn on top of the data points. The LP-spectral indices were found to be ${\rm \alpha}$ = 2.66 $\pm$ 0.09 and ${\rm \beta}$ = 0.16 $\pm$ 0.05. In the energy range of 0.2$-$300 GeV, the total flux and energy flux values were found to be (3.19 $\pm$ 0.26)$\times$10$^{-8}$ cm$^{-2}$ s$^{-1}$ and (1.89 $\pm$ 0.12)$\times$10$^{-5}$ MeV cm$^{-2}$ s$^{-1}$, respectively. Comparing these results with the spectral parameters derived for 4FGL J1903.8+0531 in the 4FGL catalog are ${\rm \alpha}$ = 2.72 $\pm$ 0.16, ${\rm \beta}$ = 0.26 $\pm$ 0.08, and energy flux = (1.94 $\pm$ 0.35)$\times$10$^{-5}$ MeV cm$^{-2}$ s$^{-1}$, we conclude that the parameters of both analyses are compatible with each other.
Discussion
==========
We present a study of emission from the SNR 3C 396 using the [*Suzaku*]{} and [*Fermi*]{}$-$LAT data. In this section, we briefly discuss the implication of our results.
Thermal X-ray emission
----------------------
The [*Suzaku*]{} X-ray spectra of the SNR shell are well described by a thermal plasma model with an electron temperature of $kT_{\rm e}$ $\sim$ 1.1 keV and an ionization time-scale of $\tau$ $\sim$ $2.3\times10^{11}$ cm$^{-3}$ s, indicating that the plasma is still ionizing. We clearly detected K-shell line of Al from the shell. The overabundant Al was found in several core-collapse SNRs (e.g. G350.1$-$0.3: @Ya14; Kes 17: @Wa16). Our abundance patterns are not consistent with the ones reported in @Su11. This may be due to the fact that we extracted spectra from a larger region in comparison to @Su11. The reason for this difference may be, because more recent atomic data is implemented in this work.
{width="70.00000%"}
The absorption column density, $N_{\rm H}$ $\sim$ 5.8$\times10^{22}$ cm$^{-2}$, inferred from the TBABS model is slightly larger than that obtained by @Ha99 [@Ol03; @Su11] from their WABS model [@Mo83] fit to [*ASCA*]{} and [*Chandra*]{} data, because we used the lower metallicity of the solar abundance set.
We also examined the X-ray spectral properties of the PWN with the XIS data. The spectrum can be described by an NEI model with an electron temperature $\sim$0.93 keV, ionization time-scale $\sim$ $3.6\times10^{11}$ cm$^{-3}$ s and power-law model with a photon index $\sim$1.97. The thermal emission might come from the front and back of the SNR shell as previously explained by @Ol03 or its origin might be physically related to the origin of the non-thermal emission. However, the nature of the thermal emission needs to be further studied.
{width="70.00000%"}
In the following, we estimate the properties of the remnant using the normalization obtained with the VVNEI model (see Table 1) in the analysis of the XIS data, considering the distance of 6.2 kpc. We estimate the volume for the SNR shell to be $V\sim 1.58\times10^{58}fd_{6.2}^{3}$ ${\rm cm^{3}}$, where $f$ is the volume filling factor ($0<f<1$) and $d_{6.2}$ is the distance in units of 6.2 kpc. Assuming $n_{\rm e}=1.2n_{\rm H}$, we found an ambient gas density of $n_{\rm e}$ $\sim$ 1.15$f^{-1/2}d_{6.2}^{-1/2}$ ${\rm cm}^{-3}$ and age of $\sim 6340f^{1/2}d_{6.2}^{1/2}$ yr. Our estimated SNR age is in agreement with previous age estimates (i.e. $\sim$7000 yr: @Ha99 and $\sim$3000 yr: @Su11). Using $M_{\rm X}$=1.4$m_{\rm H}n_{\rm e}V$, we calculate the total X-ray-emitting mass, $M_{\rm X}$ $\sim 21.4 f^{1/2}d_{6.2}^{5/2}~{M\sun}$, which is lower than that derived by @Su11 in their [*Chandra*]{} analysis ($M_{\rm X}$ $\sim 70 f^{1/2}d_{6.2}^{5/2}~{M\sun}$).
As a next step, we derive the electron density from the eastern and western regions of 3C 396 using XIS data. The obtained density is higher in the western ($\sim$2.43 ${\rm cm}^{-3}$) than in the eastern region ($\sim$0.93 ${\rm cm}^{-3}$).
In this analysis, the derived electron densities of the thermal X-ray plasma were found to be about 1$-$2 cm$^{-3}$, which is consistent with the result found by @Su11 ($\sim$1 cm$^{-3}$).
We discuss the SNR morphology. For this, we checked the morphology of the radio synchrotron emission. The radio emission is much brighter in the west than in the east, with only a faint tail extending from the prominent western shell toward east (e.g. @Su11 [@Cr16]). 3C 396 is a young SNR that contains a PWN at its centre, which is related to non-thermal X-ray emission. This characteristics differentiates this SNR from MM SNRs, which show centrally enhanced thermal X-ray emission (e.g. @Rh98). Additionally, unlike in most of the MM SNRs, there are significant electron temperature variations within 3C 396.
We compare the thermal X-ray luminosity with those of some MM SNRs similar to @Ba16. They plotted the radius versus thermal X-ray luminosity (0.3$-$10.0 keV) of MM SNRs with associations to GeV gamma rays and concluded that the value of the X-ray luminosity become smaller when SNRs evolve or when their radii become larger. Assuming a distance of 6.2 kpc, we derived an unabsorbed thermal X-ray luminosity in the 0.3$-$10.0 keV energy band of $\sim$ $1.2\times10^{35}$ erg s$^{-1}$ for 3C 396 using XIS data. In Fig. \[Fig5\], we give a plot as Fig3(b) of @Ba16, including the SNR 3C 396. Compared with other remnants, namely 3C 391, the thermal X-ray luminosity of 3C 396 is rather low despite its small radius. This could be due to absorption of soft X-ray emission from a lower temperature plasma, which is generated by shock-cloud interactions and cannot be identified in our analysis. Also, in Fig. \[Fig5\], the other SNRs are MM, some of which show RPs (e.g. @Oz09).
Gamma-ray emission
------------------
In the MeV to GeV energy range, gamma rays are produced by hadronic and/or leptonic processes. In the hadronic process, gamma rays are produced through the $\pi^0$ decay channel originating from interactions of accelerated protons with molecular material surrounding the SNR. The leptonic gamma-ray emission mechanism, on the other hand, requires accelerated electrons and ambient photons (e.g from the cosmic microwave background (CMB)) to produce gamma rays through inverse Compton (IC) scattering and non-thermal bremsstrahlung processes. The IC scattering process is accompanied by the synchrotron emission observed in radio and X-ray wavelengths.
As seen on the TS map given in Fig. \[Fig3\] left panel, the distribution of gamma rays in the energy range of 1$-$300 GeV overlaps with the western and northern parts of 3C 396. The best-fit position is pointing in a direction that coincides with dense MCs reported by @Su11 and @Ki16. This implies that the origin of GeV gamma rays may be due to the interaction of accelerated protons emanating from the SNR’s western or northern shell with the material inside the MCs.
From the point source flux observed by [*Chandra*]{}, @Ol03 discussed that the pulsar powering the PWN is a typical young pulsar of $L_{\rm sd} \sim 7.2 \times 10^{36}$ erg s$^{-1}$. Combined this with the fluxes obtained from [*Suzaku*]{} ($2.09^{+0.08}_{-0.11} \times 10^{-12}$ erg cm$^{-2}$ s$^{-1}$ in 1$-$10 keV and [*Fermi*]{} ($1.89 \times 10^{-5}$ MeV cm$^{-2}$ s$^{-1}$ in 0.2 – 300 GeV), we obtain the X-ray and GeV $\gamma$-ray efficiencies of the PWN to be $\eta_{\rm PWN,X} \approx 10^{-2.9}$ and $\eta_{\rm PWN, GeV} \approx 10^{-1.8}$. The flux upper limit in TeV obtained from Fig. A.1 of @Ab18 is 0.3 per cent of the flux of the Crab Nebula, which was measured by H.E.S.S. to be F($>$1 TeV) $= 2.26 \times 10^{-11}$ cm$^{-2}$ s$^{-1}$ [@Ah06] and gives the upper limit on the TeV $\gamma$-ray efficiencies $\eta_{\rm TeV} < 10^{-3.6}$.
The obtained efficiencies can be compared with those of the other young PWNe. Based on the spectral modeling by @TaTa10 [@TaTa11; @TaTa13], the efficiencies range as $10^{-3.9} \la \eta_{\rm TT, keV} \la 10^{-1.7}$, $10^{-4.7} \la \eta_{\rm TT, GeV} \la 10^{-3.7}$, and $10^{-5.0} \la \eta_{\rm TT, TeV} \la 10^{-3.4}$. The PWN has a typical X-ray efficiency and the upper limit on the TeV $\gamma$-ray efficiency is not strong enough to rule out the leptonic emission from the PWN. The observed GeV efficiency is fairly high compared with the leptonic emission model of young PWNe. In addition, the lower limit on the GeV to TeV flux ratio $\eta_{\rm GeV} / \eta_{\rm TeV} \ga 10^{2.2}$ is also much higher than that of the GeV PWNe candidates $\la 10$ (Fig. 12 of [@Ac13]). We conclude that the observed GeV photons do not likely come from the PWN.
Assuming the distance of 6.2 kpc, the 0.2$-$300 GeV luminosity of $1.4\times10^{35}$ erg s$^{-1}$ is close to the bright end of the SNRs observed by [*Fermi*]{} (Figs. 17 and 18 of [@Ac16]). Young SNRs are often dim in the 1$-$100 GeV energy band. Their faint GeV emission can be best explained by the IC scattering off the CMB photons by relativistic electrons [@Ac16]. The high ($\ga 10^{35}$ erg s$^{-1}$) GeV luminosity of 3C 396 is much more difficult to be reproduced by this mechanism. However, in the energy range of 1$-$100 GeV it is consistent with the high luminosity values found in old SNRs interacting with MCs which emit gamma rays likely through the hadronic mechanism.
The bremsstrahlung process may be effective in places where the density of the ambient medium is high ($n$ $>$ 100 cm$^{-3}$). In addition, hadronic interactions could dominate over the bremsstrahlung process, if the ratio of energy densities of protons to electrons ($W_{\rm p}/W_{\rm e}$) is greater than 10, which corresponds to the case when the gamma-ray emitting region is located far from the acceleration region [@Ta18]. Therefore, in order to reveal the dominating gamma-ray emission mechanism, multi-wavelength data, e.g. radio data, need to be investigated together with the GeV results we present in this paper.
Conclusions
===========
We have examined the X-ray emission from the SNR 3C 396 by studying its spectra with [*Suzaku*]{}. We also investigated GeV gamma-ray emission from the SNR as observed by [*Fermi*]{}-LAT. The conclusions of our work can be described as follows.
& Using [*Suzaku*]{} data, we found that the shell spectrum is clearly thermal ($\sim$1.12 keV), with no sign of a non-thermal component. The slightly enhanced abundances of Al and Ca found in the SNR shell suggest that the X-ray plasma is likely to be ejecta origin. The PWN emission is characterized by a power-law model with a photon index of $\sim$1.97 and thermal emission with an electron temperature of $\sim$0.93 keV.
& In the energy range of 1$-$300 GeV, we detected an excess of gamma rays associated with 3C 396 / 4FGL J1903.8+0531 and calculated the best-fit position to be R.A.(J2000) = 285$^{\circ}\!\!$.957 $\pm$ 0$^{\circ}\!\!$.027 and decl.(J2000) = 5$^{\circ}\!\!$.475 $\pm$ 0$^{\circ}\!\!$.021.
& In the energy range of 0.2$-$300 GeV, 3C 396 / 4FGL J1903.8+0531 was detected with a significance of 18$\sigma$ at the best-fit position, where it was fit with a log-parabola-type spectrum having indices of $\alpha$=2.66 and $\beta$=0.16.
& If gamma rays are produced through the leptonic PWN scenario, TeV gamma-ray emission should be observed from 3C 396, which have not been detected so far. In addition, assuming a distance of 6.2 kpc, the luminosity of 3C 396 / 4FGL J1903.8+0531 was found to be too high ($\ga 10^{35}$ erg s$^{-1}$) to be explained by the IC emission model. So, alternative gamma-ray emission mechanisms may be the bremsstrahlung process and the hadronic model. The location and the spectral features of 3C 396 / 4FGL J1903.8+0531 indicate that this SNR might be expanding into a dense MC at northern and western sides of the SNR’s shell and accelerated protons might be penetrating these dense clouds creating hadronic gamma rays. However, the leptonic contribution (i.e. bremsstrahlung emission) to the total gamma-ray emission could be significant.
& In our second paper, we will concentrate on spectral modelling of 3C 396 / 4FGL J1903.8+0531, especially using radio data, to find out if the dominating gamma-ray emission mechanism in 3C 396 is leptonic (i.e. bremsstrahlung) or hadronic in origin.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Dr. Aya Bamba, and all the [*Suzaku*]{} team members. We also thank the anonymous referee for useful comments and suggestions. AS is supported by the Scientific and Technological Research Council of Turkey (TÜBİTAK) through the BİDEB-2219 fellowship program. This work is supported in part by grant-in-aid from the Ministry of Education, Culture, Sports, Science, and Technology (MEXT) of Japan, No.18H01232(RY), No.16K17702(YO), No.17K18270(ST) and 16J06773, 18H01246 (SK). We are also grateful to the [*Fermi*]{} team for making the [*Fermi*]{}-LAT data and analysis tools available to the public.
$~$
[*Facility*]{}: [*Suzaku*]{} and [*Fermi*]{}-LAT.
[99]{}
Abdalla H. et al. (H.E.S.S Collaboration), 2018, [A&A, 612, A1]{}
Abdo A. A. et al., 2008, [Science, 322, 1218]{}
Acero F. et al. 2013, [ApJ, 773, 77]{}
Acero F. et al., 2016, [ApJS, 224, 8]{}
Aharonian F. A. et al., 2001, [A&A, 375, 1008]{}
Aharonian F. et al., 2005, [A&A, 439, 635]{}
Aharonian F. et al., 2006, [A&A, 457, 899]{}
Arnaud K. A., 1996, [in ASP Conf. Ser. 101, Astronomical Data Analysis Software and Systems V, ed. G. H. Jacoby & J. Barnes (San Francisco, CA: ASP), 17]{}
Bamba A., Sawada M., Nakano Y., Terada Y., Hewitt J., Petre R., Angelini L., 2016, [PASJ, 68, 5]{}
Becker, R. H. & Helfand, D. J., 1987, [AJ, 94, 1629]{}
Bochow A., 2011, [PhD Thesis, Ruperto-Carola University of Heidelberg]{}
Caswell J. L., Murray J. D., Roger R. S., Cole D. J., Cooke D. J., 1975, [A&A, 45, 239]{}
Caswell J. L., Haynes R. F., Milne D. K., Wellington K. J., 1982, [MNRAS, 200, 1143]{}
Cruciani A. et al., 2016, [MNRAS, 459, 4224]{}
Ergin T., Sezer A., Yamazaki, R., 2016, [Supernova Remnants: An Odyssey in Space after Stellar Death, Chania, Greece, 20]{}
The Fermi-LAT collaboration, 2019, arXiv e-prints, [arXiv:1902.10045]{}
Foster A. R., Ji L., Smith R. K., Brickhouse N. S., 2012, [ApJ, 756, 128]{}
Harrus I. M., Slane P. O., 1999, [ApJ, 516, 811]{}
Hewitt J. W., Rho J., Andersen M., Reach W. T., 2009, [ApJ, 694, 1266]{}
Ishisaki Y. et al., 2007, [PASJ, 59, 113]{}
Kilpatrick C. D., Bieging J. H., Rieke G. H., 2016, [ApJ, 816, 1]{}
Koyama K. et al., 2007, [PASJ, 59, 23]{}
Kushino A., Ishisaki Y., Morita U., Yamasaki N. Y., Ishida M., Ohashi T., Ueda Y., 2002, [PASJ, 54, 327]{}
Lee H.-G., Moon D.-S., Koo B.-C., Lee J.-J., Matthews K., 2009, [ApJ, 691, 1042]{}
Morrison R., McCammon D., 1983, [ApJ, 270, 119]{}
Nobukawa K. K. et al., 2018, [ApJ, 854, 87]{}
Olbert C. M., Keohane J. W., Arnaud K. A., Dyer K. K., Reynolds S. P., Safi-Harb S., 2003, [ApJ, 592, L45]{}
Ozawa M., Koyama K., Yamaguchi H., Masai K., Tamagawa T., 2009, [ApJL, 706, L71]{}
Ranasinghe S., Leahy D. A., 2018, [AJ, 155, 204]{}
Rho J., Petre R., 1998, [ApJ, 503, L167]{}
Scaife A. et al., 2007, [MNRAS, 377, L69]{}
Slane P. et al., 2012, [ApJ, 749, 131]{}
Slane P. O. 2017, [preprint (arXiv:1703.09311v1)]{}
Smith R. K., Brickhouse N. S., Liedahl D. A., Raymond J. C., 2001, [ApJ, 556, L91]{}
Su Y., Chen Y., Yang J., Koo B.-C., Zhou X., Lu D.-R., Jeong I.-G., DeLaney T., 2011, [ApJ, 727, 43]{}
Tanaka S. J., Takahara F., 2010, [ApJ, 715, 1248]{}
Tanaka S. J., Takahara F., 2011, [ApJ, 741, 40]{}
Tanaka S. J., Takahara F., 2013, [MNRAS, 429, 2945]{}
Tanaka T. et al. 2018, [ApJ, 866, L26]{}
Tawa N. et al., 2008, [PASJ, 60, 11]{}
Temim T., Slane P., Castro D., Plucinsky P. P., Gelfand J., Dickel J. R., 2013, [ApJ, 768, 61]{}
Temim T., Slane P., Kolb C., Blondin J., Hughes J.P., Bucciantini N., 2015, [ApJ, 808, 100]{}
Uchiyama H., Nobukawa M., Tsuru T. G., Koyama K., 2013, [PASJ, 65, 19]{}
Washino R., Uchida H., Nobukawa M., Tsuru T. G., Tanaka T., Kawabata N. K., Koyama K., 2016, [PASJ, 68, S4]{}
Wilms J., Allen A., McCray R., 2000, [ApJ, 542, 914]{}
Yasumi M., Nobukawa M., Nakashima S., Uchida H., Sugawara R., Tsuru T. G., Tanaka T., Koyama K., 2014, [PASJ, 66, 68]{}
[^1]: E-mail: [[email protected]]{} (AS)
[^2]: [[email protected]]{} (TE)
[^3]: [[email protected]]{} (NC)
[^4]: [[email protected]]{} (ST)
[^5]: Very recently, @RaLe18 updated the distance estimate as 8.5 kpc corresponding to the velocity of 69.4 km s$^{-1}$.
[^6]: <https://heasarc.nasa.gov/lheasoft>
[^7]: <http://www.atomdb.org>
[^8]: <http://fermi.gsfc.nasa.gov/ssc/data/analysis/software>
[^9]: <http://fermipy.readthedocs.io/en/latest/index.html>
[^10]: SNR’s radio location at R.A.(J2000) = 285$^{\circ}\!\!$.975 and decl.(J2000) = 5$^{\circ}\!\!$.521
[^11]: TS = -2ln($L_{\rm max,0}$/$L_{\rm max,1}$), where $L_{\rm max,0}$ is the maximum likelihood value for a model without an additional source and $L_{\rm max,1}$ is the maximum likelihood value for a model with the additional source at a given location.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Automation through IoT brings with it a whole new set of philosophical and ethical implications that we barely began to address. However, it is widely considered by many scholars as the panacea to overcome the majority of societal issues. The case of energy efficiency as an action for tackling climate change is not different: demand-response proposals or occupancy-driven energy management systems crowd the current research agenda on energy efficiency. However, there are still very few studies that have reported the effects of automation in the mid or long term beyond energy reduction (e.g. emotional feelings derived to interact with automation, complacency to the devices or perceived value of the automation throughout the time). In this workshop article, we report scientific evidence of a study conducted in ten workplaces during more than one year where we found that automating some electronic devices of common use (i.e. moving away or preventing subjects from the control of these devices) in favour of comfort and energy efficiency, is associated with a reduction of the users’ confidence in science and technology as a mean to solve all environmental current problems and reduce the willingness of people to act in favor of the environment.'
author:
- 'Diego Casado-Mansilla'
- Pablo Garaizar
- 'Anne M. Irizar'
- Diego López de Ipiña
bibliography:
- 'sample-base.bib'
title: 'On the Side Effects of Automation in IoT: Complacency and Comfort vs. Relapse and Distrust'
---
Case-Study
==========
We carried out an experimental intervention of one year designed to test the effectiveness of the persuasive techniques in the mid and long-term because of some scholars raised their concern about the feasibility of persuasion to maintain the target behaviour throughout the time. In this study, we instrumented the electrical capsule-based coffee machines of ten different workplaces distributed between two big cities of Spain (Madrid and Bilbao) in order to record the energy consumption being drawn. The reasons why we selected these appliances were: *1)* they are pretty common in work environments and are an element of shared use; *2)* they consume large amounts of energy compared to other work appliances such as monitors or laptops. More than eighty people were recruited within the offices following a snowball procedure and their participation was voluntary (we raffled an energy monitoring system among participant who completed the whole study).
The study followed a between-group design approach. Thus, three different strategies to cope with energy inefficiency were tested among the participant groups (see Figure \[fig:treatments\]). *1) Persuasive feedback*: a combination of real-time ambient feedback and subtle visual hints to support the user’s decision-making about when to switch off the appliance; *2) Energy-dashboard* : participants were provided with a Web site to track their energy consumption associated with the appliance (i.e. self-monitoring and rational information through comparisons with historic energy data); and *3) Automation* : the coffee makers were modified to autonomously switch the appliances off whenever they were not in use (i.e. the rationale behind automation was providing a sense of comfort to the users relieving them from the task of switching the appliance on and off after preparing a hot drink).
The random assignment of the experimental conditions among the ten workplaces remained as follows: Dashboard (3 workplaces), Automation (3 workplaces) and Persuasive feedback (4 workplaces).
Phases and Procedure
--------------------
The conducted study was divided into two phases: pre-pilot and post-piloting. During the former phase, the baseline of the energy wasted due to the misuse of capsule-based coffee machines (i.e. leaving the appliances in standby mode when not in use) was calculated. Furthermore, we asked participants ro respond an online questionnaire comprised of: 4 questions to obtain the socio-economic profile of each participant; 24 Likert-type questions pertaining to a questionnaire that evaluates the pro-environmental attitudes of the participants (Environmental Attitudes Inventory (EAI) [@Milfont24EAI]; and 12 Likert-type questions belonging to the Pro-Environmental Readiness to Change Questionnaire (PE-RTC) [@Tribble08PERB] that assesses the state or phases of change in which each of the participants was in relation to the intention to change their pro-environmental behavior: ’Pre-contemplation’, ’Contemplation’ or ’Action’. At the end of the experiment (post-pilot), we calculated again the amount of energy being wasted in each coffee-maker and participants were requested to re-answer the 36 questions of the two online questionnaires: EAI and PE-RTC. Once the questionnaires’ data was cleaned (e.g. remove outliers or uncompleted entries), we concluded that 81 participants appropriately responded in the pre-pilot phase and 48 participants did that accordingly at the end of the post-experimental phase. Hence, we removed the pre-pilot answers of 33 people in order to compare the results without introducing bias in the paired tests. Finally, because of the energy findings and for triangulation purposes (i.e. a technique that facilitates validation of data through cross verification from two or more sources), we wanted to grasp detailed qualitative information from users to really understand the causes of the persuasive treatment being more effective than automation or dashboard and the users’ reflections upon their assigned conditions. Because of the scope of this workshop, in the following, we only present the data resulting from the interaction with the Automated coffee-maker.
Results
=======
At the end of the study, we found that the IoT-based Persuasive treatment helped to save most energy than the other two treatments reducing the energy waste by 44.53% [@Casado16Thesis]. The Automation treatment also helped to reduce energy waste by 14.19%. Finally, the Dashboard approach did not lead to a reduction of energy waste remaining with a similar percentage as the beginning of the experiment.
Statistical Analyses over EAI and PE-RTC Questionnaires
-------------------------------------------------------
In this section, the statistical analysis of the paired comparison between the pre and post-test responses for each of the participants associated with the automation-based coffee maker is presented. On the other hand, analysis of (co)variance is also provided using ANCOVA. The aim was to carry out an exploratory study to observe which factors and dimensions of pro-environmental attitudes (EAI) and intentions of pro-environmental change (PE-RTC) could be influenced by the fact of subjecting a group of users to the automation condition.
### Automation as a whole
Sixteen people from three different working-groups interacted with the automated coffee machine. Comparing their responses between the pre-post piloting in the different constructs and scales of the questionnaires, we found statistical significance in the paired T-test on one of the scales of the EAI which evaluates the ’Trust in science and technology to solve all environmental problems’: *t (15) = 1.711, p = 0.0538*; the Effect Size (ES) using the Cohen’s coefficient *d = 0.427* with a confidence interval (CI) *CI = \[- 0.0916, 0.934\]*. According to Cohen’s power analysis criteria, this effect can be considered as medium [@Cohen88Power]. Thus, the people who interacted with the automated coffee machine decreased their confidence in technology.
A sample of related questions in this sub-scale of the EAI inventory are:
- Science and technology will eventually solve our problems with pollution, overpopulation and diminishing resources.
- Modern science will solve our environmental problems.
- We cannot keep counting on science and technology to solve our environmental problems.
Beyond this general analysis, we wanted to group working groups by socio-economic affinity by using hierarchical clustering. Having observed the similarities of the participant groups, we came out with three affinity groups, namely A, B and C. With this groups, we applied block design [@Calinski00Block] in which the experimenter divides subjects into subgroups called blocks, such that the variability within blocks is less than the variability between blocks (the previous approach where all people assigned to automation condition were studied together). The results of the groupings, the conditions they were assigned to within groups are presented in Table \[tab:asig\_experimental\].
### Automation within blocks of affinity
Similarly, as the previous analysis studying the overall people assigned to one treatment, in Block A we found that users under automation condition showed lower reliability in science and technology: *t(5) = 2.169, p = 0.0411*. ES was considered medium according to Cohen’s criteria *d = 0.547, CI = \[- 0.893,1.987\]*. Furthermore, applying ANCOVA over the three conditions, we found a difference in ’Confidence in science and technology’: *F$_{2, 20}$ = 2.872, p = 0.800* with an effect size $\eta_{p}^{2} =$ 0.223 with 22.3% of the variance explained by the automation experimental condition (posthoc Tukey was applied ) and 3.9% due to the answers in the covariate: the Pre-experimental phase (the covariate is linearly related to the dependent variable and is not related to the condition). Block B did not provide relevant differences in the studied constructs, however in Block C the subjects who interacted with the automatic coffee machines (automation) were found to decrease their active involvement in favor of the environment - the ’Action’ state of change - (in this case, we applied Wilcoxon for non-parametric data): $Z=$2.041, $p<$0.0312 with a large effect size measured in the Rosenthal coefficient $r =$ 0.510. Furthermore, in Block C we found a significant difference between the three experimental conditions in this ’Action’ construct after applying ANCOVA analysis, being automation the condition which marked the difference: $F_{2,9} = $11.264, $p = $0.0035 with an effect size $\eta_{p}^{2} =$0.714 which denotes that the automation experimental condition explains 71.4% of the total variance, while that of the pre-test questionnaire (which is the covariate) just explains 7.7% of the dependent variable.
Analysis of Results
-------------------
The case of the automation condition aroused the greatest interest at a conclusive level. For the parametric tests carried out in the non-clustered approach, it was observed that subjects under this condition reduced their confidence in technology as a driver of pro-environmental change. When sampling noise was eliminated and the subjects were grouped in affinity blocks, it was observed that the confidence in the technology also decayed in Block A. Furthermore, the case of Block C was of relevance: the 8 people subjected to the automation condition reduced their state of Action (e.g. they were less active to act in favor of the environment). This result leads us to think that the condition has caused a significant reduction in pro-environmental attitudes and intentions.
Finally, taking into account the results from ANCOVA method, in Block A we observed that people subjected to automation presented at the end of the experiment the least perception of the use of technology as the main remedy for environmental problems. Besides, in group C the subjects under the automation condition significantly reduced their ’Action’ state which entails that they seemed to be less active doing actions in favor of the environment.
Discussion
==========
Keeping people away from the decision-making about turning the coffee machine on and off has not been free from critical and sympathized voices. On the one hand, the complacency induced by not having to think about how to act, knowing that the device operates by self-efficiently, was the main motivation of a large number of subjects interviewed at the beginning of the experiment. However, the same subjects showed certain rejection signs and frustration for not being able to manipulate the device at certain times (especially when the perception was that leaving the device switched on could cause energy waste). These feelings seem to be explanatory of the main finding of the attitudinal study: people under the automation condition showed less confidence in technology as the main pillar for climate change and environmental issue. Similarly, the people subjected to automation treatment were the ones among the three with the least confidence in science and technology.
In view of all the data presented, it seems that removing the subjects from the control of a simple electronic device of shared use in favor of comfort, reduced the users’ confidence in the technology as a means to improve the overall environment in the future. Accordingly, it was observed that comfort due to automation may generate a rebound effect with respect to the passivity of acting in favor of the environment. As an example, in group C the subjects under the automation condition significantly reduced their ’Action’ state. Applying this finding to the case of reducing the waste of energy in the workplace by applying automation; we claim that reducing the involvement of users in making simple decisions related to whether a machine should stay switched off or on may reduce their perception of energy expenditure, and therefore, may reduce their actions in favor of energy efficiency in such context.
Conclusions and Future work
===========================
In this article, we have witnessed an unexpected rebound effect caused by automation. According to the data presented, the fully automated management of processes focused on energy efficiency tends to generate a rebound effect causing, on the one hand, passivity to act in favor of the environment and, on the other, widespread distrust of science and technology. The results from the pre-post questionnaires and the interviews with participants of the study have shown that people may start feeling some sort of complacency by using an automated IoT device. However, when people do not fully understand certain processes of the automation and they lack control because of it, users might present initial states of reluctance to the automation. Taking this experiment as a baseline, there appears some future lines of research. For instance, to confirm this rebound effect and continue exploring why automation is contrary to some utilization factors of the pro-environmental attitudes. Furthermore, further research should be done through new experiments in groups with similar characteristics to Block A and C to assess if the results can be replicated and are specific for certain profiles.
This research is funded by HORIZON 2020 - (RIA)-696129-GREENSOUL. We also acknowledge the support of the Spanish government for SentientThings under Grant No.: TIN2017-90042-R.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'N-body simulations are used to model the tidal disruption of the Sagittarius (Sgr) dwarf galaxy with constraints set by the positions and velocities of M giants in the Sgr tidal arms recently revealed by the Two Micron All-Sky Survey (2MASS). The simulated Sgr dwarf is placed on a variety of orbits within a Milky Way potential parameterized by variable circular velocities, halo flattenings and radial profiles. Two hundred separate test particle orbits have been used to explore a wide range of model Milky Way potentials and dwarf galaxy characteristics. The family of models is delimited by the data to a relatively narrow allowed range of parameters, and then input into N-body simulations. We present our best-fitting model, and discuss the orbital period, apoGalacticon distance, current space velocity, mass-to-light ratio, and other characteristics of the Sgr dwarf. In addition, we discuss the implications of this model for the flattening of the Galactic halo.'
author:
- 'David R. Law, Steven R. Majewski, Michael F. Skrutskie'
- 'Kathryn V. Johnston'
title: Modeling the Tidal Tails of the Sagittarius Dwarf Galaxy
---
Introduction
============
Since the discovery of the Sgr dwarf by Ibata et al. (1994) many groups (e.g., Johnston, Hernquist, & Bolte 1996, Ibata et al. 1997, Ibata & Lewis 1998, G[' o]{}mez-Flechoso, Fux, & Martinet 1999, Johnston et al. 1999, Helmi & White 2001) have sought to model the Sgr - Milky Way interaction with respect to a modest patchwork of observational constraints. Recently, Majewski et al. (2003a, hereafter “Paper I”) have shown that the extensive length of the Sgr tidal tails can be traced by M giant stars visible in the all-sky view of the system provided by the 2MASS database. Spectroscopy of Sgr candidate stars has allowed determination of radial velocities throughout the trailing tail (Majewski et al. 2003b, hereafter “Paper II”), and these substantial new constraints can be used to develop more refined models of the Sgr system.
In this contribution, we briefly describe some of the major results of such modeling. A comprehensive description of this new Sgr disruption model can be found in Law, Johnston, & Majewski (2003, hereafter “Paper III”).
Modeling the Sgr System
=======================
Following previous work by Johnston et al. (1996, 1999) the Milky Way potential is represented numerically by a Miyamoto-Nagai (1975) disk, Hernquist spheroid, and a logarithmic halo. The total mass and radial profile are fixed by requiring that the rotation curve of this model Galaxy be consistent with HI & CO tangent point observations (e.g., Honma & Sofue 1997).
The Sgr dwarf itself is represented by $10^5$ self-gravitating particles (representing both the dark and light matter components of the satellite), which are initially distributed according to a Plummer (1911) model. This satellite is evolved through the simulated Galactic potential for five orbital periods using a self-consistent field code (Hernquist & Ostriker 1992). The present-day simulated dwarf is constrained to be located at $(l,b) = (5.6^{\circ},-14.2^{\circ})$ at a solar distance of $D_{\rm Sgr} = 24$ kpc (Paper I, Ibata et al. 1995) and have a radial velocity of $v_{\rm LOS,Sgr} = 171$ km s$^{-1}$ (Ibata et al. 1997). The direction of the dwarf’s space velocity vector is determined by requiring that the dwarf orbit in the orbital plane observed in Paper I.
Subject to these requirements, test-particle orbits (i.e. orbits calculated for a test particle with the observed kinematical characteristics of Sgr) and N-body simulations are performed for simulated satellites with a variety of orbital speeds. These simulations can be additionally constrained using the 2MASS M giant distance and radial velocity data presented in Papers I and II. Fig. 1 compares the M giant data (Panels a-b, filled squares) with the model Sgr dwarf whose tidal tails best reproduce the observations (Panels c-d). Note the close agreement between model and observed debris distances and radial velocities along the trailing debris tail ($\Lambda_{\odot} = 0^{\circ}$ - $100^{\circ}$)[^1]. This best-fit model is characterized by a period of 0.75 Gyr with apoGalacticon 52 kpc, periGalacticon 14 kpc, and a present space velocity of $(U,V,W) = (237.2, -43.4, 218.9)$ km s$^{-1}$.
Although we do not attempt to model the Sgr core in detail, it is nonetheless possible to use the width of the Sgr debris stream to estimate such global characteristics as the bound mass of the dwarf. The simulated dwarf which appears to best fit the width of streams shown in Fig. 1 has a present mass of $M_{\rm Sgr} = 3 \times 10^8 M_{\odot}$ and a mass-to-light ratio $M_{\rm Sgr}/L_{\rm Sgr} = 21$.
Discussion
==========
As demonstrated in the previous section, the tidal tails of this model provide a good fit to the all-sky view of M giants presented in Papers I and II. It is therefore possible to use this model to determine what range of Milky Way models permit simulated satellites to reproduce observations. Particularly, N-body simulations can be used to constrain the flattening of the Galactic halo (e.g., Ibata et al. 2001). Fitting an orbital plane to leading and trailing M giant debris separately, we determine that the orbital pole of Sgr debris has precessed by $1.7^{\circ} \pm 2.4^{\circ}$ over about $300^{\circ}$ of orbital longitude. Repeating this calculation for N-body simulations in model dark halos with a variety of flattenings, we calculate pole precessions of $2.2^{\circ} \pm 1.6^{\circ}$, $3.5^{\circ} \pm 1.7^{\circ}$, and $5.6^{\circ} \pm 1.4^{\circ}$ for flattenings in the halo potential of $q = 1, 0.95$, and $0.90$ respectively. It therefore appears likely that the halo of the Milky Way can be described by an almost spherical potential.
Although this model provides a good match to the distances and velocities of trailing Sgr debris given in Papers I and II, it does not fit recent data obtained by Majewski et al. (2003c, hereafter “Paper IV”) in the region of the Sgr leading arm. Fig. 1 (Panel a, filled triangles) plots these new data, which has velocities slower than that of the model by up to 200 km s$^{-1}$ in the range $\Lambda_{\odot} = 300^{\circ}$ - $200^{\circ}$. There is no simple modification of the velocity of the model satellite that serves to reproduce this new trend, and this may be an indication of such other effects as dynamical friction. However, simulations suggest that including corrections from Chandrasekhar’s formulation of dynamical friction should not have a substantial effect on the observed velocities of leading tidal debris for model satellites with mass $M_{\rm Sgr, 0} \leq 10^{10} M_{\odot}$, and we find that accurately reproducing the observed trend is difficult even for satellites with initial masses greater than this.
This inconsistency and implications of the best-fit model for the size and shape of the Milky Way are discussed at greater length in Paper III.
DRL acknowledges support from the Local Organizing Committee, a U.Va. Small Research Fellowship, and the U.Va. Echols Program. The authors also acknowledge NASA/JPL contracts 1228235 (SRM/KVJ) and 1234021 (MFS).
G[' o]{}mez-Flechoso, M. A., Fux, R., & Martinet, L. 1999, , 347, 77
Helmi, A. & White, S. D. M. 2001, , 323, 529
Hernquist, L. & Ostriker, J. P. 1992, , 386, 375
Honma, M. & Sofue, Y. 1997, , 49, 453
Ibata, R. A., Gilmore, G., & Irwin, M. J. 1994, Nature, 370, 194
Ibata, R. A., Gilmore, G. & Irwin, M. J. 1995, , 277, 781
Ibata, R. A. & Lewis, G. F. 1998, , 500, 575
Ibata, R.A., Lewis, G. F., Irwin, M.J., Totten, E., & Quinn, T. 2001, , 551, 294
Ibata, R. A., Wyse, R. F. G., Gilmore, G., Irwin, M. J., & Suntzeff, N. B. 1997, , 113, 634
Johnston, K. V., Hernquist, L., & Bolte, M. 1996, , 465, 278
Johnston, K. V., Majewski, S. R., Siegel, M. H., Reid, I. N., & Kunkel, W. E. 1999, , 118, 1719
Law, D. R., Johnston, K. V., & Majewski, S. R. 2003, in prep. (“Paper III”)
Majewski, S. R., Skrutskie, M. F., Weinberg, M. D. & Ostheimer, J. C. 2003a, ApJ submitted (“Paper I”)
Majewski, S. R. et al. 2003b, ApJ submitted (“Paper II”)
Majewski, S. R. et al. 2003c, in prep. (“Paper IV”)
Miyamoto, M. & Nagai, R. 1975, , 27, 533
Plummer, H. C. 1911, , 71, 460
[^1]: We use the orbital longitude coordinate system in the Sgr orbital plane defined in Paper I.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Approaches for kinship verification often rely on cosine distances between face identification features. However, due to gender bias inherent in these features, it is hard to reliably predict whether two opposite-gender pairs are related. Instead of fine tuning the feature extractor network on kinship verification, we propose a comparator network to cope with this bias. After concatenating both features, cascaded local expert networks extract the information most relevant for their corresponding kinship relation. We demonstrate that our framework is robust against this gender bias and achieves comparable results on two tracks of the RFIW Challenge 2020. Moreover, we show how our framework can be further extended to handle partially known or unknown kinship relations.'
author:
-
- |
Anonymous FG2020 submission\
Paper ID\
bibliography:
- 'ref.bib'
title: '**A Multi-Task Comparator Framework for Kinship Verification**'
---
INTRODUCTION
============
Kinship relationship between two people is usually determined using the persons’ physical features, which can be divided into DNA, body and facial features. In contrast to the very reliable DNA-analysis, facial- and body features are used to obtain an initial and quick estimate of whether two people are related or not.
Image-based kinship verification [@georgopoulos2018modeling; @fang2010towards; @guo2012kinship; @yan2014discriminative; @yan2014prototype; @xia2012understanding; @dibeklioglu2013like; @wang2014leveraging; @dawson2018same; @wang2018cross; @robinson2018visual; @rachmadi2018paralel; @rachmadi2019image; @dahan2017kin; @wang2017kinship; @robinson2016fiw; @laiadi2019kinship; @hu2014large; @xia2012toward; @zhou2011kinship; @zhou2012gabor; @patil2019deep; @robinson2017recognizing] relies only on information present in facial images to estimate whether they are related. Due to the inherent flexibility of only needing a face image compared to more invasive DNA-sample, kinship verification with visual media has an abundance of practical uses: e.g., forensic investigations, genealogical studies, social media-based analysis and photo library management. As proposed in the RFIW Challenge 2020 [@robinson2020recognizing], one can state three problems concerning kinship verification:
1. Determine whether two persons are consanguine given a kinship relation.
2. Decide whether a person is the child of given parents.
3. Identifying relatives of a person in a gallery.
Lately, the emergence of bigger image kinship verification datasets, including CornellKinFace [@fang2010towards], KinFaceW [@lu2013neighborhood; @lu2015fg], TSKinFace [@qin2015tri], and FIW[@robinson2016fiw], has given more and more attention to kinship-related tasks and allowed the development of more reliable data-based approaches. Kinship verification from face images focuses on consanguinity kinship, which can be divided into three groups:
- Same-generation pairs: brother-brother *BB*, brother-sister *SIBS* and sister-sister *SS*
- First-generation pairs: father-son *FS*, father-daughter *FD*, mother-son *MS* and mother-daughter *MD*
- Second-generation pairs: grandfather-grandson *GFGS*, grandfather-granddaughter *GFGD*, grandmother-grandson *GMGS* and grandmother-granddaughter *GMGD*
{width="0.95\columnwidth"}
\[fig:network\_small\]
\[fig:histogram\]
{width="94.00000%"}
\[fig:network\]
As illustrated in Fig. \[fig:network\_small\], typical kinship verification approaches consist of a convolutional neural network, which extracts facial features for each image separately. These features are then fed into a kinship comparator in order to distinguish between *kin* or *non-kin*. Several methods [@robinson2018visual; @laiadi2019kinship; @robinson2020recognizing] rely on metrics like cosine distance between extracted features to determine kinship. However, as shown in the histogram in Fig.\[fig:histogram\], *kin* and *non-kin* pairs from opposite-gender kinship relations are hardly separable compared to same-gender kinship relations, which is due to the high influence of gender on the feature.
Motivated by this finding and in contrast to training the feature extractor on kinship recognition [@robinson2018visual; @dahan2017kin; @rachmadi2019image2; @aspandi2019heatmap; @patil2019deep], we propose a comparator framework, which is robust against this gender bias as we demonstrate later. Our neural kinship comparator framework is not only capable of solving typical kinship-related tasks benefiting from separated local expert networks for each kinship relation but can also be further extended with an attention module to predict the kinship relation and leverage it for tasks with unknown kinship relation.
RELATED WORK
============
According to Georgopoulos et. al. [@georgopoulos2018modeling] approaches for kinship verification can be divided into five categories:
Invariant descriptors based methods [@guo2012kinship; @fang2010towards] are focusing on how to represent local facial parts. Subspace learning-based approaches [@yan2014discriminative; @yan2014prototype] learn a kinship invariant subspace capitalizing on techniques like factor analysis and transfer learning. Metric learning-based methods [@wang2009information; @davis2007information; @ding2016robust] involve learning a distance measure or feature transformation, and are used to reduce the feature distance between *kin* pairs while extended the distance for *non-kin* pairs. Approaches using contextual and dynamic features [@xia2012understanding; @dibeklioglu2013like], applying texture descriptors and using geometric information [@wang2014leveraging] have also been studied for kinship verification.
Apart from these traditional methods, deep learning-based approaches lately achieved state-of-the-art performance in kinship verification. Widely used architectures, e.g., VGG16 [@dawson2018same], GAN & ResNet [@wang2018cross], SphereFace [@robinson2018visual], SPCNN [@rachmadi2018paralel], ShallowResNet [@rachmadi2019image] and VGGFace [@dahan2017kin; @wang2017kinship; @robinson2016fiw] have been used for this task. Laiadi et. al [@laiadi2019kinship] proposed a novel approach feeding the cosine similarity, which is computed from deep (VGG-Face descriptor) and tensor (BSIF- and LPQ-tensor using MSIDA method) features, through an extreme learning machine in order to verify kinship.
[lLLLLLLLLLLLL]{} &\
(lr)[2-13]{} & & & &\
(lr)[3-5]{} (lr)[6-9]{} (lr)[10-13]{} User & & *BB* & *SIBS* & *SS* & *FD* & *FS* & *MD* & *MS* & *GFGD* & *GFGS* & *GMGD* & *GMGS*\
vuvko & **78.1** \* & **80.2** \* & **77.3** \* & **80.4** \* & 75.2 & 80.8 & **77.7** \* & 74.4 & 77.9 & 69.4 & **75.8** \* & 59.8\
DeepBlueAI & 76.1 & 76.5 & 74.6 & 76.9 & 74.4 & 80.8 & 75.1 & 73.9 & 72.5 & 72.7 & 67.3 & **67.6** \*\
ustc-nelslip & 75.9 & 75.1 & 72.0 & 74.4 & 75.5 & **81.8** \* & 74.7 & 75.2 & **78.6** \* & 69.0 & **75.8** \* & 67.0 \
haoxl & 75.5 & 74.8 & 71.1 & 74.0 & 75.5 & 81.2 & 74.7 & 75.2 & 72.9 & 64.9 & 63.2 & 64.3\
lemoner20 & 75.4 & 75.0 & 72.2 & 74.5 & 75.4 & 80.7 & 74.0 & 75.0 & 72.0 & 66.9 & 61.7 & 65.4\
Early & 74.2 & 74.6 & 72.9 & 74.3 & 73.4 & 78.5 & 72.3 & 74.4 & 65.7 & 68.6 & 52.4 & 64.8\
**ours** & 73.6 & 66.4 & 76.0 & 65.3 & **76.9** \* & 80.1 & 76.7 & **78.2** \* & 70.0 & 73.4 & 63.9 & 60.3\
bestone & 73.2 & 69.2 & 62.4 & 67.1 & 75.4 & 81.2 & 75.4 & 75.4 & 73.1 & 69.4 & 64.7 & 62.0\
danbo3004 & 72.6 & 71.3 & 70.9 & 72.0 & 72.4 & 78.1 & 71.5 & 72.0 & 71.1 & 69.8 & 53.2 & 56.4\
tenelven & 72.3 & 72.2 & 71.4 & 73.4 & 70.1 & 77.1 & 70.0 & 71.2 & 69.8 & **74.7** \* & 63.2 & 67.0 \
\[tab:track1\]
MULTI-TASK KINSHIP COMPARATOR FRAMEWORK
=======================================
Face Feature Extractor {#sec:faceextractor}
----------------------
We embed face images into a deep feature space using the adapted ResNet-50 [@he2016identity] with the ArcFace layer according to [@deng2019arcface]. To pretrain the model with softmax cross-entropy on the refined MS-Celeb-1M dataset [@MS1M] we add an 85164-dimensional fully connected layer, which is dropped later on together with the ArcFace layer. This bottleneck architecture together with the ArcFace layer ensures a well-generalizing identity feature vector $\boldsymbol{f} \in \mathbb{R}^{512}$.
Kinship Comparator {#sec:track1}
------------------
Generally, we can describe the kinship verification task as follows: Given the triplets $\left(\boldsymbol{x}_1,\,\boldsymbol{x}_2, \,y\right)$ consisting of two images $\boldsymbol{x}_1,\,\boldsymbol{x}_2 \in \mathbb{R}^{112\times112\times3}$ and a kinship relation $y$, the goal is to determine whether $\boldsymbol{x}_1$ and $\boldsymbol{x}_2$ are related as encoded in $y$, which is denoted by the probability $z$ at the output of our framework. Both face images are embedded independently by the feature extractor described in subsection \[sec:faceextractor\] yielding the corresponding feature vectors $\boldsymbol{f}_1$ and $\boldsymbol{f}_2$.
For our kinship comparator, depicted in Fig. \[fig:network\], we concatenate both features $\boldsymbol{f}_{\text{c}} = \left[\boldsymbol{f}_1,\, \boldsymbol{f}_2\right]$. Next, the concatenated feature vector is fed into the first out of eleven local expert networks. By building a local expert for every kinship relation $y$ we allow every local expert to focus only on parts of the features relevant of its corresponding kinship relation. For same-gender kinship relations (*BB*, *SS*, *FS*, *MD*, *GFGS* and *GMGD*) the local expert can deduce from detecting separate genders that both input images cannot be related. Similarily, by detecting same gender a opposite-gender kinship relation (*SIBS*, *FD*, *MS*, *GFGD* and *GMGS*) can be excluded. Moreover, different facial features are shared from mother or father of a child affirming our proposed architecture.
Every local expert is an identical fully connected neural network with two layers. The first layer consists of 192 neurons with leaky ReLU [@maas2013rectifier] as activation function. We conclude the local expert network with a fully connected layer consisting of a single neuron and sigmoid activation function to obtain a probability $z_{2,i}$ for the $i$-th local expert between 0 and 1. While the input of the first local expert is the concatenated feature vector $\boldsymbol{f}_{\text{c}}$, the remaining local experts use the output of the previous local experts $z_{1,i-1}$. Due to this architecture, the information is first refined in every 192-dimensional layer and then the information most relevant for its specific kinship relation is extracted in the second 1-dimensional layer. Mathematically, the output of the $i$-th local expert $z_{2,i}$ can be formulated as follows: $$\begin{aligned}
&\boldsymbol{z}_{1,1} = \max \left(\boldsymbol{W}_1 \boldsymbol{f}_{\text{c}} + \boldsymbol{b}_1, 0.2 \cdot \left(\boldsymbol{W}_1 \boldsymbol{f}_{\text{c}} + \boldsymbol{b}_1\right)\right)\\
&\boldsymbol{z}_{1,i} = \max \left(\boldsymbol{W}_1 \boldsymbol{z}_{1,i-1} + \boldsymbol{b}_1, 0.2 \cdot \left(\boldsymbol{W}_1 \boldsymbol{z}_{1,i-1} + \boldsymbol{b}_1\right)\right) \label{eq:2}\\
&z_{2,i} = \operatorname{sigmoid}\left(\boldsymbol{W}_2 \boldsymbol{z}_{1,i} + b_2\right)\end{aligned}$$ with $\boldsymbol{W}_1$, $\boldsymbol{W}_2$ and $\boldsymbol{b}_1$, $b_2$ denoting the trainable weight matrix and bias (vector) of the first and second layer, respectively. Note that, (\[eq:2\]) is only valid for $i > 1$.
By concatenating the outputs of all local experts $\boldsymbol{z}_2 = \left[z_{2, 1},\,\cdots,\, z_{2,11}\right]$ we obtain a probability of $\boldsymbol{x}_1$ and $\boldsymbol{x}_2$ being related for every kinship relation $y$. Since the kinship relation $y$ is given for the kinship verification task, we can observe its corresponding probability at the output $z$ of the framework by performing a scalar multiplication of $\boldsymbol{z}_2$ with the one-hot encoding of $y$. We can also interpret this one-hot encoding, which is generated by the task-dependent kinship relation encoder depicted in Fig. \[fig:network\], as relying entirely on the output of the local expert selected by $y$. Later, we will show that this disentanglement between the predictors $z_{2,i}$ and the selection of the predictors according to an ideally given kinship relation offers a variety of opportunities for future extensions. Another benefit of this structure is the joint training of all local experts without restricting the capabilities of the framework performing multiple tasks.
Extension for Tri-Subject Verification {#sec:track2}
--------------------------------------
For the tri-subject verification task, a quadruple $\left(\boldsymbol{x}_1,\,\boldsymbol{x}_2,\,\boldsymbol{x}_3, \,y'\right)$ is given with $\boldsymbol{x}_1$, $\boldsymbol{x}_2$ and $\boldsymbol{x}_3$ indicating the image of the father, mother and child, respectively, and $y'$ denoting the gender of the child. In accordance with subsection \[sec:track1\], this task can be performed by our framework by splitting the sample into two separate triplets $\left(\boldsymbol{x}_1,\,\boldsymbol{x}_3, \,\textit{FC}\right)$ and $\left(\boldsymbol{x}_2,\,\boldsymbol{x}_3, \,\textit{MC}\right)$ with *C* being a placeholder for whether the child is the parents’ son *S* or daughter *D*. Feeding these triplets into our framework, we obtain two separate probabilities $z_{\textit{FC}}$ and $z_{\textit{MC}}$ indicating how likely it is that the child is related to the father and mother, respectively.
EXPERIMENTS
===========
Training Details {#sec:train}
----------------
Training is divided into two stages: We pretrain the face feature extractor with softmax cross-entropy loss for face identification using the MS-Celeb-1M dataset, which contains over $5.8$M images of over 85k identities [@MS1M]. As preprocessing, we align all faces with facial landmarks predicted by the MTCNN [@zhang2016MTCNN] and crop them afterwards to $112 \times 112$ pixels. Our face feature extractor achieves an accuracy of $99.63\,\%$ on the LFW benchmark [@LFW].
While keeping the weights of the face feature extractor constant, we train the kinship comparator using pairs generated from the RFIW training dataset [@robinson2020recognizing] consisting of $\approx 249$k *kin* pairs (after duplicating and swapping $\boldsymbol{x}_2$ with $\boldsymbol{x}_2$ for all same-generation pairs *BB*, *SS* and *SIBS*). In order to generate meaningful *non-kin* pairs, we randomly swap $\boldsymbol{x}_2$ with $\boldsymbol{x}_2$ from a different family with the same kinship label $y$ in every epoch. By doing so, we ensure not only high variety among the *non-kin* pairs but also that gender and age of the *non-kin* pairs match the kinship relation $y$. We preprocess the faces identically as when pretraining the face feature extractor. As data augmentation, we perform left-right flipping, and random contrast, brightness and saturation with a probability of $50\,\%$. The kinship comparator is trained on binary sigmoid cross-entropy loss for 4 epochs using the ADAM-optimizer [@ADAM] with a batch size of 200 and an initial learning rate of $0.001$, which we decrease to $0.0005$ after the second epoch. In order to improve generalization we add $20\,\%$ dropout on the concatenated feature vector $\boldsymbol{f}_{\text{c}}$ and train with an additional regularization loss on the $L^2$-Norm of all trainable weights of the kinship comparator with a factor of $2 \cdot 10^{-4}$.
Results: RFIW Track 1 - Kinship Verification
--------------------------------------------
We report our performance evaluating on the RFIW challenge dataset [@robinson2020recognizing] containing $\approx 40$k image pairs with their kinship labels. In order to decide whether the probability $z$ is sufficient to classify a pair as related, we compute the threshold which yields the best average accuracy on the validation set, which consists of $\approx 129$k pairs in total (after creating *non-kin* pairs as described in subsection \[sec:train\]). Table \[tab:track1\] illustrates the accuracy for each kinship relation $y$. Even though our approach does not yield the best average accuracy, our average accuracy is only $2.5$% lower than second place. Moreover, we outperform all other approaches for the kinship relation father-daughter *FD* and mother-son *MS*, and further obtain second-best accuracy on three more kinship relations (*SIBS*, *MD* and *GFGS*). This indicates that our approach stands out especially in opposite-gender kinship relations (*FD*, *MS* and *SIBS*), which is also affirmed by the better separable distributions of *kin* and *non-kin* pairs using the output $z$ as shown in Fig. \[fig:histogram\_logit\] compared to the cosine distance in Fig. \[fig:histogram\]. The inferior performance on other opposite-gender kinship relations (*GFGD* and *GMGS*) can be explained by the substantially smaller amount of training data ($\approx 4$k pairs for grandparent-grandchild relations compared to $61\,\text{k} - 94$k pairs for *SIBS*, *FD* and *MS*).
Table \[tab:track1\_param\] shows the accuracy of our framework for different activation functions, dropout probabilities and hidden layer sizes on the RFIW kinship verification validation dataset with the *non-kin* pairs generated as mentioned in subsection \[sec:train\]. It can be seen that a dropout of 20% - 40% together with leaky ReLU as activation function yields the best results. However, a higher amount of neurons in the hidden layer seems very likely to boost the performance on the challenge dataset even further.
[cCCC]{} Activation Function & & &\
ReLU & 20% & 192 & 76.8\
PReLU & 20% & 192 & 75.6\
Tanh & 20% & 192 & 72.5\
LReLU & 0% & 192 & 75.5\
LReLU & 10% & 192 & 76.6\
LReLU & 30% & 192 & 77.7\
LReLU & 40% & 192 & 77.1\
LReLU & 20% & 64 & 74.5\
LReLU & 20% & 128 & 77.5\
LReLU & 20% & 256 & 77.2\
LReLU & 20% & 512 &**79.6**\
LReLU & 20% & 1024 & 78.8\
LReLU & 20% & 192 & 77.5\
\[tab:track1\_param\]
Results: RFIW Track 2 - Tri-Subject Verification
------------------------------------------------
As mentioned in subsection \[sec:track2\], by splitting the tri-subject verifications into two verification problems we obtain two probabilities $z_{\textit{FC}}$ and $z_{\textit{MC}}$. Since according to Table \[tab:track1\] there is only a small difference between same-gender and opposite-gender parent-child pairs, we take the average of both probabilities and use the threshold obtaining the best accuracy on the RFIW tri-subject validation set. The results on the RFIW tri-subject challenge dataset are depicted by Table \[tab:track2\] showing that our method yields comparable results.
[lLLL]{} &\
(lr)[2-4]{} User & & &\
ustc-nelslip & 0.79 & 0.78 & 0.80\
lemoner20 & 0.78 & 0.76 & 0.80\
DeepBlueAI & 0.77 & 0.76 & 0.77\
Early & 0.77 & 0.76 & 0.77\
**ours** & 0.73 & 0.72 & 0.74\
Ferryman & 0.72 & 0.70 & 0.74\
willgo & 0.68 & 0.66 & 0.70\
\[tab:track2\]
CONCLUSIONS AND FUTURE WORKS
============================
In this paper, we present a novel framework for multi-task kinship recognition, which achieves top accuracy for five out of eleven kinship relations compared to over 20 state-of-the-art methods on the RFIW kinship verification challenge dataset. The key advantage of our architecture is the joint training of a local expert network for each kinship relation. This not only allows every expert to extract the information necessary to reliably predict its corresponding kinship but also shares and refines information among the experts. Our approach performs especially well on opposite-gender pairs, which is affirmed by the reduction of gender bias originally present in face identification features. Moreover, we demonstrate that our framework achieves comparable performance on the tri-subject verification task.
The future work is twofold: First, as already indicated by the results Table \[tab:track1\_param\] a tuning of the local expert networks can further increase the performance. Entirely local experts with both fully connected layers being separated from each other have also shown their potential in our experiments, but tend to overfit due to the missing consecutive refinement in the shared first layer. For the tri-subject verification task, a more sophisticated pooling operation based on the confidence of each probability would be capable of fusing both probabilities more reliably.
Next, we plan to demonstrate the full potential of our approach by evaluating the performance additionally for partly known kinship relations. For instance, instead of using the kinship relation a more realistic scenario of knowing only the gender of both input images could be considered. Besides, the kinship relation can be unknown as in track 3 of the RFIW challenge. Even though first experiments have shown that average/max pooling of $\boldsymbol{z}_2$ does not yield satisfactory results, we propose to use the task-dependent kinship relation encoder as an attention module, which predicts the kinship relation based on the concatenated feature vector $\boldsymbol{f}_{\text{c}}$ (indicated by the dashed arrow in Fig. \[fig:network\]). The first results indicate that the kinship relation can be correctly identified with an accuracy of at least $65\,\%$.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We present functional-type a posteriori error estimates in isogeometric analysis. These estimates, derived on functional grounds, provide guaranteed and sharp upper bounds of the exact error in the energy norm. [Moreover, since these estimates do not contain any unknown/generic constants, they are fully computable, and thus provide quantitative information on the error.]{} By exploiting the properties of non-uniform rational B-splines, we present efficient computation of these error estimates. The numerical realization and the quality of the computed error distribution are addressed. The potential and the limitations of the proposed approach are illustrated using several computational examples.'
address:
- 'Johannes Kepler University, Altenberger Strasse 69, A-4040 Linz, Austria'
- 'Dornacher Strasse 6/21, A-4040 Linz, Austria'
author:
- 'Stefan K. Kleiss'
- 'Satyendra K. Tomar'
date: 'September 8, 2014, (First version April 21, 2013)'
title: |
Guaranteed and Sharp a Posteriori Error Estimates\
in Isogeometric Analysis
---
Introduction {#sec_intro}
============
The geometry representations in finite element methods (FEM) and computer aided design (CAD) have been developed independent of each other, and are optimized for the purposes within their respective fields. As a consequence, the representations are different from each other, and a transfer of geometry information from CAD to FEM programmes (and vice versa) requires a transformation of geometry data. These transformations are, in general, not only costly, but also prone to approximation errors, and may require manual input.
*Isogeometric analysis* (IGA), introduced by Hughes et al. [@Hughes05_4135], see also [@Coott09_IGA], aims at closing this gap between FEM and CAD. The key observation is that it is a widespread standard in CAD to use geometry representations based on non-uniform rational B-splines (NURBS), and that these NURBS basis functions have properties which make them suitable as basis functions for FEM. Instead of transforming the geometry data to a conventional FEM representation, the original geometry description is used directly, and the underlying NURBS functions are used as basis for the discrete solution. This way, the geometry is represented *exactly* in the sense that the geometry obtained from CAD is not changed. Thus, the need for data transformation is eliminated, and furthermore, the exact representation from the coarsest mesh is preserved throughout the refinement process. IGA has been thoroughly studied and analyzed (see, e.g., [@Bazi06_1031; @Beirao05_271; @Cott07_4160; @Hughes10_301; @Takacs11_3568]), and its potential has been shown by successful applications to a wide range of problems (see, e.g., [@Bazi07_173; @Bazi08_3; @Buffa10_1143; @Elgu08_33; @Niel2011_3242]).
As mentioned above, the most widely used [spline]{} representations in CAD are based on NURBS. The straightforward definition of NURBS basis functions leads to a tensor-product structure of the basis functions, and thus of the discretization. Since naive mesh refinement in a tensor-product setting has global effects, the development of local refinement strategies for isogeometric analysis is a subject of current active research. Such local refinement techniques include, for example, T-splines [@Bazi10_229; @LiEtAl12_63; @Scott11_126; @Scott12_206; @Sede04_276], truncated hierarchical B-splines (THB-splines) [@Giannelli2012_485; @GiannelliJS-14], polynomial splines over hierarchical T-meshes (PHT-splines) [@Deng2008_76; @Wang11_1438], and locally-refineable splines (LR-splines) [@Dokken2013_331; @JohannessenKD-14].
The issue of adaptive, local refinement is closely linked to the question of efficient a posteriori error estimation (see, e.g., [@Ains00_book; @Repin08_book] for a general overview on error estimators). In the light of adaptive refinement, an error estimator has to identify the areas where further refinement is needed due to the local error being significantly larger than in the rest of the domain. Hence, an accurate indication of the error distribution is essential. [Another important objective in computing a posteriori error estimates is to address the *quality assurance*, i.e., to quantify the error in the computed solution with certain degree of *guarantee*.]{} However, a posteriori error estimation in isogeometric analysis is still in an infancy stage. To the best of the authors’ knowledge, the only published results are [@DedeS-12; @Doerfel10_264; @Johannessen11; @KuruVZB-14; @Vuong11_3554; @Wang11_1438; @Xu11_2021; @XuMDG-13; @ZeeV-11].
A posteriori error estimates based on hierarchical bases, proposed by Bank and Smith [@BankSmith_93], have been used in [@Doerfel10_264; @Vuong11_3554]. The reliability and efficiency of this approach is subjected to the saturation assumption on the (enlarged) underlying space and the constants in the strengthened Cauchy inequality. As the authors remarked, the first assumption is critical and its validity depends on the considered example. Moreover, an accurate estimation of constants in the strengthened Cauchy inequality requires the solution of generalized minimum eigenvalue problem. [As noted in [@Johannessen11 Page 41], this approach delivers *less than satisfactory* results.]{}
Residual-based a posteriori error estimates have been used in [@Johannessen11; @Wang11_1438; @Xu11_2021; @XuMDG-13]. This approach requires the computation of constants in Clement-type interpolation operators. Such constants are mesh (element) dependent, often generic/unknown or incomputable for general element shape; and the global constant often over-estimates the local constants, and thus the exact error. [This fact has been explicitly stated by the authors in [@Johannessen11 Pages 42-43] and in [@Wang11_1438 Remark 1]]{}.
Goal-oriented error estimation approach has been studied in [@DedeS-12; @KuruVZB-14; @ZeeV-11]. The results presented in these studies show that neither the estimates of this approach are *guaranteed* to be an upper bound, nor the efficiency indices of the estimates are sharp. Moreover, this approach also requires the solution of an adjoint problem, the cost of which can not be entirely neglected.
The approach of Zienkiewicz-Zhu type a posteriori error estimates is based on post-processing of approximate solutions, and depend on the superconvergence properties of the underlying basis. To the best of authors’ knowledge, superconvergence properties for B-splines (NURBS) functions are not yet known.
Summarily, in general situations, the reliability and efficiency of these methods often depend on undetermined constants, which is not suitable for quality assurance purposes. In this paper, we present *functional-type a posteriori error estimates* for isogeometric discretizations. These error estimates, which were introduced in [@Repin97_201; @Repin99_4311; @Repin00_481] and have been studied for various fields (see [@Repin08_book] and the references therein), provide guaranteed, sharp and fully computable bounds (without any generic undetermined constants). These estimates are derived on purely functional grounds (based on integral identities or functional analysis) and are thus applicable to any conforming approximation in the respective space. For elliptic problems with the weak solution $u\in H_0^1(\Omega)$, these error bounds involve computing an auxiliary function $y \in H(\Omega,\operatorname{div})$. [In order to get a sharp estimate, this function $y$ is computed by solving a *global* problem. This could be perceived as a drawback when compared to error estimation techniques which rely on local computations and are thus apparently cheaper. However, as briefly explained above, our emphasis is not only on adaptivity, but also on *quantifying the error in the computed solution* (and thus guaranteeing the quality of the computed solution). Therefore, the associated cost should be weighed against the stated objectives. To the best of authors’ knowledge, there is no other, particularly cheaper, method available which can fulfill these objectives in general situations. In this paper, we will elaborate how such estimates can be computed efficiently by a proper set-up of the global problem.]{}
Two aspects motivate the application of functional-type error estimates in IGA. Firstly, unlike the standard Lagrange basis functions, NURBS basis functions of degree $p$ are, in general, globally $C^{p-1}$-continuous. Hence, NURBS basis functions of degree $p\geq 2$ are, in general, at least $C^1$-continuous, and therefore, [their gradients are]{} automatically in $H(\Omega,\operatorname{div})$. Thereby, we avoid constructing complicated functions in $H(\Omega,\operatorname{div})$, in particular for higher degrees (see, e.g., [@Buffa11_1407; @BuffaEtAl11_818; @EvaEtAl13_671]). Secondly, since the considered problem is solved in an isogeometric setting, an efficient implementation of NURBS basis functions is readily available, which can be used to construct the above mentioned function $y$. Hence, applying the technique of functional-type a posteriori error estimation in a setting that relies only on the use of already available NURBS basis functions is greatly appealing.
The remainder of this paper is organized as follows. In Section \[sec\_prelims\], we define the model problem, and recall the definition and some important properties of B-spline and NURBS basis functions. [In Section \[sec\_FuncEE\], we first recall functional-type a posteriori error estimates and known implementation issues. Then, we derive a quality criterion and the local error indicator.]{} In Section \[sec\_effcomp\], we discuss a cost-efficient realization of the proposed error estimator using an illustrative numerical example. Further numerical examples are presented in Section \[sec\_numex\], and finally, conclusions are drawn in Section \[sec\_Conc\].
Preliminaries {#sec_prelims}
=============
In order to fix notation and to provide an overview, we define the model problem and recall the definition and some aspects of isogeometric analysis in this section.
Model Problem
-------------
Let $\Omega \subset {\mathbb{R}}^2$ be an open, bounded and connected Lipschitz domain with boundary $\partial \Omega$. We shall consider the following model problem:
Find the scalar function $u:\overline{\Omega} \rightarrow {\mathbb{R}}$ such that$$\begin{array}{rcl@{\qquad}l}
- \operatorname{div}( A \nabla u) &=& f & \text{in~} \Omega,\\
u &=& u_D & \text{on~} \Gamma_D = \partial \Omega,
\end{array}
\label{e_defpde}$$ where $A$, $f$ and $u_D$ are given data. We assume that $A$ is a symmetric positive definite matrix and has a positive inverse $A^{-1}$, and that there exist constants $c_1,c_2>0$ such that $$\begin{aligned}
c_1 |\xi|^2 \leq A\xi \cdot \xi \leq c_2 |\xi|^2,\quad \forall \xi \in {\mathbb{R}}^2.
\label{e_Abounds}\end{aligned}$$ Then, the norms $$\| v \|_A^2 = \int_\Omega A v \cdot v {\;dx}, \quad
\| v \|_{\bar{A}}^2 = \int_\Omega A^{-1} v \cdot v {\;dx}, \label{e_normaabar}$$ are equivalent to the $L^2$-norm $\| v \|^2 = \int_\Omega v \cdot v {\;dx}$. The weak form of problem can be written as follows:
Find $u \in V_g$, such that $$\begin{aligned}
a(u,v) = f(v), \quad \forall v\in V_0,
\label{e_defvarp}\end{aligned}$$ where $V_0 \subset H^1(\Omega)$ contains the functions which vanish on $\Gamma_D$, and $V_g \subset H^1(\Omega)$ contains the functions satisfying the Dirichlet boundary conditions $u=u_D$ on $\Gamma_D$. We assume that the problem data $A$, $f$ and $u_D$ are given such that the bilinear form $a(\cdot,\cdot)$ is bounded, symmetric and positive definite, and that $f(\cdot)$ is a bounded linear functional. The energy norm of a function $v$ is given by $\| \nabla v \|_A = \sqrt{a(v,v)}$. Note that we have considered the Dirichlet problem only for the sake of simplicity. Functional-type error estimates can be easily generalized to problems with mixed boundary conditions, see, e.g., [@Lazarov2009_952; @Repin08_book]. We discretize the problem in the standard way by choosing a finite-dimensional manifold $V_h \subset V_g$ and looking for a *discrete solution* $u_h \in V_h$. This leads to a linear system of equations of the form $$\begin{aligned}
\underline{K}_h \underline{u}_h = \underline{f}_h,
\label{e_Khuhfh}\end{aligned}$$ where $\underline{K}_h$ is the stiffness matrix induced by the bilinear form $a(\cdot,\cdot)$, $\underline{f}_h$ is the load vector, and $\underline{u}_h$ is the coefficient vector of the [discrete solution $u_h$]{}.
B-Splines, NURBS and Isogeometric Analysis {#sec_Bsplines}
------------------------------------------
We briefly recall the definition of B-spline basis functions and NURBS mappings. We only provide the basic definitions and properties relevant for the scope of this paper. For detailed discussions of NURBS basis functions, geometry mappings and their properties, we refer to, e.g., [@Coott09_IGA; @Cott07_4160; @Hughes05_4135; @Piegl1997] and the references therein. The following standard definitions and statements can also be found there.
Let $p$ be a non-negative *degree* and let $s = (s_1,\ldots,s_{m})$ be a *knot vector* with $s_i \leq s_{i+1}$ for all $i$. We consider only *open knot vectors*, i.e., knot vectors $s$ where the multiplicity of a knot is at most $p$, except for the first and last knot which have multiplicity $p+1$. For simplicity, we assume that $s_1 = 0$ and $s_{m}=1$, which can be easily achieved by a suitable scaling. The $n = m-p-1$ univariate *B-spline basis functions* $B_{i,p}^{s}: (0,1)\rightarrow {\mathbb{R}}$, $i=1,\ldots,n$, are defined recursively as follows: $$\begin{aligned}
B_{i,0}^{s}(\xi) &=&
\left\{\begin{array}{c@{\quad}l@{\quad}l}
1 & \mathrm{for} & s_i \leq \xi < s_{i+1} \\
0 & \mathrm{else}
\end{array}\right.\\
B_{i,p}^{s}(\xi) &=&
\frac{\xi-s_{i}}{s_{i+p}-s_{i}} B_{i,p-1}^{s}(\xi)
+\frac{s_{i+p+1}-\xi}{s_{i+p+1}-s_{i+1}} B_{i+1,p-1}^{s}(\xi).\end{aligned}$$ Whenever a zero denominator appears in the definition above, the corresponding function $B^s_{i,p}$ is zero, and the whole term is considered to be zero. For open knot vectors, the first and last basis function are interpolatory at the first and the last knot, respectively. The derivatives of B-spline basis functions are given by the following formula: $$\begin{aligned}
\partial_\xi B_{i,p}^s(\xi)
&=&
\frac{p}{s_{i+p}-s_i} B_{i,p-1}^s(\xi) - \frac{p}{s_{i+p+1}-s_{i+1}} B_{i+1,p-1}^s(\xi).\end{aligned}$$
B-spline basis functions of degree $p$ are, in general, globally $C^{p-1}$-continuous. In the presence of repeated knots, the continuity reduces according to the multiplicity, i.e., if a knot appears $k$ times, the continuity of a B-spline basis function of degree $p$ at that knot is $C^{p-k}$.
Let $\{B_{i,p}^s\}_{i=1}^{n_1}$ and $\{ B_{j,q}^t \}_{j=1}^{n_2}$ be two families of B-spline basis functions defined by the degrees $p$ and $q$, and the open knot vectors $$s =(s_1,\ldots,s_{n_1+p+1}),\
t =(t_1,\ldots,t_{n_2+q+1}),$$ respectively. We denote the set of all double-indices $(i,j)$ by $${\mathcal{I}}_R = \{(i,j):\ i\in\{1,\ldots,n_1\}, j\in\{1,\ldots,n_2\}\}.$$ Let $w_{(i,j)}$, $(i,j)\in{\mathcal{I}}_R$, be positive *weights*. The *bivariate NURBS basis functions* $R_{(i,j)}(\xi_1,\xi_2)$, $(i,j)\in{\mathcal{I}}_R$ are defined as follows: $$\begin{aligned}
R_{(i,j)}(\xi_1,\xi_2) &=&
\frac{ { w_{(i,j)} {\;}B_{i,p}^s(\xi_1) {\;}B^t_{j,q}(\xi_2) } }{ \sum_{(k,\ell)\in{\mathcal{I}}_R } w_{(k,\ell)} B_{k,p}^s(\xi_1) {\;}B^t_{\ell,q}(\xi_2) }.\end{aligned}$$ The continuity of the B-spline basis functions is inherited by the NURBS basis functions. Note that B-splines can be seen as a special case of NURBS with all weights being equal to one. Hence, we will not distinguish between these two and we will only use the term *NURBS* in the remainder of the paper.
The set of functions $$\hat{V}_h = \operatorname{span}\{ R_{(i,j)},\ (i,j)\in{\mathcal{I}}_R \},$$ associated with the *parameter domain* $\hat{\Omega} = (0,1)^2$, is uniquely determined by the degrees $p$ and $q$, the knot vectors $s$ and $t$, and the weights $w$. To reflect the associated polynomial degrees in respective dimensions, we will also use the notation ${\mathcal{S}}^{p,q}_{h}$ for $\hat{V}_h$, which denotes the NURBS function of degree $p$ and $C^{p-1}$-continuity in the first coordinate, degree $q$ and $C^{q-1}$-continuity in the second coordinate, and where the parameter $h$ is the characteristic cell size (non-vanishing knot-span) of the mesh for $\hat{V}_h$.
Given the set of functions $\hat{V}_h$ and a *control net* of *control points* $P_{(i,j)} \in {\mathbb{R}}^2$, where $(i,j)\in {\mathcal{I}}_R$, the two-dimensional *NURBS-surface* $G:\hat{\Omega}\rightarrow \Omega$ is defined by $$G(\xi_1,\xi_2) =
\sum_{(i,j)\in{\mathcal{I}}_R} R_{(i,j)}(\xi_1,\xi_2){\;}P_{(i,j)}.
\label{glg_NURBS_Surf}$$ We refer to $\Omega = G(\hat{\Omega})$ as the *physical domain*. We assume that the geometry mapping is continuous and bijective (i.e., not self-penetrating), which are natural assumptions for CAD-applications.
In isogeometric analysis, the isoparametric principle is applied by using the same basis functions for the discrete solution $u_h$ which are used for representing the geometry. For detailed discussion, we refer the reader to, e.g., [@Coott09_IGA; @Cott07_4160; @Hughes05_4135]. The discrete solution $u_h$ on the physical domain $\Omega$ is represented as follows: $$\begin{aligned}
u_h(x) = \sum_{(i,j)\in{\mathcal{I}}_R} u_{(i,j)}\ \big({R}_{(i,j)} \circ G^{-1}\big)(x),\end{aligned}$$ where $u_{(i,j)}\in {\mathbb{R}}$ are real-valued coefficients which form the coefficient vector $\underline{u}_h$. The discrete functions space is thus defined by $$V_h = \operatorname{span}\{ R_{(i,j)} \circ G^{-1},\ (i,j)\in {\mathcal{I}}_R\}.$$
The initial mesh, and thereby the basis functions on this initial mesh, are assumed to be given via the geometry representation of the computational domain, i.e., the initial discretization is already determined by the problem domain. The exact representation of the geometry on the initial (coarsest) level is preserved in the process of mesh refinement.
As mentioned in the introduction, the straightforward definition of NURBS basis functions, leads to a tensor-product structure of the discretization, which is the focus of this paper. Nevertheless, the error estimator presented herein is also applicable to local refinement techniques (e.g., T-splines, THB-splines, PHT-splines, LR-splines, see Section \[sec\_intro\]) since it is derived purely on functional grounds.
Functional-type a Posteriori Error Estimates {#sec_FuncEE}
============================================
[In the first two parts of this section, we will discuss the well-known theoretical upper bound for the error in the energy norm (see, e.g., [@Repin97_201; @Repin99_4311; @Repin00_481; @Repin08_book]), and we recall how to minimize this upper bound in order to get a sharp error estimate (see, e.g., [@Kraus11_1175; @Lazarov2009_952]). Thereafter, in Section \[sec\_qualcrit\], we will derive a quality criterion from the discussed theory.]{} We will comment on the realization in the isogeometric context in Section \[sec\_effcomp\].
Guaranteed Upper Bound for the Error {#sec:Majorant}
------------------------------------
The starting point for the proposed method is the following main result, which gives an upper bound for the error in the energy norm. It can be found, e.g., in [@Repin99_4311; @Repin00_481; @Repin08_book].
\[thm\_est1\] Let $C_\Omega$ be the constant in the Friedrich’s type inequality $\| v \| \leq C_{\Omega} \| \nabla v \|_A,\ \forall v \in V_0$. Let $u$ be the exact solution of the problem , and let $u_h\in V_h$ be an approximate solution. Then, the following estimate holds: $$\| \nabla u - \nabla u_h \|_A
\leq
\| A \nabla u_h - y \|_{\bar{A}} + C_\Omega \| \operatorname{div}y + f \|,
\label{e_DefMoplusRaw}$$ where $y$ is an arbitrary vector-valued function in $H(\Omega,\operatorname{div})$, and the norms are as defined in .
The constant $C_\Omega$ depends only on the domain $\Omega$ and the coefficient matrix $A$ (but not on the underlying mesh), see, e.g., [@Lazarov2009_952; @Repin08_book]. Note that $C_\Omega$ can be computed either numerically or, if one can find a domain $\Omega_\square \supset \Omega$, where $\Omega_\square$ is a square domain with side-length $\ell$, then $C_\Omega \leq c_2 \tfrac{\ell}{\pi \sqrt{d}}$, where $d$ is the dimension and $c_2$ is the constant in .
Note that, if we choose $y$ via the (unknown) exact solution $y=A\nabla u$, both sides of coincide. Hence, the estimate is sharp in the sense that, for any fixed $u_h$, we can find a function $y$ such that the upper bound is as close to the exact error as desired. The estimate given in Theorem \[thm\_est1\] is a guaranteed and fully computable upper bound for any conforming approximation $u_h \in V_{g}$.
In the following, we describe some approaches to construct the function $y$ and discuss their relative merits. For this reason, we consider a numerical example, referred to as *Example \[ex:sin\_6pix\_3piy\]* in the remainder, whose solution is a smoothly varying function in both directions.
\[ex:sin\_6pix\_3piy\] **Sinus function in a unit square:** In this numerical example, the computational domain is the unit square $\Omega = (0,1)^2$ and $u_h \in {\mathcal{S}}^{2,2}_{h}$, i.e., a piecewise quadratic function in both directions. The coefficient matrix is the identity matrix, i.e., $A=I$, and the exact solution is given by $$u = \sin(6\pi x) \sin(3\pi y).$$ The right-hand-side $f$ and the (homogeneous) boundary conditions $u_D$ are determined by the prescribed exact solution $u$.
### Post-processing of {#sec:PostProc}
It is possible to obtain good error *indicators* by constructing a function $y$ by some post-processing of the discrete solution $u_h$, see [@Lazarov2009_952; @Repin08_book] and the references therein. Since $u_{h} \in C^{p-1}$, we have $\nabla u_{h} \in (C^{p-2})^{2} \subset H(\Omega,\operatorname{div})$ for $p \ge 2$. Choosing $y = \nabla u_{h}$ will thus result in $$\begin{aligned}
\Vert \nabla u - \nabla u_h \Vert \leq C_{\Omega} \Vert \Delta u_{h} + f \Vert.
\label{eq:yh_graduh}\end{aligned}$$ Once we have calculated $\eta_Q := \Vert \Delta u_{h} + f \Vert_{Q}$ for each cell $Q$ of the mesh, we can compare the local errors and choose a criterion for selecting cells which will be marked for further refinement. Typically, one chooses a threshold $\Theta$ and marks all cells $Q$ for refinement, where the local error is above this threshold. There are several possibilities for determining $\Theta$, e.g., the bulk-criterion proposed in [@Doerfler96_1106]. For simplicity, we choose a percentage $\psi$ and mark a cell $Q$ for refinement, if $$\eta_Q > \Theta ,\text{~where~} \Theta = (100-\psi)\text{-percentile~of~}\{\eta_Q\}_Q.
\label{e_0228a}$$ The $\alpha$-percentile of a set ${\mathcal{A}}=\{a_1,\ldots,a_\nu\}$ denotes the value $\bar{a}$ below which $\alpha$ percent of all values $a_i$ fall. For example, if we choose $\psi = 20\%$ in , then $\Theta$ is chosen such that $n_Q>\Theta$ holds for 20% of all cells $Q$.
![Convergence of exact error and the majorant for Example \[ex:sin\_6pix\_3piy\].\[fig\_Conv\_EstErr\]](Figures/E1_YhGradUh_EstErr)
To show the efficiency of the estimator , in Figure \[fig\_nexSU\_Ex\], we present the cells marked for refinement by the exact error. The cells marked for refinement by the majorant given in are presented in Figure \[fig\_nexSU\_graduh\]. We see that starting from the mesh $32 \times 32$, the majorant is able to nicely capture the refinement pattern of exact error. However, from a closer look at the convergence of the exact error and the majorant, see Figure \[fig\_Conv\_EstErr\], we find that though such an estimate is a guaranteed upper bound and very cheap to compute, it over-estimates the exact error, and its convergence is slower than the exact error (due to a lack of proper scaling, different operators acting on $u_{h}$ on both sides). [^1]
### Global minimization {#sec:Minimization}
In order to obtain a sharp estimate (and not just an indicator), therefore, one has to find a function $y$ which minimizes the right-hand-side of . For minimizing the estimate numerically, we first rewrite the estimate in the following form $$\| \nabla u - \nabla u_h \|_A^2
\leq
(1+\beta) \| A \nabla u_h - y \|_{\bar{A}}^2 + (1+\tfrac{1}{\beta}) C_\Omega^2 \| \operatorname{div}y + f \|^2
\ =:\ M_\oplus^2(y,\beta),
\label{e_DefMoplus}$$ where $\beta > 0$ is a free parameter [@Lazarov2009_952; @Repin08_book]. Note that the upper bound in holds true for *any* fixed $y\in H(\Omega,\operatorname{div})$ and $\beta>0$. Hereinafter, for simplicity, we will refer to $M_\oplus^2(y,\beta)$ as the *majorant*. Introducing $$\begin{array}{r@{\ =\ }l@{\qquad}r@{\ =\ }l}
a_1 & 1+\beta, &
a_2 & (1+\tfrac{1}{\beta})C_\Omega^2,\\
B_1 & \displaystyle \| A\nabla u_h - y \|_{\bar{A}}^2, &
B_2 & \displaystyle \| \operatorname{div}y + f\|^2,
\end{array}\label{e_defaB}$$ we can briefly write the majorant as $$M_\oplus^2(y,\beta) = a_1 B_1 + a_2 B_2.$$ The *efficiency index*, defined by $$\begin{aligned}
I_{\text{eff}} = \frac{M_\oplus(y,\beta)}{\| \nabla u - \nabla u_h \|_A},
\label{eq:Def_Ieff}\end{aligned}$$ indicates how close the calculated majorant is to the exact error. The closer $I_{\text{eff}}$ is to 1, the better the estimate. Therefore, obtaining a *sharp* estimate requires to find $y\in H(\Omega,\operatorname{div})$ and $\beta >0$ as solutions to the global minimization problem $$\min_{y\in H(\Omega,\operatorname{div}),\ \beta >0} M_\oplus^2(y,\beta).$$ The technique for finding such minimizing parameters $y$ and $\beta$ will be discussed in Sections \[sec\_minmaj\] and \[sec\_NumRealIgA\]. Before proceeding further, we give the following Lemma \[lem\_Moconv\], which can be found in [@Repin08_book Prop. 3.10]. It provides an analytical result on the sharpness of the bound $M_\oplus^2(y,\beta)$. For later reference, we also sketch the proof.
A sequence of finite-dimensional subspaces $\{ Y_j \}_{j=1}^\infty$ of a Banach-space $Y$ is called *limit dense in $Y$*, if for any $\varepsilon >0$ and any $v\in Y$, there exists an index $j_\varepsilon$, such that $\inf_{p_k\in Y_k} \|p_k - v \|_{Y} < \varepsilon$ for all $k>j_\varepsilon$ .
\[lem\_Moconv\] Let the spaces $\{Y_j\}_{j = 1}^\infty$ be limit dense in $H(\Omega,\operatorname{div})$. Then $$\lim_{j\to\infty}\ \inf_{y_j\in Y_j, \beta > 0}\ M_\oplus^2(y_j,\beta) = \| \nabla u - \nabla u_h \|_A^2.$$
Recall that the $H(\Omega,\operatorname{div})$-norm $\|\cdot\|_{\operatorname{div}}$ is defined by $\|v\|_{\operatorname{div}}^2 = \|v\|^2 + \|\operatorname{div}v\|^2$. Let $\varepsilon > 0 $ be arbitrarily small, but fixed. Let $j_\varepsilon$ be the index such that, for all $k > j_\varepsilon$, there exists a $p_k \in Y_k$ with $\|A \nabla u -p_k\|_{\operatorname{div}} < \varepsilon$. Then, $$\inf_{y_j\in Y_j, \beta > 0}\ M_\oplus^2(y_j,\beta)
\leq
\ M_\oplus^2(p_k,\varepsilon)
=
(1+\varepsilon) {\| A \nabla u_h - p_k \|_{\bar{A}}^2}
+ ( 1 + \tfrac{1}{\varepsilon}) C_\Omega^2 { \| f + \operatorname{div}p_k \|^2 }.
\label{e_0212c}$$ Since $\| A v \|_{\bar{A}} = \| v\|_A$, we can write $$\begin{aligned}
\| A \nabla u_h - p_k \|_{\bar{A}}
&\leq &
\| A \nabla u_h - A \nabla u \|_{\bar{A}} + \| A \nabla u - p_k \|_{\bar{A}} \\
&=&
\| \nabla u_h - \nabla u \|_{A} + \| A \nabla u - p_k \|_{\bar{A}}.\end{aligned}$$ The norm $\| \cdot \|_{\bar{A}}$ is equivalent to the $L^2$-norm, so there exists a constant $c_A$, such that the second term in the right-hand side can be bounded by $$\| A \nabla u - p_k \|_{\bar{A}}
\leq c_A \| A \nabla u - p_k \|
\leq c_A \| A \nabla u - p_k \|_{\operatorname{div}}
\leq c_A \varepsilon.$$ Hence, we obtain the following estimate for the first term in : $${\| A \nabla u_h - p_k \|_{\bar{A}}} \leq
\| \nabla u - \nabla u_h \|_A + {\mathcal{O}}(\varepsilon).
\label{e_0212a}$$ Since $f = -\operatorname{div}A \nabla u$, we can bound the second term in as follows: $$\| \operatorname{div}p_k + f \|
=
\| \operatorname{div}p_k - \operatorname{div}A \nabla u \|
\leq
\| p_k - A \nabla u \|_{\operatorname{div}} \
\leq \ \varepsilon.
\label{e_0212b}$$ With and , we can rewrite as $$\begin{aligned}
M_\oplus^2(p_k,\varepsilon)
\leq (1+\varepsilon) ( \| \nabla u - \nabla u_h \|_A^2 + {\mathcal{O}}(\varepsilon) )
+ (1+\tfrac{1}{\varepsilon}) C_\Omega^2 \varepsilon^2
= \| \nabla u - \nabla u_h \|_A^2 + {\mathcal{O}}(\varepsilon).\end{aligned}$$ Hence, the bound $M_\oplus^2(p_k,\varepsilon) \rightarrow \| \nabla u - \nabla u_h \|_A^2 $ as $\varepsilon \rightarrow 0$.
Steps Involved in Minimizing {#sec_minmaj}
-----------------------------
As mentioned above, we need to find parameters $y$ and $\beta$ which minimize the majorant. To do this, we apply an interleaved iteration process in which we alternately fix one of the variables and minimize with respect to the other. This process, which we summarize in the following, has been described, e.g., in [@Kraus11_1175; @Lazarov2009_952].
Step 1
: Minimization with respect to $y$: Assume that $\beta >0$ is given and fixed, either by an initial guess or as a result of Step 2 below. We view the majorant $M_\oplus^2(y)$ as a quadratic function of $y$ and calculate its Gateaux-derivative $M_\oplus^2(y)'$ with respect to $y$ in direction $\tilde{y}$. Setting $M_\oplus^2(y)' = 0$, we obtain $$\begin{aligned}
a_1 \int_\Omega A^{-1} y \cdot \tilde{y} {\;dx}+ a_2 \int_\Omega \operatorname{div}y\ \operatorname{div}\tilde{y} {\;dx}&=&
a_1 \int_\Omega \nabla u_h \cdot \tilde{y} {\;dx}- a_2 \int_\Omega f\ \operatorname{div}\tilde{y} {\;dx},
\label{e_minwrty}\end{aligned}$$ where $a_1 = 1+\beta$ and $a_2 = (1+\tfrac{1}{\beta})C_\Omega^2$, as defined in . In order to solve , we choose a finite-dimensional subspace $Y_h\subset H(\Omega,\operatorname{div})$ and search for a solution $y_h \in Y_h$. Testing in all directions $\tilde{y}\in Y_h$ leads to a linear system of equations which we write as $$\begin{aligned}
\underline{L}_h \underline{y}_h = \underline{r}_h.
\label{e_Lhyhrh}\end{aligned}$$ Here, $\underline{L}_h$ and $\underline{r}_h$ are the matrix and the vector induced by the left hand side and the right hand side of equation , respectively. By solving , we obtain the coefficient vector $\underline{y}_h$ for the discrete function $y_h$ minimizing $M_\oplus^2(y)$ in $Y_h\subset H(\Omega,\operatorname{div})$. Note that this process requires non-negligible cost as we need to assemble $\underline{L}_h$ and $\underline{r}_h$ and solve the system .
Step 2
: Minimization with respect to $\beta$: Assume that $y_h$ is given from Step 1. By direct calculation, we see that $M_\oplus^2(\beta)$ is minimized with respect to $\beta$ by setting $$\begin{aligned}
\beta &=& C_{\Omega} \sqrt{\frac{B_2}{B_1}},
\label{e_minwrtb}\end{aligned}$$ where $B_1$ and $B_2$ are as defined in . Note that the evaluation of $B_1$ and $B_2$ (and thus $\beta$) requires only the evaluation of integrals, and thus involves negligible cost.
Steps 1 and 2 are repeated iteratively. We will refer to one loop of applying Step 1 and Step 2 as one *interleaved iteration*. Once we have computed minimizers $y_h$ and $\beta$, the computation of the majorant $M_\oplus^2(y_h,\beta)$ is straight-forward as it requires only the evaluation of the integrals.
Note that the matrix $\underline{L}_h$ can be written as $$\underline{L}_h = a_1 \underline{L}_h^1 + a_2 \underline{L}_h^2,
\label{e_0319a}$$ where $\underline{L}_h^1$ and $\underline{L}_h^2$ correspond to the terms $\int_\Omega A^{-1} y \cdot \tilde{y} {\;dx}$ and $\int_\Omega \operatorname{div}y\ \operatorname{div}\tilde{y} {\;dx}$ in , respectively. Since the matrices $\underline{L}_h^1$ and $\underline{L}_h^2$ in do not change in the interleaved iteration process, they need to be assembled only once. Analogously to , we can write $\underline{r}_h$ as $$\begin{aligned}
\underline{r}_h = a_1 \underline{r}_h^1 - a_2 \underline{r}_h^2,
\label{eq:vec_rh}\end{aligned}$$ where $\underline{r}_h^1$ and $\underline{r}_h^2$ correspond to the terms $\int_\Omega \nabla u_h \cdot \tilde{y}_h {\;dx}$ and $\int_\Omega f \ \operatorname{div}\tilde{y} {\;dx}$ in , respectively. The terms $\underline{r}_h^1$ and $\underline{r}_h^2$ also need to be assembled only once since they also do not change in the interleaved iteration process. The full matrix $\underline{L}_h$ and vector $\underline{r}_h$, however, do change in each iteration, because of the change in $\beta$ and $y_h$. Based on past numerical studies, see, e.g., [@Kraus11_1175; @Lazarov2009_952], and the results presented in Sections \[sec\_effcomp\] and \[sec\_numex\], it has been found that for linear problems, one or two such interleaved iterations are enough for obtaining a sufficiently accurate result.
To recapitulate, we summarize the steps for computing the majorant in Algorithm \[algo:Mplus\].
$u_h$, $f$, $C_\Omega$, $Y_h$ $M_\oplus$ $\beta:=$ initial guess Assemble and store $\underline{L}_h^1$, $\underline{L}_h^2$, $\underline{r}_h^1$, $\underline{r}_h^2$ $\underline{L}_h := (1+\beta) \underline{L}_h^1 + (1+\tfrac{1}{\beta}) C_\Omega^2 \underline{L}_h^2$ $\underline{r}_h := (1+\beta) \underline{r}_h^1 - (1+\tfrac{1}{\beta}) C_\Omega^2 \underline{r}_h^2$ Solve $\underline{L}_h \underline{y}_h = \underline{r}_h$ for $\underline{y}_h$ $B_1 := \| A \nabla u_h - y_h \|^2_{\bar{A}}$ $B_2 := \| \operatorname{div}y_h + f \|^2$ $\beta := C_\Omega \sqrt{ B_2 / B_1 }$ $M_\oplus(y,\beta) := \sqrt{(1+\beta)B_1 + (1+\tfrac{1}{\beta})C_\Omega^2 B_2}$
\[rem\_yglobal\] Note that the space $H(\Omega,\operatorname{div})$, where the auxiliary quantity $y$ is sought, is a global space, and for a general complicated problem, it is not immediately clear how to locally compute $y$ without global effect. That being said, a local version of our estimator can be devised for specific problems and data (like equilibration of flux approach), however, that will restrict its generality, which is not very appealing to us. Therefore, in the remainder of the paper, we will focus on computing the majorant from the global minimization problem.
Quality Indicator and Local Error Indicator {#sec_qualcrit}
-------------------------------------------
So far, we have defined the majorant and discussed how we minimize (numerically) the majorant over $Y_h$. Another important question, especially in the light of adaptive, local refinement, is whether a calculated majorant does correctly capture the error distribution. From the proof of Lemma \[lem\_Moconv\], we recall the following observation: $$a_1 B_1 \to \| \nabla u - \nabla u_h \|_A^2
\text{~and~}
a_2 B_2 \to 0,
\text{~as~} y_h \in H(\Omega,\operatorname{div}) \to A\nabla u.
\label{e_1stCompConv}$$ From this, we deduce the following quality indicator.
The distribution of the exact error is captured correctly, if $$a_1 B_1 > C_\oplus \ a_2 B_2 \label{e_reli}$$ with some constant $C_\oplus > 1$.
This criterion is easy to check, since the terms appearing in are evaluated in the process of minimizing $M_\oplus^2(y,\beta)$. It was found from extensive numerical studies (see examples presented in Sections \[sec\_effcomp\] and \[sec\_numex\]) that an accurate distribution of the error is obtained for $C_\oplus \ge 5$.
\[rem\_Cplus\] For the choice of $C_{\oplus} \ge 5$, we have $a_2 B_2 < a_1 B_1 /5$, and therefore, $\Vert \nabla u - \nabla u_h \Vert_A \le \sqrt{1.2~ a_1 B_1}$. One can see from all the tables in Sections \[sec\_effcomp\] and \[sec\_numex\], that whenever this criterion is satisfied, we have $I_{\text{eff}} \le 1.2$ (the ratio of $\sqrt{a_1 B_1}/\Vert \nabla u - \nabla u_h \Vert_A$ appears to be of the same magnitude as $\sqrt{1 + 1/C_{\oplus}}$. Note that this criterion does not require $a_2 B_2$ to be close to zero, but just less than $a_1 B_1 /5$. Since these approximations (of the original problem and the auxiliary problem in $H(\Omega, \operatorname{div})$) are monotonically convergent, the approximation at any level will only improve at the next refinement level, and this is why the results get better for any further refinement. Clearly, all the terms are fully computable, and thus, usable in an algorithm.
[We define the local error indicator $\eta_Q$ on a cell $Q$ as the restriction of the first component of the majorant to the cell $Q$, i.e., by $$\begin{aligned}
\eta_Q^2(y_h) = \int_Q (\nabla u_h - A^{-1}y_h ) (A\nabla u_h - y_h ){\;dx}.
\label{e_errest}\end{aligned}$$ The factor $(1+\beta)$ is omitted, since this scalar factor is the same for all cells of the domain. As remarked in the observation , the first component will converge to the exact error, thus providing a good indicator for the error distribution. A more detailed discussion of this indicator can be found in [@Repin08_book Sec. 3.6.4].]{} For refinement based on $\eta_{Q}$, we again use the criterion .
Efficiency and Computational Cost of the Proposed Estimator in the Isogeometric Context {#sec_effcomp}
=======================================================================================
We now discuss the efficiency and the computational cost of the proposed estimator based on the global minimization steps presented in Section \[sec\_minmaj\]. Through out this Section, we again consider Example \[ex:sin\_6pix\_3piy\] from Section \[sec:Majorant\]. All the computations for this example and the examples presented in Section \[sec\_numex\] are performed in [MATLAB]{}$^\circledR$ on an HP workstation Z420 with Intel Xeon CPU E5-1650, 3.2 GHz, 12 Cores and 16 GB RAM, and the linear systems and are solved using the in-built direct solver. The right-hand-side $f$ and the boundary conditions $u_D$ are determined by the prescribed exact solution $u$. We study the efficiency of the majorant based on *straight forward* computational procedure, as discussed in Section \[sec\_proc0\], and based on *cost-efficient* procedure, as discussed in Section \[sec\_NumRealIgA\], which coarsens the mesh and increases the polynomial degree simultaneously. This alternative cost-efficient procedure will then be used in Section \[sec\_numex\] for further numerical examples. In all the numerical results of Example \[ex:sin\_6pix\_3piy\] in this Section, the initial guess for $\beta$ is $0.01$.
In the tables, we indicate the mesh-size by the number of interior knot spans of the knot vectors $s$ and $t$, respectively. By this, we mean the number of knot spans without counting the vanishing knot spans at the beginning and the end of the open knot vectors. For example, if $$\begin{aligned}
s&=&(0,0,0,0.25,0.5,0.75,1,1,1)\\
t&=&{(0,0,0,0,0.5,0.5,1,1,1,1)},\end{aligned}$$ then the mesh-size is $4 \times {3}$[, since the empty knot span $(0.5,0.5)$ in $t$ is also counted as an interior knot span]{}.
We compare the timings for assembling and for solving the linear systems and , as well as the total time for assembling and solving. In the presented tables, these timings are shown in the columns labeled assembling-time, solving-time, and sum, respectively. The label *pde* indicates that the column corresponds to solving the partial differential equation , i.e., to assembling $\underline{K}_h$ and solving for $\underline{u}_h$. The label *est* indicates that the timings correspond to the estimator, i.e, assembling $\underline{L}_h$ and solving for $\underline{y}_h$. [In the column labeled $\frac{\emph{est}}{\emph{pde}}$, we present the ratio of these timings. Note that these ratios were computed *before* rounding the numbers, i.e., taking the ratios of the reported numbers may result in slightly different values.]{}
[The computed efficiency indices $I_{\text{eff}}$ (see ) are presented in tables.]{} In order to check the quality criterion discussed in Section \[sec\_qualcrit\], we present the values of $a_1B_1$ and $a_2 B_2$ and see whether the inequality is fulfilled or not. To indicate the quality of the error distribution captured by the majorant, we plot which cells are marked for refinement based on the exact local error and the criterion , and compare this to the refinement marking based on the criterion applied to the computed error estimate [(plotted in magenta)]{}.
Straightforward Procedure {#sec_proc0}
-------------------------
\[ex:case\_0\] **(Straightforward Procedure)** For the first choice for $\hat{Y}_h$, we use the same mesh as for $\hat{V}_h$, and choose $$\hat{Y}_h = {\mathcal{S}}^{p+1,p}_{h} \otimes {\mathcal{S}}^{p,p+1}_{h}.
\label{e_Yh0}$$ The function space $Y_h$ is then defined by the well known Piola transformation [@BoffiBF-13].
We consider the same setting as presented in Example \[ex:sin\_6pix\_3piy\] in Section \[sec:Majorant\]. In Table \[nexSUeff\_0\], we present the computed efficiency indices obtained with this choice of $Y_h$, which show that upper bound approaches $1$ (representing exact error) as the mesh is refined. The dashed line in Table \[nexSUeff\_0\] indicates that the criterion is fulfilled with $C_\oplus = 5$ (actually $4.94$) starting from the mesh $64 \times 64$. The cells marked by the error estimator are shown in Figure \[fig\_nexSU\_0\]. When comparing these plots [to those presented in Figure \[fig\_nexSU\_Ex\]]{}, we see that the error distribution is captured accurately starting from the mesh $32\times 32$.
The timings presented in Table \[nexSUtime\_0\], however, show that the computation of the error estimate is costlier (about $4.5$ times) than assembling and solving the original problem. This is not surprising, since, when $N_u$ denotes the number of degrees of freedom (DOF) of $u_h$, the number of DOF of $y_h$, which is vector-valued, is asymptotically $2N_u$. This results in higher assembly time and the solution time for the linear system (where a direct solver is used). Clearly, this straightforward approach is not cost-efficient. In the next section, therefore, we discuss some cost-efficient approaches for computing $y_{h}$.
mesh-size
------------------ ------------ ---------- ----------
$8 \times 8$ [ 3.43 ]{} 2.62e+01 1.17e+02
$16 \times 16$ [ 1.92 ]{} 6.07e-01 6.19e-01
$32 \times 32$ [ 1.41 ]{} 2.29e-02 9.71e-03
$64 \times 64$ [ 1.20 ]{} 1.15e-03 2.33e-04
$128 \times 128$ [ 1.10 ]{} 6.51e-05 6.54e-06
$256 \times 256$ [ 1.05 ]{} 3.87e-06 1.95e-07
$512 \times 512$ [ 1.03 ]{} 2.36e-07 5.94e-09
: Efficiency index and components of the majorant in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_0\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{3,2}_{h} \otimes {\mathcal{S}}^{2,3}_{h}$.\[nexSUeff\_0\]
------------------ -------- -------- -------- -------- ------------ --------- --------- ------------ -------- -------- ------------ --
mesh-size
$u_h$ $y_h$
$8 \times 8$ 100 220 0.04 0.17 [ 4.39 ]{} $<$0.01 $<$0.01 [ 5.16 ]{} 0.04 0.17 [ 4.40 ]{}
$16 \times 16$ 324 684 0.14 0.59 [ 4.25 ]{} $<$0.01 0.01 [ 5.39 ]{} 0.14 0.60 [ 4.26 ]{}
$32 \times 32$ 1156 2380 0.46 2.17 [ 4.70 ]{} 0.01 0.03 [ 4.71 ]{} 0.47 2.20 [ 4.70 ]{}
$64 \times 64$ 4356 8844 1.82 8.51 [ 4.68 ]{} 0.03 0.20 [ 6.15 ]{} 1.85 8.70 [ 4.70 ]{}
$128 \times 128$ 16900 34060 7.38 34.19 [ 4.63 ]{} 0.15 0.87 [ 5.70 ]{} 7.54 35.06 [ 4.65 ]{}
$256 \times 256$ 66564 133644 33.30 149.78 [ 4.50 ]{} 0.84 5.66 [ 6.78 ]{} 34.14 155.44 [ 4.55 ]{}
$512 \times 512$ 264196 529420 191.11 766.10 [ 4.01 ]{} 3.77 33.92 [ 9.00 ]{} 194.88 800.03 [ 4.11 ]{}
------------------ -------- -------- -------- -------- ------------ --------- --------- ------------ -------- -------- ------------ --
: Number of DOF and timings in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_0\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{3,2}_{h} \otimes {\mathcal{S}}^{2,3}_{h}$. \[nexSUtime\_0\]
Alternative Cost-Efficient Procedure {#sec_NumRealIgA}
------------------------------------
Recall that the cost of Step 1 of the algorithm presented in Section \[sec\_minmaj\] depends on the choice of $Y_h\subset H(\Omega,\operatorname{div})$. As shown in Lemma \[lem\_Moconv\], we can make the estimate as sharp as we desire by choosing a suitably large space $Y_h$. However, the larger $Y_h$ is chosen, the more costly setting up and solving the system becomes. Clearly, it is highly desirable to keep the cost for error estimation below the cost for solving the original problem.
As discussed above, choosing $\hat{Y}_h$ as in does not result in a cost-efficient method. Apart from the fact that $y_h$ is vector-valued while $u_h$ is scalar, another aspect contributes to the high cost for the procedure presented in Section \[sec\_proc0\]. Recall that, by choosing $\hat{Y}_h$ as in , we have $$\begin{aligned}
y_1 & \in {\mathcal{S}}^{p+1,p}_{h},\\
y_2 & \in {\mathcal{S}}^{p,p+1}_{h},\end{aligned}$$ i.e., the components of $y_h$ are in different spline spaces. Hence, we have to compute different basis functions for $y_1$ and $y_2$ (note that this can be a costly procedure for higher polynomial degrees). Furthermore, when assembling, for example, the matrix $\underline{L}_h^1$, we need to compute integrals over products of basis functions of the form $$\int_\Omega R_i R_j {\;dx}.$$ With $\hat{Y}_h$ as in , the product $R_i R_j$ of basis functions of $y_1$ is different to the product of basis functions of $y_2$, hence, the integrals have to be evaluated independently for $y_1$ and $y_2$.
\[ex:case\_1\] In the light of these observations, and since $(C^{p-2})^d \subset H(\Omega, \operatorname{div}),\ \forall p \geq 2$, we study the following alternative choice for $\hat{Y}_h$. $$\hat{Y}_h = {\mathcal{S}}^{p+1,p+1}_{h} \otimes {\mathcal{S}}^{p+1,p+1}_{h}.
\label{e_Yh1}$$ Thereby, we choose a function space $\hat{Y}_h$ on the parameter domain and, analogously to the relation of $\hat{V}_h$ and $V_h$, we define the function space $Y_h$ by the push-forward $$Y_h = \hat{Y}_h \circ G^{-1}.$$
We refer to this setting as *Case \[ex:case\_1\]* in the remainder of the paper. With this choice, $y_1$ and $y_2$ are contained in the same spline spaces. Hence, the basis functions need to be computed only once, and any computed function values can be used for both components of $y_h$. The computed efficiency indices are presented in Table \[nexSUeff\_1\], which show that we obtain even better (i.e., sharper) upper bounds for the exact error with $\hat{Y}_h$ as in than with the choice . When we compare the plots of the cells marked by the error estimator in Figure \[fig\_nexSU\_1\] to the plots in Figure \[fig\_nexSU\_Ex\], we see that the error distribution is again captured accurately starting from the mesh $32\times 32$. The dashed line in Table \[nexSUeff\_1\] indicates that the criterion is fulfilled with $C_\oplus = 5$ starting from the mesh $64\times 64$. The timings obtained with this method are presented in Table \[nexSUtime\_1\]. This approach reduced the total time needed for computing the majorant from a factor of about $4.5$ to a factor of approximately $3$ compared to the time for assembling and solving the original problem. Nevertheless, a factor of $3$ in the timings is still not very appealing, and demands further reduction in the cost.
mesh-size
------------------ ------------ ---------- ----------
$8 \times 8$ [ 2.77 ]{} 8.08e+01 1.24e+01
$16 \times 16$ [ 1.71 ]{} 5.75e-01 3.96e-01
$32 \times 32$ [ 1.32 ]{} 2.14e-02 7.05e-03
$64 \times 64$ [ 1.16 ]{} 1.11e-03 1.78e-04
$128 \times 128$ [ 1.08 ]{} 6.39e-05 5.08e-06
$256 \times 256$ [ 1.04 ]{} 3.83e-06 1.53e-07
$512 \times 512$ [ 1.02 ]{} 2.35e-07 4.69e-09
: Efficiency index and components of the majorant in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_1\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{3,3}_{h} \otimes {\mathcal{S}}^{3,3}_{h}$.\[nexSUeff\_1\]
------------------ -------- -------- -------- -------- ------------ --------- --------- ------------- -------- -------- ------------
mesh-size
$u_h$ $y_h$
$8 \times 8$ 100 242 0.04 0.11 [ 2.78 ]{} $<$0.01 $<$0.01 [ 1.51 ]{} 0.04 0.11 [ 2.76 ]{}
$16 \times 16$ 324 722 0.12 0.34 [ 2.86 ]{} $<$0.01 0.01 [ 5.33 ]{} 0.12 0.35 [ 2.90 ]{}
$32 \times 32$ 1156 2450 0.46 1.35 [ 2.94 ]{} 0.01 0.05 [ 7.69 ]{} 0.47 1.40 [ 3.01 ]{}
$64 \times 64$ 4356 8978 1.77 5.30 [ 2.99 ]{} 0.03 0.27 [ 8.02 ]{} 1.80 5.57 [ 3.09 ]{}
$128 \times 128$ 16900 34322 7.39 21.89 [ 2.96 ]{} 0.16 1.45 [ 9.26 ]{} 7.55 23.34 [ 3.09 ]{}
$256 \times 256$ 66564 134162 33.00 94.69 [ 2.87 ]{} 0.84 8.83 [ 10.54 ]{} 33.84 103.52 [ 3.06 ]{}
$512 \times 512$ 264196 530450 191.59 498.20 [ 2.60 ]{} 3.83 61.45 [ 16.06 ]{} 195.42 559.65 [ 2.86 ]{}
------------------ -------- -------- -------- -------- ------------ --------- --------- ------------- -------- -------- ------------
: Number of DOF and timings in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_1\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{3,3}_{h} \otimes {\mathcal{S}}^{3,3}_{h}$.\[nexSUtime\_1\]
\[rem\_C0FEM\] Note that the use of equal degree polynomials for both the components of $\hat{Y}_{h}$ is only possible because of extra continuity readily available from NURBS basis functions. A counter-part is not possible in FEM case simply because the derivatives of FEM basis functions (with $C^{0}$-continuity) is only in $L^{2}$, and hence, one can not avoid using proper subspaces of $H(\Omega,\operatorname{div})$, e.g., Raviart-Thomas space (with unequal degree polynomials in both the dimensions for both the components). It is further important to note from a close inspection of Tables \[nexSUeff\_0\] and \[nexSUeff\_1\] that the results from equal degree components of vector-valued quantity outperformed the results from unequal degree case.
In order to further reduce the computational cost, we reduce the number of DOFs of $y_h$ by coarsening the mesh by a factor $K$ in each dimension. The number of DOFs of $y_h$ is thus reduced to $2 N_u/K^2$ (asymptotically). The larger $K$ is chosen, the greater the reduction of DOFs will be. At the same time, if the coarsening is done too aggressively, sharp features might not be detected properly on coarse meshes. We counter the reduction in accuracy due to mesh-coarsening by increasing the polynomial degree of $y_h$ by some positive integer $k$, i.e., we choose $$\hat{Y}_h = {\mathcal{S}}^{p+k,p+k}_{Kh} \otimes {\mathcal{S}}^{p+k,p+k}_{Kh}.
\label{e_Yhk}$$ Note that, if desired, one could also choose different factors $K_1$ and $K_2$ and different degree increases $k_1$ and $k_2$ for the first and second component, respectively.
\[rem\_choiceYh\] With the choices of $\hat{Y}_h$ as in , we take advantage of the following specific property of univariate NURBS basis functions. For $C^{p-1}$-continuity, increasing the polynomial degree by $k$ only adds a total of $k$ additional basis functions. In other words, the global smoothness can be increased at the cost of only a few additional DOFs. Coarsening the mesh by a factor $K$, however, will also reduce the number of DOFs by the same factor $K$ (asymptotically). Moreover, as we will see from the three cases of Example \[ex:sin\_6pix\_3piy\], asymptotically we get better efficiency indices with higher degree $p$ and coarser meshes as compared to lower degree $p$ and finer meshes. This phenomenon is similar to the $p$ finite element discretization for problems with smooth solutions, where increasing the polynomial degree for a fixed mesh size $h$ is much more advantageous than decreasing the mesh size $h$ for a fixed (low) polynomial degree. Nevertheless, such a low cost construction for higher degree $p$ is not possible in FEM discretizations.
Note that Case \[ex:case\_1\] discussed above fits into this framework, since Case \[ex:case\_1\] corresponds to the choice $K = k = 1$.
\[ex:case\_2\] For the next setting, we apply moderate mesh-coarsening by choosing $$K = k = 2\text{\ (i.e.,\ }\hat{Y}_h = {\mathcal{S}}^{p+2,p+2}_{2h} \otimes {\mathcal{S}}^{p+2,p+2}_{2h}).$$
Similar to *Case \[ex:case\_1\]*, the function space $Y_h$ is defined by the push-forward $$Y_h = \hat{Y}_h \circ G^{-1}.$$ This setting will be referred to as *Case \[ex:case\_2\]* in the remainder of the paper. The computed efficiency indices along with the magnitudes of the terms $a_1 B_1$ and $a_2 B_2$ for Case \[ex:case\_2\] are presented in Table \[nexSUeff\_2\], and the marked cells are plotted in Figure \[fig\_nexSU\_2\]. The dashed line indicates that criterion is fulfilled with $C_\oplus = 5$, and that a good upper bound of the error is computed and the correct error distribution is captured on meshes starting from $64\times 64$. On coarse meshes, however, the efficiency index is larger than in Case \[ex:case\_1\], which is due to the boundary effects. The timings presented in Table \[nexSUtime\_2\] show that, even though Case \[ex:case\_2\] is faster than Case \[ex:case\_1\], this approach still costs roughly as much as solving the original problem. This is due to the costlier evaluation of the higher degree basis functions, as well as the increased support and overlap of the basis functions, which results in more non-zero entries in $\underline{L}_h$ than in $\underline{K}_h$.
mesh-size
------------------ ------------- ---------- ----------
$8 \times 8$ [ 14.19 ]{} 1.59e+03 8.53e+02
$16 \times 16$ [ 8.49 ]{} 1.97e+01 4.32e+00
$32 \times 32$ [ 1.82 ]{} 3.05e-02 2.41e-02
$64 \times 64$ [ 1.16 ]{} 1.12e-03 1.76e-04
$128 \times 128$ [ 1.04 ]{} 6.14e-05 2.24e-06
$256 \times 256$ [ 1.01 ]{} 3.72e-06 3.32e-08
$512 \times 512$ [ 1.00 ]{} 2.31e-07 5.13e-10
: Efficiency index and components of the majorant in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_2\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{4,4}_{2h} \otimes {\mathcal{S}}^{4,4}_{2h}$.\[nexSUeff\_2\]
------------------ -------- -------- -------- -------- ------------ --------- --------- ------------ -------- -------- ------------
mesh-size
$u_h$ $y_h$
$8 \times 8$ 100 128 0.03 0.05 [ 1.39 ]{} $<$0.01 $<$0.01 [ 1.16 ]{} 0.04 0.05 [ 1.39 ]{}
$16 \times 16$ 324 288 0.14 0.18 [ 1.29 ]{} $<$0.01 $<$0.01 [ 0.92 ]{} 0.14 0.18 [ 1.28 ]{}
$32 \times 32$ 1156 800 0.54 0.59 [ 1.10 ]{} 0.01 0.02 [ 2.32 ]{} 0.55 0.61 [ 1.11 ]{}
$64 \times 64$ 4356 2592 1.91 2.33 [ 1.22 ]{} 0.04 0.08 [ 2.09 ]{} 1.95 2.40 [ 1.23 ]{}
$128 \times 128$ 16900 9248 7.46 9.54 [ 1.28 ]{} 0.19 0.51 [ 2.75 ]{} 7.64 10.05 [ 1.32 ]{}
$256 \times 256$ 66564 34848 33.93 39.02 [ 1.15 ]{} 0.90 2.59 [ 2.88 ]{} 34.82 41.60 [ 1.19 ]{}
$512 \times 512$ 264196 135200 196.23 177.98 [ 0.91 ]{} 4.08 15.91 [ 3.90 ]{} 200.31 193.89 [ 0.97 ]{}
------------------ -------- -------- -------- -------- ------------ --------- --------- ------------ -------- -------- ------------
: Number of DOF and timings in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_2\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{4,4}_{2h} \otimes {\mathcal{S}}^{4,4}_{2h}$.\[nexSUtime\_2\]
\[ex:case\_3\] To further improve the timings, we coarsen the mesh more aggressively by a factor of $4$ and, at the same time, increase the polynomial degree of $y_h$ by 4, as compared to $u_h$, i.e., $$K = k = 4 \text{\ (i.e.,\ }\hat{Y}_h = {\mathcal{S}}^{p+4,p+4}_{4h} \otimes {\mathcal{S}}^{p+4,p+4}_{4h}).$$
Again, similar to *Case \[ex:case\_1\]*, the function space $Y_h$ is defined by the push-forward $$Y_h = \hat{Y}_h \circ G^{-1}.$$ We refer to this setting as *Case \[ex:case\_3\]* in the remainder of the paper. This aggressive coarsening notably affects the efficiency index on coarse meshes, see Table \[nexSUeff\_3\]. On fine meshes, however, the efficiency indices are close to 1 in all presented cases. The number of DOFs of $y_h$ in Case \[ex:case\_3\] is only $N_u/8$ (asymptotically). The timings presented in Table \[nexSUtime\_3\] show that this setting results in a method which can be performed significantly faster (at almost half of the cost) than solving the original problem. The more aggressive reduction of DOF outweighs the additional costs mentioned above, even though the polynomial degree is now increased by 4.
\[rmk\_Coplus\] In all Cases for Example \[ex:sin\_6pix\_3piy\], criterion is fulfilled with $C_\oplus = 5$ on meshes of size $64 \times 64$ and finer. This is indicated by the dashed lines in Tables \[nexSUeff\_0\], \[nexSUeff\_1\], \[nexSUeff\_2\] and \[nexSUeff\_3\], and is clear from Figures \[fig\_nexSU\_0\]-\[fig\_nexSU\_3\]. Therefore, Example \[ex:sin\_6pix\_3piy\] and the examples discussed in Section \[sec\_numex\] show that $C_\oplus = 5$ is a good choice for checking criterion numerically, even though this choice may be conservative in some cases.
mesh-size
------------------ ------------- ---------- ----------
$8 \times 8$ [ 11.28 ]{} 5.38e+02 1.01e+03
$16 \times 16$ [ 36.43 ]{} 2.83e+02 1.60e+02
$32 \times 32$ [ 12.63 ]{} 2.04e+00 5.81e-01
$64 \times 64$ [ 1.17 ]{} 1.13e-03 1.88e-04
$128 \times 128$ [ 1.01 ]{} 5.98e-05 3.79e-07
$256 \times 256$ [ 1.00 ]{} 3.70e-06 1.24e-09
$512 \times 512$ [ 1.00 ]{} 2.31e-07 5.32e-12
: Efficiency index and components of the majorant in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_3\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{6,6}_{4h} \otimes {\mathcal{S}}^{6,6}_{4h}$.\[nexSUeff\_3\]
------------------ -------- ------- -------- ------- ------------ --------- --------- ------------ -------- ------- ------------
mesh-size
$u_h$ $y_h$
$8 \times 8$ 100 128 0.04 0.03 [ 0.76 ]{} $<$0.01 $<$0.01 [ 1.09 ]{} 0.04 0.03 [ 0.76 ]{}
$16 \times 16$ 324 200 0.14 0.10 [ 0.69 ]{} $<$0.01 $<$0.01 [ 0.61 ]{} 0.14 0.10 [ 0.69 ]{}
$32 \times 32$ 1156 392 0.54 0.31 [ 0.57 ]{} 0.01 $<$0.01 [ 0.34 ]{} 0.55 0.31 [ 0.57 ]{}
$64 \times 64$ 4356 968 1.90 1.19 [ 0.63 ]{} 0.04 0.01 [ 0.26 ]{} 1.94 1.20 [ 0.62 ]{}
$128 \times 128$ 16900 2888 7.49 4.86 [ 0.65 ]{} 0.16 0.14 [ 0.84 ]{} 7.66 4.99 [ 0.65 ]{}
$256 \times 256$ 66564 9800 33.90 20.15 [ 0.59 ]{} 0.91 0.82 [ 0.91 ]{} 34.81 20.98 [ 0.60 ]{}
$512 \times 512$ 264196 35912 194.25 84.70 [ 0.44 ]{} 4.10 5.45 [ 1.33 ]{} 198.35 90.15 [ 0.45 ]{}
------------------ -------- ------- -------- ------- ------------ --------- --------- ------------ -------- ------- ------------
: Number of DOF and timings in Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_3\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{6,6}_{4h} \otimes {\mathcal{S}}^{6,6}_{4h}$.\[nexSUtime\_3\]
We now comment on the interleaved iterations. The results in the Tables \[nexSUeff\_0\]-\[nexSUeff\_3\] were obtained by applying only two interleaved iterations, as described in Section \[sec\_minmaj\]. As mentioned there, a sufficiently accurate result can be obtained already after the first such iteration. To illustrate this, we present the efficiency indices for Case \[ex:case\_3\] in Table \[nexSUbetaloops\], which were obtained after one, two, and four interleaved iterations, respectively. The efficiency index does vary notably on the coarser meshes, but since all of these values greatly overestimate the exact error, they do not correctly capture the error distribution. On meshes, where the criterion is fulfilled with $C_\oplus = 5$, and thus the error distribution is correctly recovered, the differences due to more interleaved iterations are insignificant.
------------------ ------- ------- -------
mesh-size
$8 \times 8$ 11.84 11.28 11.25
$16 \times 16$ 80.31 36.43 33.78
$32 \times 32$ 17.36 12.63 10.11
$64 \times 64$ 1.20 1.17 1.17
$128 \times 128$ 1.01 1.01 1.01
$256 \times 256$ 1.00 1.00 1.00
$512 \times 512$ 1.00 1.00 1.00
------------------ ------- ------- -------
: Comparison of $I_\text{eff}$ for different numbers of interleaved iterations, Example \[ex:sin\_6pix\_3piy\], Case \[ex:case\_3\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}, \hat{Y}_h = {\mathcal{S}}^{6,6}_{4h} \otimes {\mathcal{S}}^{6,6}_{4h}$.\[nexSUbetaloops\]
\[rmk\_balance\] The observations discussed above illustrate that one has to balance the sharpness of the majorant on the one hand, and the required computational effort on the one hand. Note that in typical practical applications, the exact solution (and thus the sharpness of the majorant) is not known. Therefore, to address the balance between sharpness and required computational effort, we propose the following strategy. If the mesh is coarse and the total computational cost for the error estimate is moderate, we apply no (or only moderate) coarsening. When the original mesh is fine (problem size being large), we coarsen the mesh aggressively, and thereby, profit from the fast computation of the estimate. While exercising this strategy it is important to enforce the criterion with $C_\oplus \ge 5$.
Numerical Examples {#sec_numex}
==================
In this section, we present further numerical examples which illustrate the potential of the proposed a posteriori error estimator. We will present the results corresponding to the three settings discussed in Section \[sec\_effcomp\], namely *Case \[ex:case\_1\]*, *Case \[ex:case\_2\]* and *Case \[ex:case\_3\]* with the choices $K = k = 1$, $K = k = 2$, and $K = k = 4$, respectively. As in Example \[ex:sin\_6pix\_3piy\], the initial guess for $\beta$ is $0.01$.
As discussed in Section \[sec\_Bsplines\], the parameter domain in all presented examples is the unit square $\hat{\Omega} = (0,1)^2$. The mesh-sizes in the two coordinate directions, which will be presented in the tables, are determined by the respective initial meshes, which in turn, are determined by the geometry mappings. The figures plotted in black represent the computations based on the exact error, and the figures plotted in magenta represent the computations based on the majorant. [The data presented in the tables is as described in the beginning of Section \[sec\_effcomp\].]{}
We first consider an example with reduced continuity $C^{p-m},\ m>1$.
\[ex:sin\_6pix\_3piy\_p4\] **Sinus function in a unit square with $p=q=4$ and $C^{1}$-continuity:** We consider the same exact solution and the same physical domain as in Example \[ex:sin\_6pix\_3piy\], i.e., $$\begin{aligned}
u = \sin( 6 \pi x) \sin(3 \pi y), \quad \Omega = (0,1)^2.\end{aligned}$$ However, we now use B-splines of degree $p=q=4$ to represent $\Omega$, and we add a triple knot at the coordinates $x = 0.5$ and $y = 0.5$. The initial knot vectors are thus given by $$s = t = (0,0,0,0,0,0.5,0.5,0.5,1,1,1,1,1),$$ and the geometry mapping is only $C^1$-continuous at the coordinate $0.5$.
[|c|r|rr|]{} mesh-size & & &\
\
$18 \times 18$ & [ 1.84 ]{} & 1.04e-03 & 9.00e-04\
$34 \times 34$ & [ 1.40 ]{} & 1.78e-06 & 7.23e-07\
$66 \times 66$ & [ 1.20 ]{} & 5.09e-09 & 1.00e-09\
$130 \times 130$ & [ 1.10 ]{} & 1.77e-11 & 1.74e-12\
$258 \times 258$ & [ 1.05 ]{} & 6.61e-14 & 3.25e-15\
\
$18 \times 18$ & [ 15.43]{} & 7.95e-02 & 5.75e-02\
$34 \times 34$ & [ 6.04 ]{} & 1.14e-05 & 3.53e-05\
$66 \times 66$ & [ 1.76 ]{} & 7.52e-09 & 5.69e-09\
$130 \times 130$ & [ 1.16 ]{} & 1.87e-11 & 3.01e-12\
$258 \times 258$ & [ 1.04 ]{} & 6.54e-14 & 2.49e-15\
\
$18 \times 18$ & [ 132.77 ]{} & 7.38e+00 & 2.76e+00\
$34 \times 34$ & [ 148.41 ]{} & 1.86e-02 & 9.53e-03\
$66 \times 66$ & [ 6.42 ]{} & 5.49e-08 & 1.21e-07\
$130 \times 130$ & [ 1.13 ]{} & 1.83e-11 & 2.39e-12\
$258 \times 258$ & [ 1.01 ]{} & 6.34e-14 & 3.78e-16\
\
\
The computed efficiency indices are presented in Table \[nexSUp4\_eff\]. The dashed lines, which correspond to criterion being fulfilled with $C_\oplus = 5$, again show that more aggressive mesh-coarsening requires a finer initial mesh. By this criterion, we get a good quality of the estimate and the indicated error distribution starting from the mesh $66\times 66$ in Case \[ex:case\_1\], and from $130 \times 130$ in Cases \[ex:case\_2\] and \[ex:case\_3\].
We present the cells marked for refinement by the exact error in Figure \[fig\_nexSUp4\_ex\], and the cells marked by the error estimator in Figure \[fig\_nexSUp4\_est\]. Figure \[fig\_nexSUp4\_est\_a\] shows that the error distribution is already captured on the mesh $18\times 18$ in Case \[ex:case\_1\]. In Case \[ex:case\_2\], we obtain a good indication of the error distribution from the mesh $66\times 66$, i.e., before criterion with $C_\oplus = 5$ is fulfilled. Once the error distribution is captured correctly on a certain mesh, it is also captured on all finer meshes (as in Example \[ex:sin\_6pix\_3piy\]). Hence, we do not show all plots for all meshes and cases, but only the first meshes, on which the error distribution is captured correctly. Also, we omit the presentation of the timings, since the overall behavior is as in Example \[ex:sin\_6pix\_3piy\].
In the next example, we consider the case of non-trivial PDE coefficient matrix $A$.
\[ex:sin\_6pix\_3piy\_NTA\] Let the matrix $A$ be of the form of $$\left(
\begin{array}{cc}
e^{b_{11} x + b_{12} y} & 0 \\
0 & e^{b_{21} x + b_{22} y}
\end{array}
\right),$$ which is positive definite for $b_{ij} \in \mathbb{R}^{+}$, $i, j = \{1, 2\}$. This will result in the PDE operator to be of the form of $$e^{b_{11} x + b_{12} y} \dfrac{\partial^{2}}{\partial x^{2}}
+ e^{b_{21} x + b_{22} y} \dfrac{\partial^{2}}{\partial y^{2}}
+ b_{11} e^{b_{11} x + b_{12} y} \dfrac{\partial}{\partial x}
+ b_{22} e^{b_{21} x + b_{22} y} \dfrac{\partial}{\partial y}.$$ To have this PDE operator with full generality, we take $b_{11} = 0.1, b_{12} = 0.8, b_{21} = 0.4, b_{22} = 0.7$. With this generality, to have a good comparison of the efficiency indices with the examples considered so far, we again choose the exact solution to be $u = \sin (6 \pi x) \sin (3 \pi y)$. The right hand side function is accordingly calculated and the solution has homogeneous Dirichlet boundary values.
Note that in this case the constant $C_{\Omega}$ has to be accordingly modified. For the unit square domain, and the matrix $A$ given above, its value is taken as $\dfrac{c_{2}}{\pi \sqrt{2}}$, where $$c_{2} = \max \{e^{b_{11} x + b_{12} y}, e^{b_{21} x + b_{22} y}\}.$$
[|c|r|rr|]{}mesh-size & & &\
\
$8 \times 8$ & [ 3.64 ]{} & 2.02e+02 & 5.73e+01\
$16 \times 16$ & [ 6.00 ]{} & 1.16e+01 & 7.75e+00\
$32 \times 32$ & [ 2.50 ]{} & 6.82e-02 & 9.65e-02\
$64 \times 64$ & [ 1.74 ]{} & 2.70e-03 & 1.98e-03\
$128 \times 128$ & [ 1.37 ]{} & 1.31e-04 & 4.83e-05\
$256 \times 256$ & [ 1.19 ]{} & 7.04e-06 & 1.31e-06\
$512 \times 512$ & [ 1.09 ]{} & 4.05e-07 & 3.77e-08\
\
$8 \times 8$ & [ 38.29 ]{} & 2.34e+04 & 5.36e+03\
$16 \times 16$ & [ 14.09 ]{} & 9.07e+01 & 1.61e+01\
$32 \times 32$ & [ 4.64 ]{} & 2.02e-01 & 3.67e-01\
$64 \times 64$ & [ 1.62 ]{} & 2.53e-03 & 1.55e-03\
$128 \times 128$ & [ 1.15 ]{} & 1.10e-04 & 1.57e-05\
$256 \times 256$ & [ 1.04 ]{} & 6.20e-06 & 2.16e-07\
$512 \times 512$ & [ 1.01 ]{} & 3.77e-07 & 3.27e-09\
\
$8 \times 8$ & [ 32.36 ]{} & 2.81e+03 & 1.77e+04\
$16 \times 16$ & [ 122.17 ]{} & 6.73e+03 & 1.30e+03\
$32 \times 32$ & [ 23.20 ]{} & 1.15e+01 & 2.71e+00\
$64 \times 64$ & [ 1.64 ]{} & 2.58e-03 & 1.61e-03\
$128 \times 128$ & [ 1.03 ]{} & 9.79e-05 & 2.41e-06\
$256 \times 256$ & [ 1.00 ]{} & 5.95e-06 & 7.76e-09\
$512 \times 512$ &\
The computed efficiency indices are presented in Table \[nex\_NTA\]. The dashed lines correspond to criterion being fulfilled with $C_\oplus = 5$. We see that the proposed estimator is robust with respect to the non-trivial PDE coefficient matrix $A$, and its performance is asymptotically similar to the case with the matrix $A$ being identity [^2]. Also, the presentation of the timings is again omitted since the overall behavior is as in Example \[ex:sin\_6pix\_3piy\].
In the next example, we consider a domain with a curved boundary (requiring a NURBS mapping for exact representation) and a problem whose solution has sharp peaks.
\[ex:q\_annulus\] **Domain with curved boundary:** Consider the domain of a quarter annulus. In polar coordinates, $\Omega$ is defined by $(r,\phi) \in (1,2) \times (0,\tfrac{\pi}{2})$. The circular parts of the domain boundary are represented exactly by the NURBS geometry mapping of degree 2, i.e., we have $p=q=2$. We set $A = I$, and we prescribe the exact solution$$u = (r-1)(r-2)\phi (\phi-\tfrac{\pi}{2}) e^{-\alpha (r \cos\phi-1)^2}.$$ We test our method with two values of $\alpha$, namely, $$\begin{aligned}
\text{Example~\ref{ex:q_annulus}.a:} \quad \alpha = 20, \qquad
\text{Example~\ref{ex:q_annulus}.b:} \quad \alpha = 50.\end{aligned}$$ In both examples, this function has zero Dirichlet boundary values and a peak at $x = 1$, the sharpness of which is determined by the value of $\alpha$. The exact solutions are depicted in Figure \[fig\_nexQAex\].
\[tr\]\[tr\][0]{}\[tr\]\[tr\]\[br\]\[br\][0.2]{}\[tr\]\[tr\][0]{}\[tr\]\[tr\][1]{}\[tr\]\[tr\][2]{}\[tl\]\[tl\][0]{}\[tl\]\[tl\][1]{}\[tr\]\[tr\][2]{}\
[|c|r|rr|]{}mesh-size & & &\
\
$16 \times 8$ & [ 1.83 ]{} & 9.98e-04 & 3.59e-04\
$32 \times 16$ & [ 1.29 ]{} & 2.08e-05 & 6.51e-06\
$64 \times 32$ & [ 1.13 ]{} & 1.04e-06 & 1.44e-07\
$128 \times 64$ & [ 1.07 ]{} & 5.95e-08 & 4.00e-09\
$256 \times 128$ & [ 1.03 ]{} & 3.58e-09 & 1.20e-10\
$512 \times 256$ & [ 1.02 ]{} & 2.20e-10 & 3.67e-12\
\
$16 \times 8$ & [ 13.99 ]{} & 4.44e-02 & 3.51e-02\
$32 \times 16$ & [ 4.17 ]{} & 2.00e-04 & 8.43e-05\
$64 \times 32$ & [ 1.31 ]{} & 1.20e-06 & 3.66e-07\
$128 \times 64$ & [ 1.06 ]{} & 5.91e-08 & 3.36e-09\
$256 \times 128$ & [ 1.01 ]{} & 3.51e-09 & 4.60e-11\
$512 \times 256$ & [ 1.00 ]{} & 2.17e-10 & 6.96e-13\
\
$16 \times 8$ & [ 24.87 ]{} & 1.09e-01 & 1.42e-01\
$32 \times 16$ & [ 56.02 ]{} & 2.92e-02 & 2.22e-02\
$64 \times 32$ & [ 10.42 ]{} & 7.81e-05 & 2.16e-05\
$128 \times 64$ & [ 1.11 ]{} & 6.21e-08 & 6.61e-09\
$256 \times 128$ & [ 1.00 ]{} & 3.49e-09 & 1.02e-11\
$512 \times 256$ & [ 1.00 ]{} & 2.17e-10 & 3.27e-14\
\
[|c|r|rr|]{} mesh-size & & &\
\
$16 \times 8$ & [ 3.02 ]{} & 2.94e-02 & 1.78e-02\
$32 \times 16$ & [ 1.92 ]{} & 3.57e-04 & 1.83e-04\
$64 \times 32$ & [ 1.34 ]{} & 9.15e-06 & 3.22e-06\
$128 \times 64$ & [ 1.16 ]{} & 4.67e-07 & 7.56e-08\
$256 \times 128$ & [ 1.08 ]{} & 2.67e-08 & 2.12e-09\
$512 \times 256$ & [ 1.04 ]{} & 1.60e-09 & 6.32e-11\
\
$16 \times 8$ & [ 13.84 ]{} & 3.45e-01 & 6.49e-01\
$32 \times 16$ & [ 16.76 ]{} & 2.58e-02 & 1.53e-02\
$64 \times 32$ & [ 3.16 ]{} & 4.10e-05 & 2.80e-05\
$128 \times 64$ & [ 1.25 ]{} & 5.04e-07 & 1.24e-07\
$256 \times 128$ & [ 1.05 ]{} & 2.61e-08 & 1.33e-09\
$512 \times 256$ & [ 1.01 ]{} & 1.56e-09 & 1.89e-11\
\
$16 \times 8$ & [ 17.20 ]{} & 4.24e-01 & 1.11e+00\
$32 \times 16$ & [ 76.95 ]{} & 3.24e-01 & 5.41e-01\
$64 \times 32$ & [ 83.72 ]{} & 3.02e-02 & 1.83e-02\
$128 \times 64$ & [ 4.19 ]{} & 4.64e-06 & 2.44e-06\
$256 \times 128$ & [ 1.04 ]{} & 2.59e-08 & 1.02e-09\
$512 \times 256$ & [ 1.00 ]{} & 1.55e-09 & 2.22e-12\
\
In Tables \[nexQAeff\_a\] and \[nexQAeff\_b\], the efficiency index $I_{\text{eff}}$ and the magnitudes of $a_1 B_1$ and $a_2 B_2$ are presented for both the cases of $\alpha$. The dashed lines indicate the mesh-size after which criterion with $C_\oplus = 5$ is fulfilled. The distribution of the marked cells is depicted in Figures \[fig\_nexQAmark\_a\] and \[fig\_nexQAmark\_b\]. As before, we observe that the error distribution is represented correctly if the criterion is fulfilled with $C_\oplus = 5$. When comparing Tables \[nexQAeff\_a\] and \[nexQAeff\_b\], as well as Figures \[fig\_nexQAmark\_a\] and \[fig\_nexQAmark\_b\], we notice that the more aggressive the mesh coarsening, and sharper the peak, the more refinements are needed before criterion is fulfilled and the error distribution is captured correctly. Since the timings in Example \[ex:q\_annulus\].a and Example \[ex:q\_annulus\].b show the same behavior as in Example \[ex:sin\_6pix\_3piy\], both regarding assembling-time and solving-time, we omit the presentation of these numbers. Clearly, Case \[ex:case\_3\] outperforms Cases \[ex:case\_1\] and \[ex:case\_2\] in terms of cost-efficiency.
In the next example, we test the proposed estimator in a basic adaptive refinement scheme.
\[ex:adap\_ref\] **Adaptive Refinement:** The exact solution for this example is given by $$u = ( x^2-x)(y^2-y) e^{-100 |(x,y)- (0.8,0.05)|^2 -100 |(x,y)- (0.8,0.95)|^2}.$$ The computational domain is again the unit square $\Omega = (0,1)^2$, and is represented by B-splines of degree $p=q=2$. The function $u$, which is illustrated in Figure \[fig\_nexAda\_exsol\], has zero Dirichlet boundary values and has two peaks at the coordinates $(0.8,0.05)$ and $(0.8,0.95)$.
In this example, we test a very basic adaptive refinement procedure using tensor-product B-splines. The discussion of isogeometric local refinement schemes is out of the scope of this paper (see Section \[sec\_intro\] for an overview on local refinement methods).
\[tl\]\[tl\][$1$]{}\[tl\]\[tl\][$0$]{}\[tr\]\[tr\][$1$]{}\[tr\]\[tr\][$0$]{}![Exact solution, Example \[ex:adap\_ref\].\[fig\_nexAda\_exsol\]](Figures/Eada_ExSol_3b "fig:"){height="5cm" width="5cm"}
![Error convergence, Example \[ex:adap\_ref\], Cases as in Table \[nexAda\_Ieff\].\[fig\_nexAda\_errconv\]](Figures/Eada_ErrConv_1024)
mesh-size Case
-------------------- ------------ ---------- ---------- ----------------
$16 \times 16$ [ 3.77 ]{} 9.39e-05 3.49e-05 \[ex:case\_1\]
$25 \times 26$ [ 2.06 ]{} 8.62e-07 8.11e-07 \[ex:case\_1\]
$38 \times 44$ [ 1.69 ]{} 4.30e-08 2.35e-08 \[ex:case\_1\]
$64 \times 74$ [ 1.47 ]{} 2.79e-09 1.19e-09 \[ex:case\_1\]
$92 \times 136$ [ 2.82 ]{} 8.19e-10 4.87e-10 \[ex:case\_2\]
$184 \times 256$ [ 1.30 ]{} 2.05e-11 4.55e-12 \[ex:case\_2\]
$341 \times 492$ [ 1.11 ]{} 1.45e-12 1.47e-13 \[ex:case\_2\]
$652 \times 934$ [ 1.84 ]{} 2.55e-13 1.07e-13 \[ex:case\_3\]
$1304 \times 1868$ [ 1.09 ]{} 7.40e-15 3.63e-16 \[ex:case\_3\]
: Efficiency index, components of the majorant and applied cases in Example \[ex:adap\_ref\], $\hat{V}_{h} = {\mathcal{S}}^{2,2}_{h}$. \[nexAda\_Ieff\]
\
\
We apply adaptive refinement based on a marking with $\psi = 25\%$, starting on an initial mesh $16\times 16$. On the first four steps, we apply Case \[ex:case\_1\], then Case \[ex:case\_2\] on the next three steps, and thereafter Case \[ex:case\_3\]. The efficiency indices and the applied cases are shown in Table \[nexAda\_Ieff\]. In Figure \[fig\_nexAda\_meshes\], the meshes and the marked cells are shown for steps 4, 7, and 9. Clearly, the correct areas of the domain are identified and marked for refinement. Since the solution of the problem is sufficiently regular, the error plots in Figure \[fig\_nexAda\_errconv\] show that the adaptive refinement converges with the same rate as the uniform refinement, but with a better constant. However, due to the tensor-product structure of the mesh, many superfluous DOFs are inserted outside of the marked areas, which worsens the rate of convergence for given total DOFs.
Next, we consider a classical example for a posteriori error estimation and adaptive refinement studies.
\[ex:l\_shaped\] **L-shaped domain:** We consider the Laplace equation $$\Delta u = 0 \label{nexLSlaplace}$$ with Dirichlet boundary conditions on the L-shaped domain $\Omega = (-1,1)^2 \backslash [0,1]^2$. In this example, we use a bilinear geometry mapping, i.e., $p=q=1$. The function $$u(r,\phi) = r^{\frac{2}{3}} \sin( (2\phi-\pi)/3 )$$ solves and is used to prescribe Dirichlet boundary conditions. The solution has a singularity at the re-entrant corner at $(0,0)$.
We compare uniform refinement and adaptive refinement in the tensor-product setting. In this example we set $\psi = 10\%$, and to avoid the pollution near the singularity, we only use Case \[ex:case\_1\] in the majorant computations.
mesh-size
------------------- -------------- ---------- ---------- --
$16 \times 8$ [ 1.1785 ]{} 5.67e-02 1.71e-03
$32 \times 16$ [ 1.1401 ]{} 3.44e-02 8.98e-04
$64 \times 32$ [ 1.1116 ]{} 2.09e-02 4.72e-04
$128 \times 64$ [ 1.0898 ]{} 1.28e-02 2.49e-04
$256 \times 128$ [ 1.0729 ]{} 7.87e-03 1.32e-04
$512 \times 256$ [ 1.0593 ]{} 4.86e-03 7.01e-05
$1024 \times 512$ [ 1.0485 ]{} 3.01e-03 3.73e-05
: Efficiency index and components of the majorant in Example \[ex:l\_shaped\], Case \[ex:case\_1\], uniform refinement, $\hat{V}_{h} = {\mathcal{S}}^{1,1}_{h}, \hat{Y}_h = {\mathcal{S}}^{2,2}_{h} \otimes {\mathcal{S}}^{2,2}_{h}$.\[nexLSeff\]
mesh-size
------------------ -------------- ---------- ---------- --
$16 \times 8$ [ 1.1785 ]{} 5.67e-02 1.71e-03
$22 \times 11$ [ 1.1839 ]{} 2.68e-02 8.66e-04
$30 \times 16$ [ 1.1749 ]{} 1.37e-02 4.32e-04
$39 \times 23$ [ 1.1622 ]{} 7.22e-03 2.24e-04
$55 \times 37$ [ 1.1635 ]{} 3.52e-03 1.10e-04
$87 \times 60$ [ 1.1634 ]{} 1.75e-03 5.41e-05
$133 \times 101$ [ 1.1525 ]{} 9.25e-04 2.69e-05
: Efficiency index and components of the majorant in Example \[ex:l\_shaped\], Case \[ex:case\_1\], adaptive refinement, $\hat{V}_{h} = {\mathcal{S}}^{1,1}_{h}, \hat{Y}_h = {\mathcal{S}}^{2,2}_{h} \otimes {\mathcal{S}}^{2,2}_{h}$.\[nexLSeff\_ada\]
The magnitudes of the components $a_1 B_1$ and $a_2B_2$, which are presented in Table \[nexLSeff\] for uniform refinement, and in Table \[nexLSeff\_ada\] for adaptive refinement, show that the criterion with $C_\oplus = 5$ is fulfilled on all the considered meshes.
The error plots presented in Figure \[fig\_Lada\_errconv\] show the expected faster convergence on the adaptively refined mesh, even though we are only using tensor-product splines. In Figure \[fig\_Lada\_meshes\], meshes and marked cells are shown for steps 2 and 6, again indicating that the error indicator correctly identifies the corner singularity.
![Error convergence, Example \[ex:l\_shaped\], Case \[ex:case\_1\].\[fig\_Lada\_errconv\]](Figures/Lada90_Errors)
\
In our final example, we consider an advection dominated advection diffusion equation to see the performance of the estimator for sharp boundary layers.
\[ex:adv\_diff\] **Advection dominated advection diffusion equation:** We consider the advection diffusion equation with Dirichlet boundary conditions on the unit square $\Omega = (0,1)^2$, with $p=q=2$, i.e., $$\begin{aligned}
\begin{array}{r@{\ =\ }ll}
-\kappa \Delta u + b\cdot \nabla u & 0 \quad &\text{in~}\Omega,\\
u & u_D \quad &\text{on~} \partial \Omega,
\end{array}\end{aligned}$$ where $$\begin{aligned}
\kappa = 10^{-6}, \qquad
b = (\cos \tfrac{\pi}{3}, \sin \tfrac{\pi}{3})^T, \qquad
u_D = \left\{\begin{array}{ll}
1,\text{~if~}y=0 \\
0,\text{~else}
\end{array}\right. .\end{aligned}$$
We use the standard streamline upwind Petrov-Galerkin (SUPG) scheme for the stabilization. The stabilization parameter $\tau$ is set to $\tau(Q) = h_b(Q)/2|b|$, where $h_b(Q)$ is the diameter of the cell $Q$ in direction of the flow $b$, and $|b|$ is the magnitude of the vector $b$. For advection diffusion problems, we have to adapt the majorant. Since the principle method is the same, we refer the reader to [@Repin08_book Section 4.3.1] for a detailed discussion. In this special case, where $A = \kappa I$ with $\kappa \ll |b|$, and with constant velocity vector $b$, the majorant $M_{\oplus,\text{adv}}^{2}$ for the advection diffusion problem is given by $$M_{\oplus,\text{adv}}^{2} = (1+\beta) \| A \nabla u_h - y \|_{\bar{A}}^2 + (1+\tfrac{1}{\beta}) C_{\Omega}^2 \| \operatorname{div}y + f - b\cdot\nabla u_h \|^2.$$ The strong advection and the discontinuous boundary conditions result in sharp layers. In Figure \[fig\_E4\_lay\], the expected positions of the layers are indicated by dashed lines.
[|c|rr|]{} mesh-size & &\
\
$16 \times 16$ & 1.98e-07 & 3.18e-10\
$64 \times 64$ & 6.45e-07 & 1.15e-09\
$256 \times 256$ & 2.28e-06 & 4.33e-09\
\
$16 \times 16$ & 1.83e-06 & 9.66e-10\
$64 \times 64$ & 6.50e-06 & 3.65e-09\
$256 \times 256$ & 1.86e-05 & 1.24e-08\
\
$16 \times 16$ & 3.24e-06 & 1.29e-09\
$64 \times 64$ & 2.07e-05 & 6.52e-09\
$256 \times 256$ & 6.86e-05 & 2.38e-08\
[|c|rr|rrr|rrr|rrr|]{}mesh-size & & & &\
& $u_h$ & $y_h$ & & & & & & & & &\
\
$16 \times 16$ & 324 & 722 & 0.25 & 0.39 & [ 1.56 ]{} & $<$0.01 & 0.01 & [ 6.38 ]{} & 0.25 & 0.40 & [ 1.59 ]{}\
$64 \times 64$ & 4356 & 8978 & 3.25 & 5.32 & [ 1.63 ]{} & 0.03 & 0.26 & [ 8.64 ]{} & 3.28 & 5.58 & [ 1.70 ]{}\
$256 \times 256$ & 66564 & 134162 & 51.22 & 94.15 & [ 1.84 ]{} & 0.85 & 8.84 & [ 10.35 ]{} & 52.07 & 102.99 & [ 1.98 ]{}\
\
$16 \times 16$ & 324 & 288 & 0.21 & 0.14 & [ 0.67 ]{} & $<$0.01 & $<$0.01 & [ 0.50 ]{} & 0.21 & 0.14 & [ 0.67 ]{}\
$64 \times 64$ & 4356 & 2592 & 3.26 & 2.10 & [ 0.64 ]{} & 0.03 & 0.06 & [ 2.01 ]{} & 3.29 & 2.16 & [ 0.66 ]{}\
$256 \times 256$ & 66564 & 34848 & 50.83 & 35.58 & [ 0.70 ]{} & 0.85 & 2.30 & [ 2.70 ]{} & 51.68 & 37.87 & [ 0.73 ]{}\
\
$16 \times 16$ & 324 & 200 & 0.26 & 0.10 & [ 0.39 ]{} & $<$0.01 & $<$0.01 & [ 0.58 ]{} & 0.26 & 0.10 & [ 0.40 ]{}\
$64 \times 64$ & 4356 & 968 & 3.41 & 1.21 & [ 0.35 ]{} & 0.04 & 0.01 & [ 0.26 ]{} & 3.44 & 1.22 & [ 0.35 ]{}\
$256 \times 256$ & 66564 & 9800 & 52.40 & 19.83 & [ 0.38 ]{} & 1.02 & 0.91 & [ 0.89 ]{} & 53.42 & 20.74 & [ 0.39 ]{}\
The magnitudes of $a_1B_1$ and $a_2B_2$ presented in Table \[nexADeff\] indicate that the criterion with $C_\oplus = 5$ is fulfilled on all the considered meshes. The distribution of the marked cells presented in Figures \[fig\_E4\_064\] and \[fig\_E4\_256\] provides the visual indication that the expected layers are accurately detected by the error estimator.
For this example, the timings presented in Table \[nexADtime\] show that, unlike the previous examples, assembling and solving the system for the estimator is faster than for the original problem not only in Case \[ex:case\_3\] (less than $1/2$ of the original cost), but also in Case \[ex:case\_2\] (about $2/3$ of the original cost). This is due to the SUPG stabilization which is costlier than computing the additional term $b\cdot \nabla u_h$ in the majorant $M_{\oplus,\text{adv}}^{2}$.
Conclusion {#sec_Conc}
==========
We have proposed a method for cost-efficient computation of guaranteed and sharp a posteriori error estimates in IGA. This method relies only on the use of NURBS basis functions, without the need for constructing complicated basis functions in $H(\Omega,\operatorname{div})$. We have discussed different settings which allow the user to balance the sharpness of the bound and accurate error distribution on the one hand, and the required computational cost of the error estimator on the other hand (see Remark \[rmk\_balance\]). For the presented settings, we have derived a quality criterion, which is easy to check numerically and which indicates whether the computed estimate is sharp or not (see Remark \[rmk\_Coplus\]). Two properties of NURBS basis functions are exploited. Firstly, the basis functions are, in general, automatically in $H(\Omega,\operatorname{div})$ due to their high smoothness. Without this property, we could not use NURBS of equal degree for both components as basis functions for the minimizing function $y_h$. Secondly, increasing the polynomial degree of NURBS basis functions adds only few DOFs. This fact is necessary for keeping the computational cost of the majorant as low as possible (see Remark \[rem\_choiceYh\]). It is important to note that none of these properties are possible in FEM discretizations based on $C^{0}$ basis functions.
Apart from the topical interest of a posteriori error estimation and adaptivity, the presented method should also be of interest in parametrization of computational domain. For example, for $r$- refinement in IGA, i.e., to optimize the placement of inner control points, the proposed estimator can be used to accurately detect the regions with large error and then use the optimization algorithm to reposition the control points. Such a problem of $r$- refinement has been studied in [@Xu11_2021; @XuMDG-13]. Finally, in this paper, we have only considered tensor-product NURBS discretizations. While the extension of this method to locally refined isogeometric discretizations and also to three dimensions is, in theory, straightforward, the actual performance and efficiency of the error estimator on such methods and meshes is the subject of further studies.
Acknowledgements {#acknowledgements .unnumbered}
================
The authors are grateful to Prof. Sergey I. Repin, V.A. Steklov Institute of Mathematics, St. Petersburg, for helpful discussions. Authors are also thankful to unknown referees for their helpful comments. The support from the Austrian Science Fund (FWF) through the project P21516-N18 is gratefully acknowledged.
[10]{}
M. Ainsworth and J.T. Oden. . Wiley Interscience, 2000.
R.E. Bank and R.K. Smith. A posteriori error estimates based on hierarchical bases. , 30(4):921–935, 1993.
Y. Bazilevs, L. Beir[ã]{}o da Veiga, J.A. Cottrell, T.J.R. Hughes, and G. Sangalli. Isogeometric analysis: approximation, stability and error estimates for [$h$]{}-refined meshes. , 16(7):1031–1090, 2006.
Y. Bazilevs, V.M. Calo, J.A. Cottrell, J.A. Evans, T.J.R. Hughes, S. Lipton, M.A. Scott, and T.W. Sederberg. Isogeometric analysis using [T]{}-splines. , 199(5-8):229–263, 2010.
Y. Bazilevs, V.M. Calo, J.A. Cottrell, T.J.R. Hughes, A. Reali, and G. Scovazzi. Variational multiscale residual-based turbulence modeling for large eddy simulation of incompressible flows. , 197(1-4):173–201, 2007.
Y. Bazilevs, V.M. Calo, T.J.R. Hughes, and Y. Zhang. Isogeometric fluid-structure interaction: theory, algorithms, and computations. , 43:3–37, 2008.
L. Beirão da Veiga, A. Buffa, J. Rivas, and G. Sangalli. Some estimates for $h$- $p$- $k$- refinement in isogeometric analysis. , 118:271–305, 2011.
D. Boffi, F. Brezzi, and M. Fortin. . Springer, Berlin, 2013.
A. Buffa, C. de Falco, and G. Sangalli. so[G]{}eometric [A]{}nalysis: [S]{}table elements for the 2[D]{} [S]{}tokes equation. , 65(11-12):1407–1422, 2011.
A. Buffa, J. Rivas, G. Sangalli, and R. Vázquez. Isogeometric discrete differential forms in three dimensions. , 49(2):818–844, 2011.
A. Buffa, G. Sangalli, and R. V[á]{}zquez. Isogeometric analysis in electromagnetics: [B]{}-splines approximation. , 199(17-20):1143–1152, 2010.
J. Cottrell, T.J.R. Hughes, and Y. Bazilevs. . Wiley, Chichester, 2009.
J.A. Cottrell, T.J.R. Hughes, and A. Reali. Studies of refinement and continuity in isogeometric structural analysis. , 196:4160–4183, 2007.
L. Dede and H.A.F.A. Santos. B-spline goal oriented error estimators for geometrically nonlinear rods. , 49:35–52, 2012.
J. Deng, F. Chen, X. Li, C. Hu, W. Tong, Z. Yang, and Y. Feng. Polynomial splines over hierarchical [T]{}-meshes. , 70:76–86, 2008.
T. Dokken, T. Lyche, and K.F. Pettersen. Polynomial splines over locally refined box-partitions , 30(3):331–356, 2013.
M.R. D[ö]{}rfel, B. J[ü]{}ttler, and B. Simeon. Adaptive isogeometric analysis by local [$h$]{}-refinement with [T]{}-splines. , 199(5-8):264–275, 2010.
W. Dörfler. A convergent adaptive algorithm for [P]{}oisson’s equation. , 33(3):1106–1124, 1996.
T. Elguedj, Y. Bazilevs, V.M. Calo, and T.J.R. Hughes. , 197(33–40):2732–2762, 2008.
J.A. Evans and T.J.R. Hughes. sogeometric [D]{}ivergence-conforming [B]{}-splines for the [D]{}arcy-[S]{}tokes-[B]{}rinkman equations. , 23(04):671–741, 2013.
C. Giannelli, B. Jüttler, and H. Speleers. -splines: The truncated basis for hierarchical splines. , 29(7):485–498, 2012.
C. Giannelli, B. Jüttler, and H. Speleers. Strongly stable bases for adaptively refined multilevel splines spaces. , 40:459–490, 2014.
T.J.R. Hughes, J. Cottrell, and Y. Bazilevs. Isogeometric analysis: [CAD]{}, finite elements, [NURBS]{}, exact geometry and mesh refinement. , 194(39-41):4135–4195, 2005.
T.J.R. Hughes, A. Reali, and G. Sangalli. Efficient quadrature for [NURBS]{}-based isogeometric analysis. , 199(5-8):301–313, 2010.
K.A. Johannessen. An adaptive isogeometric finite element analysis. Master’s thesis, Norwegian University of Science and Technology, 2009.
K.A. Johannessen, T. Kvamsdal and T. Dokken. Isogeometric analysis using LR B-splines. , 269:471–514, 2014.
J.K. Kraus and S.K. Tomar. Algebraic multilevel iteration method for lowest order [R]{}aviart-[T]{}homas space and applications. , 86(10):1175–1196, 2011.
G. Kuru, C.V. Verhoosel, K.G. van der Zee and E.H. van Brummelen. Goal-oriented isogeometric analysis with hierarchical splines. , 270:270–292, 2014.
R. Lazarov, S.I. Repin, and S.K. Tomar. Functional a posteriori error estimates for discontinuous [G]{}alerkin approximations of elliptic problems. , 25(4):952–971, 2009.
X. Li, J. Zheng, T.W. Sederberg, T.J.R. Hughes, and M.A. Scott. On linear independence of [T]{}-spline blending functions. , 29(1):63–76, 2012.
P.N. Nielsen, A.R. Gersborg, J. Gravesen, and N.L. Pedersen. Discretizations in isogeometric analysis of [N]{}avier-[S]{}tokes flow. , 200(45-46):3242–3253, 2011.
L. Piegl and W. Tiller. . Monographs in Visual Communications. Springer Berlin Heidelberg, 2. edition, 1997.
S.I. Repin. A posteriori error estimation for nonlinear variational problems by duality theory. , 99:927–935, 2000. See also: [*Zapiski Nauchnych Seminarov POMI*]{}, 243:201–214, 1997.
S.I. Repin. A posteriori error estimates for approximate solutions to variational problems with strongly convex functionals. , 97:4311–4328, 1999.
S.I. Repin. A posteriori error estimation for variational problems with uniformly convex functionals. , 69(230):481–500, 2000.
S.I. Repin. . Walter de Gruyter, Berlin, Germany, 2008.
M.A. Scott, M.J. Borden, C.V. Verhoosel, T.W. Sederberg, and T.J.R. Hughes. Isogeometric finite element data structures based on [Bézier]{} extraction of [T]{}-splines. , 88:126–156, 2011.
M.A. Scott, X. Li, T.W. Sederberg, and T.J.R. Hughes. Local refinement of analysis-suitable [T]{}-splines. , 213-216:206–222, 2012.
T.W. Sederberg, D.L. Cardon, G.T. Finnigan, N.S. North, J. Zheng, and T. Lyche. -spline simplification and local refinement. , 23(3):276–283, 2004.
T. Takacs and B. Jüttler. Existence of stiffness matrix integrals for singularly parametrized domains in isogeometric analysis. , 200:3568–3582, 2011.
K.G. van der Zee and C.V. Verhoosel. Isogeometric analysis-based goal oriented error estimation for free boundary problems. , 47:600–609, 2011.
A.-V. Vuong, C. Giannelli, B. Jüttler, and B. Simeon. A hierarchical approach to adaptive local refinement in isogeometric analysis. , 200(49-52):3554–3567, 2011.
P. Wang, J. Xu, J. Deng, and F. Chen. Adaptive isogeometric analysis using rational [PHT]{}-splines. , 43(11):1438–1448, 2011.
G. Xu, B. Mourrain, R. Duvigneau, and A. Galligo. Parameterization of computational domain in isogeometric analysis: Methods and comparison. , 200(23-24):2021–2031, 2011.
G. Xu, B. Mourrain, R. Duvigneau, and A. Galligo. Optimal analysis-aware parameterization of computational domain in 3D isogeometric analysis. , 45:812–821, 2013.
[^1]: We also studied a patch-wise interpolation approach. Unfortunately, this approach is neither a cheap one (to compute $y$) nor does it result in desired efficiency indices in the proximity of $1$, and therefore, we do not present its results.
[^2]: Some deviation could be attributed to the fact that we used same number of quadrature points for the evaluation of the matrices in both the cases, which is not sufficient when the PDE coefficients are of exponential form.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'The quantization scheme is suggested for a spatially inhomogeneous 1+1 Bianchi I model. The scheme consists in quantization of the equations of motion and gives the operator (so-called quasi-Heisenberg) equations describing an explicit evolution of a system. Some particular gauge suitable for quantization is proposed. The Wheeler-DeWitt equation is considered in the vicinity of zero scale factor and it is used to construct a space, where the quasi-Heisenberg operators act. Spatial discretization as a UV regularization procedure is suggested for the equations of motion.'
author:
- |
Cherkas S.L. [^1]\
[Institute for Nuclear Problems, Bobruiskaya str. 11, Minsk, 220050, Belarus]{}\
Kalashnikov V.L.[^2]\
[Institut für Photonik, Technische Universität Wien,]{}\
[ Gusshausstrasse 27/387, Vienna A-1040, Austria]{}\
title: '**Quantization of the inhomogeneous Bianchi I model: quasi-Heisenberg picture** '
---
Introduction
============
Spatially homogeneous minisuperspace models [@hr; @kie; @rov] are often used as a testbed for the quantum gravity [@wheel; @witt; @CP; @w; @shest; @hal]. Inhomogeneous 1+1 Bianchi I model belongs to the so-called midisuperspace models [@bar] and has more rich properties, in particular, it admits an existence of gravitational waves. This model can be reduced to the Gowdy one [@gowdy] for which the solution of the Wheeler-DeWitt equation has been obtained in a closed form [@mizng; @berger]. However, the solution of the Wheeler-DeWitt equation does not resolve the problem of the gravity quantization completely. An interpretation of the Wheeler-DeWitt equation encounters the absence of a variable, which would play the role of time, and all approaches to the quantum gravity face, as a rule, this challenge. Some approaches regarding the relation of the Wheeler-DeWitt equation to dynamics have been suggested. For instance, the approach of Ref.[[@hal1]]{} explores the notion of “arrival time" from the non-relativistic quantum mechanics to build the incoherent histories. Also, there exists a more straightforward approach consisting in quantization of the equations of motion [@prep1; @prep2; @gen]. The result of quantization is the so-called quasi-Heisenberg operators.
Below we apply this approach to the quantization of the Bianchi I model. The reason why we investigate the Bianchi I model instead the Gowdy one is that the former has a Hamiltonian, which is diagonal on the momentums. Besides, it divides naturally a spatial geometry into the scale factor and the remaining conformal geometry [@York], while the Gowdy model suggests other separation. Also, the Wheeler-DeWitt equation will be used below since the quasi-Heisenberg operators act in a space of solutions of this equation.
Nonuniform Bianchi model
=========================
Let us consider the metric given by the interval
$$ds^2=e^{2\alpha}\left(d\eta^2-e^{-4 B}dx^2-e^{2
B+2\sqrt{3}\,V}dy^2-e^{2 B-2\sqrt{3}\,V}dz^2\right),$$
where the functions $\alpha, B, V$ depend on the spatial coordinate $x$ and the conformal time $\eta$. The spatially homogeneous metric of such a type has been considered in Ref. [@mizn].
Substitution of this metric into the Einstein equations allows obtaining a set of five independent equations. Two of them are the Hamiltonian and the momentum constraints H=e\^[2]{}(-\^[2]{}+B\^[2]{}+V\^[2]{}) +e\^[2+4 B]{}((\_x )\^2 \
+\_[xx]{}+(\_x B)\^2+\_[xx]{}B+\_x\_x B+(\_x V)\^2)=0,\[hamcon\]\
P= e\^[2]{}(-\_x\^+\_x B B\^+\_x B\^+\_x V V\^+\_xB\^+\_x\^)=0, \[pconcon\] and the rest are the equations of motion \^-e\^[4B]{}(\_[xx]{} +(\_[x]{})\^2+(\_x V)\^2+ (\_x B)\^2+\_[xx]{}B+4\_xB\_x)\
+\^[ 2]{}+ V\^[ 2]{}+B\^[ 2]{} =0, \[11\]\
B\^+e\^[4B]{}(\_[xx]{}B+2(\_[x]{}B)\^2+6(\_x V)\^2-2(\_x)\^2+2\_x\_xB+2\_[xx]{})\
+2 B\^\^=0, \[12\]\
V\^+2 V\^\^-e\^[4B]{}(\_[xx]{} V+2\_xV\_x+4\_xB\_xV)=0. \[10\]
The full Hamiltonian $H=\int \mathcal H dx$ has to be zero during an evolution of system.
The relevant Hamiltonian and the momentum constraints written in the terms of momentums $\pi_V(x)\equiv\frac{\delta H}{\delta {
V^\prime(x) }}=e^{2\alpha}V^\prime$, $\pi_B(x)\equiv\frac{\delta
H}{\delta { B^\prime(x) }}=e^{2\alpha}B^\prime$ and $p_\alpha(x)\equiv-\frac{\delta H}{\delta \alpha
^\prime(x)}=e^{2\alpha}\alpha^\prime$ take the form
H=e\^[-2]{}(-p\_\^[ 2]{}+\_B\^[ 2]{}+\_V\^[ 2]{}) \
+e\^[2+4 B]{}((\_x )\^2+\_[xx]{}+(\_x B)\^2+\_[xx]{}B+\_x\_x B+(\_x V)\^2),
P=-p\_\_x+\_B\_x B+\_V\_x V+\_x \_B+\_x p\_.
Using the Poisson brackets {F(x),G(x\^)}=(-+ \
+- +- )d\[pois\]
one can obtain the constraint algebra: $$\begin{aligned}
\{\mathcal H(x),\mathcal H(x^\prime)\}=(\mathcal P(x)e^{4
B(x)}+\mathcal P(x^\prime)e^{4
B(x^\prime)})\delta^\prime(x^\prime-x),\nonumber\\
\{\mathcal P(x),\mathcal P(x^\prime)\}=(\mathcal P(x)+\mathcal
P(x^\prime))\delta^\prime(x^\prime-x),\nonumber\\
\{\mathcal H(x),\mathcal P(x^\prime)\}=\frac{2}{3}(\mathcal
H(x)+\mathcal
H(x^\prime))\delta^\prime(x^\prime-x)-\frac{1}{3}\mathcal
H^\prime(x)\delta(x^\prime-x).\end{aligned}$$
It is also possible to find the evolution of constraints by calculation of their Poisson brackets with the Hamiltonian $H$:
$$\begin{aligned}
\ptl_\eta\mathcal P(\eta,x)=\{H,\mathcal
P(\eta,x)\}=\frac{1}{3}\ptl_x \mathcal H(\eta,x),\label{sv1}\\
\ptl_\eta\mathcal H=\ptl_x\left(e^{4B}\mathcal
P\right).\label{sv2}\end{aligned}$$
Quantization
============
The quantization procedure consists in the formulation of initial conditions for the quasi-Heisenberg operators. Thereafter it is permissible for operators to evolve in accordance with the equations of motion, considered as the operator equations. To determine the initial commutation relations, we will use the Dirac quantization procedure [@dirac; @han; @git]. Besides the constraints, an additional gauge condition is needed.
Let us take the following gauge at the initial moment of time $$\begin{aligned}
\mathcal A= \alpha-\alpha_0, \label{condal}\\
\mathcal B=\ptl_x \pi_B,\label{condB}\end{aligned}$$ where $\alpha_0$ should be tended to $-\infty$.
For quantization by means of the Dirac brackets [@dirac], one has to calculate the matrix $M_{ij}(x,x^\prime)=\{\Phi_i(x),\Phi_j(x^\prime)\}$, where a set of constraints is $$\Phi_{i}=(\mathcal H,\mathcal P, \mathcal A,\mathcal B).$$
At the shell of constraints $\Phi_i(x)=0$, the matrix $M_{ij}(x,x^\prime)$ in the vicinity $\alpha_0\rightarrow-\infty$ has the form M(x,x\^)=(
[cccc]{} 0&0&&0\
0&0&-&\_B\^(x-x\^)\
-&-&0&0\
0&-\_B\^(x-x\^)&0&0
).
Due to antisymmetry of the Poison brackets, $M_{ij}(x,x^\prime)$ obeys the identity $M_{ij}(x,x^\prime)=-M_{j\,i}(x^\prime,x)$. The inverse matrix satisfying $\int
M_{ij}(x,x^{\prime\prime})M_{j\,k}^{-1}
(x^{\prime\prime},x^{\prime})dx^{\prime\prime}=\delta_{ik}\delta(x-x^\prime)$ is given by $$\bm M^{-1}(x,x^\prime)=\left(
\begin{array}{cccc}
0&0&-e^{2\alpha_0}\frac{\delta(x-x^\prime)}{p_\alpha(x)}&e^{2\alpha_0}\frac{\theta(x-x^\prime)}{3p_\alpha(x)\pi_B}\\
0&0&0&-\frac{\Delta(x-x^\prime)}{\pi_B}\\
e^{2\alpha_0}\frac{\delta(x-x^\prime)}{p_\alpha(x)}&0&0&0\\
-e^{2\alpha_0}\frac{\theta(x^\prime-x)}{3p_\alpha(x^\prime)\pi_B}&\frac{\Delta(x^\prime-x)}{\pi_B}&0&0
\end{array}
\right),$$ where $\theta(x)$ is an antiderivative of the Dirac delta-function: $\theta^\prime(x)=\delta(x)$, and $\Delta(x)$ is an antiderivative of $\theta(x)$: $\Delta^\prime(x)=\theta(x)$. It is assumed, that the antiderivatives have the following symmetry properties: $\theta(-x)=-\theta(x)$, $\Delta(-x)=\Delta(x)$.
For the Dirac quantization, one needs to calculate the Dirac brackets which give the commutation relations for the corresponding operators at an initial moment of time after multiplication by $-i$.
Calculation of the Dirac brackets $$\begin{aligned}
\{G(x),F(x^\prime)\}_D=\{G(x),F(x^\prime)\}~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
\\-\sum_{i,j}\int
\{G(x),\Phi_i(x^{\prime\prime}\}
M^{-1}_{ij}(x^{\prime\prime},x^{\prime\prime\prime})\{\Phi_j(x^{\prime\prime\prime}),F(x^\prime)\}dx^{\prime\prime}dx^{\prime\prime\prime}\end{aligned}$$ leads to $$\begin{aligned}
\{\pi_V(x),V(x^\prime)\}_D=\delta(x-x^\prime),\nonumber \\
\{\pi_B(x),B(x^\prime)\}_D=0,\nonumber\\
\{\pi_\alpha(x),\alpha(x^\prime)\}_D=0,\nonumber\\
\{\pi_B(x),V(x^\prime)\}_D=0,\nonumber\\
\{\pi_B(x),p_\alpha(x^\prime)\}_D=0,\nonumber\\
\{B(x),V(x^\prime)\}_D=-\frac{\pi_V(x)}{3p_\alpha(x)\,\pi_B}\delta(x-x^\prime),\nonumber\\
\{B(x),p_\alpha(x^\prime)\}_D=-\frac{1}{\pi_B}(\ptl_{x^{\prime}}p_\alpha(x^\prime)\theta(x^\prime-x)+p_\alpha(x^\prime)
\delta(x^\prime-x)),\nonumber\\
\{\pi_B(x),\alpha(x^\prime)\}_D=0,\nonumber\\
\{\pi_\alpha(x),V(x^\prime)\}_D=\frac{\pi_V(x)}{p_\alpha(x)}\delta(x^\prime-x).\end{aligned}$$
From the foregoing it follows that $\alpha$ and $\pi_B$ are initially $c$-numbers in the gauge considered.
In fact we use some time-dependent gauge, which is known only at an initial moment of time, and it is permissable for the commutation relations in the model under consideration to evolve.
Operator realization of the commutation relations at the initial moment of time (corresponding to $\alpha_0\rightarrow -\infty$) may be written as $$\hat \pi_V(x)=-i\frac{\delta}{\delta V(x)},$$
$$\hat p_\alpha(x)=\sqrt{\hat \pi_V^2(x)+\pi_B^2},$$
$$\hat B(x)=-\frac{1}{\pi_B}\left(\int_0^x S(\hat
\pi_V(x^\prime)\ptl_{x^\prime}V(x^\prime))dx^\prime
+\frac{1}{3}\hat p_\alpha(x)\right),$$
where the symbol $S$ denotes symmetrization of the noncommutative operators, i.e. $S(\hat A \hat B)=\frac{1}{2}(\hat A\hat B+\hat
B\hat A )$ or $S(\hat A\hat B\hat C)=\frac{1}{6}(\hat A\hat B\hat
C+\hat B\hat A\hat C+\hat A\hat C\hat B+\dots)$.
Thus the equation of motion (\[11\]),(\[12\]),(\[10\]) should be considered as the operator equations of motion with the initial conditions V(0,x)=V(x), V\^(0,x)=-i e\^[-2\_0]{}, \
B(0,x)=-(\_0\^x S(-i \_[x\^]{}V(x\^))dx\^+),\
B\^(0,x)=e\^[-2\_0]{}\_B, (0,x)=\_0, \^(0,x)=e\^[-2\_0]{}, \[rlz2\] where $\pi_B$ is some constant and $\alpha_0$ should be tended to $-\infty$.
Let us suggest a space, where these operators act. Let’s consider the Wheeler-DeWitt equation in the vicinity of $\alpha\rightarrow -\infty$ $$\left(\frac{\delta^2}{\delta \alpha(x)}-\frac{\delta^2}{\delta^2
V(x)}+\pi_B^2\right)\Psi[\alpha,V]=0, \label{witt}$$ where we take into account that $\pi_B$ is some constant.
Solution of Eq. (\[witt\]) is of the form of the wave packet $$\Psi[\alpha,V]=\int C[\pi_V]\,e^{\int\left(-i
\alpha(x)\sqrt{\pi_B^2+\pi_V^2(x) }+i\pi_V(x) V(x)\right)d
x}\,\mathcal D \pi_V,$$ $\mathcal D \pi_V$ denotes functional integration over $\pi_V(x)$.
A mean value of an arbitrary operator can be evaluated as [@prep1; @prep2; @gen] $$\begin{aligned}
<\Psi|\hat A[\alpha,-i\frac{\delta}{\delta V},V]|\Psi>=i
\int \biggl( \Psi^*[\alpha,V] {\hat D^{1/4}}\hat A\,{\hat
D^{-1/4}}\frac{\delta}{\delta \alpha(x)
}\Psi[\alpha,V]\nonumber\\~~~~~~~~~~~~~~~~~~-\left(\frac{\delta}{\delta
\alpha(x) }\Psi^*[\alpha,V]\right){\hat D^{-1/4}}\hat A\,{\hat
D^{1/4}}\,\Psi[\alpha,V]\biggr)\mathcal D
V\biggr|_{\,\alpha(x)=\alpha_0\rightarrow -\infty }, \label{mean1}\end{aligned}$$ where $\hat D(x)=-\frac{\delta^2}{\delta V^2(x)}+\pi_B^2$ and $\mathcal D V$ denotes functional integration over $V(x)$.
In many cases, it is more convenient to use the momentum representation $\hat \pi_V(x)=\pi_V(x)$, $\hat V(x)=i\frac{\delta
}{\delta \pi_V(x)} $, where the wave function $\psi$ is $${\psi}[\alpha,\pi_V]=C[\pi_V]\exp\left({-i\int\alpha(x)\sqrt{\pi_V^2(x)+\pi_B^2}\,dx}\right).$$ Then, a mean value of an operator becomes $$\begin{aligned}
<\psi|\hat A[\alpha,\pi_V(x),i\frac{\delta}{\delta \pi_V(x)}|\psi>=~~~
~~~~~~~~~~~~~~~~~~~
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~\nonumber\\\int
C^*[\pi_V]e^{-i \int \alpha(x)\sqrt{\pi_V^2(x)+\pi_B^2} \,dx}\hat
A\,e^{i \int
\alpha(x)\sqrt{\pi_V^2(x)+\pi_B^2}\,dx}C[\pi_V]\,\mathcal D
\pi_V\biggl|_{\,\alpha(x)=\alpha_0\rightarrow -\infty }.
\label{mean2}\end{aligned}$$
Thus, one has an exact quantization scheme consisting of the Wheeler-DeWitt equation in the vicinity of small scale factor (\[witt\]), the operator initial conditions (\[rlz2\]) for the equations of motion and the expressions (\[mean1\]), (\[mean2\]) for calculation of the mean values of operators.
Discretization of the operator equations
========================================
At least two generations of physicists can not overcome divergencies arising in rigorous operator formulation of the ordinary QFT in the 4-dimensional Minkowsky space, although some success for the 1+1 and 1+3 dimensional models has been reached [@jaffe]. Here we consider a discretization as a method of regularization of the functional operator equations, which eliminates the infinite quantities. An alternative is to use the discretized action initially, however, we do not consider such a possibility here. The discretization consists in choosing of some spatial box of length $L$ and granulation of it by points $x_i$ separated by distance $\Delta x$. Periodicity condition is implied at $x_1=x_N$.
Continuous oscillators of $V$-field should be replaced by the discrete momentums $ \pi_V(x)\rightarrow \pi_{Vj}/\sqrt{\Delta
x}$, $V(x)\rightarrow V_j /\sqrt{\Delta x}$ [@tirr], where $\pi_{Vj}$ and $V_j$ posses an ordinary commutation relation $[\pi_{Vn},V_m]=-i\delta_{nm}$ including the Kronecker symbol $\delta_{m\,n}$. Really in this case for $\pi_V(x_n)$ and $V(x_m)$, we have $[\pi_{V}(x_n),V(x_m)]=-i\delta_{nm}/\Delta x$, that turns into the Dirac delta-function $\delta_{nm}/\Delta x\rightarrow \delta(x_n-x_m)$ in the limit $\Delta x\rightarrow 0$. However, it is more convenient to use straightforwardly the quantities $\pi_{Vj}= \pi_V(x_j)$, $V_j=V(x_j)$ for which the commutation relations $[\pi_{Vn},V_m]=-i\delta_{nm}/\Delta x$ are satisfied. These commutation relations can be realized by the operators $\hat V_m=V_m $, $\hat\pi_{Vn}=-\frac{i}{\Delta x} \frac{\ptl}{\ptl V_n}$ or $\hat
\pi_{Vm}=\pi_{Vm} $, $\hat V_n=\frac{i}{\Delta x} \frac{\ptl}{\ptl
\pi_{Vn}}$.
The discrete Wheeler-DeWitt equation in the vicinity of $\alpha_j=\alpha_0\rightarrow - \infty$ has the following form in the momentum representation:
$$\left(-\frac{1}{(\Delta x)^2}\frac{\ptl^2}{\ptl^2
\alpha_i}+\pi_{Vi}^2+\pi_B^2\right)\psi(\alpha_i,\dots\alpha_N,\pi_{V1},\dots
\pi_{VN})=0,$$
with a solution (\_1…\_N,\_[V1]{},…,\_[VN]{})= \
C(\_[V1]{},…...,\_[VN]{})([-ix \_[j=1]{}\^N\_j]{}), where $\Delta x$ is the discretization length.
A mean value of an arbitrary operator can be evaluated as $$\begin{aligned}
<\psi|\hat A(\alpha_1\dots\alpha_N,\pi_{V1},
\dots\pi_{VN},\frac{i}{\Delta x}\frac{\ptl}{\ptl \pi_{V1}},\dots ,\frac{i}{\Delta x}\frac{\ptl}{\ptl \pi_{VN}})|\psi>
=~~~~~~~~~~~~~~~~~~~~~\nonumber\\\int
C^*(\pi_{V1}\dots\pi_{VN})e^{-i\Delta x \sum_{j=1}^N
\alpha_j\sqrt{\pi_{Vj}^2+\pi_B^2} }\hat A\,e^{i \Delta x
\sum_{j=1}^N \alpha_j\sqrt{\pi_{Vj}^2+\pi_B^2}}\nonumber
\\C(\pi_{V1}\dots\pi_{VN})d\pi_{V1}\dots
d\pi_{VN} \biggl|_{\,\alpha_1\dots\alpha_N=\alpha_0\rightarrow
-\infty }.\end{aligned}$$
For the equations of motion, one has to take some operator ordering and then to come to its discrete version. A question arises about the conservation of constraints during evolution. Constraints can be violated by both noncommutativity of operators and discretization of the equations of motion. The last question is analogous to that regarding the energy conservation in a system of the decretized magnetic hydrodynamic equations [@popov]. Special discretization schemes were suggested for the magnetic hydrodynamic equations, which conserve the energy [@popov].
Let us propose a solution of both problems. Let’s consider the discretized operator Hamiltonian, the momentum constraints and, besides, choose the symmetrical ordering of operators:
\_j=S( e\^[2\_j]{}(-\_j\^+ B\_j\^+ B\_j\^\
+ V\_j\^++)),\[pcond\]\
\_j=S(e\^[2\_j]{}(-\_j\^[2]{}+B\_j\^[2]{}+V\_j\^[2]{}) +e\^[2\_j+4 B\_j]{}(( )\^2\
++( )\^2+\
++( )\^2)).\[hamcond\]
Here, we need more careful definition of the symmetrization [@masl; @kar]. Let’s there is an arbitrary function $f(x_1, x_2, \cdots x_n)$ of $n-$ variables. Then, one can define a formal Fourier transform f(\_1,\_2…\_n)=f(x\_1,x\_2,…x\_n)e\^[-i(x\_1\_1+x\_2\_2…x\_n\_n )]{}d x\_1…d x\_n. A symmetrized function of noncommutative operators $\hat A_1, \dots \hat A_n $ is defined as S(f(A\_1,A\_2…A\_n))=f(\_1,\_2,…\_n)e\^[i(A\_1\_1+A\_2\_2…A\_n\_n )]{}d \_1…d \_n.
Our idea is to use the discretized version of Eqs. (\[sv1\]),(\[sv2\]) as the equations of motion:
\_\_j=,\[eqHd\]\
\_\_j=.\[eqPd\]
Using the formula for differentiation of a symmetrized function [@kar]: S(f(A\_1(t),A\_2(t)…A\_n(t)))=S(\_j\^n [\_j]{} f(A\_1(t),A\_2(t)…A\_n(t))), \[differ\] (here $\partial_j f(x_1,x_2\dots x_n)$ denotes a partial derivative of a function $f$ over the $j-$argument) allows calculating the time derivatives in the left hand side of Eqs. (\[eqHd\],\[eqPd\]) and rewriting Eqs. (\[eqHd\],\[eqPd\]) in the form of S( e\^[2\_j]{}(B\_j\^B\_j\^+V\_j\^V\_j\^- \_j\^\_j\^+ e\^[4B\_j]{}(7 B\_j\^2+7B\_[j+1]{}\^2+2 B\_[j-1]{}+3V\_j\^2\
-2B\_j(2+7B\_[j+1]{}-4\_j+4\_[j+1]{})+2B\_[j+1]{}(1-4\_j+4\_[j+1]{})-6V\_jV\_[j+1]{}\
+3V\_[j+1]{}\^2-4\_j+\_j\^2 +2\_[j+1]{}-2\_j\_[j+1]{}+\_[j+1]{}\^2 +2\_[j-1]{})(2 B\_j\^+\_j\^)\
+\_j\^(B\_j\^[2]{}+V\_j\^[2]{}-\_j\^[2]{})+ e\^[4 B\_j]{}((-2+7B\_j-7B\_[j+1]{}+4\_j-4\_[j+1]{})B\_j\^\
+(1-7B\_j+7 B\_[j+1]{}-4\_j+4\_[j+1]{})B\_[j+1]{}\^+B\_[j-1]{}\^+ 3V\_jV\_j\^-3V\_[j+1]{}V\_j\^\
-3V\_jV\_[j+1]{}\^+3V\_[j+1]{}V\_[j+1]{}\^- 2\_j\^+4B\_j\_j\^-4B\_[j+1]{}\_j\^+\_j\_j\^-\_[j+1]{}\_j\^+ \_[j+1]{}\^\
-4B\_j\_[j+1]{}\^+4B\_[j+1]{}\_[j+1]{}\^-\_j\_[j+1]{}\^+\_[j+1]{} \_[j+1]{}\^+\_[j-1]{}\^) ) )\
=, \[discrh\]\
S(e\^[2\_j]{}(-3B\_j\^[2]{}-3V\_j\^[2]{}+2B\_[j+1]{}\^\_j\^-\_j\^[2]{}+2\_j\_j\^[2]{}-2\_[j+1]{}\_j\^[2]{}\
+3V\_j\^(V\_[j+1]{}\^+2(-V\_j+V\_[j+1]{})\_j\^)+\_j\^\_[j+1]{}\^+ B\_j\^(3B\_[j+1]{}\^+(-4-6B\_j+6B\_[j+1]{}-4\_j\
+4\_[j+1]{})\_j\^+2\_[j+1]{}\^)+ (-1-3B\_j+3B\_[j+1]{})B\_j\^-2\_jB\_j\^+2\_[j+1]{}B\_j\^+B\_[j+1]{}\^\
+3(-V\_j+V\_[j+1]{})V\_j\^-\_j\^+(\_j-\_[j+1]{})\_j\^+\_[j+1]{}\^ ))=[\_[j+1]{}-\_[j]{}]{}, \[discrp\] where $\hat{\mathcal H}_{j}$ and $\hat{\mathcal P}_{j}$ are given by (\[pcond\]), (\[hamcond\]). Moreover, one can consider Eqs. (\[discrh\]), (\[discrp\]) without the right hand side as the equations of motion, since $\hat{\mathcal H}_j$, $\hat{\mathcal
P_j}$ equal to zero initially at $\alpha_0\approx -\infty$ in accordance with the initial conditions. One must be sure only that the time derivatives $\hat{\mathcal H}^\prime_j$, $\hat{\mathcal
P}^\prime_j$ equal to zero, as well.
As the third equation, the discretized version of Eq. (\[10\]) can be taken V\^\_j+S(2 V\_j\^\_j\^-e\^[4B]{}( \
+2+4))=0.\[last\]
Thus we have three operator equations (\[discrh\]),(\[discrp\]),(\[last\]) (or equivalently (\[discrp\]),(\[last\]) without the right-hand side) which for commuting quantities and in the continuous limit $\Delta
x \rightarrow 0$ are fully equivalent to the classical equations (\[11\]), (\[12\]), (\[10\]).
The equations should be solved with the following initial condition V\_j(0)=, V\_j\^(0)=e\^[-2\_0]{}\_[Vj]{}, \
B\_j(0)=-(+\_[k=1]{}\^j S(\_[Vk]{}(-))),\
B\_j\^(0)=e\^[-2\_0]{}\_B, \_j(0)=\_0, \^\_j=e\^[-2\_0]{}. \[in2\]
In the vicinity of small scale factors, the time derivatives are much larger than the spatial ones so that the operator equations take a simple form
\_j\^+\_j\^[ 2]{}+ V\_j\^[ 2]{}+B\_j\^[ 2]{} =0, \[11a\]\
B\_j\^+2 B\_j\^\_j\^=0, \[12a\]\
V\_j\^+2 V\_j\^\_j\^=0, \[10a\] with the constraint $-\hat\alpha_j^{\prime 2}+\hat B_j^{\prime
2}+\hat V_j^{\prime 2}=0$. All quantities, i.e. $\hat
\alpha_j^{\prime},\hat B_j^{\prime}, \hat V_j^{\prime}$ in these equations commute with each other.
The solution of Eqs. (\[11a\]),(\[11a\]),(\[12a\]) with the initial conditions (\[in2\]) is given as $$\begin{aligned}
\hat V_j(\eta)=\hat V_j(0)+\frac{
\pi_{Vj}}{2\sqrt{\pi_B^2+\pi_{Vj}^2}}\ln\left(1+2e^{-2 \alpha_0
}\sqrt{\pi_B^2+\pi_{Vj}^2}\,\eta\right),\\
\hat B_j(\eta)=\hat B_j(0)+\frac{
\pi_{B}}{2\sqrt{\pi_B^2+\pi_{Vj}^2}}\ln\left(1+2e^{-2 \alpha_0
}\sqrt{\pi_B^2+\pi_{Vj}^2}\,\eta\right),\\
\hat \alpha_j(\eta)=\alpha_0+\frac{1}{2}\ln\left(1+2e^{-2 \alpha_0
}\sqrt{\pi_B^2+\pi_{Vj}^2}\,\eta\right).\end{aligned}$$
Although the evolution in the vicinity of $\alpha\approx\alpha_0\approx - \infty$ is relatively simple, the evolution governed by the general equations (\[discrh\]), (\[discrp\]), (\[last\]), when the fields begin to oscillate is very complicated and needs the numerical investigation that is beyond the aims of the present article.
Discussion and Conclusion
=========================
It should be said about a choice of a state of system considered. We do not identify the state as an “initial" because there is only the state $C(\pi_{V1},\dots...,\pi_{VN})$, which describes all evolution of system and allows calculating the mean values of the quasi-Heisenberg operators. This state is not related to the notion of “vacuum" since there is no field oscillators in the limit $\alpha\rightarrow -\infty$. In Ref. [@mizn] solution of the Wheeler-DeWitt equation for the Gowdy model is investigated. Although the another quantities were introduced in [@mizn] instead of logarithm of scale factor $\alpha$ and $B$-field, namely T=B+,\
=6(B-) the asymptotic of the Wheeler-DeWitt equation in the vicinity $T\rightarrow -\infty$ has the form containing only the momentums by analogy with Eq. (\[witt\]) and admits solutions of the plane wave type. Asymptotic of the Wheeler-DeWitt equation in the vicinity $T\rightarrow \infty$ is of the oscillator type [@mizn]. Quasiclassical treatment of the Wheeler-DeWitt equation with regard to an evolution of universe allows an interpretation as a scattering problem [@mizn] i.e. transition of the packet of plane waves at $T\rightarrow -\infty$ to the number of gravitons (i.e exited oscillators) at $T\rightarrow \infty$. It should be noted that there were no the state at $T\rightarrow
-\infty$ which give no gravitons at $T\rightarrow \infty$. The later work [@berger] about a quantization of the Gowdy model was in some sense a step back because it considered graviton creation from vacuum in a style of Refs. [@Par69; @SexUrb69; @ZelSta71; @frol]. It was suggested [@berger] that at some time $T_0$ where the field oscillators already have formed there is no gravitons and later the gravitons appear from vacuum during an evolution. However, our opinion is that there is no physical background for the existence of such a time $T_0$ where the universe was empty.
In our evolutionary picture, we have also the state describing evolution of universe as a packet of the “plane waves". In a future this will give some gravitons under the vacuum. More exactly if one calculated the correlators $<\hat V(\eta,x)\hat
V(\eta,x^\prime)>$ he will find that they are analogous to the correlators of the QFT corresponding to some gravitons under vacuum. There is no wave packet which gives a pure vacuum of universe in the future, i.e. appearance of matter is inevitable in this model but the matter is not created from vacuum, because at any time the universe is not empty.
As an example of the initial state one may take C(\_[V1]{},…...,\_[VN]{})=C\_N(-b\_[j=1]{}\^N \_[Vj]{}\^2), where $C_N$ is the normalization constant. This state implies that the momentums are random and independent in spatial points. This state does not look similar to a vacuum state of the QFT, as the fields (and their momentums) in a QFT vacuum state are highly correlated in the nearest spatial points.
As an example of the mean value calculation, one can take the evolution of a “Hubble constant" of system at small $\eta$: =<>=<(-\_j)>=U([1]{}/[4]{},[3]{}/[4]{},b \_B\^2), where $U(a,b,z)$ is the confluent hypergeometric function.
For this state, the universe is expanded uniformly in a mean. It would be interesting to calculate the evolution of correlators $<\hat V_j(\eta)\hat
V_n(\eta)>$. Initially the field $\hat V_j$ is uncorrelated for different $j$: $<\hat V_j(\eta)\hat V_n(\eta)>\sim \delta_{jn}$ but then some correlation should arise and some analogy with the vacuum of QFT in this sense should appear so that the fields become highly correlated in the nearest spatial points.
The question arises: why do the commutation relations related to the field $V$ have the usual form initially but do become to be broken after an evolution and, probably, look differently (the difference have to be related to the Hubble constant) at the present time? The answer is that it is a consequence of the chosen initial gauge. The gauge can be different, but the gauge considered is simplest and allows obtaining the commutation relations explicitly at the initial moment of time. As a matter of fact, the gauge choice remains to be complex issue.
To summarize, failure of the QFT in flat spacetime to deal with such an inherently non-linear theory as gravity and existence of “problem of time" insist on an invention of some new quantization procedures. The considered quasi-Heisenberg quantization scheme may provide a calculational framework for the investigation of the quantum evolution of a system. The goal of further investigation may be the vacuum energy problem, more exactly, its possible zero value in the quantization scheme considered, that may result from compensation of the zero point fluctuations of gravitational waves by the quantum fluctuations of the scale factor. Thereby, the fluctuations do not contribute to the mean evolution of a system. It should be noted that this will be purely quantum effect, since it is absent in classics [@riple]. An additional issue is calculation of the field correlators in order to determine their correspondence to the correlators of ordinary QFT in late times.
[99]{}
J. B. Hartle, S.W. Hawking, [*Phys. Rev. D*]{}, [**28**]{} (1983), 2960. C.Kiefer, B.Sandhoefer, [*Quantum Cosmology*]{}, gr-qc/0804.0672.
C. Rovelli, [*Phys. Rev. D*]{} [**42**]{} (1990), 2638–2646.
J.A. Wheeler, [Superspace and Nature of Quantum Geometrodynamics.]{} In: DeWitt, C., Wheeler,J.A. (eds.) Battelle Rencontres, Benjamin, New York, 1968.
B.S. DeWitt, [*Phys. Rev.*]{} [**160**]{} (1967), 1113.
A. Ashtekar, J. Stachel (eds), [*Conceptual problems of quantum gravity.*]{} Birkhäuser, Boston, 1991.
D.L. Wiltshire, [*An introduction to quantum cosmology,*]{} arXiv:gr-qc/0101003.
T.P. Shestakova, C. Simeone, [*Grav. Cosmol.*]{} [**10**]{} (2004), 161.
J.J. Halliwell, [*Introductory Lectures on Quantum Cosmology,*]{} arXiv:0909.2566.
F. Barbero, E.J.S. Villasenor, [*Living Rev.*]{}, [**6**]{} (2010).
R.H. Gowdy, [*Ann. Phys. (N.Y.)*]{}, [**83**]{} (1974), 203 .
C.W. Mizner, [*Phys. Rev.*]{}, [**8**]{} (1973), 3271.
B.K. Berger, [*Ann. Phys.,*]{} [**83**]{} (1974), 458.
J.J.Halliwell, [*J. Phys., Conf. Ser.*]{}, [**306**]{} (2011), 012023.
S.L.Cherkas, V.L. Kalashnikov, [*Quantum evolution of the Universe from $\tau=0$ in the constrained quasi-Heisenberg picture.* ]{}, Proc. VIIIth International School-seminar “The Actual Problems of Microworld Physics”, (Gomel, July 25-August 5) Vol. 1. pp. 208. JINR, Dubna (2006), arXiv:gr-qc/0502044.
S.L.Cherkas, V.L. Kalashnikov, [*Grav.Cosmol.*]{} [**12**]{}(2006), 126, arXiv:gr-qc/0512107.
S.L. Cherkas, V.L. Kalashnikov, [*Gen. Rel. Grav.*]{}, [**44**]{} (2012), 3081-3102.
J. W. Jr. York, [*Phys. Rev. Lett.*]{} [**26**]{} (1971), 1656 .
C.W. Mizner, K.S. Torn and J.A. Wheeler. [*Gravitation*]{}.(W. H. Freeman & Company, New-York) Vol. 2., 1973.
P.A.M. Dirac, [*Lectures on Quantum Mechanics,*]{} Yeshiva University, N.Y., 1964.
A. Hanson, T. Regge, C. Teitelboim, [*Constraint Hamiltonian Systems,*]{} Contributi del Centro Linceo Interdisc. di Scienze Matem. e loro Applic **22** (1976).
D.M. Gitman, I.V. Tyutin, [*Quantization of Fields with Constraints.*]{} Springer, Berlin, 1990.
A. Jaffe, [*Constructive quantum field theory*]{}, in Mathematical Physics 2000, A. S. Fokas, A. Grigorian, T. Kibble, B. Zegarlinsky, Eds., Proceedings of the XIII International Congress on Mathematical Physics, Imperial College, London 17-22 July 2000, Imp. Coll. Press, London, 2000,111-127,posted at http://www.arthurjaffe.com/Assets/pdf/CQFT.pdf.
E.M. Henley, W. Thirring, [*Elementary quantum field theory*]{}, McGraw-Hill Book Company, Inc., New York, 1962.
A.A. Samarskii, Yu. P. Popov, [*Raznostnye metody reshenia zadach gazovoi dinamiki*]{}, Nauka, Moskow 1992, \[in Russian\].
V.P. Maslov, [*Operator Methods,*]{} Nauka, Moscow (1973) \[in Russian\]
M. V. Karasev, V. P. Maslov, [*Nonlinear Poisson Brackets: Geometry and Quantization,*]{} American Mathematical Society, Providence, 1993.
L. Parker, [ *Phys. Rev.,*]{} [**183**]{} (1969), 1057.
R. U. Sexl, H. K. Urbantke, [*Phys. Rev.,*]{} [**179**]{} (1969), 1247.
Ya. Zel’dovich, A. Starobinsky, [*Zh. Eksp. Teor. Fiz.,*]{}, [**61**]{} (1971), 2161-2175 , \[ [*Sov. Phys.- JETP,*]{} [**34**]{} (1972), 1159. \] A.A. Grib, S.G. Mamaev, [*Yad.Fiz.*]{} [**10**]{} (1969), 1276.
S. L. Cherkas, V. L. Kalashnikov, *Nonlinear Phenomena in Complex Systems* **15** (2012), 253, arXiv:1206.5976.
[^1]: E-mail:[email protected]
[^2]: E-mail:[email protected]
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We construct a six-dimensional gauge-Higgs unification model with the enlarged gauge group of $E_6$ on $S^2/Z_2$ orbifold compactification. The standard model particle contents and gauge symmetry are obtained by utilizing a monopole background field and imposing appropriate parity conditions on the orbifold. In particular, a realistic Higgs potential suitable for breaking the electroweak gauge symmetry is obtained without introducing extra matter or assuming an additional symmetry relation between the SU(2) isometry transformation on the $S^2$ and the gauge symmetry. The Higgs boson is a KK mode associated with the extra-dimensional components of gauge field. Its odd KK-parity makes it a stable particle, and thus a potential dark matter candidate in the model. We also compute the KK masses of all fields at tree level.'
author:
- 'Cheng-Wei Chiang'
- Takaaki Nomura
title: 'A six-dimensional gauge-Higgs unification model based on $E_6$ gauge symmetry'
---
Introduction \[sec:intro\]
==========================
Precision electroweak measurements suggest that the mass of the Higgs boson in the standard model (SM) should be of ${\cal O}(100)$ GeV. However, this leads to a serious problem, the so-called hierarchy problem, because the Higgs boson mass generally suffers from quadratic divergence at the quantum level. It is thus unnatural for the Higgs boson to be so light if the theory cutoff scale is high, unless some mechanism is introduced for stabilization. Such a problem generally calls for some symmetry ([*e.g.*]{}, supersymmetry) to control the scalar sector and leads to physics beyond the SM.
In the late 70’s, an alternative method to stabilize the Higgs boson mass had been proposed. The basic idea was to embed the Higgs field as the extra-dimensional components of gauge field in a higher dimensional space, with an enlarged gauge symmetry broken down to the SM gauge group in 4D spacetime [@Manton:1979kb; @Forgacs:1979zs; @Fairlie:1979at]. This idea of gauge-Higgs unification had recently been revived . A desirable feature of such models is that the gauge origin ensures that the Higgs mass in the bulk is protected from quadratic divergence. Moreover, by compactifying the theory on orbifolds, unwanted fields can be projected out from low-energy spectrum. The compactification scale would be taken as TeV scale [@Antoniadis:1990ew; @Antoniadis:1993jp]. A simple implementation of the idea in 5D, however, encounters the difficulty of a small Higgs mass due to the absense of a tree-level Higgs potential. One is then led to consider 6D models because a quartic Higgs interaction term can arise from the gauge kinetic term [@Scrucca:2003ut]. The Higgs mass can also be enhanced through the introduction of a warped spacetime [@Contino:2003ve; @Hosotani:2005nz] or by choosing a suitable bulk matter content [@Haba:2004qf]. However, the quadratic mass term here is still radiatively generated and possibly divergent. A more successful 6D model based on the SO(12) gauge group was proposed, where a monopole background exists to break the higher dimensional symmetry and results in a negative squared mass [@Nomura:2008sx]. Nevertheless, a set of symmetries relating the SU(2) isometry transformation on $S^2$ to the gauge transformation of the gauge fields has to be imposed in order to carry out dimensional reduction of the gauge sector. This approach of dimensional reduction is known as the coset space dimensional reduction, and leads to a stronger constraint on the four dimensional Lagrangian after dimensional reduction [@Manton:1979kb; @Forgacs:1979zs; @Kapetanakis:1992hf].
We consider a gauge-Higgs unification model defined on the 6D spacetime where the extra spatial dimensions are compactified on a 2-sphere $S^2$. The gauge symmetry in the model constructed here is assumed to be $E_6$. With a background field configuration and suitable boundary conditions of $S^2$ on the fields, we obtain the full SM particle contents as the zero modes in the model. In particular, no relation between extra-dimensional isometry and gauge symmetries is needed. We are able to identify a Higgs boson doublet coming from the two extra-spatial components of the gauge fields in the adjoint representation. Unwanted modes are either projected out by compactification or given masses due to the interaction with the background field. The Higgs potential in the effective 4D theory has the desired form to break the electroweak symmetry. The compactification scale is fixed with the input of the $W$ boson mass. A mass relation between the Higgs and $W$ bosons is obtained. The Weinberg angle is the same as the usual SU(5) grand-unified theory (GUT). Moreover, the Higgs particle is a Kaluza-Klein (KK) mode with an odd KK-parity. It is stable under a $Z_2$ symmetry and thus a potential dark matter candidate. Discussions about the dark matter candidate are also given in other gauge-Higgs unification models [@Panico:2008bx; @Carena:2009yt; @Hosotani:2009jk; @Haba:2009xu].
This paper is organized as follows. In Section \[sec:model\], we describe the 6D model compactified on the $S^2/Z_2$ orbifold. A consistent set of parity assignments of fields in both representations is given, followed by reviewing the branching of the $E_6$ group and the reduction of its fundamental and adjoint representations. We then work out the details of obtaining the SM particle contents as the zero modes of gauge and fermion fields in the model. In Section \[sec:higgs\], we identify a KK mode of an appropriate representation of the extra-dimensional components of gauge field as the Higgs field in the SM. After obtaining the required commutation relations of gauge generators, we compute the tree-level Higgs potential. The result is then used to obtain a relation between the Higgs mass and the $W$ boson mass. In Section \[KKmass\], we discuss the KK mode mass spectra for fermions and gauge bosons in the existence of the background gauge field. We find that the Higgs boson in the model is a potential candidate of dark matter due to its odd KK-parity. Our findings are summarized in Section \[sec:summary\].
Model \[sec:model\]
===================
In this section, we develop the model based on E$_6$ gauge symmetry in six-dimensional spacetime with $S^2/Z_2$ extra space. On the orbifold $S^2/Z_2$, a set of non-trivial boundary conditions is imposed to restrict the gauge symmetry and massless particle contents in four-dimensional spacetime. We also introduce in this model a background gauge field, which corresponds to a Dirac monopole configuration, to obtain chiral fermions in four dimensions. We then show how the E$_6$ gauge symmetry is reduced to the SM gauge symmetry with some extra U(1)’s, [*i.e.*]{}, SU(3) $\times$ SU(2) $\times$ U(1)$_Y$ $\times$ U(1)$_X$ $\times$ U(1)$_Z$, and how the massless gauge bosons and the SM Higgs boson in four dimensions are obtained in the model. We note in passing that all gauge groups of lower ranks ([*e.g.*]{}, SO(10), SO(11), SU(6), etc) either cannot give a Higgs field in the right representation or do not support SM chiral fermions in four dimensions.
Action in six-dimensional spacetime
-----------------------------------
We start by considering the E$_6$ gauge symmetry group in six-dimensional spacetime, which is assumed to be a direct product of the four-dimensional Minkowski spacetime $M^4$ and the compactified two-sphere orbifold $S^2/Z_2$, [*i.e.*]{}, $M^4 \times S^2/Z_2$. The two-sphere has a radius of $R$. We denote the six-dimensional spacetime coordinates by $X^M = (x^{\mu}, y^{\theta}=\theta, y^{\phi}=\phi)$, where $x^{\mu}$ and $\{ \theta, \phi \}$ are the $M^4$ coordinates and spherical coordinates of $S^2$, respectively. The spacetime index $M$ runs over $\mu \in \{0,1,2,3 \}$ and $\alpha \in \{ \theta, \phi \}$. The orbifold $S^2/Z_2$ is defined by the identification of $(\theta,\phi)$ and $(\pi - \theta,-\phi)$ [@Maru:2009wu]. The two fixed points are $(\pi/2,0)$ and $(\pi/2,\pi)$. The spacetime metric of $M^6$ is g\_[MN]{} =
\_ & 0\
0 & -g\_
, where $\eta_{\mu \nu} = \mbox{diag}(1,-1,-1,-1)$ and $g_{\alpha \beta} = R^2 \mbox{diag}(1,\sin^2 \theta)$ are the metrics associated with $M^4$ and $S^2$, respectively. The action in six-dimensional spacetime is then S\_6 = dx\^4 d R\^2 where $D_{M}$ ($M=0,1,2,3,\theta,\phi$) are covariant derivatives, $\Gamma^{\mu,a}$ are the Dirac gamma matrices in six dimensions, and $e^{\alpha}_a$ are the vielbeins on the two-sphere. Explicitly, $$\begin{array}{lll}
D_{\mu} = \partial_{\mu} - iA_{\mu},
&
D_{\theta} = \partial_{\theta} -i A_{\theta},
&
D_{\phi} = \partial_{\phi} -i \frac{\Sigma_3}{2} \cos \theta -iA_{\phi}, \\
\Gamma_{\mu} = \gamma_{\mu} \otimes \mathbf{I}_2,
&
\Gamma_4 = \gamma_{5} \otimes \sigma_1,
&
\Gamma_5 = \gamma_{5} \otimes \sigma_2, \\
e^1_{\theta} = R,
&
e^2_{\phi} = R \sin \theta,
&
e^1_{\phi} = e^2_{\theta} = 0,
\end{array}$$ where $\sigma_i \, (i=1,2,3)$ are the Pauli matrices, $\mathbf{I}_d$ is the $d \times d$ identity matrix, and $\Sigma_3$ is defined as $\Sigma_3=\mathbf{I}_4 \otimes \sigma_3$. The gauge field strength is $F_{MN}=\partial_M A_N -\partial_N A_M -i[A_M,A_N]$. Note that the covariant derivative $D_{\phi}$ has a spin connection term $i \frac{\Sigma_3}{2} \cos \theta$ for fermions because of the nonzero curvature of the two-sphere. This term generally induces a fermion mass in the four-dimensional effective action after integrating out the extra space. This mass term, as we will see in Section \[sec:fermions\], can be avoided by introducing a background gauge field $A^B_{\phi} \equiv {\tilde A}^B_\phi \sin\theta$ that corresponds to a Dirac monopole [@RandjbarDaemi:1982hi] \[background\] [A]{}\^B\_ = - Q , (-: 0 < , +: ) where $Q$ is proportional to the generator of a U(1) subgroup of the original gauge group E$_6$.
Boundary conditions on the two-sphere orbifold \[sec:bc\]
---------------------------------------------------------
On the two-sphere orbifold, one can consider two parity operations $P_1: \, (\theta,\phi) \to (\pi-\theta,\phi)$ and $P_2: \, (\theta,\phi) \to (\pi-\theta,2\pi-\phi)$, which are related to each other by an azimuthal translation $\phi \to \phi+2\pi$. We impose the following boundary conditions on both gauge and fermion fields under the two parity operations: \[boundary-condition1\] A\_ (x,-,-) &=& P\_1 A\_(x,,) P\_1 ,\
\[boundary-condition2\] A\_[,]{}(x,-,-) &=& - P\_1 A\_[,]{}(x,,) P\_1 ,\
\[boundary-condition3\] (x,-,-) &=& \_5 P\_1 (x,,) ,\
\[boundary-condition4\] A\_ (x,-,2-) &=& P\_2 A\_(x,,) P\_2 ,\
\[boundary-condition5\] A\_[,]{}(x,-,2-) &=& - P\_2 A\_[,]{}(x,,) P\_2 ,\
\[boundary-condition6\] (x,-,2-) &=& \_5 P\_2 (x,,) . These boundary conditions are determined by requiring the invariance of the six -dimensional action under the transformation $(\theta,\phi) \rightarrow (\pi-\theta,-\phi)$.
The projection matrices $P_{1,2}$ act on the gauge group representation space and have eigenvalues $\pm 1$. They assign different parities for different representation components. For fermion boundary conditions, the sign in front of $\gamma_5$ can be either $+$ or $-$ since the fermions always appear in bilinear forms in the action. The 4-dimensional action is then restricted by these parity assignments and our choice of the background gauge field.
Gauge group reduction \[sec:group\]
-----------------------------------
We consider the following gauge group reduction \[group-red\] E\_6 && SO(10) U(1)\_Z\
&& SU(5) U(1)\_X U(1)\_Z\
&& SU(3) SU(2) U(1)\_Y U(1)\_X U(1)\_Z . The background gauge field in Eq. (\[background\]) is chosen to belong to the U(1)$_Z$ group. This choice is needed in order to obtain chiral SM fermions in four dimensions to be discussed later. There are two other symmetry reduction schemes. One can prove that the results in those two schemes are effectively the same as the one considered here once we require the correct U(1) combinations for the hypercharge and the background field.
We then impose the parity assignments with respect to the fixed points, Eqs. (\[boundary-condition1\])-(\[boundary-condition6\]). The parity assignments for the fundamental representation of E$_6$ is chosen to be \[d27\] [**27**]{} &=& (1,2)(-3,-2,-2)\^[(+,+)]{}+(1,2)(3,2,-2)\^[(-,-)]{}+(1,2)(-3,3,1)\^[(+,-)]{}\
&& + (1,1)(6,-1,1)\^[(+,+)]{}+(1,1)(0,-5,1)\^[(-,-)]{}+(1,1)(0,0,4)\^[(-,+)]{}\
&& + (3,2)(1,-1,1)\^[(-,+)]{}+(3,1)(-2,2,-2)\^[(+,-)]{} + (|[3]{},1)(-4,-1,1)\^[(+,+)]{}\
&& +(|[3]{},1)(2,3,1)\^[(+,+)]{}+(|[3]{},1)(2,-2,2)\^[(-,+)]{}, where, for example, $(+,-)$ means that the parities under $P_1$ and $P_2$ are (even,odd). By the requirement of consistency, we find that the components of $A_{\mu}$ in the adjoint representation have the parities under $A_{\mu} \rightarrow P_1 A_{\mu} P_1$ $(P_2 A_{\mu} P_2)$ as follows: $$\begin{aligned}
\label{d78}
{\bf 78}|_{A_{\mu}} &=& \underline{(8,1)(0,0,0)^{(+,+)}+(1,3)(0,0,0)^{(+,+)}} \nonumber \\
&& + \underline{(1,1)(0,0,0)^{(+,+)} +(1,1)(0,0,0)^{(+,+)}+(1,1)(0,0,0)^{(+,+)}} \nonumber \\
&& + (3,2)(-5,0,0)^{(-,+)}+(\bar{3},2)(5,0,0)^{(-,+)}+ (3,2)(1,4,0)^{(+,-)}+ (\bar{3},2)(-1,-4,0)^{(+,-)} \nonumber \\
&& +(3,1)(4,-4,0)^{(-,-)}+(\bar{3},1)(-4,4,0)^{(-,-)}+(1,1)(-6,-4,0)^{(-,-)}+(1,1)(6,4,0)^{(-,-)} \nonumber \\
&& + (3,2)(1,-1,-3)^{(+,+)}+ (\bar{3},2)(-1,1,3)^{(+,+)}+(3,1)(4,1,3)^{(-,+)}+(\bar{3},1)(-4,-1,-3)^{(-,+)} \nonumber \\
&& + (3,1)(-2,-3,3)^{(+,-)}+ (\bar{3},1)(2,3,-3)^{(+,-)}+(1,2)(-3,3,-3)^{(-,-)}+(1,2)(3,-3,3)^{(-,-)} \nonumber \\
&& +(1,1)(-6,1,3)^{(-,+)}+(1,1)(6,-1,-3)^{(-,+)}+(1,1)(0,-5,-3)^{(+,-)}+(1,1)(0,5,3)^{(+,-)}, \nonumber \\\end{aligned}$$ where the underlined components correspond to the adjoint representations of SU(3) $\times$ SU(2) $\times$ U(1)$_Y$ $\times$ U(1)$_X$ $\times$ U(1)$_Z$, respectively. We note that the components with parity $(+,+)$ can have massless zero modes in four dimensions. Such components include the adjoint representations of SU(3) $\times$ SU(2) $\times$ U(1)$^3$, $(3,2)(1,-1,-3)$ and its conjugate. The latter components seem problematic since they do not appear in the low-energy spectrum. In fact, these components acquire masses due to the background field from the term proportional to $F_{\mu \phi} F^{\mu}_{\ \phi}$ && Tr\
&& Tr . For the components of $A_{\mu}$ with nonzero U(1)$_Z$ charge, we have $$A_{\mu}^i Q_i+ A_{i\mu} Q^{i} \in A_{\mu} ~,$$ where $Q_i \, (Q^i = Q_i^{\dagger})$ are generators corresponding to distinct components in Eq. (\[d78\]) that have nonzero U(1)$_Z$ charges, and $A_{i\mu} \, (A_{\mu}^i = A_{i \mu}^{\dagger})$ are the corresponding components of $A_{\mu}$. We then find the term $$\begin{aligned}
\frac{1}{\sin^2 \theta}Tr[[A_{\mu},A^B_\phi][A^{\mu},A^B_\phi]]
&=& \frac{(\cos \theta \mp 1)^2}{\sin^2 \theta}
Tr[[A_{\mu}^i Q_i+A_{i \mu} Q^i,Q][ A^{i \mu} Q_i+A_{i}^{\mu} Q^i,Q]] \nonumber \\
&=& -2 |q|^2 \frac{(\cos \theta \mp 1)^2}{\sin^2 \theta} A^{i \mu} A_{i \mu} ~,\end{aligned}$$ where $q$ is the $Q$ charge of the relevant component. Use of the facts that $A_{\phi}^B$ belongs to U(1)$_Z$ and that $Tr[Q_i Q^i]=2$ has been made in the above equation. A mass is thus associated with the lowest modes of those components of $A_{\mu}$ with nonzero U(1)$_Z$ charges: && d Tr.|\_[lowest]{}\
&& - + m\^2\_B A\_\^i(x) A\^[i]{}(x) , where the subscript ‘lowest’ means that only the lowest KK modes are kept. Here the lowest KK modes of $A_{\mu}$ correspond to the term $A_{\mu}(x)/\sqrt{4 \pi}$ in the KK expansion. In summary, any representation of $A_\mu$ carrying a nonzero U(1)$_Z$ charge acquires a mass $m_B$ from the background field contribution after one integrates over the extra spatial coordinates. More explicitly, \[eq:nonSMgaugeMass\] m\^2\_B = d 0.39 for the zero mode. Therefore, the $(3,2)(1,-1,-3)$ representation and its conjugate are elevated in mass to disappear from the low-energy spectrum. In the end, the correct symmetry reduction is achieved since only the components of 4-dimensional gauge field $A_{\mu}$ in the adjoint representation of SU(3) $\times$ SU(2) $\times$ U(1)$_Y$ $\times$ U(1)$_X$ $\times$ U(1)$_Z$ are allowed to have zero modes. A general discussion about the KK mode masses of $A_{\mu}$ will be given in Section \[KKmass\].
Scalar field contents in four dimensions \[sec:scalar\]
-------------------------------------------------------
The scalar contents in four dimensions are obtained from the extra-dimensional components of the gauge field $\{ A_{\theta}, A_{\phi} \}$ after integrating out the extra spatial coordinates. The kinetic term and potential term of $\{A_{\theta}, A_{\phi} \}$ are obtained from the gauge sector containing these components $$\begin{aligned}
\label{action-scalar}
S_{\rm scalar}
&=& \int dx^4 d \Omega \Bigl( \frac{1}{2 g^2} Tr[ F_{\mu \theta} F^{\mu}_{\ \theta} ] + \frac{1}{2 g^2 \sin^2 \theta} Tr[ F_{\mu \phi} F^{\mu}_{\ \phi} ] \nonumber \\
&& \qquad \qquad -\frac{1}{2 g^2 R^2 \sin^2 \theta} Tr[ F_{\theta \phi} F_{\theta \phi} ] \Bigr) \nonumber \\
&\rightarrow& \int dx^4 d \Omega \Bigl( \frac{1}{2 g^2} Tr[(\partial_{\mu} A_{\theta}-i[A_{\mu},A_{\theta}])^2]
+ \frac{1}{2 g^2} Tr[(\partial_{\mu} A_{\theta}-i[A_{\mu},\tilde{A}_{\phi}])^2 ] \nonumber \\
&& \qquad \qquad -\frac{1}{2 g^2 R^2} Tr \biggl[ \biggl( \frac{1}{\sin \theta} \partial_{\theta} (\sin \theta \tilde{A}_{\phi} + \sin \theta \tilde{A}^B_{\phi}) -\frac{1}{\sin \theta} \partial_{\phi} A_{\theta}
- i[A_{\theta},\tilde{A}_{\phi}+\tilde{A}^B_{\phi}] \biggr)^2 \biggr] ~, \nonumber \\\end{aligned}$$ where we have taken $A_{\phi} = \tilde{A}_{\phi} \sin \theta + \tilde{A}_{\phi}^B \sin \theta$. In the second step indicated by the arrow in Eq. (\[action-scalar\]), we have omitted terms which do not involve $A_{\theta}$ and $\tilde{A}_{\phi}$ from the right-hand side of the first equality. It is known that one generally cannot obtain massless modes for physical scalar components in four dimensions [@Maru:2006; @Dohi:2010vc]. One can see this by noting that the eigenfunction of the operator $\frac{1}{ \sin \theta } \partial_{\theta} \sin \theta$ with zero eigenvalue is not normalizable [@Maru:2006]. In other words, these fields have only KK modes. However, an interesting feature is that it is possible to obtain a negative squared mass when taking into account the interactions between the background gauge field $\tilde{A}_{\phi}^B$ and $\{A_{\theta}, \tilde{A}_{\phi} \}$. This happens when the component carries a nonzero U(1)$_Z$ charge, as the background gauge field belongs to U(1)$_Z$. In this case, the $(\ell=1,m=1)$ modes of these real scalar components are found to have a negative squared mass in four dimensions. They can be identified as the Higgs fields once they are shown to belong to the correct representation under the SM gauge group. Here the numbers $(\ell,m)$ are the angular momentum quantum number on $S^2/Z_2$, and each KK mode is characterized by these numbers. One can show that the $(\ell=1,m=0)$ mode has a positive squared mass and is not considered as the Higgs field. A discussion of the KK masses with general $(\ell,m)$ will be given in Section \[KKmass\] .
With the parity assignments with respect to the fixed points, Eqs. (\[boundary-condition2\]) and (\[boundary-condition5\]), we have for the $A_{\theta}$ and $A_{\phi}$ fields \[78scalar\] [**78**]{}|\_[A\_[,]{}]{} & = & (8,1)(0,0,0)\^[(-,-)]{}+(1,3)(0,0,0)\^[(-,-)]{}\
&& +(1,1)(0,0,0)\^[(-,-)]{} +(1,1)(0,0,0)\^[(-,-)]{}+(1,1)(0,0,0)\^[(-,-)]{}\
&& + (3,2)(-5,0,0)\^[(+,-)]{}+(|[3]{},2)(5,0,0)\^[(+,-)]{}+ (3,2)(1,4,0)\^[(-,+)]{}+ (|[3]{},2)(-1,-4,0)\^[(-,+)]{}\
&& +(3,1)(4,-4,0)\^[(+,+)]{}+(|[3]{},1)(-4,4,0)\^[(+,+)]{}+(1,1)(-6,-4,0)\^[(+,+)]{}+(1,1)(6,4,0)\^[(+,+)]{}\
&& + (3,2)(1,-1,-3)\^[(-,-)]{}+ (|[3]{},2)(-1,1,3)\^[(-,-)]{}+(3,1)(4,1,3)\^[(+,-)]{}+(|[3]{},1)(-4,-1,-3)\^[(+,-)]{}\
&& + (3,1)(-2,-3,3)\^[(-,+)]{}+ (|[3]{},1)(2,3,-3)\^[(-,+)]{}+(1,2)(-3,3,-3)\^[(+,+)]{}+(1,2)(3,-3,3)\^[(+,+)]{}\
&& +(1,1)(-6,1,3)\^[(+,-)]{}+(1,1)(6,-1,-3)\^[(+,-)]{}+(1,1)(0,-5,-3)\^[(-,+)]{}+(1,1)(0,5,3)\^[(-,+)]{} .\
Components with $(+,-)$ or $(-,+)$ parity do not have KK modes since they are odd under $\phi \rightarrow \phi+2\pi$ and the KK modes of gauge field are specified by integer angular momentum quantum numbers $\ell$ and $m$ on the two-sphere. We then concentrate on the components which have either $(+,+)$ or $(-,-)$ parity and nonzero U(1)$_Z$ charges as the candidate for the Higgs field. These include $\{ (1,2)(3,-3,3) + {\rm h.c.} \}$ and $\{(3,2)(1,-1,-3) + {\rm h.c.} \}$ with parities $(+,+)$ and $(-,-)$, respectively. The representations $(1,2)(-3,3,-3)$ and $(1,2)(3,-3,3)$ have the correct quantum numbers for the SM Higgs doublet. Therefore, we identify the $(1,1)$ mode of these components as the SM Higgs fields in four dimensions.
Chiral fermions in four dimensions \[sec:fermions\]
---------------------------------------------------
We introduce fermions as the Weyl spinor fields of the six-dimensional Lorentz group SO(1,5). They can be written in terms of the SO(1,3) Weyl spinors as \[chiralR\] \_+ =
\_R\
\_L
,\
\[chiralL\] \_- =
\_L\
\_R
. In general, fermions on the two-sphere do not have massless KK modes because of the positive curvature of the two-sphere. The massless modes can be obtained by incorporating the background gauge field (\[background\]) though, for it can cancel the contribution from the positive curvature. In this case, the condition for obtaining a massless fermion mode is \[massless-condition\] Q = , where $Q$ comes from the background gauge field and is proportional to the U(1)$_Z$ generator [@RandjbarDaemi:1982hi; @Maru:2009wu; @Dohi:2010vc]. We observe that the upper \[lower\] component on the RHS of Eq. (\[chiralR\]) \[(\[chiralL\])\] has a massless mode for the $+$ $(-)$ sign on the RHS of Eq. (\[massless-condition\]).
In our model, we choose the fermions as the Weyl fermions $\Psi_-$ belonging to the [**27**]{} representation of E$_6$. The [**27**]{} representation is decomposed as in Eq. (\[d27\]) under the group reduction, Eq. (\[group-red\]). In this decomposition, we find that our choice of the background gauge field of U(1)$_Z$ is suitable for obtaining massless fermions since all such components have U(1)$_Z$ charge 1. In the fundemantal representation, the U(1)$_Z$ generator is Q\_Z = (-2,-2,-2,-2,1,1,1,1,4,1,1,1,1,1,1,-2,-2,-2,1,1,1,1,1,1,-2,-2,-2) ,\
according to the decomposition Eq. (\[d27\]). By identifying $Q=3Q_Z$, we readily obtain the condition $$Q \Psi_- = \frac{1}{2} \Psi_-.$$ Therefore, the chiral fermions $\psi_L$ in four dimensions have zero modes.
Next, we consider the parity assignments for the fermions with respect to the fixed points of $S^2/Z_2$. The boundary conditions are given by Eqs. (\[boundary-condition3\]) and (\[boundary-condition6\]). It turns out that four ${\bf 27}$ fermion copies with different boundary conditions are needed in order to obtain an entire generation of massless SM fermions. They are denoted by $\Psi^{(1,2,3,4)}$ with the following parity assignments \_\^[(i)]{} (x,-,-) &=& \_5 P\_1 \_\^[(i)]{}(x,,) ,\
\_\^[(i)]{} (x,-,2-) &=& \_5 P\_2 \_\^[(i)]{}(x,,) , where $\gamma_5$ is the chirality operator, and $(\xi,\eta) = (+,+)$, $(-,-)$, $(+,-)$ and $(-,+)$ for $i = 1,2,3,4$, respectively. From these fermions we find that $\psi_{1,2,3,4}$ have the parity assignments \_[\_L\^[(1)]{}]{} &=& (1,2)(-3,-2,-2)\^[(-,-)]{}+(1,2)(3,2,-2)\^[(+,+)]{}+(1,2)(-3,3,1)\^[(-,+)]{}\
&& + (1,1)(6,-1,1)\^[(-,-)]{}+ +(1,1)(0,0,4)\^[(+,-)]{}\
&& + (3,2)(1,-1,1)\^[(+,-)]{}+(3,1)(-2,2,-2)\^[(-,+)]{} + (|[3]{},1)(-4,-1,1)\^[(-,-)]{}\
&& +(|[3]{},1)(2,3,1)\^[(-,-)]{}+(|[3]{},1)(2,-2,2)\^[(+,-)]{}\
[**27**]{}\_[\_L\^[(2)]{}]{} &=& (1,2)(-3,-2,-2)\^[(+,+)]{}+(1,2)(3,2,-2)\^[(-,-)]{}+(1,2)(-3,3,1)\^[(+,-)]{}\
&& + + (1,1)(0,-5,1)\^[(-,-)]{}+(1,1)(0,0,4)\^[(-,+)]{}\
&& + (3,2)(1,-1,1)\^[(-,+)]{}+(3,1)(-2,2,-2)\^[(+,-)]{} +\
&& + +(|[3]{},1)(2,-2,2)\^[(-,+)]{}\
[**27**]{}\_[\_L\^[(3)]{}]{} &=& (1,2)(-3,-2,-2)\^[(-,+)]{}+(1,2)(3,2,-2)\^[(+,-)]{}+(1,2)(-3,3,1)\^[(-,-)]{}\
&& + (1,1)(6,-1,1)\^[(-,+)]{}+(1,1)(0,-5,1)\^[(+,-)]{}+(1,1)(0,0,4)\^[(+,+)]{}\
&& + +(3,1)(-2,2,-2)\^[(-,-)]{} + (|[3]{},1)(-4,-1,1)\^[(-,+)]{}\
&& +(|[3]{},1)(2,3,1)\^[(-,+)]{}+(|[3]{},1)(2,-2,2)\^[(+,+)]{}\
[**27**]{}\_[\_L\^[(4)]{}]{} &=& (1,2)(-3,-2,-2)\^[(+,-)]{}+(1,2)(3,2,-2)\^[(-,+)]{}+\
&& + (1,1)(6,-1,1)\^[(+,-)]{}+(1,1)(0,-5,1)\^[(-,+)]{}+(1,1)(0,0,4)\^[(-,-)]{}\
&& + (3,2)(1,-1,1)\^[(-,-)]{}+(3,1)(-2,2,-2)\^[(+,+)]{} + (|[3]{},1)(-4,-1,1)\^[(+,-)]{}\
&& +(|[3]{},1)(2,3,1)\^[(+,-)]{}+(|[3]{},1)(2,-2,2)\^[(-,-)]{} , where the underlined components have even parities and U(1)$_Z$ charge 1. One can readily identify one generation of SM fermions, including a right-handed neutrino, as the zero modes of these components.
A long-standing problem in the gauge-Higgs unification framework is the Yukawa couplings of the Higgs boson to the matter fields. This is because the couplings here all arise from gauge interactions. It is therefore extremely difficult to derive the observed rich fermion mass spectrum purely from the gauge coupling. In order to have flavor-dependent Yukawa couplings, one promising solution is to consider SM matter fields localized at orbifold fixed points and make use of nonlocal interactions with Wilson lines [@Csaki:2002ur].
Higgs potential \[sec:higgs\]
=============================
Higgs sector \[sec:Higgssector\]
--------------------------------
The Lagrangian for the Higgs sector is derived from the gauge sector that contains extra-dimensional components of the gauge field $\{A_{\theta}, \tilde{A}_{\phi} \}$, as given in Eq. (\[action-scalar\]), by considering the lowest KK modes of them. The kinetic term and potential term are, respectively, L\_[K]{} &=& d . ( Tr\[(\_ A\_-i\[A\_,A\_\])\^2\] + Tr\[(\_ A\_-i\[A\_,\_\])\^2 \] ) |\_ ,\
V &=& d . Tr |\_ .\
Consider the $(1,1)$ mode of the $ \{ (1,2)(3,-3,3) + {\rm h.c.} \}$ representation in Eq. (\[78scalar\]) as argued in the previous section. The gauge fields are given by the following KK expansions \[expansion1\] A\_ &=& - \[ \_1(x) \_ Y\_[11]{}\^-( , ) + \_2(x) \_ Y\_[11]{}\^-( , ) \] + ,\
\[expansion2\] \_ &=& \[ \_2(x) \_ Y\_[11]{}\^-( ,)-\_1(x) \_ Y\_[11]{}\^-( ,)\] + , where $\cdots$ represents higher KK mode terms [@Maru:2009wu]. The function $Y_{11}^- = -1/\sqrt{2} [Y_{11}+Y_{1-1}]$ is odd under $(\theta,\phi) \rightarrow (\pi/2-\theta,-\phi)$ . We will discuss their higher KK modes and masses in the existence of the background gauge field in Section \[KKmass\]. With Eqs. (\[expansion1\]) and (\[expansion2\]), the kinetic term becomes L\_[K]{}(x) = ( Tr\[D\_ \_1(x) D\^ \_1(x)\] + Tr\[D\_ \_2(x) D\^ \_2(x)\] ) , where $D_{\mu} \Phi_{1,2} = \partial_{\mu} \Phi_{1,2} -i[A_{\mu},\Phi_{1,2}]$ is the covariant derivative acting on $\Phi_{1,2}$. The potential term, on the other hand, is V &=& d Tr , where $\partial_{\theta} (\sin \theta \tilde{A}_{\phi}^B) = Q \cos \theta$ from Eq. (\[background\]) is used. Expanding the square in the trace, we get \[potential\] V &=& d Tr , where terms that vanish after the $d\Omega$ integration are directly omitted. In the end, the potential is simplified to V = Tr , where use of $\tilde{A}_{\phi}^B = -Q (\cos \theta \mp 1) / \sin \theta$ has been made and $\mu_1 = 1-\frac{3}{2} \ln 2$ and $\mu_2 = \frac{3}{4}(1-2\ln2)$.
We now take the following linear combination of $\Phi_1$ and $\Phi_2$ to form a complex Higgs doublet, $$\begin{aligned}
\label{okikae1}
\Phi(x) &=& \frac{1}{\sqrt{2}} (\Phi_1(x)+i\Phi_2(x)) ~, \\
\label{okikae2}
\Phi(x)^{\dagger} &=& \frac{1}{\sqrt{2}} (\Phi_1(x)-i\Phi_2(x)) ~.\end{aligned}$$ It is straightforward to see that $$\begin{aligned}
[\Phi_1(x),\Phi_2(x)] = i [\Phi(x), \Phi^{\dagger}(x)] ~.\end{aligned}$$ The kinetic term and the Higgs potential now become \[kinetic-t\] L\_[K]{} &=& Tr\[D\_ \^(x) D\^ (x) \] ,\
\[potential-t\] V &=& Tr . To further simplify the above expressions, we need to find out the algebra of the gauge group generators. Note that the E$_6$ generators are chosen according to the decomposition of the adjoint representation given in Eq. (\[d78\]) &&{ Q\_i, Q\_, Q\_Y, Q\_X, Q\_Z,\
&& Q\_[ax (-5,0,0)]{}, Q\^[ax(5,0,0)]{}, Q\_[ax(1,4,0)]{}, Q\^[ax(-1,-4,0)]{},\
&& Q\_[a(4,-4,0)]{}, Q\^[a(-4,4,0)]{}, Q\_[(-6,-4,0)]{}, Q\_[(6,4,0)]{},\
&& Q\_[ax(1,-1,-3)]{}, Q\^[ax(-1,1,3)]{}, Q\_[a(4,1,3)]{}, Q\^[a(-4,-1,-3)]{},\
&& Q\_[a(-2,-3,3)]{}, Q\^[a(2,3,-3)]{}, Q\_[x(3,-3,3)]{}, Q\^[x(-3,3,-3)]{},\
&& Q\_[(-6,1,3)]{}, Q\_[(6,-1,-3)]{}, Q\_[(0,-5,-3)]{},Q\_[(0,5,3)]{} } , where the generators are listed in the corresponding order of the terms in Eq. (\[d78\]) and the indices \[generators\] && i=1,...,8: Q\_i: ,\
&& =1,2,3: Q\_: ,\
&& Q\_[X,Y,Z]{}: ,\
&& x=1,2: ,\
&& a=1,2,3: . Here we take the standard normalization for generators, $Tr[Q Q^{\dagger}] = 2$. The Higgs fields are in the representations of $(1,2)(3,-3,3)$ and $(1,2)(-3,3,-3)$. We write $$\label{Higgs}
\Phi(x) = \phi^{x} Q_{x(3,-3,3)} \quad (\Phi^{\dagger}(x) = \phi_x Q^{x(-3,3,-3)}) ~.$$ Likewise, the gauge field $A_{\mu}(x)$ in terms of the $Q$’s in Eq. (\[generators\]) is \[gauge\] A\_(x) = A\_\^i Q\_i+A\_\^ Q\_+B\_ Q\_Y+C\_ Q\_X+E\_ Q\_Z. The commutation relations between the generators $Q_{\alpha}$, $Q_{X,Y,Z}$, $Q_{x(3,-3,3)}$ and $Q^{x(-3,3,-3)}$ are summarized in Table. \[commutators\].
-- ----------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------
$\left[ Q_{\alpha},Q_{x(3,-3,3)} \right] = \frac{1}{\sqrt{6}} (\sigma_{\alpha} )^y_x Q_{y(3,-3,3)}$ $\left[ Q_{\alpha},Q^{ x(-3,3,-3) } \right] = - \frac{1}{\sqrt{6}} (\sigma_{\alpha}^* )^y_x Q^{y(-3,3,-3) }$
$\left[ Q_{x(3,-3,3)},Q_{y(3,-3,3)} \right] = 0$ $\left[ Q_Z, Q_{x(3,-3,3)} \right] = \frac{1}{2} Q_{x(3,-3,3)}$
$\left[ Q_X, Q_{x(3,-3,3)} \right] = -\frac{1}{2} \sqrt{\frac{3}{5}} Q_{x(3,-3,3)}$ $\left[ Q_Y, Q_{x(3,-3,3)} \right] = \frac{1}{\sqrt{10}} Q_{x(3,-3,3)}$
-- ----------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------
: Commutation relations of $Q_{\alpha}$, $Q_{X,Y,Z}$, $Q_{x(3,-3,3)}$ and $Q^{x(-3,3,-3)}$, where $\sigma_i$ are the Pauli matrices.[]{data-label="commutators"}
Finally, we obtain the Lagrangian associated with the Higgs field by applying Eqs. (\[Higgs\], \[gauge\]) to Eqs. (\[kinetic-t\], \[potential-t\]) and carrying out the trace. Furthermore, to obtain the canonical form of kinetic terms, the Higgs field, the gauge field, and the gauge coupling need to be rescaled in the following way: \[notation\] &&\
&& A\_ A\_\
&& = g\_2 , where $g_2$ denotes the SU(2) gauge coupling. The Higgs sector is then given by \_[Higgs]{} = |D\_ |\^2 - V() where \[cova\] D\_ &=& ,\
\[H-potential\] V &=& - \^ + (\^ )\^2 , where $\chi=7+9\mu_1+9\mu_2$. We have omitted the constant term in the Higgs potential. Comparing the potential derived above with the standard form $\mu^2\phi^\dagger\phi + \lambda (\phi^\dagger\phi)^2$ in the SM, we see that the model has a tree-level $\mu^2$ term that is negative and proportional to $R^{-2}$. Moreover, the quartic coupling $\lambda = 3 g^2 / (40 \pi R^2)$ is related to the six-dimensional gauge coupling $g$ and grants perturbative calculations because it is about $0.16$, using the value of $R$ to be extracted in the next section. Therefore, the order parameter in this model is controlled by a single parameter $R$, the compactification scale.
In fact, the $(1,1)$ mode of the $\{(3,2)(1,-1,-3) + {\rm h.c.}\}$ representation also has a negative squared mass term because it has the same $Q_z$ charge as the $\{(1,2)(3,-3,3) + {\rm h.c.}\}$ representation. Therefore, it would induce not only electroweak symmetry breaking but also color symmetry breaking. This undesirable feature can be cured by adding brane terms ( F\_\^a F\^[a]{} )\^2 ( -2 ) , where $a$ denotes the group index of the $\{(3,2)(1,-1,-3) + {\rm h.c.}\}$ representation. These brane terms preserve the $Z'_2$ symmetry which corresponds to the symmetry under the transformation $(\phi \rightarrow \phi+\pi)$. With an appropriate choice of the dimensionless constant $\alpha$, the squared mass of the $(1,1)$ can be lifted to become positive and sufficiently large.
Spontaneous symmetry breaking and Higgs mass
--------------------------------------------
Due to a negative mass term, the Higgs potential in Eq. (\[H-potential\]) can induce the spontaneous symmetry breakdown: SU(2) $\times$ U(1)$_Y$ $\rightarrow$ U(1)$_{\rm EM}$ in the SM. The Higgs field acquires a vaccum expectation value (VEV) =
0\
v
v = . One immedialtey finds that the $W$ boson mass $$m_W = \frac{g_2}{2} v = \frac{1}{6}\sqrt{ \frac{5 \chi}{2} } \frac{1}{R}
\simeq \frac{0.53}{R},$$ from which the compactification scale $R^{-1} \simeq 152$ GeV is inferred. Moreover, the Higgs boson mass at the tree level is $$m_H = \sqrt{\frac{3}{20 \pi}} \frac{g v}{R} = 3 \sqrt{\frac{2}{5}} m_W
= \frac{\sqrt{\chi}}{2} \frac{1}{R} ~,$$ which is about $152$ GeV, numerically very close to the compactification scale. Since the hypercharge of the Higgs field is $1/2$, the U(1)$_Y$ gauge coupling is derived from Eq. (\[cova\]) as $$g_Y = \frac{g}{\sqrt{10 \pi R^2}} ~.$$ The Weinberg angle is thus given by $$\begin{aligned}
\sin^2 \theta_W = \frac{g_Y^2}{g_2^2+g_Y^2}
= \frac{3}{8} ~,\end{aligned}$$ and the $Z$ boson mass $$\begin{aligned}
m_Z = \frac{m_W}{\cos \theta_W} = m_W \sqrt{\frac{8}{5}} ~,\end{aligned}$$ both at the tree level. These relations are the same as the SU(5) GUT at the unification scale. This is not surprising because this part only depends on the group structure.
KK mode spectrum of each field \[KKmass\]
=========================================
In this section, we compute the KK mass spectra of both fermion and gauge fields in the existence of the background gauge field. The masses are basically conrtrolled by the compactification radius $R$ of the two-sphere. They receive two kinds of contributions: one arising from the angular momentum in the $S^2$ space, and the other coming from the interactions with the background field.
KK masses of fermions
---------------------
The KK masses for fermions have been given in Refs. [@RandjbarDaemi:1982hi; @Maru:2009wu; @Dohi:2010vc]. We give them in terms of our notation here: \[KKmass-fermion\] M\_[m]{}\^[KK]{}(\_L) = , where $q$ is proportional to the U(1)$_Z$ charge of fermion and determined by the action of $Q=3Q_Z$ on fermions as $Q \Psi = q \Psi = 3q_Z \Psi$. Note that the mass does not depend on the quantum number $m$. The lightest KK mass, corresponding to $\ell = 1$ and $q_Z = 1/6$, is about 214 GeV at the tree level. The range of $\ell$ is (+: \_[R(L)]{} \_[+(-)]{}, - : \_[L(R)]{} \_[-(+)]{} ) . We thus can have zero mode for $Q \Psi = \pm \frac{1}{2} \Psi$, where this condition is given in Eq. (\[massless-condition\]).
KK masses of $A_{\mu}$
----------------------
For the four-dimensional gauge field $A_{\mu}$, its kinetic term and KK mass term are obtained from the terms \[FF\] L=d Tr . Taking terms quadratic in $A_{\mu}$, we get L\_[quad]{} &=& d Tr , where $\tilde{A}^B_{\phi}$ is the background gauge field given in Eq. (\[background\]). The KK expansion of $A_{\mu}$ is A\_ = \_[m]{} A\_\^[m]{}(x) Y\_[m]{}\^(,) where $Y_{\ell m}^{\pm}(\theta,\phi)$ are the linear combinations of spherical harmonics satisfying the boundary condition $Y_{\ell m}^{\pm}(\pi-\theta,-\phi) = \pm Y_{\ell m}^{\pm}(\theta,\phi)$. Their explicit forms are [@Maru:2009wu] \[modef1\] Y\_[m]{}\^+(, ) && \[Y\_[m]{}(, ) + (-1)\^ Y\_[-m]{}(, )\] m =0\
\[modef2\] Y\_[m]{}\^-(, ) && \[Y\_[m]{}(, ) - (-1)\^ Y\_[-m]{}(, )\] m =0\
\[modef3\] Y\_[0]{}\^[+(-)]{}() && {
[l]{} Y\_[0]{}() m=0 =\
0 m=0 =.
. Note that we do not have KK mode functions that are odd under $\phi \rightarrow \phi + 2 \pi$ since the KK modes are specified by the integer angular momentum quantum numbers $\ell$ and $m$ of gauge field $A_M$ on the two-sphere. Thus, the components of $A_{\mu}$ and $A_{\theta,\phi}$ with $(+,-)$ or $(-,+)$ parities do not have corresponding KK modes. Applying the KK expansion and integrating about $d \Omega$, we obtain the kinetic and KK mass terms for the KK modes of $A_{\mu}$ \[masstermAm\] L\_M &=& - + A\_\^[m]{}(x) A\^[m ]{}(x)\
&& + A\_\^[m]{}(x) A\^[m ]{}(x) , where we have used $Tr[Q_i Q^i]=2$ and $[A_{\mu}(x),Q_Z] = q_Z (A_{\mu}^i(x) Q_i - A_{i \mu}(x) Q^i )$. Therefore, the KK masses of $A_\mu$ are \[KKmass-gauge\] M\_[m]{}\^[KK]{}(A\_) &=& ,\
(m\^B\_[m]{})\^2 &=& 9 q\_Z\^2 d (Y\_[m]{}\^)\^2 , where $m^B_{\ell m}$ corresponds to the contribution from the background gauge field. Note that Eq. (\[KKmass-gauge\]) agrees with Eq. (\[eq:nonSMgaugeMass\]) when $\ell = 0$. Also, since the SM gauge bosons have $q_Z = 0$, their KK masses are simply $\sqrt{\ell(\ell+1)}/R$ at the tree level.
KK masses of $A_{\theta,\phi}$
------------------------------
The kinetic and KK mass terms of $A_{\theta}$ and $A_{\phi}$ are obtained from the terms in the higher dimensional gauge sector \[scalar\] L &=& d { ( Tr\[(\_ A\_-i\[A\_,A\_\])\^2\] + Tr\[(\_ A\_-i\[A\_,\_\])\^2 \] )\
&& - Tr } .\
The first line on the right-hand side of Eq. (\[scalar\]) corresponds to the kinetic terms, and the second line corresponds to the potential term. Applying the background gauge field Eq. (\[background\]), the potential becomes L\_V = - d Tr For $A_{\theta}$ and $A_{\phi}$ we use the following KK expansions to obtain the KK mass terms, \[expansion3\] A\_(x,,) &=& \_[m (0)]{} ,\
\[expansion4\] A\_(x,,) &=& \_[m (0)]{} , where the factor of $1/\sqrt{\ell(\ell+1)}$ is needed for normalization. These particular forms are convenient in giving diagonalized KK mass terms [@Maru:2009wu]. Applying the KK expansions Eq. (\[expansion3\]) and Eq. (\[expansion4\]), we obtain the kinetic term L\_[K]{} = \_[m (0)]{} Tr where only terms quadratic in $\partial_{\mu} \Phi$ are retained. The potential term L\_V &=& - \_[m (0)]{} d Tr ,\
where only terms diagonal in $(\ell,m)$ are consider. Using the relation $\frac{1}{\sin \theta} \partial_{\theta} (\sin \theta \partial_{\theta} Y_{\ell m}) + \frac{1}{\sin^2 \theta} \partial_{\phi}^2 Y_{\ell m} = -\ell(\ell+1)Y_{\ell m}$, the potential term is simplified as L\_V &=& - \_[m(0)]{} d Tr . To obtain the mass term, we focus on terms quadratic in $\Phi_{1,2}$: \[massterm\] L\_M &=& - d Tr .\
Note that we have dropped the term proportional to $[\Phi_1,\tilde{A}_{\phi}^B] [\Phi_2,\tilde{A}_{\phi}^B]$ because this term vanishes after turning the field into the linear combinations of $\Phi$ and $\Phi^\dagger$, Eqs. (\[okikae1\]) and (\[okikae1\]): Tr\[\[\_1,\_\^B\]\[\_1,\_\^B\]\] && Tr\[\[(+\^),Q\] \[(-\^),Q\] \]\
&& Tr\[(-\^)(+\^)\]\
&& Tr\[\^\] - Tr\[\^ \] =0 Integrating the second term of Eq. (\[massterm\]) by part, we obtain L\_M &=& - ( (+1) Tr\[(\_2\^[m]{})\^2\] +2i Tr\[Q \[\_1\^[m]{},\_2\^[m]{}\] \]\
&& -2 i Tr\[Q \[\_1\^[m]{},\_2\^[m]{}\] \] d Y\_[m]{}\^ \_ Y\_[m]{}\^\
&& - \[\_1\^[m]{},Q\]\^2 d (\_ Y\_[m]{}\^)\^2\
&& - \[\_2\^[m]{},Q\]\^2 d ) .\
Therefore, the KK masses depend on the U(1)$_Z$ charges of the scalar fields.
For components with zero U(1)$_Z$ charge, we write $\Phi_{1(2)}(x)$ as $\phi_{1(2)}(x) Q$ where $Q$ is the corresponding generator of E$_6$ in Eq. (\[d78\]) with zero U(1)$_Z$ charge. Taking the trace, we have the following kinetic and KK mass terms instead: L = \_[m(0)]{} ( \_ \_1\^[m]{}(x) \^ \_1\^[m]{}(x) + \_ \_2\^[m]{}(x) \^ \_2\^[m]{}(x) + (+1) \_2\^[m]{}(x) \_2\^[m]{}(x) ) where we have made the substitution $\phi_i \rightarrow g \phi_i$. Note that $\phi_1$ is considered as a massless Nambu-Goldstone (NG) boson in this case.
For components with nonzero U(1)$_Z$ charge, we use Eq. (\[okikae1\]) and (\[okikae2\]) and write $\Phi(x)$ as $\phi^i(x)Q_i$ where $Q_i$ is the corresponding generator of $E_6$ in Eq. (\[d78\]) with nonzero U(1)$_Z$ charge. The commutator between $Q$ and $\Phi$ is = 3\[Q\_Z,Q\_i\] \^i = 3 q\_Z \^i , where we have used $Q=3Q_Z$ and that $q_Z$ is a constant determined by the U(1)$_Z$ charge of the corresponding component. Finally, the Lagrangian becomes L &=& \_[m(0)]{} { \_ \^\_[m]{} \^ \_[m]{}\
&& - }.\
where the subscript $i$ is omitted for simplicity. The KK masses of the complex scalar field $\phi$ are then $$\begin{aligned}
\label{KKmass-scalar}
M_{\ell m}^{KK}(\phi) &=& \frac{1}{R} \sqrt{\frac{\ell(\ell+1)}{2}+(m_{\ell m}^B)^2} ~, \nonumber \\
(m_{\ell m}^B)^2 &=& -3 q_Z +3 q_Z \int d \Omega \frac{\cos \theta \mp 1}{\sin \theta} Y_{\ell m}^{\pm} \partial_{\theta} Y_{\ell m}^{\pm}
+\frac{9 q_Z^2}{ 2\ell(\ell+1)} \int d \Omega \frac{(\cos \theta \mp1)^2}{\sin^2 \theta} (\partial_{\theta} Y_{\ell m}^{\pm})^2 \nonumber \\
&& \qquad + \frac{9 q_Z^2}{2\ell(\ell+1)} \int d \Omega \frac{(\cos \theta \mp 1)^2}{\sin^2 \theta} \frac{(\partial_{\phi} Y_{\ell m}^{\pm})^2}{\sin^2 \theta} ~.\end{aligned}$$ The squared KK mass $\left( M_{\ell m}^{KK} \right)^2$ is always positive except for the lowest mode ($\ell=1,m=1$). In fact, the squared KK mass of the $(1,1)$ mode agrees with the coefficient of quadratic term in the Higgs potential (\[H-potential\]).
Dark matter candidate
---------------------
In our model, each KK particle is associated with a KK parity derived from an additional $Z_2'$ discrete symmetry of $(\theta,\phi) \rightarrow (\theta,\phi + \pi)$, corresponding to the exchange of the two fixed points on the orbifold [@Maru:2009wu]. The KK-parity is given by $(-1)^m$, and is conserved as a consequence of the $Z_2'$ symmetry of the Lagrangian in six-dimensional spacetime. Therefore, the lightest KK particle with an odd $m$ will be stable.
A comparison among Eqs. (\[KKmass-fermion\]), (\[KKmass-gauge\]) and (\[KKmass-scalar\]) indicates that the lightest KK particles are the $(\ell=1,m=1)$ modes of the scalar components with non-zero U(1)$_Z$ charges since their masses receive a negative contribution from the background gauge field. They include the components $\{(3,2)(1,-1,-3)^{(-,-)} + {\rm h.c.} \}$ and $\{(1,2)(-3,3,-3)^{(+,+)} + {\rm h.c.} \}$ in Eq. (\[78scalar\]) since the other components either have zero U(1)$_Z$ charge or are odd under $\phi \rightarrow \phi + 2 \pi$. At the tree level, both of them have the same and negative KK squared mass since their U(1)$_Z$ charges are same ($q_Z = 1/2$). As argued at the end of Section \[sec:Higgssector\], the squared mass of the former representation can be lifted by brane terms to be sufficiently large to avoid color symmetry breaking. Its mass depends on the parameter $\alpha$ in the brane terms. The latter representation actually gives the Higgs field that has a mass about 152 GeV. We assume that the mass of the $(1,1)$ mode of the $\{(3,2)(1,-1,-3)^{(-,-)} + {\rm h.c.} \}$ components is heavier than the Higgs boson mass since a colored particle is not suitable for dark matter candidate. Therefore, the model has the $(1,1)$ mode of the $\{(1,2)(-3,3,-3)^{(+,+)} + {\rm h.c.} \}$ representation as the lightest and stable KK particle, which is simply the Higgs boson.
Summary \[sec:summary\]
=======================
The gauge-Higgs unification is an attractive idea because it can unify the SM gauge bosons and Higgs boson under a higher dimensional spacetime symmetry. The gauge invariance prevents the Higgs boson in the bulk from receiving radiative corrections that diverge with the cutoff scale, thus easing the gauge hierarchy problem. However, one still encounters the difficulty in getting an appropriate Higgs potential to break the electroweak symmetry in five dimensional spacetime. Extra particles are generally needed in order to generate such a potential and a sufficiently large Higgs mass radiatively. When one goes to six spacetime dimensions and considers the $S^2/Z_2$ orbifold, it is possible to render a suitable Higgs potential by incorporating a background gauge field in the extra-dimensional components. To fully achieve that, nevertheless, one has to assume a special symmetry that relates the SU(2) isometry transformation in $S^2$ to the gauge transformation.
We consider in this paper a six-dimensional gauge-Higgs unification model, in which the gauge group is enlarged to E$_6$ and the extra space is the $S^2/Z_2$ orbifold. By specifying two sets of parity transformation properties for the fields and employing a Dirac monopole configuration for the background gauge field, we have a successful symmetry reduction to the SM gauge group plus two extra U(1)’s. In our model, the background gauge field $A_{\phi}^B$ plays important roles in several aspects. First, it renders massless chiral fermions by canceling the spin-connection term in the covariant derivative. Secondly, it elevates the masses of unwanted representations of $A_\mu$ to roughly the compactification scale in four dimensional spacetime. Finally, from the gauge kinetic term, it gives rise to a negative mass square term for the Higgs potential at tree level.
At the low energy, we obtain only the SM particles. The SM gauge bosons all originate from a single adjoint representation of the E$_6$ group. The chiral fermions, including a right-handed neutrino, are derived from four copies of fundamental representation, each of which have a distinct parities under the two parity transformations. We also obtain exactly one complex Higgs doublet from the extra dimensional components of the gauge field.
We have computed the Higgs potential in this model. The squared mass is related to the compactification radius, and the quartic coupling to the E$_6$ gauge coupling. The radius of the compactified two-sphere is extracted to be around (152 GeV)$^{-1}$. The Higgs boson mass is predicted to be about 150 GeV at tree level. Due to the gauge group structure, we obtain $\sin^2\theta_W = 3/8$, the same as in the SU(5) GUT at the unification scale.
Through KK expansions, we have calculated the mass spectra of the gauge and fermion fields. In general, these masses involve two contributions: one related to the angular momentum eigenvalues in the extra dimensions $\ell(\ell+1)$, and the other due to the interactions between the KK modes and the background gauge field. Finally, the model can have a dark matter candidate due to the KK parity under the $Z'_2$ symmetry. It is the Higgs boson of the model.
Acknowledgments {#acknowledgments .unnumbered}
===============
This research was supported in part by the National Science Council of Taiwan, R.O.C. under Grant No. NSC 97-2112-M-008-002-MY3 and the NCTS.
[99]{}
N. S. Manton, Nucl. Phys. B [**158**]{}, 141 (1979). P. Forgacs and N. S. Manton, Commun. Math. Phys. [**72**]{}, 15 (1980). D. B. Fairlie, Phys. Lett. B [**82**]{}, 97 (1979). G. R. Dvali, S. Randjbar-Daemi and R. Tabbash, Phys. Rev. D [**65**]{}, 064021 (2002) \[arXiv:hep-ph/0102307\]. N. Arkani-Hamed, A. G. Cohen and H. Georgi, Phys. Lett. B [**513**]{}, 232 (2001) \[arXiv:hep-ph/0105239\]. C. Csaki, C. Grojean and H. Murayama, Phys. Rev. D [**67**]{}, 085012 (2003) \[arXiv:hep-ph/0210133\].
C. A. Scrucca, M. Serone and L. Silvestrini, Nucl. Phys. B [**669**]{}, 128 (2003) \[arXiv:hep-ph/0304220\]. I. Antoniadis, K. Benakli and M. Quiros, New J. Phys [**3**]{}, 20(2001) \[arXiv:hep-th/0108005\].
L. J. Hall, Y. Nomura and D. Tucker-Smith, Nucl. Phys. B [**639**]{}, 307 (2002) \[arXiv:hep-ph/0107331\]. G. Burdman and Y. Nomura, Nucl. Phys. B [**656**]{}, 3 (2003) \[arXiv:hep-ph/0210257\]. N. Haba and Y. Shimizu, Phys. Rev. D [**67**]{}, 095001 (2003) \[Erratum-ibid. D [**69**]{}, 059902 (2004)\] \[arXiv:hep-ph/0212166\]. K. w. Choi, N. y. Haba, K. S. Jeong, K. i. Okumura, Y. Shimizu and M. Yamaguchi, JHEP [**0402**]{}, 037 (2004) \[arXiv:hep-ph/0312178\]. Y. Sakamura, Phys. Rev. D [**76**]{}, 065002(2007) . \[arXiv:0705.1334 \[hep-ph\] \]
A.D. Medina, N.R. Shah, and C.E.M. Wagner, Phys. Rev. D [**76**]{}, 095010 (2007). \[arXiv:0706.1281 \[hep-ph\] \]
C.S. Lim, and N. Maru. Phys. Lett. B [**653**]{}, 320 (2007) \[arXiv:0706.1397 \[hep-ph\] \].
Y. Hosotani, K. Oda, T. ohnuma and Y. Sakamura, Phys. Rev. D [**78**]{}, 096002 (2008) \[Erratum-ibid. D [**79**]{}, 079902 (2009)\] \[arXiv:0806.0480 \[hep-ph\]\].
I. Antoniadis, Phys. Lett. B [**246**]{}, 377 (1990)
I. Antoniadis and K. Benakli, Phys. Lett. B [**326**]{}, 69 (1994) \[arXiv:hep-th/9310151\].
Y. Hosotani, Phys. Lett. B [**126**]{}, 309 (1983); Annals Phys. [**190**]{}, 233 (1989).
C.A. Scrucca, M. Serone, L. Silvestrini, and A. Wulzer, JHEP [**0402**]{}, 049 (2004) \[arXiv:hep-th/0312267\].
R. Contino, Y. Nomura and A. Pomarol Nucl. Phys. B [**671**]{}, 148 (2003) \[arXiv:hep-ph/0306259\].
Y. Hosotani and M. Mabe, Phys. Lett. B [**615**]{}, 257 (2005)\[arXiv:hep-ph/0503020\]
N. Haba, Y. Hosotani, Y. Kawamura and T. Yamashita Phys. Rev. D [**70**]{}, 015010 (2004) \[arXiv:hep-ph/0401183\].
T. Nomura and J. Sato, Nucl. Phys. B [**811**]{}, 109 (2009) \[arXiv:0810.0898 \[hep-ph\]\].
D. Kapetanakis and G. Zoupanos, Phys. Rept. 219 (1992) 1.
G. Panico, E. Ponton, J. Santiago and M. Serone, Phys. Rev. D [**77**]{}, 115012 (2008) \[arXiv:0801.1645 \[hep-ph\]\].
M. Carena, A. D. Medina, N. R. Shah and C. E. M. Wagner, Phys. Rev. D [**79**]{}, 096010 (2009) \[arXiv:0901.0609 \[hep-ph\]\].
Y. Hosotani, P. Ko and M. Tanaka, Phys. Lett. B [**680**]{}, 179 (2009)\[arXiv:0908.0212 \[hep-ph\] \]
N. Haba, S. Matsumoto, N. Okada and T. Yamashita, JHEP [**1003**]{}, 064 (2010) \[arXiv:0910.3741 \[hep-ph\]\].
S. Randjbar-Daemi, A. Salam and J. A. Strathdee, Nucl. Phys. B [**214**]{}, 491 (1983) C. S. Lim, N. Maru and K. Hasegawa, J.Phys.Soc.Jap.77 (2008) 074101. \[arXiv:0605180\[hep-th\]\]
N. Maru, T. Nomura, J. Sato and M. Yamanaka Nucl. Phys. B [**830**]{}, 414 (2010) \[arXiv:0810.0898 \[hep-ph\]\]
H. Dohi and K. Oda, arXiv.1004.3722 (2010)
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Strong scaling relations between host galaxy properties (such as stellar mass, bulge mass, luminosity, effective radius etc) and their nuclear supermassive black hole’s mass point towards a close co-evolution. In this work, we first review previous efforts supporting the fundamental importance of the relation between supermassive black hole mass and stellar velocity dispersion ($M_{\rm BH}$-$\sigma_{\rm e}$). We then present further original work supporting this claim via analysis of residuals and principal component analysis applied to some among the latest compilations of local galaxy samples with dynamically measured supermassive black hole masses. We conclude with a review of the main physical scenarios in favour of the existence of a $M_{\rm BH}$-$\sigma_{\rm e}$ relation, with a focus on momentum-driven outflows.'
author:
-
bibliography:
- 'bib.bib'
title: 'The case for the fundamental $M_{\rm BH}$-$\sigma$ relation'
---
Introduction
============
Observational evidence suggests that most local galaxies host a central supermassive black hole (henceforth simply ‘black hole’, not to be confused with an ‘ordinary’ stellar mass black hole). Indeed, galaxies for which high-resolution data can be acquired show stellar kinematic patterns strongly suggesting the presence of a central massive dark object . The central black hole masses, inferred from dynamical measurements of the motions of stars and/or gas in the host galaxies, appear to scale with galaxy-wide properties (or perhaps *bulge*-wide properties), such as stellar mass [@1998AJ....115.2285M; @2004ApJ...604L..89H] and stellar velocity dispersion [@2000ApJ...539L..13G; @2002ApJ...578...90F; @2002ApJ...574..740T; @2009ApJ...698..198G; @2013ApJ...764..184M; @2015MNRAS.446.2330S]. The existence of such correlations is remarkable, as the black hole’s (sub-parsec scale) sphere of influence is orders of magnitude smaller than the scale of it’s host galaxy (kilo-parsec scale). These correlations suggest a close link (a “co-evolution”) between black holes and host galaxies .
The existence of massive black holes at the centre of galaxies also lends further support to the widely-accepted paradigm that quasars, and more generally Active Galactic Nuclei (AGN), are powered by matter accreting onto a central black hole. The release of gravitational energy from an infalling body of mass $m$ approaching the Schwarzschild radius $R_s=2GM/c^2$ of a compact object of mass $M$, is in fact one of the most efficient known processes to release enough energy to explain the large luminosities in AGN. As discussed by [@1997iagn.book.....P], the emission from release of gravitational energy increases with the compactness of the source $M/r$. Assuming that most of the gravitational energy $E$ powering the emission from an accreting black hole originates from within a few times $R_s$, say $r=5R_s$, one could set $E=GMm/5R_s$, implying $E=0.1mc^2$. The latter efficiency $\eta \sim 0.1$ of energy conversion in units of the rest-mass energy, is orders of magnitude higher than the efficiency in stellar fusion ($\eta \sim 0.008$). Theoretical models have also suggested that the energy/momentum release from the central black hole, routinely known as “AGN feedback”, could have profound consequences on the fate of its host galaxy, potentially driving out a galaxy’s gas reservoir, quenching star formation, and shaping the above-mentioned scaling relations .
The most prominent and studied scaling relations relate the black hole mass $M_{\rm BH}$ to the stellar velocity dispersion $\sigma_{\rm e}$ [@2000ApJ...539L...9F] and the (stellar) mass of the host bulge, $M_{\rm bulge}$ (and by extension the luminosity of the bulge $L_{\rm bulge}$, see [@2003ApJ...589L..21M]). Other types of correlations have been proposed in the literature, such as correlations with the bulge light concentration $c_{\rm bulge}$ [@2001ApJ...563L..11G] and even the mass of the surrounding dark matter halo $M_{\rm halo}$ [@2002ApJ...578...90F]. This review will focus on the $M_{\rm BH}$-$\sigma_{\rm e}$ relation, where $\sigma_{\rm e}$ is the stellar velocity dispersion inferred from spectral absorption lines (see [@2010gfe..book.....M], Chapter 2).
The $M_{\rm BH}$-$\sigma_{\rm e}$ relation has attracted the attention of the astronomical community since its discovery [@2000ApJ...539L...9F], as it is believed to be closely connected to the galaxy/halo gravitational potential well, and thus may be related to the above-mentioned AGN feedback process [@2004ApJ...600..580G], as further discussed in Section 3.2. The relation is of the form:
$$\label{eq:badmsigma}
\log\frac{M_{\rm BH}}{M_\odot} = \alpha + \beta \log\frac{\sigma_{\rm e}}{200kms^{-1}} \, .$$
[@2000ApJ...539L...9F] initially retrieved a normalization and slope of, respectively, $\alpha=8.14\pm1.80$ and $\beta=4.80\pm0.54$, whereas more recent work [e.g., @2007ApJ...663...53T] suggests $\alpha = 8.21 \pm 0.06$ and $\beta = 3.83 \pm 0.21$. There is some debate in the literature concerning the exact shape of the $M_{\rm BH}$-$\sigma_{\rm e}$ and its dependence on, for example, morphological type or even environment (see, e.g., [@2007ApJ...662..808L], [@2006MNRAS.365.1082W] and [@2008MNRAS.386.2242H] for more details). It has been noted (e.g. [@2012Natur.491..729V]) that several overmassive black holes exist on this relation, hosted by galaxies that have undergone fewer than usual mergers, in tension with semi-analytic models [@2015MNRAS.446.2330S]. However, these outliers could simply be the result of incorrect modelling of the galactic bulge/disc [@2016MNRAS.457..320S].
Several groups have noted that the $M_{\rm BH}$-$\sigma_{\rm e}$ relation only weakly evolves with redshift (if at all) [e.g. @2009arXiv0908.0328G; @2013ApJ...764...80S; @2015ApJ...805...96S]. Supporting work by other groups base their conclusions on direct estimates of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation on high redshift quasar samples [@Woo_2008], and studies based on comparing the cumulative accretion from AGN with the local black hole mass density, retrieved from assigning to all local galaxies a black hole mass via the $M_{\rm BH}$-$\sigma_{\rm e}$ relation (e.g. [@2009ApJ...694..867S], [@2012ApJ...761....5Z]).
On the assumption that all local galaxies host a central black hole, scaling relations could in principle allow us to assign black hole masses to all local galaxies without a direct dynamical mass measurements, thus generating large-scale black hole mass statistical distributions such as black hole mass functions or correlation functions (see [@2009NewAR..53...57S] and [@2015ApJ...798...54G] for more focused reviews on this topic). For example, a number of groups have used luminosity, as performed by [@2004MNRAS.354.1020S], [@1999MNRAS.307..637S] and [@2004MNRAS.351..169M], or even Sersic index, as performed by [@2007MNRAS.378..198G]), to generate black hole mass functions. This procedure of course relies on two assumptions: firstly, that the observer has correctly identified the surrogate observable of black hole mass, and secondly that the established scaling relation is reliable. For example, the $M_{\rm BH}$-$\sigma_{\rm e}$ and $M_{\rm BH}$-$L_{\rm bulge}$, probably the most commonly used relations, have led to different black hole mass function estimates ([@2007ApJ...670..249L], [@2007ApJ...663...53T]).
An important question is whether the same black hole-galaxy scaling relations hold for both active and inactive galaxies. Several groups suggest that this is indeed the case [e.g. @2015ApJ...813...82R; @caglar2019llama; @2019MNRAS.485.1278S]. It is important to stress that the samples of nearby (inactive) galaxies on which the black hole-host galaxy relations are based, still remain relatively small, only comprising around $\sim70$-$80$ objects. A key difficulty relies of course in acquiring sufficiently high-resolution data to allow for dynamical black hole mass measurements (see [@1999AdSpR..23..925F], [@2005SSRv..116..523F] and for reviews of observational challenges).
Indeed, there is a growing body of work ([@2007ApJ...660..267B], [@2015ApJS..218...10V], [@2016MNRAS.460.3119S], [@2017MNRAS.466.4029S], [@2019MNRAS.485.1278S], [@2019arXiv191010175S]) supporting the view that that current dynamical black hole mass samples may indeed be “biased-high”, possibly due to angular resolution selection effects (see [@2013degn.book.....M]), with meaningful consequences for any study based on the “raw” relations. Interestingly, [@2016MNRAS.460.3119S] showed that, via aimed Monte Carlo simulations, irrespective of the presence of an underlying resolution bias, the raw and “de-biased” scaling relations would still share similar slopes and overall statistical properties (e.g., very similar residuals around the mean), with (noticeable) differences arising only in the normalization between observed and de-biased scaling relations. In particular, the $M_{\rm BH}$-$\sigma_{\rm e}$ was shown to be more robust and the least affected by possible angular resolution effects.
The main aim of this work is threefold: i) to review the evidence in favour of the primary importance of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation above other black hole scaling relations, ii) to provide further support to velocity dispersion as the main host galaxy property driving the connection between black holes an their hosts, and iii) to review the main theoretical scenarios that give rise to the $M_{\rm BH}$-$\sigma_{\rm e}$ relation, with a focus on momentum-driven outflows. In sections \[section:residuals\] and \[section:pca\] we will describe original evidence based on residuals and principle component analysis (respectively) in support of the primary role played by $M_{\rm BH}$-$\sigma_{\rm e}$. In section \[section:theory\] we include a description of the theoretical scenarios behind the physical origin of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation. We then conclude in section \[section:conclusion\].
Where cosmological parameters are required, we set $h=0.7$, $\Omega_m=0.3$ and $\Omega_\Lambda=0.7$.
The case for Velocity dispersion {#section:VD}
================================
Review of previous work
-----------------------
Standard regression analyses showed that the $M_{\rm BH}$-$\sigma_{\rm e}$ has the lowest *intrinsic scatter* of any black hole scaling relation; e.g. [@2009ApJ...698..198G], [@2016ApJ...818...47S] and [@2016ApJ...831..134V]. This alone suggests $\sigma_{\rm e}$ is different from other variables. [@2012MNRAS.419.2497B] came to the conclusion that $M_{\rm BH}$ was fundamentally driven by $\sigma_{\rm e}$ due to its relative tightness. This work also tested the possibility for multi-dimensional relations, concluding that the introduction of additional variables barely reduced the scatter with respect to the $M_{\rm BH}$-$\sigma_{\rm e}$, suggesting that stellar velocity dispersion remains a fundamental driving parameter. The amount of scatter characterizing diverse black hole scaling relations has been studied by several groups ([@2003ApJ...589L..21M], [@2007ApJ...669...67H]). [@2003ApJ...589L..21M] and [@2007ApJ...669...67H] explored the addition of the effective radius $R_{\rm e}$ to $\sigma_{\rm e}$ to create a “fundamental plane” in the black hole scaling relations, further discussing in [@2007ApJ...669...45H] how this relation naturally arises in their simulations, as a (tilted) correlation between black hole mass and bulge binding energy. This conclusion was supported by [@2016ApJ...818...47S], who argued for a multidimensional relation deriving from the bulge kinetic energy ($M_{\rm bulge}\sigma_{\rm e}^2$), as originally suggested by [@2005IJMPD..14.1861F].
[@2019MNRAS.490..600D] presented a systematic study of black hole scaling relations on an improved sample of local black holes, confirming that “the correlation with the effective velocity dispersion is not significantly improved by higher dimensionality”. The authors concluded that the $M_{\rm BH}$-$\sigma_{\rm e}$ is fundamental over multidimensional alternatives, independent of bulge decompositions. This is in line with [@2016ApJ...831..134V], who claimed that the $M_{\rm BH}$-$M_{\rm bulge}$ is mostly a projection of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation.
On more general grounds it has been suggested that, in terms of galactic scaling relations, velocity dispersion may be statistically more significant and relevant than other galaxy observables (e.g., [@2011MNRAS.412..684B], [@2011MNRAS.412L...6B]). [@2005AJ....129...61B] analysed the color-magnitude-velocity dispersion relation of a early-type galaxy sample of the Sloan Digital Sky Survey (SDSS), concluding that color-magnitude relations are entirely a consequence of the combination of more fundamental correlations with velocity dispersion.
[@2007ApJ...660..267B] noted that the $M_{\rm BH}$-$\sigma_{\rm e}$ and $M_{\rm BH}$-$L_{\rm bulge}$ predict different abundances of black holes, with the former predicting a smaller number of more massive black holes. Interestingly, the combined $\sigma_{\rm e}$-$L$ relation (for the dynamically measured black hole sample, e.g. [@2002MNRAS.335..965Y]) is inconsistent with the same relation from the SDSS, with smaller $L_{\rm bulge}$ for given $\sigma_{\rm e}$ (regardless of the band used to estimate luminosity). This suggests that the dynamical sample of local black holes may be biased towards objects with higher velocity dispersion when compared to local galaxies of similar luminosity, which obviously calls into question the accuracy of the raw $M_{\rm BH}$-$\sigma_{\rm e}$ and $M_{\rm BH}$-$L_{\rm bulge}$ relations. While unable to identify the source of the bias, modelling of this effect by [@2007ApJ...660..267B] and [@2016MNRAS.460.3119S] suggested that the bias in the $M_{\rm BH}$-$\sigma_{\rm e}$ is likely to be small, whereas the $M_{\rm BH}$-$L_{\rm bulge}$ is likely to predict over-massive black holes at a fixed galaxy (total) luminosity/stellar mass.
Residuals analysis {#section:residuals}
------------------
We start by revisiting the residual analysis on the black hole scaling relations following the method outlined in [@2016MNRAS.460.3119S], [@2017MNRAS.466.4029S] and [@2019MNRAS.485.1278S]. Residuals in pairwise correlations [@2012MNRAS.422.1825S] allow for a statistically sound approach to probe the relative importance among variables in the scaling with black hole mass. Residuals are computed as $$\Delta(Y|X)\equiv\log Y-\langle \log Y|\log X \rangle \,
\label{eq|resid}$$ where the residual is computed in the $Y$ variable (at fixed $X$) from the log-log-linear fit of $Y(X)$ vs $X$, i.e. $\langle \log Y|\log X \rangle$. For each pair of variables, each residual is computed 200 times, and at each iteration five objects at random are removed from the original sample. From the full ensemble of realizations, we then measure the mean slope and its 1$\sigma$ uncertainty.
Our results are shown in Figures \[fig|FigureResiduals\] and \[fig|FigureResidualsETGS\], which show the residuals extracted from the recent homogeneous sample calibrated by [@2019MNRAS.490..600D]. Figure \[fig|FigureResiduals\] shows that black hole mass strongly correlates with velocity dispersion at fixed galaxy luminosity with a Pearson coefficient $r\sim0.7$ (top left panel), and even more so at fixed effective radius with $r\sim 0.8$ (bottom left panel), while the corresponding correlations with stellar luminosity or effective radius are significantly less strong with $r\sim0.4$ at fixed velocity dispersion (right panels). Figure \[fig|FigureResidualsETGS\] shows the residuals restricting the analysis to only early type galaxies (red circles). The residuals appear quite similar in both slopes and related Pearson coefficients. These results further support the findings by [@2016MNRAS.460.3119S] (shown, for comparison, in figure \[fig:FS16\]) that velocity dispersion is more fundamental than effective radius and stellar mass, and that even disc-dominated galaxies follow similar scaling relations.
The total slope of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation can be estimated as $M_{\rm BH}\propto\sigma^{\beta}M_*^{\alpha}\propto\sigma^{\beta+\alpha\,\gamma}$, where $\gamma$ comes from $M_*\propto\sigma^{\gamma}$. Since SDSS galaxies tend towards $\gamma\approx 2.2$ [@2017MNRAS.466.4029S], and the residuals in Figure \[fig|FigureResiduals\] yield $\beta\sim 3$ and $\alpha\sim 0.4$, one obtains a total dependence of $M_{\rm BH}\propto\sigma_{\rm e}^{5}$, consistent with models of black hole growth being regulated by AGN feedback, as further discussed in Section 3.2 (e.g. , [@1999MNRAS.308L..39F], [@2003ApJ...596L..27K], [@2004ApJ...600..580G]).
PCA analysis {#section:pca}
------------
We will now present additional *original* work in favour of the $M_{\rm BH}$-$\sigma_{\rm e}$ being more fundamental, via Principal Component Analysis (PCA; [@Jolliffe1986]), which is a powerful complementary statistical technique to the residuals analysis presented above. PCA is a mathematical procedure that diagonalises the covariance matrix of variables in a dataset, providing a set of uncorrelated linearly transformed parameters, called principal components, defined by a set of orthogonal eigenvectors. The new orientation ensures that the first principal component (PC1) contains as much as possible of the variance in the data, and that the maximum of the remaining variance is contained in each succeeding orthogonal principal component (PC2, PC3, etc.). In other words, PCA finds the optimal projection of a number of (possibly correlated) physical observables into a smaller number of uncorrelated variables, revealing which quantities are more responsible for the variance (or, in some sense, for the information) in the dataset. PCA has been widely adopted in extragalactic astronomy, for instance to search for possible *dimensionality reduction* of the parameter space necessary to describe a sample (e.g., [@LaraLopez2010; @Hunt2012]) or to study the mutual dependencies between observed gas- and metallicity-based galaxy scaling relations (e.g., [@Bothwell2016; @Hunt2016; @Ginolfi2019]). Here we use PCA as an alternative technique to explore the black hole scaling relations. In detail, by quantifying through PCA the robustness of the correlations between $M_{\rm BH}$ and, in turn, $\sigma_{\rm e}$, $L$ (total luminosity) and $R_{\rm e}$ (the bulge effective radius), we can infer which of these observables provides a more fundamental scaling relation.
### Black Hole scaling relations {#sec:2DPCA}
In the PCA analysis we continue to use the dataset from [@2019MNRAS.490..600D]. To ensure that quantities with a higher dispersion are not over-weighted, we normalize our variables to their mean values, dividing by the standard deviation of their distributions. We therefore define the new variables (for convenience, in what follows we simply define $L=L_K$): $$\log(M_{\rm BH})^{\rm PCA} = [\log(M_{\rm BH}) - 8.43]/0.99$$ $$\log(\sigma_{\rm e})^{\rm PCA} = [\log(\sigma_{\rm e}) - 2.30]/0.18$$ $$\log(L)^{\rm PCA} = [\log(L) - 10.92] /0.75$$ $$\log(R_{\rm e})^{\rm PCA} = [\log(R_{\rm e}) - 0.34]/0.69$$ We perform three different PCA on the 2D-space datasets formed by $M_{\rm BH}$ and, in turn, one among $\sigma_{\rm e}$, $L$ and $R_{\rm e}$. The resulting principal component coefficients are reported in Table 1. We account for uncertainties in our results following a commonly adopted method (see e.g., [@Bothwell2016; @Ginolfi2019]). We perform a Monte Carlo bootstrap running $10^5$ iterations, in each of which we perturb all the analysed quantities by an amount randomly extracted in a range of values defined by their respective measurement errors. Thus, the reported principal component’s coefficients and their errors are computed, respectively, from the average and the standard deviation of the values obtained over all the iterations.
In the upper panels of Figure \[fig:fig\_PCA\_1\] we show the determined mutually orthogonal eigenvectors drawn onto the planes defined by the 2D-space datasets consisting of log($M_{\rm BH}$)$^{\rm PCA}$ and, in turn, log($\sigma_{\rm e}$)$^{\rm PCA}$, log($L$)$^{\rm PCA}$ and log($R_{\rm e}$)$^{\rm PCA}$. The three datasets, projected into the principal components, are shown in the lower panels of Figure \[fig:fig\_PCA\_1\]. We find that, although in all three cases PC2 contains only a small fraction of the total variance (see Table 1), confirming that an overall good physical correlation exists among the variables, in the $M_{\rm BH}$-$\sigma_{\rm e}$ relation PC2 is minimised and the dataset can be very well described uniquely by the PC1. In detail, we find that in the $M_{\rm BH}$-$\sigma_{\rm e}$ relation 95.4 $\pm$ 0.4 % of the variance is contained into PC1 (with the little remaining information contained in PC2), while lower amount of variance are contained in the PC1 of the $M_{\rm BH}$ - $L$ relation (89.7 $\pm$ 0.5 %) and in the PC1 of the $M_{\rm BH}$ - $R_{\rm e}$ relation (88.2 $\pm$ 0.3 %).
Since in all the three cases PC2 contains a little variance, we can set it to zero to obtain a linear approximation of the correlation among our observables from the PCA projected datasets. Thus, we obtain the following PCA model predictions: $$\log(M_{\rm BH})^{\rm model}_{\rm \sigma} = 5.4 (\pm0.1) \log(\sigma_{\rm e}) - 4.01 (\pm0.09)$$ $$\log(M_{\rm BH})^{\rm model}_{\rm L} = 1.32 (\pm0.04) \log(L) - 6.0 (\pm0.1)$$ $$\log(M_{\rm BH})^{\rm model}_{\rm R_{\rm e}} = 1.42 (\pm0.07) \log(R_{\rm e}) + 7.9 (\pm0.1),$$
where the non-normalized variables are restored (as defined in the equations discussed above), and the errors on the parameters are computed propagating the uncertainties on the principal component coefficients and on the mean values of the distributions.
In the upper panels of Figure \[fig:fig\_PCA\_2\] we show a comparison between the observations in our 2D-space datasets and the PCA model relations, while in the lower panels we show the distributions of the corresponding residuals. We find that the relation for which our PCA model can better reproduce the data is the $M_{\rm BH}$ - $\sigma_{\rm e}$, with a *Gaussian* 1$\sigma$ scatter of $\sigma \sim 0.47$. For the $M_{\rm BH}$ - $L$ and $M_{\rm BH}$ - $R_{\rm e}$ relations, our PCA model yields larger scatters in the residuals, respectively $\sigma \sim 0.63$ and $\sigma \sim 0.7$. These larger scatters are linked to the lower variance contained by PC1 in the samples respect to the $M_{\rm BH}$ - $\sigma_{\rm e}$ case, and therefore a more significant loss of information when setting PC2 to zero.
Altogether, our 2D-space PCA analysis suggests that, among $\sigma_{\rm e}$, $L$ and $R_{\rm e}$, $\sigma_{\rm e}$ is the observable that better correlates with $M_{\rm BH}$. The $M_{\rm BH}$-$\sigma_{\rm e}$ is the *more fundamental* scaling relation, with more than 95% of the information contained in the PC1 and only a scatter of $\sigma \lesssim 0.5$ in the residuals between the data and the PCA model relation.
### 4D-space PCA {#sec:4DPCA}
As a complementary way of exploring the mutual dependencies among the observables in our sample, we perform a PCA in the 4D-space defined by $M_{\rm BH}$, $\sigma_{\rm e}$, $L$ and $R_{\rm e}$. We find that PC1 contains 86.1 $\pm$ 0.4 % of the variance, confirming that the full set of four observables can be approximately well described by a 2D surface. PC2 contains 9.8 $\pm$ 0.4 % of the variance, meaning that accounting for a third dimension could recover $\sim$10 % of the information, while PC3 and PC4 contains only $\sim 2 ~\%$ (see Table 2). Following the same scheme discussed in Sec. \[sec:2DPCA\], setting to zero the PC that contain less variance, we obtain the best PCA model relation that expresses $M_{\rm BH}$ in terms of the other observables in the dataset:
$$\log(M_{\rm BH})^{\rm model}_{\rm 4D} = 4.05 \log(\sigma_{\rm e}) + 0.64 \log(L) - 0.32 \log(R_{\rm e}) - 7.66.$$
Consistently with the result obtained in Section \[sec:2DPCA\], we find that the primary dependence is attributed to $\sigma_{\rm e}$, i.e., the quantity that better describes $M_{\rm BH}$. $L$ and $R_{\rm e}$ have a secondary and tertiary dependence respectively, with relatively much lower weights ($\sim 16\%$ and $\sim 10\%$, computed as the ratios between the coefficients) with respect to $\sigma_{\rm e}$ Interestingly, as shown in Figure \[fig:fig\_PCA\_3\], the residuals obtained from the 4D-space PCA model relation are worse ($\sigma \sim 0.55$) than in the 2D-space PCA model relation obtained trough the optimal projection of the $M_{\rm BH}$-$\sigma_{\rm e}$ space. This effect is likely to be ascribed to some intrinsic noise introduced when adding $L_{\rm k}$ and $R_{\rm e}$ in a 4D-space.
Theoretical Perspective {#section:theory}
=======================
As we have seen, a growing body of work is pointing to the fundamental importance of the $M_{\rm BH}$-$\sigma_{\rm e}$. A key perspective that we have so far neglected in regards to black hole scaling relations is that of the *theoretical modeller*, which we will explore in this section.
The parameters of the galaxy that correlate with $M_{\rm BH}$ tell us which physical processes are most important in setting the black hole mass. Each parameter is related to certain physical quantities. For example, velocity dispersion is naturally related to the mass of the galaxy’s spheroidal component, and by extension to its gravitational potential. In the simplest case, modelling the bulge as an isothermal density profile, gas density is $\rho \propto \sigma_{\rm e}^2$ and its weight (the product of the gas mass and gravitational acceleration) is $W \propto \sigma_{\rm e}^4$. Therefore, modelling a connection between the upper limit of the black hole mass and the weight of the gas surrounding it may indeed be a good starting point to explaining the correlation. Alternatively, if the SMBH mass were controlled by stellar processes, such as turbulence driven by stellar feedback, we would expect a strong correlation between $M_{\rm BH}$ and stellar mass. Similarly, if the rate of SMBH feeding from large-scale reservoirs were an important constraint, a correlation with the bulge size $R_{\rm e}$ or dynamical timescale $t_{\rm dyn} \simeq R_{\rm e}/\sigma_{\rm e}$ might emerge. The fact that such correlations are not seen suggests that these processes are secondary to the host’s gravitational potential.
A very promising group of models that have emerged over the past two decades are those based on AGN feedback . The common argument is that AGN luminosity transfers energy to the surrounding gas and at some point drives it away, quenching further black hole growth. These models are generally capable of explaining not only the $\sigma_{\rm e}$ relation, but also the presence of quasi-relativistic nuclear winds and large-scale massive outflows observed in many active galaxies. Other models that presume either no causal connection between galaxy and black hole growth [@Peng2007ApJ; @Jahnke2011ApJ] or those that claim the black hole to be merely a passive recipient of a fraction of the gas used to build up the bulge [@Haan2009ApJ; @Angles2013ApJ; @Angles2015ApJ] make no predictions regarding outflows and generally connect the black hole mass to the mass, rather than velocity dispersion, of the galaxy bulge.
There are several ways of transferring AGN power to the surrounding gas, e.g. radiation, winds and/or jets [@Morganti2017arXiv]. Jets are typically efficient on galaxy cluster scales, heating intergalactic gas and prevent it from falling back into the galaxy . This process, referred to as “maintenance mode” of feedback, prevents the SMBH mass from growing above the limit established by the $M_{\rm BH}$-$\sigma_{\rm e}$ relation. Jets are considered to be the primary form of feedback in AGN that accrete at low rates and have luminosities $L < 0.01 L_{\rm AGN}$ [@Merloni2007MNRAS]. The opposite type of feedback is known as “quasar mode”, and it is believed to be most efficient in more luminous AGN. Here, again, there are two possibilities in which energy can be transferred. Directly coupling AGN luminosity to the gas in the interstellar medium is possible if the gas is dusty (due to a very high opacity, see [@Fabian2008MNRAS]). On the other hand, dust evaporates when shocked to the temperatures expected within AGN outflows [@Barnes2018arXiv], potentially limiting the impact of radiation-driven outflows. A much more promising avenue is to connect the AGN with the surrounding gas via a quasi-relativistic wind . Such a model naturally produces both a $M_{\rm BH}$-$\sigma_{\rm e}$ relation similar to the observed one, and outflow properties in excellent agreement with observations, both within galaxies and on intergalactic scales in galaxy groups [@Lapi2005ApJ].
AGN wind-driven feedback
------------------------
AGN are highly variable on essentially all timescales and are known to occasionally reach the Eddington luminosity $$L_{\rm Edd} = \frac{4 \pi G M_{\rm BH} c}{\kappa_{\rm e.s.}},$$ where $\kappa_{\rm e.s} \simeq 0.346$ cm$^2$ g$^{-1}$ is the electron scattering opacity. Under such circumstances, the geometrically thin accretion disc produces a quasi-spherical wind that self-regulates to an optical depth $\tau \sim 1$ [@King2003MNRAS]. Therefore each photon emitted by the AGN will, on average, scatter only once before escaping to infinity, and the wind carries a momentum rate $$\dot{M}_{\rm w} v_{\rm w} = \tau \frac{L_{\rm AGN}}{c},$$ where $\dot{M}_{\rm w}$ is the wind mass flow rate, $v_{\rm w}$ is the wind velocity and $L_{\rm AGN} \equiv lL_{\rm Edd}$ is the AGN luminosity, where $l$ is the Eddington ratio. By writing $L_{\rm AGN} = \eta \dot{M}_{\rm BH} c^2$, we find the wind velocity to be $$v_{\rm w} = \frac{\tau \eta}{\dot{m}}c,$$ where $\dot{m} \equiv \dot{M}_{\rm w}/\dot{M}_{\rm BH}$. The value of $\dot{m}$ is highly uncertain, but should not be extremely different from unity. To see this, consider the extreme ends of the possible range of $\dot{M}_{\rm BH}$. If the accretion rate on to the accretion disc is significantly below Eddington, no wind is produced, while if the accretion rate rises above the Eddington limit, the wind moderates the accretion flow. Overall, the highest possible average accretion rate is the dynamical rate: $$\dot{M}_{\rm dyn} = f_{\rm g} \frac{\sigma^3}{G} \simeq \frac{64}{\sigma_{\rm200}}\dot{M}_{\rm Edd},$$ where $f_{\rm g} \simeq 0.16$ is the cosmological gas fraction and $\sigma \equiv 200 \sigma_{\rm200}$ km s$^{-1}$ is the velocity dispersion in the galaxy . In deriving the second equality, we used the $M_{\rm BH} - \sigma$ relation that is derived below, in eq. \[eq:mcritical\]. Therefore, in most cases, the SMBH feeding rate is not significantly higher than the Eddington rate, unless $M_{\rm BH}$ is well below the observed relation. As a result, we take $\dot{m} \sim 1$ for the rest of this section. This leads to the final expression for the AGN wind velocity $$v_{\rm w} \simeq\eta c \simeq 0.1c,$$ which is very close to the average velocity in observed winds . The kinetic power of the wind is $$\dot{E}_{\rm w} = \frac{\dot{M}_{\rm w} v_{\rm w}^2}{2} \simeq \frac{\eta}{2}L_{\rm AGN} \simeq 0.05 L_{\rm AGN}.$$
The wind rapidly reaches the interstellar medium (ISM) surrounding the AGN and shocks against it. The shock is strong, since $v_{\rm w}/\sigma \gg 1$, and the wind heats up to a temperature $$T_{\rm sh} = \frac{3m_{\rm p}v_{\rm w}^2}{16 k_{\rm b}} \simeq 10^{10} {\rm K},$$ where $m_{\rm p}$ is the proton mass, and $k_{\rm b}$ is the Boltzmann constant. The most efficient cooling process at this temperature is Inverse Compton (IC) cooling via interaction with AGN photons [@King2003ApJ; @Faucher2012MNRAS]. Most of the photons interact with electrons in the shocked wind, and a two-temperature plasma develops [@Faucher2012MNRAS]. The actual cooling timescale then depends on the timescale for energy equilibration between electrons and protons. As a result, cooling is highly inefficient and the shocked wind can expand as an approximately adiabatic bubble.
The subsequent evolution of the expanding bubble depends on the density structure of the ISM. Most of the energy stored in the hot wind bubble escapes through the low-density channels and creates a large-scale outflow [@Zubovas2014MNRAS]. Denser clouds, however, remain and are mainly affected by the direct push of the wind material. These two situations create two kinds of outflow, known as energy-driven and momentum-driven, respectively. The latter kind is responsible for establishing the $M_{\rm BH} - \sigma_{\rm e}$ relation.
The predicted relation
----------------------
Momentum-driven outflows push against the dense clouds surrounding the black hole. These clouds are the most likely sources of subsequent black hole feeding, therefore their removal quenches further black hole growth for a significant time and establishes the $M_{\rm BH}$-$\sigma_{\rm e}$ relation [@King2003ApJ; @Murray2005ApJ; @King2010MNRASa]. Considering the balance between AGN wind momentum and the weight of the gas $W_{\rm gas}$ leads to a critical AGN luminosity required for clearing the dense gas: $$L_{\rm crit} = W_{\rm gas}c \simeq \frac{4f_{\rm g}\sigma^4c}{G},$$ where the second equality assumes that the gas distribution and the background gravitational potential are isothermal, i.e. $\rho = \sigma^2/\left(2\pi G R^2\right)$ [@Murray2005ApJ]. Equating this critical luminosity with the Eddington luminosity of the black hole allows us to derive a critical mass [@King2010MNRASa]: $$\label{eq:mcritical}
M_{\rm crit} \simeq \frac{f_{\rm g}\kappa_{\rm e.s.}\sigma^4}{\pi G^2} \simeq 3.2\times10^8 \frac{f_{\rm g}}{0.16} \sigma_{\rm200}^4 \, M_\odot.$$ This value is very close to the observed one, although it has a slightly shallower slope. This discrepancy may be explained by the fact that the black hole still grows during the time while it drives the gas away [@Zubovas2012MNRASb]. As the gas is pushed away, it joins the energy-driven outflow. This outflow coasts for approximately an order of magnitude longer than the AGN phase inflating it and stalls at a distance [@King2011MNRAS] $$R_{\rm stall} \simeq \frac{v_{\rm e}^2}{\sigma} t_{\rm AGN},$$ where $t_{\rm AGN}$ is the duration of the driving phase and the energy-driven outflow velocity is [@King2005ApJ; @Zubovas2012ApJ] $$v_{\rm e} = \left(\frac{2\eta c}{3 \sigma}\frac{0.16}{f_{\rm g}}\right)^{1/3} \simeq 925 \sigma_{\rm200}^{2/3} \left(\frac{0.16}{f_{\rm g}}\right)^{1/3} {\rm km s}^{-1}.$$ By equating $R_{\rm stall}$ with either the bulge radius or the virial radius of the galaxy, we obtain the time $t_{\rm AGN}$ for which the galaxy must be active in order to quench further accretion on to the black hole and find $t_{\rm AGN} \propto R \sigma^{-1/3} \propto \sigma^{2/3}$, since $R \propto \sigma_{\rm e}$ on average [this relation arises from the Fundamental plane of galaxies, see @Djorgovski1987ApJ; @Cappellari2013MNRAS]. Note that this growth does not need to happen all at once: as long as the outflow is still progressing by the time the next episode begins, the system behaves as if it was powered by a continuously shining AGN [@Zubovas2019MNRASa].
This extra growth steepens the $M_{\rm BH}$-$\sigma_{\rm e}$ relation beyond the simpler analytical prediction and brings it more in line with observations [@Zubovas2012MNRASb]. Furthermore, it shows that galaxy radius may be an important secondary parameter determining the final black hole mass.
As a final note, the extra black hole growth while clearing the galaxy also depends on its spin. Since a rapidly spinning black hole produces more luminosity and drives a faster outflow than a slow-spinning one, the latter has to be active for longer and grow more before it clears the gas from the galaxy. Although present-day estimates of black hole spins are not robust or numerous enough to test this prediction in detail, this might become possible in the near future [@Zubovas2019MNRASc].
In general, theoretical models based on momentum-driven outflows are capable of naturally explaining the relationship between black hole mass and velocity dispersion, primarily due to the latter acting as a tracer of the host’s gravitational potential well. In addition, these models could account for *secondary*, weaker dependencies on, e.g., galaxy stellar mass or size, which may still be allowed by current data as discussed above (see, e.g., Figures \[fig|FigureResiduals\] and \[fig:fig\_PCA\_2\] and @2016MNRAS.460.3119S).
Discussion and Conclusions {#section:conclusion}
==========================
In this paper we have reviewed previous evidence for the $M_{\rm BH}$-$\sigma_{\rm e}$ being the most fundamental of all black hole-host galaxy scaling relations (among those discovered so far) and we have presented new evidence based on the statistical analysis of the sample recently compiled by [@2019MNRAS.490..600D]. Both residuals (e.g., [@2016MNRAS.460.3119S]) and PCA analyses point to $\sigma_{\rm e}$ being more fundamental than both stellar luminosity/mass or effective radius in their correlation to central black hole mass.
Theoretically, as reviewed by , the $M_{\rm BH}$-$\sigma_{\rm e}$ arises as a consequence of AGN feedback. In short, the black hole in these models is expected to grow until it becomes massive enough to drive energetic/high-momentum large-scale winds that can potentially remove residual gas, inhibiting further star formation and black hole growth. The limiting mass reached by the black hole, which ultimately depends on the potential well of the host, naturally provides an explanation for the existence of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation.
Its fundamental nature and lower inclination towards selection biases (in comparison to other scaling relations, e.g. [@2016MNRAS.460.3119S]) make the $M_{\rm BH}$-$\sigma_{\rm e}$ relation the ideal benchmark for statistical studies of black holes in a variety of contexts. The $M_{\rm BH}$-$\sigma_{\rm e}$ relation should always be the one adopted to constrain the $f_{\rm vir}$ factor used in reverberation mapping studies (see e.g. [@2006ApJ...641..689V]) to infer black hole masses from active galaxies (e.g., [@2019MNRAS.485.1278S]). The $M_{\rm BH}$-$\sigma_{\rm e}$ relation also provides more robust large-scale clustering predictions in black hole mock catalogues [@2019arXiv191010175S]. Furthermore, pulsar timing array predictions of the gravitational wave background (e.g. [@2013CQGra..30v4009K]) are strongly dependent on the normalization of the black hole scaling relations ([@2008MNRAS.390..192S], [@2016MNRAS.460.3119S]), but they should be based on the $M_{\rm BH}$-$\sigma_{\rm e}$ rather than on the $M_{\rm BH}$-$M_*$ relation (see [@2015MNRAS.451.2417R]).
The shape and scatter of the $M_{\rm BH}$-$\sigma_{\rm e}$ relation could yield important information on the evolutionary channels of black hole growth. For example, its scatter could retain memory of the merger histories of the host galaxies [@2015MNRAS.446.2330S]. More broadly speaking, global star formation and black hole growth from continuity equation argument modelling is known to peak at around $z\sim2$ (e.g. [@2009ApJ...690...20S], [@2014MNRAS.439.2736D]). This is in itself consistent with the idea that black holes and their hosts may be co-evolving, and understanding how the $M_{\rm BH}$-$\sigma_{\rm e}$ relation precisely evolves over cosmic time or change as a function of environment could set invaluable constraints on the mechanisms behind black hole growth (e.g. [@2014MNRAS.442.2304H], [@2015MNRAS.453.4112F] and [@2015MNRAS.452..575S]).
Conflict of Interest Statement {#conflict-of-interest-statement .unnumbered}
==============================
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Author Contributions {#author-contributions .unnumbered}
====================
C. Marsden is the primary author and wrote the background and review sections. F. Shankar provided supervision, edited the manuscript, and provided the calculations of the residuals. M. Ginolfi performed and wrote the section on PCA analysis, K. Zubovas wrote the section on theoretical models.
Funding {#funding .unnumbered}
=======
C. Marsden acknowledges funding from the ESPRC for his PhD. F. Shankar acknowledges partial support from a Leverhulme Trust Research Fellowship.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Stefano de Nicola and Alessandro Marconi for sharing their data in electronic format. We acknowledge extensive use of the Python libraries astropy, matplotlib, numpy, pandas, and scipy.
Figure captions {#figure-captions .unnumbered}
===============
{width="100.00000%"}
{width="100.00000%"}
{width="100.00000%"}
\[tab:table1\]
[ll]{} & \
\[-1.5ex\] $M_{\rm BH}$ - $\sigma_{\rm vel}$ & Variance \[%\]\
\[+0.5ex\] & \
\[-1.5ex\] PC 1 & 95.4 $\pm$ 0.4 %\
\[+0.5ex\] PC 2 & 4.6 $\pm$ 0.3 \
\[+0.5ex\]
$M_{\rm BH}$ - $L_{\rm k}$ & Variance \[%\]\
\[0.5ex\] & \
\[-1.5ex\] PC 1 & 89.7 $\pm$ 0.5 \
\[+0.5ex\] PC 2 & 10.3 $\pm$ 0.4 \
\[+0.5ex\]
$M_{\rm BH}$ - $R_{\rm e}$ & Variance \[%\]\
\[+0.5ex\] & \
\[-1.5ex\] PC 1 & 88.2 $\pm$ 0.3 \
\[+0.5ex\] PC 2 & 11.8 $\pm$ 0.2 \
\[+0.5ex\]
\[tab:table2\]
[llllll]{} & & & & & \
\[-1.5ex\] Principal Component & log($\sigma_{\rm vel}$)$^{\rm PCA}$ & log($L_{\rm k}$)$^{\rm PCA}$ & log($R_{\rm e}$)$^{\rm PCA}$ & log($M_{\rm BH}$)$^{\rm PCA}$ & Variance \[%\]\
\[+0.5ex\] & & & & & \
\[-1.5ex\] PC 1 & 0.491 $\pm$ 0.002 & 0.514 $\pm$ 0.001 & 0.490 $\pm$ 0.002 & 0.503 $\pm$ 0.002 & 86.1 $\pm$ 0.4\
\[+0.5ex\] PC 2 & 0.14 $\pm$ 0.01 & -0.09 $\pm$ 0.02 & -0.15 $\pm$ 0.01 & +0.09 $\pm$ 0.02 & 9.8 $\pm$ 0.5\
\[+0.5ex\] PC 3 & 0.22 $\pm$ 0.04 & -0.45 $\pm$ 0.03 & +0.38 $\pm$ 0.02 & -0.13 $\pm$ 0.05 & 2.5 $\pm$ 0.3\
\[+0.5ex\] PC 4 & 0.39 $\pm$ 0.03 & +0.25 $\pm$ 0.05 & -0.12 $\pm$ 0.05 & -0.52 $\pm$ 0.02 & 1.6 $\pm$ 0.3\
\[+0.5ex\]
![ The distribution of the residuals computed subtracting the 4D-space PCA model predictions to the observed $M_{\rm BH}$ is shown (grey histogram), along with its Gaussian fits (red dashed line). []{data-label="fig:fig_PCA_3"}](fig_PCA_3_updated.pdf){width="0.75\columnwidth"}
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We investigate the performance of the range-separated hybrid (RSH) scheme, which combines long-range Hartree-Fock (HF) and a short-range density-functional approximation (DFA), for calculating photoexcitation/photoionization spectra of the H and He atoms, using a B-spline basis set in order to correctly describe the continuum part of the spectra. The study of these simple systems allows us to quantify the influence on the spectra of the errors coming from the short-range exchange-correlation DFA and from the missing long-range correlation in the RSH scheme. We study the differences between using the long-range HF exchange (nonlocal) potential and the long-range exact exchange (local) potential. Contrary to the former, the latter supports a series of Rydberg states and gives reasonable photoexcitation/photoionization spectra, even without applying linear-response theory. The most accurate spectra are obtained with the linear-response time-dependent range-separated hybrid (TDRSH) scheme. In particular, for the He atom at the optimal value of the range-separation parameter, TDRSH gives slightly more accurate photoexcitation and photoionization spectra than standard linear-response time-dependent HF. More generally, the present work shows the potential of range-separated density-functional theory for calculating linear and nonlinear optical properties involving continuum states.'
author:
- Felipe Zapata
- Eleonora Luppi
- Julien Toulouse
date: 'March 14, 2019'
title: |
Linear-response range-separated density-functional theory\
for atomic photoexcitation and photoionization spectra
---
Introduction
============
Nowadays, time-dependent density-functional theory (TDDFT) [@RunGro-PRL-84], applied within the linear-response formalism [@GroKoh-PRL-85; @Cas-INC-95; @PetGosGro-PRL-96], is a widely used approach for calculating photoexcitation spectra (transitions from bound to bound states) of electronic systems. In spite of many successes, it is however well known that usual (semi-)local density-functional approximations (DFAs), i.e. the local-density approximation (LDA) and generalized-gradient approximations (GGAs), for the exchange-correlation potential and its associated exchange-correlation kernel do not correctly describe long-range electronic transitions, such as those to Rydberg [@CasJamCasSal-JCP-98] and charge-transfer [@DreWeiHea-JCP-03] states in atomic and molecular systems. A better description of Rydberg excitations can be obtained with exchange-correlation potential approximations having the correct $-1/r$ long-range asymptotic decay [@LeeBae-PRA-94; @TozHan-JCP-98; @CasSal-JCP-00; @SchGriGisBae-JCP-00], even though it has been shown that accurate Rydberg excitation energies and oscillator strengths can in fact be extracted from LDA calculations in small atoms [@WasMaiBur-PRL-03; @WasBur-PRL-05]. A more general solution for correcting both Rydberg and charge-transfer excitations is given by range-separated TDDFT approaches [@TawTsuYanYanHir-JCP-04; @YanTewHan-CPL-04; @PeaHelSalKeaLutTozHan-PCCP-06; @LivBae-PCCP-07; @BaeLivSal-ARPC-10; @FroKneJen-JCP-13; @RebSavTou-MP-13] which express the long-range part of the exchange potential and kernel at the Hartree-Fock (HF) level. These range-separated approaches also give reasonably accurate values for the ionization energy threshold [@YanTewHan-CPL-04; @GerAng-CPL-05a; @TsuSonSuzHir-JCP-10].
Linear-response TDDFT has also been used for calculating photoionization spectra (transitions from bound to continuum states) of atoms and molecules [@ZanSov-PRA-80; @LevSov-PRA-84; @SteDecLis-JPB-95; @SteAltFroDec-CP-97; @SteDec-JPB-97; @SteDec-JCP-00; @SteDecGor-JCP-01; @SteFroDec-JCP-05; @SteTofFroDec-JCP-06; @TofSteDec-PRA-06; @SteTofFroDec-TCA-07; @ZhoChu-PRA-09]. These calculations are less standard in quantum chemistry since they involve spatial grid methods or B-spline basis sets for a proper description of the continuum states. In this case as well, usual (semi-)local DFAs provide a limited accuracy and asymptotically corrected exchange-correlation potential approximations give more satisfactory results. More accurate still, but less common, are photoionization spectra calculated with the exact-exchange (EXX) potential [@SteDecGor-JCP-01] or the localized HF exchange potential and its associated kernel [@ZhoChu-PRA-09]. Recently, range-separated approximations have been successfully used for calculating photoexcitation and photoionization spectra of molecular systems using time-propagation TDDFT with Gaussian basis sets together with an effective lifetime model compensating for the missing continuum states [@LopGov-JCTC-13; @FerBalLop-JCTC-15; @SisAbaMauGaaSchLop-JCP-16]. However, to the best of our knowledge, range-separated approximations have not yet been used in frequency-domain linear-response TDDFT calculations of photoionization spectra.
In this work, we explore the performance of the linear-response time-dependent range-separated hybrid (TDRSH) scheme [@RebSavTou-MP-13; @TouRebGouDobSeaAng-JCP-13] for calculating photoexcitation and photoionization spectra of the H and He atoms using a B-spline basis set to accurately describe the continuum part of the spectra. The TDRSH scheme allows us to treat long-range exchange effects at the HF level and short-range exchange-correlation effects within (semi-)local DFAs. First, the dependence of the range-separated hybrid (RSH) orbital energies on the range-separation parameter is investigated, as well as the effect of replacing the long-range HF exchange nonlocal potential by the long-range EXX local potential (resulting in a scheme that we refer to as RSH-EXX). Second, oscillator strengths directly computed with the RSH and the RSH-EXX orbitals are compared with oscillator strengths obtained with the linear-response TDRSH scheme. The study of the H atom allows us to quantify the residual self-interaction error coming from the short-range exchange-correlation DFA, and the study of the He atom permits to quantify the effect of the missing long-range correlation in the RSH scheme. This work constitutes a first step for applying range-separated TDDFT to strong-field phenomena, such as high-harmonic generation or above-threshold ionization, where long-range effects and continuum states play an important role.
The outline of the paper is as follows. In Sec. \[sec:theory\], firstly, we briefly review the RSH scheme and introduce the RSH-EXX variant, and, secondly, we review the linear-response TDRSH method. In Sec. \[sec:implementation\], the basis set of B-spline functions is defined, and we indicate how the range-separated two-electron integrals are computed using an exact spherical harmonic expansion for the range-separated interaction. In Sec. \[sec:results\] results are presented and discussed. Firstly, we show the performance of the B-spline basis set for describing the density of continuum states of the H atom within the different methods. Secondly, the dependence of the orbital energies of the H and He atoms on the range-separation parameter is analyzed. Thirdly, different calculated photoexcitation/photoionization spectra for the H and He atoms are discussed and compared with exact results. In Sec. \[sec:conclusions\], conclusions and perspectives are given. Unless otherwise indicated, Hartree atomic units are used throughout the paper.
Range-separated density-functional theory {#sec:theory}
=========================================
Range-separated hybrid scheme
-----------------------------
Range-separated density-functional theory (see, e.g., Refs. ) is based on the splitting of the Coulomb electron-electron interaction $w_\text{ee}(r)=1/r$ into long-range (lr) and short-range (sr) contributions $$w_{{\ensuremath{\text{ee}}}}(r)=w_{{\ensuremath{\text{ee}}}}^{\ensuremath{\text{lr}}}(r)+w_{{\ensuremath{\text{ee}}}}^{\ensuremath{\text{sr}}}(r),$$ and the most common forms for the long-range and short-range interactions are $$\label{erflr}
w_{{\ensuremath{\text{ee}}}}^{\ensuremath{\text{lr}}}(r)=\frac{\operatorname{erf}(\mu r)}{r},$$ and $$\label{erfcsr}
w_{{\ensuremath{\text{ee}}}}^{\ensuremath{\text{sr}}}(r)=\frac{\operatorname{erfc}(\mu r)}{r}.$$ where $\operatorname{erf}$ and $\operatorname{erfc}$ are the error function and the complementary error function, respectively, and $\mu$ is a tunable range-separation parameter controlling the range of the separation. Using this decomposition, it is possible to rigorously combine a long-range wave-function approach with a complementary short-range DFA.
The simplest approach in range-separated density-functional theory consists in using a single-determinant wave function for the long-range interaction. This leads to the RSH scheme [@AngGerSavTou-PRA-05] which spin orbitals $\{\varphi_{p}({\ensuremath{\mathbf{x}}})\}$ (where ${\ensuremath{\mathbf{x}}}=({\ensuremath{\mathbf{r}}},\sigma)$ are space-spin coordinates) and orbital energies $\varepsilon_{p}$ can be determined for a given system by the following eigenvalue problem, $$\begin{aligned}
\left( -\frac{1}{2} \bm{\nabla}^2 +v_{\text{ne}}({\ensuremath{\mathbf{r}}}) + v_\text{H}({\ensuremath{\mathbf{r}}}) + v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}}) \right) \varphi_{p}({\ensuremath{\mathbf{x}}})
\nonumber\\
+ \int v_{\text{x}}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}},{\ensuremath{\mathbf{x}}}') \varphi_{p}({\ensuremath{\mathbf{x}}}') {\ensuremath{\text{d}}}{\ensuremath{\mathbf{x}}}' = \varepsilon_{p}\varphi_{p}({\ensuremath{\mathbf{x}}}),
\label{RSH}\end{aligned}$$ where $v_{\text{ne}}({\ensuremath{\mathbf{r}}})$ is the nuclei-electron potential, $v_\text{H}({\ensuremath{\mathbf{r}}})$ is the Hartree potential for the Coulomb electron-electron interaction, $$\begin{aligned}
v_\text{H}({\ensuremath{\mathbf{r}}}) = \int n({\ensuremath{\mathbf{x}}}') w_{\ensuremath{\text{ee}}}(|{\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}'|) {\ensuremath{\text{d}}}{\ensuremath{\mathbf{x}}}',\end{aligned}$$ where $n({\ensuremath{\mathbf{x}}})=\sum_i^\text{occ} |\varphi_{i}({\ensuremath{\mathbf{x}}})|^2$ are the spin densities ($i$ refers to occupied spin orbitals), $v_{\text{x}}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}},{\ensuremath{\mathbf{x}}}')$ is the nonlocal HF exchange potential for the long-range electron-electron interaction, $$\begin{aligned}
v_{\text{x}}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}},{\ensuremath{\mathbf{x}}}') = - \sum_{i}^\text{occ} \varphi_{i}^*({\ensuremath{\mathbf{x}}}') \varphi_{i}({\ensuremath{\mathbf{x}}}) w_{\ensuremath{\text{ee}}}^{\ensuremath{\text{lr}}}(|{\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}'|),\end{aligned}$$ and $v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}})$ is the short-range exchange-correlation potential $$\begin{aligned}
v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}}) = \frac{\delta \bar{E}_\text{xc}^{\ensuremath{\text{sr}}}}{\delta n({\ensuremath{\mathbf{x}}})},\end{aligned}$$ where $\bar{E}_\text{xc}^{\ensuremath{\text{sr}}}$ is the complement short-range exchange-correlation density functional. In this work, we use the short-range spin-dependent LDA exchange-correlation functional of Ref. for $\bar{E}_\text{xc}^{\ensuremath{\text{sr}}}$. The long-range and short-range potentials, $v_{\text{x}}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}},{\ensuremath{\mathbf{x}}}')$ and $v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}})$, explicitly depend on the range-separation parameter $\mu$, and consequently the spin orbitals, the orbital energies, and the density also implicitly depend on it. For $\mu=0$, $v_{\text{x}}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}},{\ensuremath{\mathbf{x}}}')$ vanishes and $v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}})$ becomes the usual full-range LDA exchange-correlation potential, and thus the RSH scheme reduces to standard Kohn-Sham LDA. For $\mu\to\infty$, $v_{\text{x}}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}},{\ensuremath{\mathbf{x}}}')$ becomes the usual full-range HF exchange potential and $v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}})$ vanishes, and thus the RSH scheme reduces to standard HF.
In the present paper, we also consider the following variant of the RSH scheme, $$\begin{aligned}
\left( -\frac{1}{2} \bm{\nabla}^2 +v_{\text{ne}}({\ensuremath{\mathbf{r}}}) + v_\text{H}({\ensuremath{\mathbf{r}}}) + v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}}) + v_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}) \right) \varphi_{p}({\ensuremath{\mathbf{x}}})
\nonumber\\
= \varepsilon_{p} \varphi_{p}({\ensuremath{\mathbf{x}}}), \;\;\;\;\;\;
\label{RSHEXX}\end{aligned}$$ in which the long-range nonlocal HF exchange potential has been replaced by the long-range local EXX [@TalSha-PRA-76; @GorLev-PRA-94; @GorLev-IJQC-95] potential $$\begin{aligned}
\label{vexx}
v_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}) = \frac{\delta E_\text{x}^{\ensuremath{\text{lr}}}}{\delta n({\ensuremath{\mathbf{x}}})},\end{aligned}$$ where $E_\text{x}^{\ensuremath{\text{lr}}}$ is the long-range exchange density functional [@TouGorSav-IJQC-06; @TouSav-JMS-06]. We will refer to this scheme as RSH-EXX. The calculation of the EXX potential is involved [@FilUmrGon-PRA-96; @Gor-PRL-99; @IvaHirBar-PRL-99], with the exception of one- and two-electron systems. Indeed, for one-electron systems, the long-range EXX potential is simply $$\begin{aligned}
v_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}) = - v^{{\ensuremath{\text{lr}}}}_\text{H}({\ensuremath{\mathbf{r}}}),\end{aligned}$$ and for systems of two electrons in a single spatial orbital, it is $$\begin{aligned}
v_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}) = - \frac{1}{2} v^{{\ensuremath{\text{lr}}}}_\text{H}({\ensuremath{\mathbf{r}}}),\end{aligned}$$ where $v^{\ensuremath{\text{lr}}}_\text{H}({\ensuremath{\mathbf{r}}}) = \int n({\ensuremath{\mathbf{x}}}') w_{\ensuremath{\text{ee}}}^{\ensuremath{\text{lr}}}(|{\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}'|) {\ensuremath{\text{d}}}{\ensuremath{\mathbf{x}}}'$ is the long-range Hartree potential. For these one- and two-electron cases, it can be shown that Eqs. (\[RSH\]) and (\[RSHEXX\]) give identical occupied orbitals but different unoccupied orbitals. More generally, for systems with more than two electrons, the HF and EXX exchange potentials give similar occupied orbitals but very different unoccupied orbitals.
Once orbitals and orbital energies are obtained from Eqs. (\[RSH\]) and (\[RSHEXX\]), the bare oscillator strengths can be calculated. They are defined as $$\begin{aligned}
\label{oscillator0}
f_{ia}^0 = \frac{2}{3} \omega_{ia}^0 \sum_{\nu=x,y,z} |d_{\nu,ia}|^2,\end{aligned}$$ where $i$ and $a$ refer to occupied and unoccupied spin orbitals, respectively, $\omega_{ia}^0 = \varepsilon_{a} - \varepsilon_{i}$ are the bare excitation energies and $d_{\nu,ia} = \int \varphi_{i}^*({\ensuremath{\mathbf{x}}}) r_\nu \varphi_{a}({\ensuremath{\mathbf{x}}}) {\ensuremath{\text{d}}}{\ensuremath{\mathbf{x}}}$ are the dipole-moment transition integrals. We will consider these bare excitation energies $\omega_{ia}^0$ and oscillator strengths $f_{ia}^0$ for a first approximation to photoexcitation/photoionization spectra.
Linear-response time-dependent range-separated hybrid scheme
------------------------------------------------------------
In the time-dependent extension of the RSH scheme within linear response (referred to as TDRSH) [@RebSavTou-MP-13; @TouRebGouDobSeaAng-JCP-13; @FroKneJen-JCP-13], one has to solve the following pseudo-Hermitian eigenvalue equation $$\begin{aligned}
\label{TDRSH}
\begin{pmatrix}
{\ensuremath{\mathbf{A}}} & {\ensuremath{\mathbf{B}}} \\
-{\ensuremath{\mathbf{B}}}^* & -{\ensuremath{\mathbf{A}}}^*
\end{pmatrix}
\begin{pmatrix}
{\ensuremath{\mathbf{X}}}_n \\ {\ensuremath{\mathbf{Y}}}_n
\end{pmatrix} = \omega_n
\begin{pmatrix}
{\ensuremath{\mathbf{X}}}_n \\ {\ensuremath{\mathbf{Y}}}_n
\end{pmatrix},\end{aligned}$$ whose solutions come in pairs: excitation energies $\omega_n>0$ with eigenvectors $({\ensuremath{\mathbf{X}}}_n,{\ensuremath{\mathbf{Y}}}_n)$, and de-excitation energies $\omega_n<0$ with eigenvectors $({\ensuremath{\mathbf{Y}}}_n^*,{\ensuremath{\mathbf{X}}}_n^*)$. The elements of the matrices ${\ensuremath{\mathbf{A}}}$ and ${\ensuremath{\mathbf{B}}}$ are $$\begin{aligned}
A_{ia,jb} = (\varepsilon_{a} -\varepsilon_{i}) \delta_{ij} \delta_{ab} + K_{ia,jb},\end{aligned}$$ $$\begin{aligned}
B_{ia,jb} = K_{ia,bj},\end{aligned}$$ where $i,j$ and $a,b$ refer to occupied and unoccupied RSH spin orbitals, respectively, and the coupling matrix ${\ensuremath{\mathbf{K}}}$ contains the contributions from the Hartree kernel $f_{\ensuremath{\text{H}}}({\ensuremath{\mathbf{r}}}_1,{\ensuremath{\mathbf{r}}}_2)=w_{\ensuremath{\text{ee}}}(|{\ensuremath{\mathbf{r}}}_1-{\ensuremath{\mathbf{r}}}_2|)$, the long-range HF exchange kernel $f_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2;{\ensuremath{\mathbf{x}}}_1',{\ensuremath{\mathbf{x}}}_2')=-w_{\ensuremath{\text{ee}}}^{\ensuremath{\text{lr}}}(|{\ensuremath{\mathbf{r}}}_1-{\ensuremath{\mathbf{r}}}_2|) \delta({\ensuremath{\mathbf{x}}}_1-{\ensuremath{\mathbf{x}}}_2') \delta({\ensuremath{\mathbf{x}}}_1'-{\ensuremath{\mathbf{x}}}_2)$, and the adiabatic short-range exchange-correlation kernel $f_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2)=\delta v_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}}_1)/\delta n({\ensuremath{\mathbf{x}}}_2)$ $$\begin{aligned}
\label{K}
K_{ia,jb} &=& {\ensuremath{\langle aj \vert}} f_{\ensuremath{\text{H}}}{\ensuremath{\vert ib \rangle}}+ {\ensuremath{\langle aj \vert}} f_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}} {\ensuremath{\vert ib \rangle}} + {\ensuremath{\langle aj \vert}} f_\text{xc}^{{\ensuremath{\text{sr}}}} {\ensuremath{\vert ib \rangle}}
\nonumber\\
&=& {\ensuremath{\langle aj \vert}} w_{\ensuremath{\text{ee}}}{\ensuremath{\vert ib \rangle}} - {\ensuremath{\langle aj \vert}} w_{\ensuremath{\text{ee}}}^{\ensuremath{\text{lr}}}{\ensuremath{\vert bi \rangle}} + {\ensuremath{\langle aj \vert}} f_\text{xc}^{{\ensuremath{\text{sr}}}} {\ensuremath{\vert ib \rangle}},\end{aligned}$$ where ${\ensuremath{\langle aj \vert}} w_{\ensuremath{\text{ee}}}{\ensuremath{\vert ib \rangle}}$ and ${\ensuremath{\langle aj \vert}} w_{\ensuremath{\text{ee}}}^{\ensuremath{\text{lr}}}{\ensuremath{\vert bi \rangle}}$ are the two-electron integrals associated with the Coulomb and long-range interactions, respectively, and ${\ensuremath{\langle aj \vert}} f_\text{xc}^{{\ensuremath{\text{sr}}}} {\ensuremath{\vert ib \rangle}} = \iint \varphi_a^*({\ensuremath{\mathbf{x}}}_1) \varphi_j^*({\ensuremath{\mathbf{x}}}_2) f_\text{xc}^{{\ensuremath{\text{sr}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2) \varphi_i({\ensuremath{\mathbf{x}}}_1) \varphi_b({\ensuremath{\mathbf{x}}}_2) {\ensuremath{\text{d}}}{\ensuremath{\mathbf{x}}}_1 {\ensuremath{\text{d}}}{\ensuremath{\mathbf{x}}}_2 $. Since we use the short-range LDA exchange-correlation density functional, for $\mu=0$ the TDRSH scheme reduces to the usual linear-response time-dependent local-density approximation (TDLDA). For $\mu\to\infty$, the TDRSH scheme reduces to standard linear-response time-dependent Hartree-Fock (TDHF).
The time-dependent extension of the RSH-EXX variant within linear response (referred to as TDRSH-EXX) leads to identical equations with the exception that the long-range HF exchange kernel $f_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{HF}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2;{\ensuremath{\mathbf{x}}}_1',{\ensuremath{\mathbf{x}}}_2')$ is replaced by the long-range frequency-dependent EXX kernel [@Gor-PRA-98; @Gor-IJQC-98] $f_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2;\omega)=\delta v_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}_1,\omega)/\delta n({\ensuremath{\mathbf{x}}}_2,\omega)$. For one-electron systems, the long-range EXX kernel is simply $$\begin{aligned}
f_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2;\omega)= -f_{\ensuremath{\text{H}}}^{\ensuremath{\text{lr}}}({\ensuremath{\mathbf{r}}}_1,{\ensuremath{\mathbf{r}}}_2),\end{aligned}$$ and, for systems with two electrons in a single spatial orbital, it is $$\begin{aligned}
f_\text{x}^{{\ensuremath{\text{lr}}},{\ensuremath{\text{EXX}}}}({\ensuremath{\mathbf{x}}}_1,{\ensuremath{\mathbf{x}}}_2;\omega)= -\frac{1}{2} f_{\ensuremath{\text{H}}}^{\ensuremath{\text{lr}}}({\ensuremath{\mathbf{r}}}_1,{\ensuremath{\mathbf{r}}}_2),\end{aligned}$$ where $f_{\ensuremath{\text{H}}}^{\ensuremath{\text{lr}}}({\ensuremath{\mathbf{r}}}_1,{\ensuremath{\mathbf{r}}}_2)=w_{\ensuremath{\text{ee}}}^{\ensuremath{\text{lr}}}(|{\ensuremath{\mathbf{r}}}_1-{\ensuremath{\mathbf{r}}}_2|)$ is the long-range Hartree kernel. For these one- and two-electron cases, TDRSH and TDRSH-EXX give rise to identical excitation energies and oscillator strengths.
Finally, we can calculate the corresponding TDRSH (or TDRSH-EXX) oscillator strengths as $$\begin{aligned}
\label{oscillator}
f_{n} = \frac{2}{3} \omega_{n} \sum_{\nu=x,y,z} \left| d_{\nu,ia} (X_{n,ia} + Y_{n,ia}) \right|^2.\end{aligned}$$ In the limit of a complete basis set, the linear-response oscillator strengths in Eq. (\[oscillator\]) always fulfill the Thomas-Reiche-Kuhn (TRK) sum rule, $\sum_n f_{n} = N$ where $N$ is the electron number. The bare oscillator strengths of Eq. (\[oscillator0\]) fulfill the TRK sum rule only in the case where the orbitals have been obtained from an effective local potential, i.e. for LDA and RSH-EXX but not for HF and RSH (see Ref. ).
Implementation in\
a B-spline basis set {#sec:implementation}
====================
In practice, each spin orbital is decomposed into a product of a spatial orbital and a spin function, $\varphi_{p}({\ensuremath{\mathbf{x}}})=\varphi_{p}({\ensuremath{\mathbf{r}}}) \delta_{\sigma_p,\sigma}$ where $\sigma_p$ is the spin of the spin orbital $p$, and we use spin-adapted equations. As we investigate atomic systems, the spatial orbitals are written in spherical coordinates, $$\varphi_{p}({\ensuremath{\mathbf{r}}})=R_{n_pl_p}(r)Y_{l_p}^{m_p}(\Omega),$$ where $Y_{l_p}^{m_p}(\Omega)$ are the spherical harmonics ($\Omega$ stands for the angles $\theta,\phi$) and the radial functions $R_{n_pl_p}(r)$ are expressed as linear combinations of B-spline functions of order $k_{\ensuremath{\text{s}}}$, $$\begin{aligned}
R_{n_p l_p}(r)=\sum_{\alpha=1}^{N_{\ensuremath{\text{s}}}}c_\alpha^{n_p l_p}\frac{B^{k_{\ensuremath{\text{s}}}}_\alpha(r)}{r},\end{aligned}$$ where $N_{\ensuremath{\text{s}}}$ is the dimension of the basis. To completely define a basis of B-spline functions, a non-decreasing sequence of $N_{\ensuremath{\text{s}}}+k_{\ensuremath{\text{s}}}$ knot points (some knot points are possibly coincident) must be given [@Boor-78]. The B-spline function $B^{k_{\ensuremath{\text{s}}}}_\alpha(r)$ is non zero only on the supporting interval $[r_\alpha,r_{\alpha+k_{\ensuremath{\text{s}}}}]$ (containing $k_{\ensuremath{\text{s}}}+1$ consecutive knot points) and is a piecewise function composed of polynomials of degree $k_{\ensuremath{\text{s}}}-1$ with continuous first $k_{\ensuremath{\text{s}}}-m$ derivatives across each knot of multiplicity $m$. We have chosen the first and the last knots to be $k_{\ensuremath{\text{s}}}$-fold degenerate, i.e. $r_1 = r_2 = \cdots = r_{k_{\ensuremath{\text{s}}}} = R_{\text{min}}$ and $r_{{N_{\ensuremath{\text{s}}}+1}} = r_{{N_{\ensuremath{\text{s}}}+2}} = \cdots = r_{{N_{\ensuremath{\text{s}}}+k_{\ensuremath{\text{s}}}}}= R_{\text{max}}$, while the multiplicity of the other knots is unity. The spatial grid spacing was chosen to be constant in the whole radial space between two consecutive non-coincident points and is given by $\Delta r = R_{\text{max}}/(N_{\ensuremath{\text{s}}}-k_{\ensuremath{\text{s}}}+1)$. In the present work, the first and the last B-spline functions were removed from the calculation to ensure zero boundary conditions at $r=R_{\text{min}}$ and $r=R_{\text{max}}$. The results presented in this paper have been obtained using the following parameters: $k_{\ensuremath{\text{s}}}=8$, $N_{\ensuremath{\text{s}}}=200$, $R_{\text{min}}=0$, and $R_{\text{max}} = 100$ bohr. Moreover, we need to use only s and p$_z$ spherical harmonics.
Working with such a B-spline representation, one must compute matrix elements involving integrals over B-spline functions. The principle of the calculation of one-electron and two-electron integrals over B-spline functions are well described by Bachau *et al.* in Ref. . We will now briefly review the computation of the standard Coulomb two-electron integrals over B-spline functions, and then we will present the calculation of the long-range or short-range two-electron integrals over B-spline functions, the latter being original to the present work.
Coulomb two-electron integrals
------------------------------
The Coulomb electron-electron interaction is given by $$\label{coulomb}
w_{\ensuremath{\text{ee}}}(|{\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}'|)=\frac{1}{\left(|{\ensuremath{\mathbf{r}}}|^2+|{\ensuremath{\mathbf{r}}}'|^2-2|{\ensuremath{\mathbf{r}}}||{\ensuremath{\mathbf{r}}}'|\cos\gamma \right)^{1/2}},$$ where ${\ensuremath{\mathbf{r}}}$ and ${\ensuremath{\mathbf{r}}}'$ are electron vector positions and $\gamma$ is the angle between them. The multipolar expansion for this interaction is $$\label{coulomb}
w_{\ensuremath{\text{ee}}}(|{\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}'|)=\sum_{k=0}^{\infty}\left[\frac{r_<^k}{r_>^{k+1}}\right]\sum_{m_k=-k}^{k}(-1)^{m_k} C_{-m_k}^k(\Omega)C_{m_k}^k(\Omega'),$$ where $r_<=\mathrm{min}(|{\ensuremath{\mathbf{r}}}|,|{\ensuremath{\mathbf{r}}}'|)$ and $r_>=\mathrm{max}(|{\ensuremath{\mathbf{r}}}|,|{\ensuremath{\mathbf{r}}}'|)$ and $C_{m_k}^k(\Omega)=\left(4\pi/(2k+1)\right)^{1/2}Y_k^{m_k}(\Omega)$ are the renormalized spherical harmonics. The Coulomb two-electron integrals, in the spatial orbital basis, can then be expressed as the sum of products of radial integrals and angular factors $$\begin{aligned}
\label{coulomb_integral}
\nonumber
\langle pq|w_{\ensuremath{\text{ee}}}|tu\rangle&=&\sum_{k=0}^{\infty}R^k(p, q; t, u)\sum_{m_k=-k}^{k}\delta_{m_k,m_p-m_t}\delta_{m_k,m_q-m_u}\\
&\times&(-1)^{m_k} c^k(l_p, m_p, l_t, m_t) c^k(l_q, m_q, l_u, m_u),\end{aligned}$$ where $R^k(p, q; t, u)$ are the two-dimensional radial Slater integrals and the angular coefficients $c^k(l_p, m_p, l_t, m_t)$ and $c^k(l_q, m_q, l_u, m_u)$ are obtained from the Gaunt coefficients [@RCowan-81; @Cer-THESIS-12]. The coefficient $c^k(l, m, l', m')$ is non zero only if $|l-l'|\leq k \leq l+l'$ and if $l+l'+k$ is an even integer, which makes the sum over $k$ in Eq. (\[coulomb\_integral\]) exactly terminate. The Slater integrals are defined as $$\begin{aligned}
\nonumber
R^k(p, q; t, u)&=&\sum_{\alpha=1}^{N_{\ensuremath{\text{s}}}}\sum_{\lambda=1}^{N_{\ensuremath{\text{s}}}}\sum_{\beta=1}^{N_{\ensuremath{\text{s}}}}\sum_{\nu=1}^{N_{\ensuremath{\text{s}}}}c_{\alpha}^{n_pl_p}c_{\lambda}^{n_ql_q}c_{\beta}^{n_tl_t}c_{\nu}^{n_ul_u}\\
& &\times R^k(\alpha, \lambda; \beta, \nu),\end{aligned}$$ where $R^k(\alpha, \lambda; \beta, \nu)$ are the Slater matrix elements given by $$\begin{aligned}
\label{slaterelement}
\nonumber
R^k(\alpha, \lambda; \beta, \nu)&=&\int_{0}^{\infty}\int_{0}^{\infty}B_{\alpha}^{k_{\ensuremath{\text{s}}}}(r)B_{\lambda}^{k_{\ensuremath{\text{s}}}}(r')\left[\frac{r_<^k}{r_>^{k+1}}\right]\\
& &\times B_{\beta}^{k_{\ensuremath{\text{s}}}}(r)B_{\nu}^{k_{\ensuremath{\text{s}}}}(r'){\ensuremath{\text{d}}}r {\ensuremath{\text{d}}}r'.\end{aligned}$$
In order to compute the Slater matrix elements $R^k(\alpha, \lambda; \beta, \nu)$, we have implemented the integration-cell algorithm developed by Qiu and Froese Fischer [@CFFischer-99]. This algorithm exploits all possible symmetries and B-spline properties to evaluate efficiently the integrals in each two-dimensional radial region on which the integrals are defined. Gaussian quadrature is used to compute the integrals in each cell.
Long-range and short-range two-electron integrals
-------------------------------------------------
A closed form of the multipolar expansion of the short-range electron-electron interaction defined in Eq. (\[erfcsr\]) was determined by Ángyán *et al.* [@Janos-06], following a previous work of Marshall [@Marshall-02] who applied the Gegenbauer addition theorem to the Laplace transform of Eq. (\[erfcsr\]). This exact expansion is $$\begin{aligned}
w_{{\ensuremath{\text{ee}}}}^{{\ensuremath{\text{sr}}}}(|{\ensuremath{\mathbf{r}}}-{\ensuremath{\mathbf{r}}}'|)&=&\sum_{k=0}^{\infty}S^k(r_>,r_<;\mu)
\nonumber\\
&&\times\sum_{m_k=-k}^{k}(-1)^{m_k} C_{-m_k}^k(\Omega)C_{m_k}^k(\Omega'), \;\;\;\end{aligned}$$ where the $\mu$-dependent radial function is written in terms of the scaled radial coordinates $\Xi=\mu \; r_>$ and $\xi=\mu\; r_<$ as $$\label{srkernel}
S^k(r_>,r_<;\mu)=\mu \; \Phi^k(\Xi,\xi),$$ with $$\begin{aligned}
\label{Phi}
\nonumber
\Phi^k(\Xi,\xi)&=&H^k(\Xi,\xi)+F^k(\Xi,\xi)\\
& &+\sum_{m=1}^k F^{k-m}(\Xi,\xi)\frac{\Xi^{2m}+\xi^{2m}}{(\xi\;\Xi)^m},\end{aligned}$$ and the introduced auxiliary functions $$\begin{aligned}
\nonumber
H^k(\Xi,\xi)&=&\frac{1}{2(\xi\;\Xi)^{k+1}}\left[\left(\Xi^{2k+1}+\xi^{2k+1}\right)\operatorname{erfc}(\Xi+\xi)\right.\\
& &\left.-\left(\Xi^{2k+1}-\xi^{2k+1}\right)\operatorname{erfc}(\Xi-\xi)\right], \end{aligned}$$ and $$\begin{aligned}
\nonumber
F^k(\Xi,\xi)&=&\frac{2}{\pi^{1/2}}\sum_{p=0}^k\left(-\frac{1}{4(\xi\;\Xi)}\right)^{p+1}\frac{(k+p)!}{p!(k-p)!}\\
& &\times\left[(-1)^{k-p} e^{-(\Xi+\xi)^2}-e^{-(\Xi-\xi)^2} \right].\end{aligned}$$
In order to arrive at a separable expression in $\Xi$ and $\xi$, Ángyán *et al.* [@Janos-06] also introduced a power series expansion of the radial function $\Phi^k(\Xi,\xi)$ in the smaller reduced variable $\xi$. However, the range of validity of this expansion truncated to the first few terms is limited to small values of $\xi$, i.e. $\xi \lesssim 1.5$, and higher-order expansions show spurious oscillations. After some tests, we decided to use the exact short-range radial function $\Phi^k(\Xi,\xi)$ without expansion in our work.
The expression of the short-range two-electron integrals $\langle pq|w^{{\ensuremath{\text{sr}}}}_{\ensuremath{\text{ee}}}|tu\rangle$ is then identical to the one in Eq. (\[coulomb\_integral\]) with the simple difference that the radial term is not given by the standard Slater matrix elements. Now, the radial kernel in Eq. (\[slaterelement\]) is changed to that of Eq. (\[srkernel\]). Due to the fact that the radial kernel is not multiplicatively separable in the variables $r_>$ and $r_<$, the integration-cell algorithm is modified in order to calculate all integrals as non-separable two-dimensional integrals. In a second step, the long-range two-electron integrals can be simply obtained by difference $$\langle pq|w^{{\ensuremath{\text{lr}}}}_{\ensuremath{\text{ee}}}|tu\rangle=\langle pq|w_{\ensuremath{\text{ee}}}|tu\rangle-\langle pq|w^{{\ensuremath{\text{sr}}}}_{\ensuremath{\text{ee}}}|tu\rangle.$$
Results and discussion {#sec:results}
======================
In this section, photoexcitation and photoionization spectra for the H and He atoms are presented. Photoexcitation and photoionization processes imply transitions from bound to bound and from bound to continuum states, respectively. For this reason, we first check the density of continuum states obtained with our B-spline basis set. After that, we show how orbital energies for the H and He atoms are influenced by the range-separation parameter $\mu$. Finally, having in mind these aspects, we discuss the different calculated spectra. All the studied transitions correspond to dipole-allowed spin-singlet transitions from the Lyman series, i.e. $1\text{s}\rightarrow n\text{p}$.
Density of continuum states {#sec:DOS}
---------------------------
In Fig. \[DOS\], the radial density of states (DOS) of a free particle in a spherical box is compared with the radial DOS of the continuum p orbitals of the H atom computed with the exact Hamiltonian or with the HF or LDA effective Hamiltonian using the B-spline basis set. The radial DOS of a free particle is given by [@BachCorDecHanMart-RepProgPhys-01] $\rho(\varepsilon)=R_\text{max}/\pi\sqrt{2\varepsilon}$ where $R_\text{max}$ is the radial size of the box, while for the different Hamiltonians using the B-spline basis set (with the same $R_\text{max}$) the radial DOS is calculated by finite differences as $\rho(\varepsilon_p)=2/(\varepsilon_{p+1} -\varepsilon_{p-1})$ where $\varepsilon_{p}$ are positive orbital energies.
As one can observe, the radial DOS computed with the LDA or the HF Hamiltonian is essentially identical to the DOS of the free particle. This can be explained by the fact that since the unoccupied LDA and HF orbitals do not see a $-1/r$ attractive potential they are all unbound and they all contribute to the continuum, similarly to the free-particle case. By contrast, for the exact Hamiltonian with the same B-spline basis set, one obtains a slightly smaller DOS in the low-energy region. This is due to the presence of the $-1/r$ attractive Coulomb potential which supports a series of bound Rydberg states, necessarily implying less unoccupied orbitals in the continuum for a given basis.
![Radial density of states (DOS) for a free particle, $\rho(\varepsilon_p)=R_\text{max}/\pi\sqrt{2\varepsilon_p}$, in a spherical box of size $R_\text{max} = 100$ bohr, and for the continuum p orbitals of the H atom computed with the exact Hamiltonian, or with the HF or LDA effective Hamiltonian using the B-spline basis set with the same $R_\text{max}$.[]{data-label="DOS"}](Fig1.pdf)
We have checked that, by increasing the size of the simulation box, together with the number of B-spline functions in the basis so as to keep constant the density of B-spline functions, the DOS of the exact Hamiltonian converges, albeit slowly, to the free-particle DOS. This must be the case since, for potentials vanishing at infinity, the global density of unbound states is independent of the potential for an infinite simulation box (only the local DOS depends on the potential, see e.g. Ref. ). From a numerical point of view, the computation of the DOS can be seen as a convergence test. With the present basis set, a huge energy range of the continuum spectrum is described correctly, and the difference between the DOS of the exact Hamiltonian and the free-particle DOS at low energies ($0.0 - 0.2$ Ha) is only about $10^{-4}$ Ha$^{-1}$. This difference is small enough to fairly compare the different methods considered in this paper.
The calculation of the DOS is also important in order to compute proper oscillator strengths involving continuum states. Because of the use of a finite simulation box, the calculated positive-energy orbitals form, of course, a discrete set and not strictly a continuum. These positive-energy orbitals are thus not energy normalized as the exact continuum states should be. To better approximate pointwise the exact continuum wave functions, the obtained positive-energy orbitals should be renormalized. Following Macías *et al.* [@Macias88], we renormalize the positive-energy orbitals by the square root of the DOS as $\tilde{\varphi}_p({\ensuremath{\mathbf{r}}})= \sqrt{\rho(\varepsilon_p)}\varphi_p({\ensuremath{\mathbf{r}}})$.
Range-separated orbital energies {#sec:orbital}
--------------------------------
 
 
In Fig. \[Horbital\] we show the 1s and the low-lying p orbital energies for the H atom calculated with both the RSH and RSH-EXX methods as a function of the range-separation parameter $\mu$.
As one observes in Fig. \[Horbital\]a, with the RSH method only the 1s ground state is bound, and the energy of this state is strongly dependent on $\mu$. At $\mu=0$, the self-interaction error introduced by the LDA exchange-correlation potential is maximal. But, when $\mu$ increases, the long-range HF exchange potential progressively replaces the long-range part of the LDA exchange-correlation potential and the self-interaction error is gradually eliminated until reaching the HF limit for $\mu\to\infty$, where one obtains the exact 1s orbital energy. The p orbitals (and all the other unoccupied orbitals) are always unbound and their (positive) energies are insensible to the value of $\mu$. One also observes that the approximate continuum of p orbitals has a DOS correctly decreasing as the energy increases, as previously seen in Fig. \[DOS\].
In Fig. \[Horbital\]b, one sees that the 1s orbital energy computed with the RSH-EXX method is identical to the 1s orbital energy obtained by the RSH scheme, as expected. However, a very different behavior is observed for the unoccupied p orbitals. Starting from the LDA limit at $\mu=0$ where all unoccupied orbitals are unbound, when the value of $\mu$ increases one sees the emergence of a series of bound Rydberg states coming down from the continuum. This is due to the introduction of an attractive $-1/r$ term in the long-range EXX potential, which supports a Rydberg series. For $\mu\to\infty$, we obtain the spectrum of the exact hydrogen Hamiltonian calculated with the B-spline basis set. Necessarily, with the finite basis used, the appearance of the discrete bound states is accompanied by a small reduction of the density of continuum states, as we already observed in Fig. \[DOS\] with the exact Hamiltonian.
Another interesting aspect that can be observed in Fig. \[Horbital\]b is the fact that the different bound-state energies reach their exact $\mu\to\infty$ values at different values of $\mu$. Thus, for a fixed small value of $\mu$, each bound-state energy is affected differently by the self-interaction error. For the compact 1s orbital, the self-interaction error is eliminated for $\mu \gtrsim 1$ bohr$^{-1}$. For the more diffuse 2p Rydberg state, the self-interaction error is essentially eliminated with $\mu \gtrsim 0.5$ bohr$^{-1}$. When we continue to climb in the Rydberg series, the orbitals become more and more diffuse and the self-interaction error is eliminated from smaller and smaller values of $\mu$.
In Fig. \[Heorbital\], the 1s and low-lying p orbital energies for the He atom are shown. Again, for the RSH method, one sees in Fig. \[Heorbital\]a that only the occupied 1s orbital is bound and all the unoccupied p orbitals are in the continuum. Similarly to the case of the H atom, at $\mu=0$ the 1s orbital energy is too high, which can essentially be attributed to the self-interaction error in the LDA exchange-correlation potential. This error decreases when $\mu$ increases and the 1s orbital energy converges to its HF value for $\mu\to\infty$. However, contrary to the case of the H atom, for this two-electron system, the 1s HF orbital energy is not equal to the opposite of the exact ionization energy but is slightly too low due to missing correlation effects. In the spirit of the optimally tuned range-separated hybrids [@LivBae-PCCP-07; @SteKroBae-JACS-09; @SteKroBae-JCP-09], the range-separation parameter $\mu$ can be chosen so that the HOMO orbital energy is equal to the opposite of the exact ionization energy, which gives $\mu=1.115$ bohr$^{-1}$ for the He atom.
As regards the RSH-EXX method, one sees again in Fig. \[Heorbital\]b that, for this two-electron system, the 1s RSH-EXX orbital energy is identical to the 1s RSH orbital energy. As in the case of the H atom, the introduction of the long-range EXX potential generates a series of bound Rydberg states, whose energies converge to the Kohn-Sham EXX orbital energies for $\mu\to\infty$. For the Rydberg states of the He atom, it turns out that the Kohn-Sham EXX orbital energies are practically identical to the exact Kohn-Sham orbital energies [@UmrSavGon-INC-98], implying that the Kohn-Sham correlation potential has essentially no effect on these Rydberg states. As we will see, contrary to the RSH case, the set of unoccupied RSH-EXX orbitals can be considered as a reasonably good first approximation for the computation of photoexcitation and photoionization spectra, even before applying linear-response theory.
Photoexcitation and photoionization\
spectra for the hydrogen atom
------------------------------------
![Photoexcitation/photoionization spectra calculated with different methods for the H atom. In [**(a)**]{} comparison of the HF, LDA, and TDLDA methods with respect to the calculation with the exact Hamiltonian. In [**(b)**]{} comparison of the RSH, RSH-EXX, and TDRSH methods (all of them with a range-separation parameter of $\mu=0.5$ bohr$^{-1}$) with respect to the calculation with the exact Hamiltonian.[]{data-label="Hspectra"}](Fig4a.pdf "fig:") ![Photoexcitation/photoionization spectra calculated with different methods for the H atom. In [**(a)**]{} comparison of the HF, LDA, and TDLDA methods with respect to the calculation with the exact Hamiltonian. In [**(b)**]{} comparison of the RSH, RSH-EXX, and TDRSH methods (all of them with a range-separation parameter of $\mu=0.5$ bohr$^{-1}$) with respect to the calculation with the exact Hamiltonian.[]{data-label="Hspectra"}](Fig4b.pdf "fig:")
![Comparison of the renormalized radial amplitude $\tilde{R}(r) = \sqrt{\rho(\varepsilon)} R(r)$ of the continuum p orbital involved in the transition energy $\omega_n=\varepsilon - \varepsilon_\text{1s}=0.8$ Ha calculated by HF, LDA, RSH, and RSH-EXX (with a range-separation parameter of $\mu=0.5$ bohr$^{-1}$) with respect to the exact calculation for the H atom.[]{data-label="p-orbital"}](Fig5.pdf)
In Fig. \[Hspectra\], photoexcitation/photoionization spectra for the H atom calculated with different methods are shown. For the calculation using the exact Hamiltonian, the spectrum is correctly divided into a discrete and a continuum part, corresponding to the photoexcitation and photoionization processes, respectively. As already discussed in Sec. \[sec:DOS\], for all calculations, the continuum states have been renormalized, or equivalently the oscillator strengths of the continuum part of the spectrum have been renormalized as $\tilde{f}_{1\text{s}\to n\text{p}} = \rho(\varepsilon_{n\text{p}}) f_{1\text{s}\to n\text{p}}$ where $\rho(\varepsilon_{n\text{p}})$ is the DOS at the corresponding positive orbital energy $\varepsilon_{n\text{p}}$. Moreover, for better readability of the spectra, following Refs. , we have also renormalized the oscillator strengths of the discrete part of the spectrum as $\tilde{f}_{1\text{s}\to n\text{p}} = n^3 f_{1\text{s}\to n\text{p}}$ where $n$ is the principal quantum number of the excited p orbital. This makes the transition between the discrete and the continuum part of the spectrum smooth. Another thing is, since we are working with a finite B-spline basis set principally targeting a good continuum, we obtain only a limited number of Rydberg states and the last Rydberg states near the ionization threshold are not accurately described. In particular, the corresponding oscillator strengths are overestimated (not shown). To fix this problem, we could for example use quantum defect theory in order to accurately extract the series of Rydberg states [@AlSharif98; @Friedrich98; @Faassen06; @Faassen09]. However, for the propose of the present work, we did not find necessary to do that, and instead we have simply corrected the oscillator strengths of the last Rydberg states by interpolating between the oscillator strengths of the first five Rydberg states and the oscillator strength of the first continuum state using a second-order polynomial function of the type $\tilde{f}_n=c_0+c_1\;\omega_n+c_2\;\omega_n^2$. This procedure was applied for all spectra having a discrete part.
Let us first discuss the spectra in Fig. \[Hspectra\]a. The LDA spectrum, calculated using the bare oscillator strengths of Eq. (\[oscillator0\]), does not possess a discrete photoexcitation part, which was of course expected since the LDA potential does not support bound Rydberg states, as seen in the $\mu=0$ limit of Fig. \[Horbital\]. The ionization threshold energy, giving the onset of the continuum spectrum, is much lower than the exact value (0.5 Ha) due to the self-interaction error in the ground-state orbital energy. At the ionization threshold, the LDA oscillator strengths are zero, in agreement with the Wigner-threshold law [@Wig-PR-48; @SadBohCavEsrFabMacRau-JPB-00] for potentials lacking a long-range attractive $-1/r$ Coulomb tail. Close above the ionization threshold, the LDA spectrum has an unphysical large peak, which corresponds to continuum states with an important local character. However, as noted in Ref. , at the exact Rydberg transition energies, the LDA continuum oscillator strengths are actually reasonably good approximations to the exact discrete oscillator strengths, which was explained by the fact that the LDA potential is approximately the exact Kohn-Sham potential shifted by a constant. Moreover, above the exact ionization energy, LDA reproduces relatively well the exact photoionization spectrum and becomes essentially asymptotically exact in the high-energy limit. This is consistent with the fact that, at a sufficiently high transition energy, the LDA continuum orbitals are very similar to the exact ones, at least in the spatial region relevant for the calculation of the oscillation strengths, as shown in Fig. \[p-orbital\].
The TDLDA spectrum differs notably from the LDA spectrum only in that the unphysical peak at around $0.3$ Ha, close above its ionization threshold, has an even larger intensity. This increased intensity comes from the contribution of the LDA exchange-correlation kernel (not shown). The LDA exchange-correlation kernel being local, its larger impact is for the low-lying LDA continuum orbitals having a local character. As the TRK sum rule must be satisfied, the higher peak in the TDLDA spectrum is followed by a decrease of the oscillator strengths faster than in the LDA spectrum, until they reach the same asymptotic behavior.
The HF spectrum in Fig. \[Hspectra\]a not only has no discrete photoexcitation part, as expected since the unoccupied HF orbitals are unbound (see the $\mu\to\infty$ limit of Fig. \[Horbital\]a), but does not even look as a photoionization spectrum. The HF unoccupied orbitals actually represent approximations to the continuum states of the H$^-$ anion, and are thus much more diffuse than the exact continuum states of the H atom, as shown in Fig. \[p-orbital\]. Consequently, the HF spectrum has in fact the characteristic shape of the photodetachment spectrum of the H$^-$ anion [@BetSal-BOOK-57; @Rau-JAA-96] (with the caveat that the initial state is the 1s orbital of the H atom instead of the 1s orbital of the H$^-$ anion). Finally, note that, for the H atom, linear-response TDHF gives of course the exact photoexcitation/photoionization spectrum.
Let us now discuss the spectra obtained with the range-separated methods in Fig. \[Hspectra\]b. The common value of the range-separation parameter $\mu=0.5$ bohr$^{-1}$ has been used [@GerAng-CPL-05a]. The RSH spectrum looks like the photodetachment spectrum of the H$^-$ anion. This is not surprising since the RSH effective Hamiltonian contains a long-range HF exchange potential. The RSH continuum orbitals are similarly diffuse as the HF continuum orbitals, as shown in Fig. \[p-orbital\]. The RSH ionization threshold energy is slightly smaller than the exact value (0.5 Ha) due to the remaining self-interaction error in the 1s orbital energy stemming from the short-range LDA exchange-correlation potential at this value of $\mu$. The RSH-EXX ionization threshold is identical to the RSH one, but, contrary to the RSH spectrum, the RSH-EXX spectrum correctly shows a discrete photoexcitation part and a continuum photoionization part. Beside the small redshift of the spectrum, the self-interaction error at this value of $\mu$ manifests itself in slightly too small RSH-EXX oscillator strengths. The RSH-EXX continuum orbitals are very similar to the exact continuum orbitals, as shown in Fig. \[p-orbital\]. Finally, at this value of $\mu$, TDRSH gives a photoexcitation/photoionization spectrum essentially identical to the RSH-EXX spectrum.
Photoexcitation and photoionization\
spectra for the helium atom
------------------------------------
![Photoexcitation and photoionization spectra calculated with different methods for the He atom. In [**(a)**]{} comparison of HF, TDHF, LDA, and TDLDA methods. In [**(b)**]{} comparison of RSH, RSH-EXX, and TDRSH methods (all of them with a range-separation parameter of $\mu=1.115$ bohr$^{-1}$).[]{data-label="Hespectra"}](Fig6a.pdf "fig:") ![Photoexcitation and photoionization spectra calculated with different methods for the He atom. In [**(a)**]{} comparison of HF, TDHF, LDA, and TDLDA methods. In [**(b)**]{} comparison of RSH, RSH-EXX, and TDRSH methods (all of them with a range-separation parameter of $\mu=1.115$ bohr$^{-1}$).[]{data-label="Hespectra"}](Fig6b.pdf "fig:")
![Photoionization cross-section profile for the He atom. Normalized cross sections are given (in Hartree atomic units) by $\sigma_{n}= (2\pi^2/c) \tilde{f}_{n}$ where $\tilde{f}_{n}$ are the renormalized oscillator strengths and $c$ is the speed of light. Conversion factors 1 Ha = 27.207696 eV and 1 bohr$^2=28.00283$ Mb are employed. The experimental data and the FCI results are from Ref. . []{data-label="exp"}](Fig7.pdf)
--------------------- ------------ -------- -- ------------ -------- -- ------------ -------- -- ------------ --------
Transition
$\omega_n$ $f_n$ $\omega_n$ $f_n$ $\omega_n$ $f_n$ $\omega_n$ $f_n$
1$^1$S $\to$ 2$^1$P 0.7799 0.2762 0.7970 0.2518 0.7766 0.3303 0.7827 0.2547
1$^1$S $\to$ 3$^1$P 0.8486 0.0734 0.8636 0.0704 0.8474 0.0857 0.8493 0.0708
1$^1$S $\to$ 4$^1$P 0.8727 0.0299 0.8872 0.0291 0.8721 0.0344 0.8729 0.0292
1$^1$S $\to$ 5$^1$P 0.8838 0.0150 0.8982 0.0148 0.8835 0.0172 0.8839 0.0148
1$^1$S $\to$ 6$^1$P 0.8899 0.0086 0.9042 0.0087 0.8897 0.0100 0.8899 0.0087
Ionization energy
$^a$From Ref. .
--------------------- ------------ -------- -- ------------ -------- -- ------------ -------- -- ------------ --------
\[tab:helium\]
In Fig. \[Hespectra\], different photoexcitation/photoionization spectra for the He atom are shown. As in the H atom case, the oscillator strengths of the discrete part of the TDHF, RSH-EXX, and TDRSH spectra have been interpolated (using again the oscillator strengths of first five Rydberg states and of the first continuum state) to correct the overestimation of the oscillator strengths for the last Rydberg transitions. The excitation energies and the (non-interpolated) oscillator strengths of the first five discrete transitions are reported in Table \[tab:helium\] and compared with exact results. The photoionization part of some of the calculated spectra are compared with full configuration-interaction (FCI) calculations and experimental results in Fig. \[exp\].
In Fig. \[Hespectra\]a, one sees that the HF spectrum looks again like a photodetachment spectrum, corresponding in this case to the He$^-$ anion. By contrast, TDHF gives a reasonable photoexcitation/photoionization spectrum. In particular, for the first discrete transitions listed in Table \[tab:helium\], TDHF gives slightly too large excitation energies by at most about 0.02 Ha (or 0.5 eV) and slightly too small oscillator strengths by at most about 0.025. The ionization energy is also slightly too large by about 0.015 Ha, as already seen from the HF 1s orbital energy in the $\mu\to\infty$ limit of Fig. \[Heorbital\]. As regards the photoionization part of the spectrum, one sees in Fig. \[exp\] that TDHF gives slightly too large photoionization cross sections.
The LDA spectrum in Fig. \[Hespectra\]a is also similar to the LDA spectrum for the H atom. The ionization threshold energy is much too low, and the spectrum lacks a discrete part and has an unphysical maximum close above the ionization threshold. Except from that, taking as reference the TDHF spectrum (which is close to the exact spectrum), the LDA spectrum is a reasonable approximation to the photoionization spectrum and, again as noted in Ref. , a reasonable continuous approximation to the photoexcitation spectrum. In comparison to LDA, TDLDA [@ZapLupTou-JJJ-XX-note] gives smaller and less accurate oscillator strengths in the lower-energy part of the spectrum but, the TRK sum rule having to be preserved, larger oscillator strengths in the higher-energy part of the spectrum, resulting in an accurate high-energy asymptotic behavior as seen in Fig. \[exp\].
Fig. \[Hespectra\]b shows the spectra calculated with RSH, RSH-EXX, and TDRSH using for the range-separation parameter the value $\mu=1.115$ bohr$^{-1}$ which imposes the exact ionization energy, as explained in Sec. \[sec:orbital\]. The RSH spectrum is similar to the HF spectrum and does not represent a photoexcitation/photoionization spectrum. By contrast, the RSH-EXX spectra is qualitatively correct for a photoexcitation/photoionization spectrum. As shown in Table \[tab:helium\], in comparison with TDHF, RSH-EXX gives more accurate Rydberg excitation energies, with a largest error of about 0.003 Ha (or 0.08 eV), but less accurate oscillator strengths which are significantly overestimated. The TDRSH method also gives a correct photoexcitation/photoionization spectrum, with the advantage that it gives Rydberg excitation energies as accurate as the RSH-EXX ones and corresponding oscillator strengths as accurate as the TDHF ones. As shown in Fig. \[exp\], TDRSH also gives a slightly more accurate photoionization cross-section profile than TDHF.
Conclusions {#sec:conclusions}
===========
We have investigated the performance of the RSH scheme for calculating photoexcitation/photoionization spectra of the H and He atoms, using a B-spline basis set in order to correctly describe the continuum part of the spectra. The study of these simple systems allowed us to quantify the influence on the spectra of the errors coming from the short-range exchange-correlation LDA and from the missing long-range correlation in the RSH scheme. For the He atom, it is possible to choose a value for the range-separation parameter $\mu$ for which these errors compensate each other so as to obtain the exact ionization energy.
We have studied the differences between using the long-range HF exchange nonlocal potential and the long-range EXX local potential. Contrary to the former, the latter supports a series of Rydberg states and the corresponding RSH-EXX scheme, even without applying linear-response theory, gives reasonable photoexcitation/photoionization spectra. Nevertheless, the most accurate spectra are obtained with linear-response TDRSH (or TDRSH-EXX since they are equivalent for one- and two-electron systems). In particular, for the He atom at the optimal value of $\mu$, TDRSH gives slightly more accurate photoexcitation and photoionization spectra than standard TDHF.
The present work calls for further developments. First, the merits of TDRSH (and/or TDRSH-EXX) for calculating photoexcitation/photoionization spectra of larger atoms and molecules, where screening effects are important, should now be investigated. Second, it would be interesting to test the effects of going beyond the LDA for the short-range exchange-correlation functional [@TouColSav-JCP-05; @GolWerStoLeiGorSav-CP-06] and adding long-range wave-function correlation [@FroKneJen-JCP-13; @HedHeiKneFroJen-JCP-13; @RebTou-JCP-16]. Third, time-propagation TDRSH could be implemented to go beyond linear response and tackle strong-field phenomena, such as high-harmonic generation and above-threshold ionization [@LabZapCocVenTouCaiTaiLup-JCTC-18].
[80]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ****, ().
, ****, ().
, in **, edited by (, , ), p. .
, , , ****, ().
, , , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , , , ****, ().
, , , ****, ().
, , , , , , , ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, , , , , , ****, ().
, , , , , , ****, ().
, in **, edited by (, , ), pp. .
, , , ****, ().
, , , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ** (, ).
, , , , , ****, ().
, **, Los Alamos Series in Basic and Applied Sciences (, , ).
, Ph.D. thesis, ().
, ****, ().
, , , ****, ().
, ****, ().
, ** (, , ).
, , , , ****, ().
, , , in **, edited by , , (, ), pp. .
, , , ****, ().
, , , ****, ().
, ** (, ).
, ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , , , , , ****, ().
, ** (, , ).
, ****, ().
, , , ****, ().
, ****, ().
, , , ****, ().
, , , , , , ****, ().
, , , , , ****, ().
, ****, ().
, , , , , , , , ****, ().
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We address the problem of anomaly detection in videos. The goal is to identify unusual behaviours automatically by learning exclusively from normal videos. Most existing approaches are usually data-hungry and have limited generalization abilities. They usually need to be trained on a large number of videos from a target scene to achieve good results in that scene. In this paper, we propose a novel few-shot scene-adaptive anomaly detection problem to address the limitations of previous approaches. Our goal is to learn to detect anomalies in a previously unseen scene with only a few frames. A reliable solution for this new problem will have huge potential in real-world applications since it is expensive to collect a massive amount of data for each target scene. We propose a meta-learning based approach for solving this new problem; extensive experimental results demonstrate the effectiveness of our proposed method. All codes are released in .'
author:
- Yiwei Lu
- Frank Yu
- Mahesh Kumar Krishna Reddy
- Yang Wang
bibliography:
- 'egbib.bib'
title: 'Few-Shot Scene-Adaptive Anomaly Detection'
---
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- Fedor Shmarov
- Nicola Paoletti
- Ezio Bartocci
- Shan Lin
- 'Scott A. Smolka'
- Paolo Zuliani
bibliography:
- 'memocode2017.bib'
title: Automated Synthesis of Safe and Robust PID Controllers for Stochastic Hybrid Systems
---
#### **Acknowledgements**
Research supported in part by EPSRC (UK) grant EP/N031962/1, FWF (Austria) S 11405-N23 (RiSE/SHiNE), AFOSR Grant FA9550-14-1-0261 and NSF Grants IIS-1447549, CNS-1446832, CNS-1445770, CNS-1445770, CNS-1553273, CNS-1536086, CNS 1463722, and IIS-1460370.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- Dan Hooper
bibliography:
- 'dyson.bib'
title: 'Life Versus Dark Energy: How An Advanced Civilization Could Resist the Accelerating Expansion of the Universe'
---
0.2in
Introduction
============
While it may be difficult to predict the detailed behavior of an advanced civilization, it is clear that the objectives of any such system would generically require, or at least benefit from, large quantities of useable energy. With this in mind, Freeman Dyson speculated in his 1960 paper [@1960Sci...131.1667D] that such civilizations would be likely to build structures around stars that are capable of collecting all or most of the light emitted, using this energy and then reemitting the waste heat in the form of high-entropy, infrared or sub-millimeter radiation [@1966ApJ...144.1216S]. Such “Dyson Spheres” are not only a staple of science fiction, but have also been the target of many astrophysical searches and other scientific investigations .
On timescales of tens of billions of years and longer, the expansion of the universe will ultimately limit the ability of an advanced civilization to accumulate and consume useable energy, a fact that has only been exacerbated by the discovery of dark energy [@Riess:1998cb; @Perlmutter:1998np]. As space expands, stars and other objects fall beyond the cosmic horizon, making it impossible for them to ever again be observed or otherwise interacted with.[^1] As dark energy comes to increasingly dominate the total energy density, our universe will enter a phase of exponential expansion, $a(t) \propto e^{H t}$, where $H=H_0 \, \Omega^{1/2}_{\Lambda, 0}$ is the asymptotic value of the Hubble constant in terms of the current Hubble constant, $H_0=67.8$ km/s/Mpc, and the abundance of dark energy, $\Omega_{\Lambda,0}=0.692$ [@Ade:2015xua]. Within approximately 100 billion years, all of the matter that is not gravitationally bound to the galaxies that make up our Local Group will become causally disconnected from the Milky Way, falling beyond the limits of our cosmic horizon [@Krauss:1999hj; @Loeb:2001dh; @Nagamine:2002wi; @Nagamine:2003ih].
In this paper, we speculate about how an advanced civilization would respond to the challenge of living in a universe that is dominated by dark energy. Here we have in mind a civilization that has reached Type III status on the Kardashev scale, which entails the ability to harness the energy produced by stars throughout an entire galaxy [@1964SvA.....8..217K]. Given the inevitability of the encroaching horizon, any sufficiently advanced civilization that is determined to maximize its ability to utilize energy will expand throughout the universe, attempting to secure as many stars as possible before they become permanently inaccessible. To this end, they could build Dyson Spheres or other such structures around the stars that are encountered, and use the energy that is collected to propel those stars toward the center of the civilization, where they will become gravitationally bound and thus protected from the future expansion of space. Broadly speaking, the validity of this conclusion relies only on two modest assumptions, namely that 1) a highly advanced civilization will attempt to maximize its access to usable energy, and that 2) our current understanding of dark energy and its impact on the future expansion history of our universe is approximately correct.
In the following, we will calculate which stars could be effectively harvested in this way. We find that very high-mass stars will often evolve beyond the main sequence before reaching their destination of the central civilization, while very low-mass stars will oftentimes generate too little energy (and thus provide too little acceleration) to avoid falling beyond the horizon. For these reasons, stars with masses in the approximate range of $M\sim (0.2-1) M_{\odot}$ will be the most attractive targets of such an effort. A civilization that begins to expand in the current epoch, traveling at a maximum speed of 10% (1%) of the speed of light, could harvest stars in this mass range out to a co-moving radius of approximately 50 Mpc (20 Mpc). Unlike more conventional Dyson Spheres, these structures would not necessarily emit in the infrared or sub-millimeter bands, but would instead use the collected energy to propel the captured stars, providing new and potentially distinctive signatures of an advanced civilization in this stage of expansion and stellar collection.
Gathering Stars from Throughout the Local Universe
==================================================
The expansion rate of our universe is described by the Friedmann equations. For the case of a spatially flat universe that is dominated by matter and dark energy (with an equation of state of $w=-1$), the first of these equations can be written as follows: $$\bigg(\frac{\dot{a}}{a}\bigg)^2 = \frac{8\pi G}{3} \bigg[\frac{\Omega_{M,0}}{a^3}+\Omega_{\Lambda,0}\bigg],
\label{friedmann}$$ where $a$ is the scale factor, $\dot{a}$ is its time derivative, $G$ is the Newtonian gravitational constant, and the current abundances of matter and dark energy are given by $\Omega_{M,0}=0.308$ and $\Omega_{\Lambda,0}=0.692$, respectively [@Ade:2015xua].
Expanding outward at a speed of $v_{\rm exp}$, a civilization could traverse the following co-moving distance as a function of time: $$d_{\rm CM}(t) = \int^t_0 \frac{v_{\rm exp} \,dt'}{a(t')}.$$ When a star is reached, a Dyson Sphere could be constructed and used to accelerate the surrounded star. If a fraction, $\eta$, of this energy is somehow[^2] transferred into the kinetic energy of the star, its speed after a time, $\Delta t$, would be given by: $$\begin{aligned}
\label{velocity}
v &=& \bigg(\frac{2 L \Delta t}{M}\bigg)^{1/2} \\
&\approx& 0.012 \,c \, \bigg(\frac{\Delta t}{1 \, {\rm Gyr}}\bigg)^{1/2}\, \bigg(\frac{L}{L_{\odot}}\bigg)^{1/2} \, \bigg(\frac{M_{\odot}}{M}\bigg)^{1/2} \, \bigg(\frac{\eta}{1}\bigg)^{1/2}, \nonumber\end{aligned}$$ where $c$ is the speed of light and $L_{\odot}$ and $M_{\odot}$ are the luminosity and mass of the Sun, respectively. Although it would be relatively straightforward to accelerate a star by using a series of reflective surfaces that direct the star’s light in one direction ([i.e. ]{}a Shkadov thruster [@1987brig.iafcR....S]), this would yield a velocity no greater than $v = L \Delta t/Mc \approx 7\times 10^{-5} \,c \, (\Delta t/1 \, {\rm Gyr})(L/L_{\odot}) \, (M_{\odot}/M)$, which is much lower than that yielded by the above expression (for $\eta=1$). By using part of the star’s mass (or the mass of another star) as a propellant, however, such high velocities could potentially be attained, while respecting both energy and momentum conservation. In our calculations that follow, we will assume that the advanced civilization in question is able to accelerate stars to the velocities described by Eq. \[velocity\].
In this study, we will focus on stars that are evolving on the main sequence, for which the luminosity is related to the star’s mass as follows [@2004sipp.book.....H]: $$\begin{aligned}
\frac{L}{L_{\odot}} &\approx& 0.028 \, \bigg(\frac{M}{0.4 \,M_{\odot}}\bigg)^{2.3}, \, \, \, \,M < 0.43 M_{\odot} \\
\frac{L}{L_{\odot}} &\approx& \bigg(\frac{M}{M_{\odot}}\bigg)^4, \, \, \,\,\,\,\,\,\, 0.43 M_{\odot} < M < 2 M_{\odot} \nonumber \\
\frac{L}{L_{\odot}} &\approx& 16 \, \bigg(\frac{M}{2 \, M_{\odot}}\bigg)^{3.5}, \, \, \, \,\, M > 2 M_{\odot}. \nonumber\end{aligned}$$ We also assume that each star remains on the main sequence for a length of time that is given by [@2004sipp.book.....H]: $$\tau_{_{\rm MS}} \approx 10 \, {\rm Gyr} \times (M_{\odot}/M)^{2.5}.
\label{lifetime}$$
Combining the above equations, we arrive at the following: $$\begin{aligned}
v &\approx& 0.0032 \,c \, \bigg(\frac{\Delta t}{1 \, {\rm Gyr}}\bigg)^{1/2} \, \bigg(\frac{M}{0.4 \,M_{\odot}}\bigg)^{0.65} \, \bigg(\frac{\eta}{1}\bigg)^{1/2}, \, \, \, \,M < 0.43 M_{\odot}\\
v &\approx& 0.012 \,c \, \bigg(\frac{\Delta t}{1 \, {\rm Gyr}}\bigg)^{1/2}\, \bigg(\frac{M}{M_{\odot}}\bigg)^{1.5} \,\bigg(\frac{\eta}{1}\bigg)^{1/2}, \, \, \,\,\,\,\,\,\, 0.43 M_{\odot} < M < 2 M_{\odot} \nonumber \\
v &\approx& 0.034 \,c \, \bigg(\frac{\Delta t}{1 \, {\rm Gyr}}\bigg)^{1/2} \, \bigg(\frac{M}{2 \, M_{\odot}}\bigg)^{1.25} \, \bigg(\frac{\eta}{1}\bigg)^{1/2}, \, \, \, \,\, M > 2 M_{\odot}. \nonumber\end{aligned}$$
![A cartoon summarizing the prospects for an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch ($t=0$). Blue, yellow and red symbols represent stars with masses of 2, 1 and 0.2$M_{\odot}$, respectively. The colored lines denote the (co-moving) distances those stars have travelled after 1 Gyr, 10 Gyr or 100 Gyr, adopting a maximum speed of 10% of the speed of light and assuming that approximately 100% of the collected energy is converted into kinetic energy of the star ($\eta=1$). The results shown apply to the case in which each star is encountered as it begins its main sequence evolution. The thick black line in the $t=100$ Gyr frame represents the horizon at that time in cosmic history. Very distant stars with either very low or high masses will not be collected, as they will either fall beyond the cosmic horizon or evolve beyond the main sequence before reaching their destination.[]{data-label="cartoon"}](cartoon-1Gyr.png "fig:"){width="3.15in"} ![A cartoon summarizing the prospects for an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch ($t=0$). Blue, yellow and red symbols represent stars with masses of 2, 1 and 0.2$M_{\odot}$, respectively. The colored lines denote the (co-moving) distances those stars have travelled after 1 Gyr, 10 Gyr or 100 Gyr, adopting a maximum speed of 10% of the speed of light and assuming that approximately 100% of the collected energy is converted into kinetic energy of the star ($\eta=1$). The results shown apply to the case in which each star is encountered as it begins its main sequence evolution. The thick black line in the $t=100$ Gyr frame represents the horizon at that time in cosmic history. Very distant stars with either very low or high masses will not be collected, as they will either fall beyond the cosmic horizon or evolve beyond the main sequence before reaching their destination.[]{data-label="cartoon"}](cartoon-10Gyr.png "fig:"){width="3.15in"} ![A cartoon summarizing the prospects for an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch ($t=0$). Blue, yellow and red symbols represent stars with masses of 2, 1 and 0.2$M_{\odot}$, respectively. The colored lines denote the (co-moving) distances those stars have travelled after 1 Gyr, 10 Gyr or 100 Gyr, adopting a maximum speed of 10% of the speed of light and assuming that approximately 100% of the collected energy is converted into kinetic energy of the star ($\eta=1$). The results shown apply to the case in which each star is encountered as it begins its main sequence evolution. The thick black line in the $t=100$ Gyr frame represents the horizon at that time in cosmic history. Very distant stars with either very low or high masses will not be collected, as they will either fall beyond the cosmic horizon or evolve beyond the main sequence before reaching their destination.[]{data-label="cartoon"}](cartoon-100Gyr.png "fig:"){width="3.15in"}
![A summary of the prospects for an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch. Stars in the red (upper left) regions will ultimately fall beyond the cosmic horizon, while those in the blue (right) regions will evolve beyond the main sequence before reaching their destination, and thus not provide useful energy. The grey dashed lines denote the length of time that is required to reach and transport the star. We show results for transport that is limited to speeds below 10%, 1% or 0.1% of the speed of light, and assume that the Dyson Spheres transfer approximately 100% of the collected energy to the kinetic energy of the star ($\eta=1$). The blue region in each frame has been calculated for the optimistic case of stars that are starting their main sequence evolution at the time that they are encountered (see Fig. \[age\]).[]{data-label="us"}](us-pt1.png "fig:"){width="3.08in"} ![A summary of the prospects for an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch. Stars in the red (upper left) regions will ultimately fall beyond the cosmic horizon, while those in the blue (right) regions will evolve beyond the main sequence before reaching their destination, and thus not provide useful energy. The grey dashed lines denote the length of time that is required to reach and transport the star. We show results for transport that is limited to speeds below 10%, 1% or 0.1% of the speed of light, and assume that the Dyson Spheres transfer approximately 100% of the collected energy to the kinetic energy of the star ($\eta=1$). The blue region in each frame has been calculated for the optimistic case of stars that are starting their main sequence evolution at the time that they are encountered (see Fig. \[age\]).[]{data-label="us"}](us-pt01.png "fig:"){width="3.08in"} ![A summary of the prospects for an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch. Stars in the red (upper left) regions will ultimately fall beyond the cosmic horizon, while those in the blue (right) regions will evolve beyond the main sequence before reaching their destination, and thus not provide useful energy. The grey dashed lines denote the length of time that is required to reach and transport the star. We show results for transport that is limited to speeds below 10%, 1% or 0.1% of the speed of light, and assume that the Dyson Spheres transfer approximately 100% of the collected energy to the kinetic energy of the star ($\eta=1$). The blue region in each frame has been calculated for the optimistic case of stars that are starting their main sequence evolution at the time that they are encountered (see Fig. \[age\]).[]{data-label="us"}](us-pt001.png "fig:"){width="3.08in"}
![As in Fig. \[us\], but for stars that are encountered after having evolved for 1, 3 or 10 billion years on the main sequence. This has little impact on low-mass stars which evolve very slowly, but limits the utility of more massive stars.[]{data-label="age"}](us-pt1-age0.png "fig:"){width="3.08in"} ![As in Fig. \[us\], but for stars that are encountered after having evolved for 1, 3 or 10 billion years on the main sequence. This has little impact on low-mass stars which evolve very slowly, but limits the utility of more massive stars.[]{data-label="age"}](us-pt1-age1.png "fig:"){width="3.08in"} ![As in Fig. \[us\], but for stars that are encountered after having evolved for 1, 3 or 10 billion years on the main sequence. This has little impact on low-mass stars which evolve very slowly, but limits the utility of more massive stars.[]{data-label="age"}](us-pt1-age3.png "fig:"){width="3.08in"} ![As in Fig. \[us\], but for stars that are encountered after having evolved for 1, 3 or 10 billion years on the main sequence. This has little impact on low-mass stars which evolve very slowly, but limits the utility of more massive stars.[]{data-label="age"}](us-pt1-age10.png "fig:"){width="3.08in"}
The ability to transport useable stars to a central civilization will be limited by two factors: 1) the speed of the encroaching horizon, and 2) the lifetime of stars on the main sequence. If an encountered star is either of very low mass (and thus slow to accelerate), very high mass (and thus short lived), and/or very far away, it may be impractical to construct a Dyson Sphere with the goal of transporting it to the central civilization. In Figs. \[cartoon\] and \[us\], we summarize the ability of an advanced civilization to transport usable stars to a central location, assuming that such efforts begin in the present epoch. Stars in the red (upper left) region of each frame in Fig. \[us\] do not produce enough luminosity to be accelerated to speeds that are fast enough to avoid falling beyond the cosmic horizon. In other words, they will not reach the central civilization within a finite time, and thus will not be the target of such efforts. On the other hand, the stars in the blue (right) regions of this figure will evolve beyond the main sequence prior to arrival at the central location, and thus will not be able to provide useful energy to the civilization in question. More specifically, such stars will reach the central civilization only after a transit time greater than that given in Eq. \[lifetime\]. Results are shown in Fig. \[us\] for transport that is limited to speeds below 10%, 1% or 0.1% of the speed of light, and in each case we assume that the Dyson Spheres transfer approximately 100% of the collected energy to the kinetic energy of their captured star ($\eta=1$). In each frame, the grey dashed lines denote the length of time that is required to reach and transport the star. For the case of $v_{\rm max} = 0.1c$, a civilization could ultimately collect and put to use stars that are currently as far away as 65 Mpc.
In calculating the results shown in Figs. \[cartoon\] and \[us\], we have optimistically assumed that each star is at the beginning of its main sequence evolution at the time that it is encountered. More realistically, such stars will span a range of ages and stages of their evolution. Such evolution has little impact on low-mass stars which evolve very slowly, but can be quite important for more massive stars. In Fig. \[age\], we show how our results change for the case of stars that are encountered after having evolved for 1, 3 or 10 billion years on the main sequence.
Integrating our results over the initial mass function of Ref. [@Kroupa:2002ky] and the cosmic star formation rate of Ref. [@Madau:2014bja], we estimate that an advanced civilization (with $v_{\rm max}=0.1 c$ and $\eta=1$) could increase the total stellar luminosity bound to the Local Group at a point in time 30 billion years in the future by a factor of several thousand relative to that which would have otherwise been available.[^3] Over a period of roughly a trillion years, the total luminosity of these stars will drop substantially, but will continue to produce substantial quantities of useable energy due to the longevity of the lightest main sequence stars.
New Observational Signatures of Advanced Civilizations
======================================================
![As in Fig. \[us\], but for the case of an advanced civilization that began to expand and collect stars 1 Gyr (left) or 5 Gyr (right) in the past. The orange region (only present in the left frame) denotes those stars which have not yet been reached in the present epoch, while those stars in the yellow regions would presently be en route toward the center of the civilization. The stars in the white regions will have already arrived at the central civilization and be providing it with useable energy.[]{data-label="minus"}](minus1-pt1.png "fig:"){width="3.08in"} ![As in Fig. \[us\], but for the case of an advanced civilization that began to expand and collect stars 1 Gyr (left) or 5 Gyr (right) in the past. The orange region (only present in the left frame) denotes those stars which have not yet been reached in the present epoch, while those stars in the yellow regions would presently be en route toward the center of the civilization. The stars in the white regions will have already arrived at the central civilization and be providing it with useable energy.[]{data-label="minus"}](minus5-pt1.png "fig:"){width="3.08in"}
In the previous section, we performed our calculations for the case of an advanced civilization that expands outward from the Milky Way (or Local Group) starting in the current epoch. It is of course possible, however, that life has already evolved elsewhere in our universe, and that civilizations far more advanced than our own may already exist within our Hubble volume. If this is the case, then they may have already begun to collect stars from their surrounding cosmological environment, altering the distribution of stars and leading to potentially observable signatures.
In Fig. \[minus\], we show results for a civilization that began its efforts to collect stars either 1 or 5 billion years ago, in each case adopting a maximum velocity of $v_{\rm max}=0.1 c$ and an efficiency of $\eta=1$. In the 1 Gyr case, the bulk of the stars within a radius of a few Mpc and with masses in the range of $M \sim (1-4) M_{\odot}$ will have already been collected, while most of the lighter stars out to a radius of $\sim$30 Mpc will presently be in transit on their way to the central civilization. For the 5 Gyr case, stars over a wide range of masses will have already been collected from throughout the surrounding several or even several tens of Mpc. From our vantage point, such a civilization would appear as a extended region, tens of Mpc in radius, with few or no perceivable stars lighter than approximately $\sim$$2 M_{\odot}$ (as such stars will be surrounded by Dyson Spheres). Furthermore, unlike traditional Dyson Spheres, those stars that are currently en route to the central civilization could be visible as a result of the propulsion that they are currently undergoing. The propellant could plausibly take a wide range of forms, and we do not speculate here about its spectral or other signatures. That being said, such acceleration would necessarily require large amounts of energy and likely produce significant fluxes of electromagnetic radiation.
Many of the past searches for Dyson Spheres have focused on detecting the presence of structures around individual stars within the Milky Way (for example, Refs. [@2009ASPC..420..415C; @2009ApJ...698.2075C; @2018arXiv180408351Z]). Here, we are instead considering galaxies and groups of galaxies in which many or most of the stars are surrounded by Dyson Spheres (and may have been removed from the galaxy), leading to very different observational strategies and signatures [@2015ApJ...810...23Z; @2015ApJS..217...25G; @2014ApJ...792...27W]. The spectrum of starlight from a galaxy that has had its useful ($M \lsim M_{\odot}$) stars harvested by an advanced civilization would be dominated by massive stars and thus peak at longer wavelengths than otherwise would have been the case. Although such measurements are very challenging, detailed spectroscopy has successfully been used to infer the approximate stellar mass function of nearby galaxies [@2012ApJ...760...71C].
Summary and Conclusions
=======================
In this paper, we have considered the likely response of a highly advanced civilization to the accelerating expansion of space caused by the presence of dark energy in our universe. If no action is taken, all stars that are not gravitationally bound to the Local Group will move beyond the cosmic horizon and become inaccessible on a timescale of approximately 100 billion years, permanently limiting the quantity of energy that could ultimately be used by such a civilization. In this paper, we argue that in order to maximize its ability to acquire useable energy, a sufficiently advanced civilization will expand rapidly outward, build Dyson Spheres or other similar structures around the stars as they are encountered, and use the collected energy to accelerate the stars away from the encroaching horizon and toward the central civilization. The most attractive targets of such a program will be those stars with masses in the approximate range of $M\sim (0.2-1) M_{\odot}$, as more massive stars are generally too short lived while lighter stars do not produce enough energy to accelerate fast enough to avoid falling beyond the horizon. For a civilization that embarks upon this task in the current epoch, stars in this mass range could be harvested out to a co-moving radius of several tens of Mpc, potentially increasing the total amount of energy that is available in the from of starlight in the distant future by a factor of several thousand.
If an advanced civilization has already embarked upon a program such as this elsewhere in the universe, it would be expected to provide a number of potentially observable signatures. Such a civilization could appear as a region up to tens of Mpc in radius in which most or all of the stars lighter than $\sim 2 M_{\odot}$ are surrounded by Dyson Spheres. Furthermore, during their acceleration and transit toward the central civilization, such stars could be accompanied by visible signals resulting from the large quantities of energy involved in their propulsion.
**Acknowledgments.** We would like to thank Sam McDermott, Jim Annis and Gordan Krnjaic for very helpful discussions. DH is supported by the US Department of Energy under contract DE-FG02-13ER41958. Fermilab is operated by Fermi Research Alliance, LLC, under contract DE- AC02-07CH11359 with the US Department of Energy.
[^1]: Strictly speaking, objects do not cross the horizon, but are increasingly redshifted as they approach this boundary. In any case, such objects become invisible and unreachable as a result of the expansion of space.
[^2]: We leave it to the advanced civilization to figure out how exactly this would be accomplished.
[^3]: In performing this estimate, we have adopted a density of stars within 1 Mpc of the Local Group that is a factor of 20 greater than the cosmological average [@Gonzalez:2013pqa].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Antonio B. Mei'
- Isaiah Gray
- Yongjian Tang
- Jüürgen Schubert
- Don Werder
- Jason Bartell
- 'Daniel C. Ralph'
- 'Gregory D. Fuchs'
- 'Darrell G. Schlom'
title: |
Local Photothermal Control of Phase Transitions for\
On-demand Room-temperature Rewritable Magnetic Patterning
---
**The ability to make controlled patterns of magnetic structures within a nonmagnetic background is essential for several types of existing and proposed technologies. Such patterns provide the foundation of magnetic memory and logic devices[@Imre:2006is], allow the creation of artificial spin-ice lattices[@Wang:2006hea; @Louis:2018bh] and enable the study of magnon propagation[@Chumak:2015fa]. Here, we report a novel approach for magnetic patterning that allows repeated creation and erasure of arbitrary shapes of thin-film ferromagnetic structures. This strategy is enabled by epitaxial Fe~0.52~Rh~0.48~ thin films designed so that both ferromagnetic and antiferromagnetic phases are bistable at room temperature. Starting with the film in a uniform antiferromagnetic state, we demonstrate the ability to write arbitrary patterns of the ferromagnetic phase by local heating with a focused laser. If desired, the results can then be erased by cooling with a thermoelectric cooler and the material repeatedly re-patterned.**
Intermetallic Fe~1-$x$~Rh~$x$~ (B2, $Pm\bar{3}m$) exhibits a hysteretic antiferromagnetic/ferromagnetic transformation which has been harnessed to produce composite multiferroics exhibiting record-breaking magnetoelectric coupling coefficients[@Cherifi:2014du], magnetocaloric refrigerators with competitive cooling capabilities[@Liu:2016kl] and novel memories based on antiferromagnetic order[@Marti:2014fl]. In this study, we design epitaxial Fe~1-$x$~Rh~$x$~ films so that both ferromagnetic and antiferromagnetic states are simultaneously bistable at room temperature. We engineer the width of the thermal hysteresis to be sufficiently narrow to enable efficient controllability, but also wide enough to robustly withstand thermal perturbations. Moderate heating by a focused laser is used to locally drive antiferromagnetic regions controllably to the ferromagnetic phase, demonstrating the patterning of arbitrary magnetic features on the submicron scale. These findings present opportunities for writing and erasing high-fidelity magnetically active nanostructures that are of interest for magnonic crystals[@Vogel:2015ky], artificial spin-ice lattices[@Nisoli:2013dd] and memory[@Weller:2014fh] and logic devices[@Cowburn:2000to].
![\[fig:sample\] **Fully-dense phase-pure untwinned epitaxial B2 Fe~0.52~Rh~0.48~/MgO(001) layers grown via molecular-beam epitaxy.** **a**, RBS spectrum; labeled iron and rhodium spectral features correspond to a magnetically bistable rhodium fraction of 0.48. **b**, BF-TEM image of the entire film cross section together with **c**, a corresponding SAED pattern. Note the weaker film and more intense substrate reflections. **d**, XRD $\theta$-$2\theta$ scan; film (violet) and substrate (orange) reflections are indexed. **e**, XRD $\omega$-rocking scan about the 001 film peak. ](Figures_01.pdf){width="47.00000%"}
Fe~1-$x$~Rh~$x$~ films are grown epitaxially on single-crystalline (001)-oriented MgO substrates using molecular-beam epitaxy. The fraction $x$ of rhodium in the film is carefully tuned to 0.48, where the hysteretic antiferromagnet/ferromagnet phase transitions are centered near room temperature. Film compositions are confirmed via Rutherford backscattering spectrometry (RBS) measurements (Fig. \[fig:sample\]**a**) using the areal ratio of corresponding iron and rhodium spectral features. Cross-sectional bright-field transmission electron microscopy (BF-TEM) images (Fig. \[fig:sample\]**b**), selected area electron diffraction (SAED) patterns (Fig. \[fig:sample\]**c**) and x-ray diffraction (XRD) $\theta$-$2\theta$ scans (Fig. \[fig:sample\]**d**) demonstrate that the films are fully-dense phase-pure untwinned epitaxial Fe~0.52~Rh~0.48~ layers with the B2 CsCl-structure and that the film lattice is rotated 45$^\circ$ in-plane with respect to the underlying B1 NaCl-structure MgO(001) substrate crystal: (001)~Fe~0.52~Rh~0.48~~ $\mid\mid$ (001)~MgO~ and \[110\]~Fe~0.52~Rh~0.48~~ $\mid\mid$ \[100\]~MgO~. The structural perfection of the films is established from the width of the $\omega$-rocking curve scans (Fig. \[fig:sample\]**e**) to be consistent with our previous report of high-quality epitaxial Fe~1-$x$~Rh~$x$~ layers[@Mei:2018dj].
![\[fig:MH\] **Designed room-temperature magnetic bistability of epitaxial Fe~0.52~Rh~0.48~/MgO(001) layers.** In-plane magnetisation $\vec{M}$ as a function of $\vec{H}$ field (clockwise) and temperature $T$ (radial) during **a** cooling and **b** heating. Positive field points along Fe~0.52~Rh~0.48~ $[110]$. Red and blue regions correspond to opposite orientations of the film magnetisation. **c**, Temperature-dependent remanent magnetisation $M_r(T)$ and **d** resistivity $\rho(T)$ during cooling (dotted) and heating (solid). ](Figures_02.pdf){width="47.00000%"}
The room-temperature magnetic bistability of our epitaxial Fe~0.52~Rh~0.48~/MgO(001) layers is established through the combination of in-plane vibrating sample magnetometry (VSM) and electrical transport measurements. Magnetisation measurements collected as a function of temperature $T$ and applied magnetic field $\vec{H}$ during cooling and heating are presented in Fig. \[fig:MH\]**a**,**b**; $\vec{H}$ is applied along Fe~0.52~Rh~0.48~ $[110]$. At room temperature, the as-deposited films exhibit a field-dependent magnetic hysteresis characteristic of FM order (Fig. \[fig:MH\]**a**) with measured saturation magnetisations $M_s \approx 4 \mu_B/$f.u. and coercive fields $H_c \approx 50$ Oe, consistent with prior reports[@Mei:2018dj].
Cooling from 300 to 275 K suppresses the hysteresis associated with the FM state (Fig. \[fig:MH\]**a**). In parallel, we observe a four-fold reduction in remanent magnetisation (Fig. \[fig:MH\]**c**) from $M_r = $ 4 to 1 $\mu_B/$f.u. and an approximate 50% increase in film resistivity (Fig. \[fig:MH\]**d**) from $\rho = $ 50 to 72 $\mu\Omega$-cm. These features are consistent with a magnetic transition, in which initially ferromagnetic Fe~0.52~Rh~0.48~ layers adopt an antiferromagnetic configuration characterised by anti-aligned neighbouring iron moments[@Shirane:1963bp], decreased carrier densities[@Mankovsky:2017dta] and increased scattering rates[@Kudrnovsky:2015bz]. With decreasing $T$ below 250 K, remanent magnetisations decrease and saturate at $\approx 0.2 \mu_B/$f.u. (Fig. \[fig:MH\]**c**); nonvanishing $M_r$ values are a common feature in Fe~1-$x$~Rh~$x$~ films and have been attributed to remanent ferromagnetism near the film/substrate interface[@Fan:2010gu].
On heating to 350 K, the field-dependent magnetic hysteresis (Fig. \[fig:MH\]**b**) and large remanent magnetisation values (Fig. \[fig:MH\]**c**) defining the FM state are restored. The dissimilar critical temperatures during heating ($\approx$ 350 K) and cooling ($\approx$ 275 K) are a hallmark of a first-order phase transition and result in a thermal hysteresis and a window of bistability over which both FM and AF states can simultaneously coexist. As a result, not only can the magnetic order of our epitaxial Fe~0.52~Rh~0.48~ layer be toggled between the AF and FM states using a commercial single-stage thermoelectric device, but the magnetic phase so defined will persist robustly at room temperature.
As the temperature is further raised from 350 to 400 K, the coercive field of the re-established FM state decreases from $H_c \sim$ 200 to 50 Oe. We attribute the enhanced coercivity at 350 K to exchange-coupling between recently-formed FM regions and the AF bulk, which decreases as the FM domains coalesce and the AF regions shrink at higher temperature. Over the same temperature range, resistivity values gradually decrease as a larger fraction of the sample transitions to the more conductive state associated with FM order (Fig. \[fig:MH\]**d**). The gradual descent in film resistivity upon heating contrasts sharply with the abrupt jump observed upon cooling and reflects different kinetics in the heating and cooling branches of the transition, analogous to the asymmetry between melting and freezing in a liquid/solid phase transition[@deVries:2014hq].
Collectively, the in-plane magnetometry and transport measurements establish that we have successfully tuned exchange interactions in epitaxial Fe~0.52~Rh~0.48~/MgO(001) films such that, under ambient conditions, the layers exhibit bistable magnetic order. Switching between AF and FM states can be achieved by heating and cooling over a practical temperature range, accessible to Peltier devices. Next, we leverage the designed magnetic bistability of our films to demonstrate magnetic patterning through the local photothermal control of exchange interactions.
{width="97.00000%"}
To interrogate the local magnetic order, we employ a microscope setup[@Bartell:2015hp] based on the anomalous Nernst effect (ANE), the thermal analogue of the anomalous Hall effect. A magnetic conductor with magnetisation $\vec{M}(\vec{r})$ subjected to a thermal gradient $\vec{\nabla} T(\vec{r})$ produces an electric field $\vec{E}_\textrm{ANE}(\vec{r})$ given by $$\vec{E}_\textrm{ANE}(\vec{r}) = -N \vec{\nabla} T(\vec{r}) \times \mu_o \vec{M}(\vec{r}),$$ in which $N$ is the anomalous Nernst coefficient of the material and $\mu_o$ is the permeability of free space. Thermal gradients of approximately $0.15$ K/nm (maximum temperature difference $\sim$4 K) are produced primarily along the out-of-plane direction of our $\sim$35-nm-thick film by focusing a mode-locked Al~2~O~3~:Ti laser pulse ($\lambda = 785$ nm wavelength) with a fluence of 0.8 mJ/cm^2^ onto a diffraction-limited spot on the sample surface (see Fig. \[fig:ANE\]**a**). The laser spot is then rastered across photolithographically defined 10 $\mu$m $\times$ 30 $\mu$m device structures; in this geometry, the resulting $\vec{E}_\textrm{ANE}$ engenders a voltage $V_\textrm{ANE}$ proportional to the in-plane component of $\vec{M}$ locally perpendicular to the device channel ($M_x$ in Figs. \[fig:ANE\]**a**-**g**).
Figure \[fig:ANE\]**b** shows a representative $V_\textrm{ANE}$ map obtained at room temperature from a device initialised in the AF phase. The weak contrast observed is a combination of unpinned uncompensated moments that rotate with applied magnetic field [@Fan:2010gu] and pinned uncompensated moments that are strongly exchange-coupled to the bulk Néel order. The uncompensated moments are consistent with the 0.2 $\mu_B$/f.u. remanent magnetisation detected in the AF regime using magnetometry (Fig. \[fig:MH\]**c**) and are characterised in detail in ref. [@Gray:ctOFlBwo].
To switch the magnetic order, we increase the laser fluence ten-fold to 8 mJ/cm^2^. Pulsed laser heating at this fluence causes a peak temperature increase in the Fe~0.52~Rh~0.48~ of $\sim$40 K (estimated from finite-element simulations), which locally induces the FM phase while globally maintaining the sample at room temperature and in the AF state. Since our films are engineered to display magnetic bistability at room temperature, the induced FM regions persist even after photoheating when the region cools back to ambient temperature. We can therefore write FM regions at high fluence and image them without perturbing the written pattern at low fluence.
We demonstrate magnetic writing using the test pattern shown in Fig. \[fig:ANE\]**c**, which includes rectangles of varying aspect ratio as well as pixel-sized dots for determining the minimum writing resolution. Figure \[fig:ANE\]**d**,**e** are ANE images collected from a patterned area with 1 kOe fields applied along the $-x$ and $+x$ directions, respectively. Written regions exhibit neighbouring positive (red) and negative (blue) contrast resembling dipoles. The contrast contains two components. The first component arises from spatial inhomogeneity in the thermal conductivity, which is higher in the FM phase of Fe~0.52~Rh~0.48~ than in the AF phase. Near an in-plane AF/FM boundary, there is an imperfect cancellation of the in-plane charge Seebeck electric field along $+y$ and $-y$. The sign of the resulting Seebeck voltage depends on which side of the thermal discontinuity the laser is focused, leading to a strong dipole-like feature.
The second component giving rise to the contrast in Fig. \[fig:ANE\]**d**,**e** is the ANE. To isolate this signal from the non-magnetic charge Seebeck response, we compute the half-difference of images Fig. \[fig:ANE\]**d**,**e**, which preserves features of the signal that switch in a magnetic field while eliminating features that are non-magnetic. The result, presented in Fig. \[fig:ANE\]**f**, shows FM regions (dark red) within an AF background (light red). The light red contrast in the unwritten regions represents unpinned uncompensated moments. The observed FM shapes are consistent with the generating pattern and exhibit features with sub-micron dimensions, comparable to the laser spot size. To show that FM regions of arbitrary shape above this resolution can be patterned, we write and image another pattern with letters spelling out the authors’ affiliation (Fig. \[fig:ANE\]**f**).
After FM writing, the film can be reset to the AF state by cooling with a Peltier device. In contrast to present magnetic writing procedures, which employ lithography, ion irradiation[@Kim:2012cm; @Bali:2014dl] and implantation[@Fassbender:2008fe], our process is fully repeatable and supports multiple write/erase cycles. Subsequent imaging reveals no permanent changes associated with the writing process. We envisage combining the processes discussed here with commercially mature heat-assisted magnetic recording technologies[@Stipe:2010ce] to achieve high-fidelity patterning with a resolution in the tens of nanometre range. This could promote novel competitive memories as well as facilitate prototyping of magnetic cellular automata[@Imre:2006is], magnonic crystals[@Chumak:2015fa] and artificial spin-ice lattices[@Louis:2018bh].
Acknowledgements
================
The authors thank K. Palmen and W. Zander for their help performing RBS measurements. A.B.M., Y.T. and D.G.S. acknowledge support in part by the Semiconductor Research Corporation (SRC) under nCORE tasks 2758.001 and 2758.003 and by the NSF under the E2CDA programme (ECCS-1740136). I.G. was supported by the Cornell Center for Materials Research with funding from the NSF MRSEC programme (DMR-1719875). J.B. acknowledges support by the AFOSR (FA9550‐14‐1‐0243). Materials synthesis was performed in a facility supported by the National Science Foundation (Platform for the Accelerated Realization, Analysis and Discovery of Interface Materials (PARADIM)) under Cooperative Agreement No. DMR-1539918. Substrate preparation was performed in part at the Cornell NanoScale Facility, a member of the National Nanotechnology Coordinated Infrastructure (NNCI), which is supported by the NSF (Grant No. ECCS-1542081).
Methods
=======
Film growth
-----------
Epitaxial Fe~1-$x$~Rh~$x$~ films are grown using molecular-beam epitaxy to thicknesses of $\sim$35 nm on single-crystalline (001)-oriented MgO substrates in a Veeco GEN10 system with a base pressure of $1\times 10^{-8}$ Torr. Iron (99.995% pure) and rhodium (99.95% pure) species are simultaneously supplied to the growth surface from independent effusion cells. Molecular fluxes are calibrated using x-ray reflectivity (XRR) and quartz crystal microbalances and configured to produce films with rhodium fractions $x$ equal to 0.48. A substrate temperature of 420 $^\circ$C (estimated from a thermocouple in indirect contact with the growth surface and concealed from incident molecular fluxes) is employed for film growth and subsequent half-hour-long *in situ* anneals. The anneal, which is performed immediately after film deposition, is designed to help order bcc Fe~1-$x$~Rh~$x$~ alloys into the B2 CsCl-structure intermetallic with iron and rhodium residing on distinct positions of the two-atom basis.
Compositional measurements
--------------------------
Fe~1-$x$~Rh~$x$~ film compositions are determined using RBS using a probe comprised of 1.4 MeV He$^+$ ions. The scattering geometry is defined by incident angle $\alpha = 7^\circ$, exit angle $\beta = 163^\circ$ and scattering angle $\theta = 170^\circ$. Spectra are integrated to a total accumulated ion dose of 15 $\mu$C. Film chemistry is determined by quantifying the area under iron and rhodium spectral features using an established procedure[@Petrov:1983fz].
Structural characterisation
---------------------------
X-ray-based measurements are performed using a four-circle Philips X’pert MRD diffractometer operated with Cu$K_{\alpha1}$ radiation of wavelength $\lambda$ = 0.15406 nm ($\Delta\lambda/\lambda = 10^{-4}$). The incident beam optics consists of a four-bounce Ge 220 monochromator and a programable 0.125-mm-thick Ni attenuator. For XRR and XRD scans, a 1/16$^\circ$ divergence slit and a Xe proportional detector is employed as receiving optics. For $\omega$-rocking curve measurements, the divergence slit is replaced with a Ge 220 triple-axis analyzer crystal, proving an angular resolution of 12 arc-sec.
BF-TEM images and SAED patterns are collected in an FEI F20 transmission electron microscope with a field-emission source operated at 200 kV. The specimen foils are prepared near the MgO(010) zone axis by cutting vertical film sections in an FEI Strata 400 DualBeam. Initial milling is done using a 30 keV Ga^+^ focused ion beam. For final polishing, the ion energy is reduced to 5 keV.
Magnetic characterisation
-------------------------
The magnetic order of as-deposited Fe~1-$x$~Rh~$x$~ layers is investigated in a Quantum Design physical property measurement system (PPMS). Temperature-dependent transport measurements are performed using the van der Pauw geometry with pressed-indium contacts by incrementally cycling the temperature in 5 K steps in the 200–395 K range. Magnetisation $\vec{M}$ vs. applied magnetic field $\vec{H}$ data is collected over the same temperature window by equipping the PPMS setup with a VSM module and orienting the sample such that Fe~1-$x$~Rh~$x$~$[110] \mid\mid \vec{H}$. At each temperature set point, the sample magnetisation is recorded while the magnetic field is swept between $\pm200$ Oe.
Magnetic imaging
----------------
To write and image ferromagnetic patterns, we use an anomalous Nernst effect microscope.[@Bartell:2015hp] In this technique the local sample magnetization is transduced into an electrical voltage via a local thermal gradient. We generate local thermal gradients using a pulsed Coherent Mira 900 Al~2~O~3~:Ti laser tuned to 780 nm wavelength. We employ 3-ps-wide pulses and a repetition rate of 76 MHz (13 ns period). The laser is focused to a diffraction-limited 650 nm-diameter spot using a 0.90 numerical aperture microscope objective. We raster the laser using a 4f optical path in combination with a voice coil-controlled fast-steering mirror. To detect $V_\textrm{ANE}$, we first collect the laser-induced voltage pulse train through a coplanar waveguide into a microwave transmission line and amplify the pulses by 40 dB with 0.1-3 GHz bandwidth. The pulse train is then sent to the radio-frequency port of a DC-12 GHz electrical mixer, where it is mixed with a 600 ps-wide pulse train from an arbitrary waveform generator that is referenced to the laser repetition rate. The mixer output voltage, $V_\textrm{ANE}$, is measured with a lock-in amplifier referenced to intensity modulation of the light with a photoelastic modulator. We estimated the temperature change induced by the laser using resistivity measurements in conjunction with time-domain finite element calculations performed in the COMSOL Multiphysics software package.
Author contributions
====================
A.B.M conceived and coordinated the study, synthesized the films and characterized their physical properties. I.G. fabricated the devices and performed the magnetic writing and imaging. Y.T. carried out the VSM analyses. J.S. conducted the RBS investigations and data analyses. D.W. performed the electron microscopy. All authors contributed to the discussion and writing of the manuscript.
[10]{} url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
. ** ****, ().
*et al.* . ** ****, ().
*et al.* ** ****, ().
, , & . ** ****, ().
*et al.* ** ****, ().
*et al.* ** ****, ().
*et al.* ** ****, ().
*et al.* . ** ****, ().
, & . ** ****, ().
*et al.* . ** ****, ().
& . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
*et al.* . ** ****, ().
, & . ** ****, ().
*et al.* . ** ****, ().
*et al.* . ** ****, ().
, , & ** ****, ().
*et al.* . **.
*et al.* ** ****, ().
*et al.* ** ****, ().
& . ** ****, ().
*et al.* . ** ****, ().
, , & . ** ****, ().
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Pau Panareda Busto, Ahsan Iqbal, and Juergen Gall, [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'egbib.bib'
title: Open Set Domain Adaptation for Image and Action Recognition
---
[Shell : Bare Demo of IEEEtran.cls for Computer Society Journals]{}
the last years, impressive results have been achieved on large-scale datasets for image classification or action recognition. Acquiring such large annotated datasets, however, is very expensive and there is a need to transfer the knowledge from existing annotated datasets to unlabelled data that is relevant for a specific application. If the labelled and unlabelled data have different characteristics, they have been sampled from two different domains. In particular, datasets that have been collected from the Internet, [e.g.]{}, from platforms for sharing videos or images, differ greatly from data that needs to be processed for an application. To address the domain shift between the labelled dataset, which is the source domain, and the unlabelled data from the target domain, various unsupervised domain adaptation approaches have been proposed. If the data from the target source is partially labelled, the problem is termed semi-supervised domain adaptation. In this work, we address unsupervised and semi-supervised domain adaptation in the context of image and action recognition.
Although the methods for domain adaptation have been advanced tremendously in the last years [@DA_Saenko10; @DA_Gopalan11; @DA_Gong12; @CNN-DA_Chopra13; @DA_Hoffman14; @CNN-DA_Ganin15; @CNN-DA_Hsu15; @CNN-DA_Ghifary16; @CNN-DA_Tzeng17; @CNN-DA_Motiian17], the evaluation protocols were restricted to a scenario where all categories in the target domain are known and match the categories in the source domain. Fig. \[fig:closed\] illustrates such a *closed set domain adaptation* setting. The assumption that all images or videos that are in the target domain belong to categories in the source domain, however, is violated in most cases. In particular if the number of potential categories is very large as it is the case for object or action categories, the target domain contains images or videos of categories that are not present in the source domain since they are not of interest for a specific application. We therefore propose a more realistic evaluation setting for unsupervised or semi-supervised domain adaptation, namely *open set domain adaptation*, which builds on the concept of open sets [@DATA_Scheirer13; @DATA_Scheirer14; @CNN_Bendale16]. As illustrated in Fig. \[fig:teaser\], the source and target domains are not anymore restricted in the open set case to share the same categories as in the closed set case, but both domains contain images or videos from categories that are not present in the other domain.
To address the problem of open set domain adaptation, we propose a generic approach that learns a linear mapping that maps the feature space of the source domain to the feature space of the target domain. It assigns a subset of images or videos of the target domain to the categories of the source domain and transforms the feature space of the source domain gradually towards the feature space of the target domain. By using a subset instead of the entire set, the approach handles images or videos in the target domain that are not related to any sample in the source domain. The approach can be applied to any feature space, which includes features extracted from images as well as features extracted from videos. The approach works in particular very well for features spaces that are extracted by convolutional networks and outperforms most end-to-end learning approaches for domain adaptation. The good performance of the approach coincides with the observation that deep convolutional networks tend to linearise manifolds of image domains [@BengioMDR13; @UpchurchGPPSBW17]. In this case, a linear mapping is sufficient to map the feature space of the source domain to the feature space of the target domain. In particular, the flexibility of the approach, which can be used for images and videos, for open set and closed set domain adaptation, as well as unsupervised and semi-supervised domain adaptation, makes the approach a versatile tool for applications. An overview of the approach for unsupervised open set domain adaptation is given in Fig. \[fig:pipeline\].
A preliminary version of this work was presented in [@Busto17]. In this work, we introduce open set domain adaptation for action recognition and provide a thorough experimental evaluation, which includes open set domain adaptation from synthetic data to real data and an evaluation of the proposed approach for standard closed set protocols. In total, we evaluate the approach on 26 *open set* and 34 *closed set* combinations of source and target domains including the *Office* dataset [@DA_Saenko10], its extension with the *Caltech* dataset [@DA_Gong12], the *Cross-Dataset Analysis* [@DATA_Tommasi14], the *Sentiment dataset* [@DA_Blitzer07], synthetic data [@DATA_Peng17], and two action recognition datasets, namely the *Kinetics Human Action Video Dataset* [@Data_Kay17] and the *UCF101 Action Recognition Dataset* [@Data_Soomro12]. Our approach achieves state-of-the-art results in all settings both for unsupervised and semi-supervised open set domain adaptation and obtains competitive results compared state-of-the-art deep leaning approaches for closed set domain adaptation.
Related Work {#sec:related}
============
Domain Adaptation
-----------------
The interest in studying domain adaptation techniques for computer vision problems increased with the release of a benchmark by Saenko et al. [@DA_Saenko10] for domain adaptation in the context of object classification. The first relevant works on unsupervised domain adaptation for object categorisation were presented by Golapan et al. [@DA_Gopalan11] and Gong et al. [@DA_Gong12], who proposed an alignment in a common subspace of source and target samples using the properties of Grassmanian manifolds. Jointly transforming source and target domains into a common low dimensional space was also done together with a conjugate gradient minimisation of a transformation matrix with orthogonality constraints [@DA_Baktashmotlagh13] and with dictionary learning to find subspace interpolations [@DA_Ni13; @DA_Shekhar13; @DA_Xu15]. Sun et al. [@DA_Sun14; @CNN-DA_Sun15] presented a very efficient solution based on second-order statistics to align a source domain with a target domain. Herath et al. [@DA_Herath17] also match second-order statistics with a joint estimation of latent spaces. To obtain an estimate of the target distribution in the latent space, Gholami et al. [@CNN-DA_Gholami17] introduce a Bayesian approximation to jointly learn a softmax classifier across-domains. Similarly, Csurka et al. [@DA_Csurka16] jointly denoise source and target samples to reconstruct data without partial random corruption. Zhang et al. [@DA_Zhang17] also align distributions, but they include geometrical differences in a joint optimisation. Sharing certain similarities with associations between domains, Gong et al. [@DA_Gong13] minimise the Maximum Mean Discrepancy (MMD) [@MATH_Gretton06] of two datasets. They assign instances to latent domains and solve it by a relaxed binary optimisation. Hsu et al. [@CNN-DA_Hsu15] use a similar idea allowing instances to be linked to all other samples.
Semi-supervised domain adaptation approaches take advantage of knowing the class labels of a few target samples. Aytar et al. [@DA_Aytar11] proposed a transfer learning formulation to regularise the training of target classifiers. Exploiting pairwise constraints across domains, Saenko et al. [@DA_Saenko10] and Kulis et al. [@DA_Kulis11] learn a transformation to minimise the effect of the domain shift while also training target classifiers. Following the same idea, Hoffman et al. [@DA_Hoffman13] considered an iterative process to alternatively minimise the classification weights and the transformation matrix. In a different context, [@DA_Busto15] proposed a weakly supervised approach to refine coarse viewpoint annotations of real images by synthetic images. In contrast to semi-supervised approaches, the task of viewpoint refinement assumes that all images in the target domain are labelled but not with the desired granularity.
The idea of selecting the most relevant information of each domain has been studied in early domain adaptation methods in the context of natural language processing [@DA_Blitzer06]. Pivot features that behave the same way for discriminative learning in both domains were selected to model their correlations. Gong et al. [@DA_Gong13ICML] presented an algorithm that selects a subset of source samples that are distributed most similarly to the target domain. Another technique that deals with instance selection has been proposed by Sangineto et al. [@DA_Sangineto14]. They train weak classifiers on random partitions of the target domain and evaluate them in the source domain. The best performing classifiers are then selected. Other works have also exploited greedy algorithms that iteratively add target samples to the training process, while the least relevant source samples are removed [@DA_Bruzzone10; @CNN-DA_Tommasi13].
During the last years, a large number of domain adaptation methods have been based on deep convolutional neural networks (CNN) [@CNN_Krizhevsky12], which learn more discriminative feature representations than hand-crafted features and substantially reduce the domain bias between datasets in object recognition tasks [@CNN_Donahue2014]. Non-adapted classifiers trained with features extracted from CNN layers outperform domain adaptation methods with shallow feature descriptors [@CNN_Donahue2014; @CNN-DA_Sun15]. Many of these deep domain adaptation architectures are inspired by the traditional methods and seek to minimise the MMD distance as a regulariser to learn features for source and target samples jointly [@CNN-DA_Ghifary14; @CNN-DA_Tzeng14; @CNN-DA_Long15; @CNN-DA_Long16; @CNN-DA_Yan17]. Recently, Carlucci et al. [@CNN-DA_Carlucci17] extend this type of networks and use intermediate layers for the alignment of distributions before batch normalisation. They learn a parameter that steers the contribution of each domain at a given layer. Ganin et al. [@CNN-DA_Ganin15] added a domain classifier network after the CNN to maximize the discriminatory loss of both domains while jointly minimising the classification loss using source data. More recently, Tzeng et al. [@CNN-DA_Tzeng17] propose a generalized framework for adversarial adaptation. In the semi-supervised setting, Mottian et al. [@CNN-DA_Motiian17] present a deep domain adaptation method that exploits the domain loss minimisation while maximizing the distances between labelled samples from different domains and classes. Other forms of data representation, such as hash codes [@CNN-DA_Venkateswara17] and scatter tensors [@CNN-DA_Koniusz17; @DA_Lu17], have also been combined with deep domain adaptation architectures to further reduce the domain bias.
Open Set Recognition
--------------------
The inclusion of *open sets* in recognition tasks appeared in the field of face recognition, where evaluation datasets contain unseen face instances as impostors that have to be rejected [@DATA_Phillips00; @DATA_Li05]. Such open set protocols are nowadays widely used for evaluating face recognition approaches [@CNN_Sun15]. The generalisation towards an open set scenario for multi-object classification was introduced by Schreier et al. [@DATA_Scheirer13], who addressed the more realistic case of a finite set of known objects mixed with many unknown ones. Based on this principle, [@OBJ_Jain14] and [@DATA_Scheirer14] propose multi-class classifiers that detect unknown instances by learning SVMs that assign probabilistic decision scores instead of class labels. More recently, Bendale and Boult [@CNN_Bendale16] adapt traditional neural networks for open set recognition tasks by introducing a new layer that estimates the probability of an object to be labelled as unseen class.
Closely related are also the works [@MATH_Zhang06] and [@MATH_Bartlett08] that add a regulariser to detect uninformative data and penalise a misclassification during training. Lately, Gavves et al. [@DA_Gavves15] present an active learning technique, whose intially trained SVMs on a subset of known classes are used as priors to further train novel object classes from other target datasets.
Open Set Domain Adaptation {#sec:method}
==========================
We present an approach that iterates between solving the labelling problem of target samples, [i.e.]{}, associating a subset of the target samples to the known categories of the source domain, and computing a mapping from the source to the target domain by minimising the distances of the assignments. The transformed source samples are then used in the next iteration to re-estimate the assignments and update the transformation. This iterative process is repeated until convergence and is illustrated in Fig. \[fig:pipeline\].
In Section \[sec:uns\], we describe the unsupervised assignment of target samples to categories of the source domain. The semi-supervised case is described in Section \[sec:sup\]. Section \[sec:transform\] finally describes how the mapping from the source domain to the target domain is estimated from the previous assignments. This part is the same for the unsupervised and semi-supervised setting.
Unsupervised Domain Adaptation {#sec:uns}
------------------------------
We first address the problem of unsupervised domain adaptation, [i.e.]{}, none of the target samples are annotated, in an open set protocol. Given a set of classes $\mathcal{C}$ in the source domain, including $|\mathcal{C}-1|$ known classes and an additional unknown class that gathers all instances from other irrelevant categories, we aim to label the target samples $\mathcal{T} = \{T_1,\dots,T_{|\mathcal{T}|}\}$ by a class $c \in \mathcal{C}$. We define the cost of assigning a target sample $T_t$ to a class $c$ by $d_{ct} = \left\Vert S_c - T_t \right\Vert^2_2$ where $T_t \in \mathbb{R}^D$ is the feature representation of the target sample $t$ and $S_c \in \mathbb{R}^D$ is the mean of all samples in the source domain labelled by class $c$. To increase the robustness of the assignment, we do not enforce that all target samples are assigned to a class as shown in Fig. \[fig:pipeline\](b). The cost of declaring a target sample as outlier is defined by a parameter $\lambda$, which is discussed in Section \[sec:experiments:lambda\].
Having defined the individual assignment costs, we can formulate the entire assignment problem by:
$$\label{eq:assignments}
\begin{aligned}
& \underset{x_{ct},o_t}{\operatorname{\text{minimise}}}
& & \sum_{t}\bigg(\sum_{c} d_{ct} x_{ct} + \lambda o_t\bigg) \\
& \text{subject to}
& & \sum_{c} x_{ct} + o_t = 1 & \forall t~\text{,} \\
& & & \sum_{t} x_{ct} \geq 1 & \forall c~\text{,} \\
& & & x_{ct}, o_t \in \{0,1\} & \forall c,t~\text{.}
\end{aligned}$$
By minimising the constrained objective function, we obtain the binary variables $x_{ct}$ and $o_t$ as solution of the assignment problem. The first type of constraints ensures that a target sample is either assigned to one class, [i.e.]{}, $x_{ct}=1$, or declared as outlier, [i.e.]{}, $o_t=1$. The second type of constraints ensures that at least one target sample is assigned to each class . We use the constraint integer program package SCIP [@MATH_Achterberg09] to solve all proposed formulations. As it is shown in Fig. \[fig:pipeline\](b), we label the targets also by the unknown class. Note that the unknown class combines all objects that are not of interest. Even if the unknowns in the source and target domain belong to different semantic classes, a target sample might be closer to the mean of all negatives than to any other positive class. In this case, we can confidentially label a target sample as unknown. In our experiments, we show that it makes not much difference if the unknown class is included in the unsupervised setting since the outlier handling discards target samples that are not close to the mean of negatives.
Semi-supervised Domain Adaptation {#sec:sup}
---------------------------------
The unsupervised assignment problem naturally extends to a semi-supervised setting when a few target samples are annotated. In this case, we only have to extend the formulation by additional constraints that enforce that the annotated target samples do not change the label, [i.e.]{}, $$\label{eq:sup}
~~~~~~~~~~~~~~~~~~~~~~x_{\hat{c}_tt} = 1 ~~~~~~~~~~~~~~~~~~~~~~~~\forall (t,\hat{c}_t) \in \mathcal{L} \text{,}$$ where $\mathcal{L}$ denotes the set of labelled target samples and $\hat{c}_t$ the class label provided for target sample $t$. In order to exploit the labelled target samples better, one can use the neighbourhood structure in the source and target domain. While the constraints remain the same, the objective function can be changed to $$\label{eq:lc}
\begin{aligned}
& & & \sum_{t}\left(\sum_{c} x_{ct} \bigg( d_{ct} + \sum_{t' \in N_t} \sum_{c'} d_{cc'} x_{c't'} \bigg) + \lambda o_t \right) \text{,} \\
\end{aligned}$$ where $ d_{cc'} = \left\Vert S_c - S_{c'} \right\Vert^2_2$. While in the cost of labelling a target sample $t$ by the class $c$ is given only by $d_{ct}$, a second term is added in . It is computed over all neighbours $N_t$ of $t$ and adds the distance between the classes in the source domain as additional cost if a neighbour is assigned to another class than the target sample $t$. The objective function , however, becomes quadratic and therefore NP-hard to solve. Thus, we transform the *quadratic assignment problem* into a mixed 0-1 linear program using the Kaufman and Broeckx linearisation [@MATH_Kaufman78]. By substituting $$\label{eq:L}
w_{ct} = x_{ct} \left( \sum_{t' \in N_t} \sum_{c'} d_{cc'} x_{c't'} \right) \text{,}$$ we derive to the linearised problem $$\label{eq:linearized}
\begin{aligned}
& \underset{x_{ct},w_{ct},o_t}{\operatorname{\text{minimise}}}
& & \sum_{t}\left(\sum_{c} d_{ct} x_{ct} + \sum_{c} w_{ct} + \lambda o_t\right)\\
& \text{subject to}
& & \sum_{c} x_{ct} + o_t = 1 & \forall t~\text{,} \\
& & & \sum_{t} x_{ct} \geq 1 & \forall c~\text{,} \\
& & & a_{ct}x_{ct}+\sum_{t' \in N_t} \sum_{c'} d_{cc'} x_{c't'} - w_{ct} \le a_{ct} & \forall s,t~\text{,} \\
& & & x_{ct}, o_t \in \{0,1\} & \forall c,t~\text{,} \\
& & & w_{ct} \ge 0 & \forall c,t~\text{,}
\end{aligned}$$ where $a_{ct} = \sum_{t' \in N_t} \sum_{c'} d_{cc'}$.
Mapping {#sec:transform}
-------
As illustrated in Fig. \[fig:pipeline\], we iterate between solving the assignment problem, as described in Section \[sec:uns\] or \[sec:sup\], and estimating the mapping from the source domain to the target domain. We consider a linear transformation, which is represented by a matrix $W \in \mathbb{R}^{D \times D}$. We estimate $W$ by minimising the following loss function: $$\label{eq:objMat}
f(W) = \dfrac{1}{2}\sum_t\sum_c x_{ct} \Vert W S_c - T_t \Vert^2_2~\text{,}$$ which can be written in matrix form: $$\label{eq:objMat2}
f(W) = \dfrac{1}{2}|| W P_S - P_T ||^2_F~\text{.}$$ The matrices $P_S$ and $P_T \in \mathbb{R}^{D\times L}$ with $L=\sum_t\sum_c x_{ct}$ represent all assignments, where the columns denote the actual associations. The quadratic nature of the convex objective function may be seen as a linear least squares problem, which can be easily solved by any available QP solver. State-of-the-art features based on convolutional neural networks, however, are high dimensional and the number of target instances is usually very large. We use therefore non-linear optimisation [@MATH_Svanberg02; @MATH_Johnson10] to optimise $f(W)$. The derivatives of are given by $$\label{eq:diff}
\dfrac{\partial f(W)}{\partial W} = W(P_SP_S^T) - P_TP_S^T~\text{.}$$ If $L < D$, [i.e.]{}, the number of samples, which have been assigned to a known class, is smaller than the dimensionality of the features, the optimisation also deals with an underdetermined linear least squares formulation. In this case, the solver converges to the matrix $W$ with the smallest norm, which is still a valid solution.
After the transformation $W$ is estimated, we map the source samples to the target domain. We therefore iterate the process of solving the assignment problem and estimating the mapping from the source domain to the target domain until it converges. After the approach has converged, we train linear SVMs in a one-vs-one setting on the transformed source samples. For the semi-supervised setting, we also include the annotated target samples $\mathcal{L}$ to the training set. The linear SVMs are then used to obtain the final labelling of the target samples as illustrated in Fig. \[fig:pipeline\_d\].
Experiments {#sec:experiments}
===========
We evaluate our method in the context of domain adaptation for image classification and action recognition. In this setting, the images or videos of the source domain are annotated by class labels and the goal is to classify the images or videos in the target domain. We report the accuracies for both unsupervised and semi-supervised scenarios, where target samples are unlabelled or partially labelled, respectively. For consistency, we use *libsvm* [@MATH_Chang01] since it has also been used in other works, [e.g.]{}, [@DA_Fernando13] and [@CNN-DA_Sun15]. We set the misclassification parameter $C = 0.001$ in all experiments, which allows for a soft margin optimisation that works best in such classification tasks [@DA_Fernando13; @CNN-DA_Sun15]. The source code and the described open set protocols are available at <https://github.com/Heliot7/open-set-da>.
Parameter configuration {#sec:experiments:lambda}
-----------------------
Our algorithm contains a few parameters that need to be defined. For the outlier rejection, we use $$\label{eq:lambda}
\lambda = \rho\big(\max_{t,c}d_{ct} + \min_{t,c}d_{ct}\big),$$ [i.e.]{}, $\lambda$ is adapted automatically based on the distances $d_{ct}$ and $\rho$, which is set to $0.5$ unless otherwise specified. While higher values of $\lambda$ closer to the largest distance barely discard any outlier, lower values almost reject all assignments. We iterate the approach until the maximum number of 10 iterations is reached or if the distance $$\label{eq:diststop}
\sqrt{\sum_{t}\sum_{c}x_{ct}\left\Vert W_k S_{c} - T_{t}\right\Vert_2^2}~~\text{}$$ is below $\epsilon = 0.01$, where $W_{k}$ denotes the estimated transformation at iteration $k$. In practice, the process converges after 3-5 iterations.
Open set domain adaptation {#open-set-domain-adaptation}
--------------------------
### Office dataset {#exp:office}
We evaluate and compare our approach on the *Office* dataset [@DA_Saenko10], which is the standard benchmark for domain adaptation with CNN features. It provides three different domains, namely *Amazon (A)*, *DSLR (D)* and *Webcam (W)*. While the *Amazon* dataset contains centred objects on white background, the other two comprise pictures taken in an office environment but with different quality levels. In total, there are 31 common classes for 6 source-target combinations. This means that there are 4 combinations with a considerable domain shift (A $\rightarrow$ D, A $\rightarrow$ W, D $\rightarrow$ A, W $\rightarrow$ A) and 2 with a minor domain shift (D $\rightarrow$ W, W $\rightarrow$ D). [Following the standard protocol and for a fair comparison with the other methods, we extract feature vectors from the fully connected layer-7 (fc7) of the AlexNet model [@CNN_Krizhevsky12]]{}.
We introduce an open set protocol for this dataset by taking the 10 classes that are also common in the *Caltech* dataset [@DA_Gong12] as shared classes. In alphabetical order, the classes 11-20 are used as unknowns in the source domain and 21-31 as unknowns in the target domain, [i.e.]{}, the unknown classes in the source and target domain are not shared. For evaluation, each sample in the target domain needs to be correctly classified either by one of the 10 shared classes or as unknown. In order to compare with a closed setting (CS), we report the accuracy when source and target domain contain only samples of the 10 shared classes. Since OS is evaluated on all target samples, we also report the numbers when the accuracy is only measured on the same target samples as CS, [i.e.]{}, only for the shared 10 classes. The latter protocol is denoted by OS$^*$(10) and provides a direct comparison to CS(10).
---------------------- ---------- ------------- ---------- ---------- ------------- ----------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10)
LSVM 87.1 70.7 72.6 77.5 53.9 57.5
DAN [@CNN-DA_Long15] 88.1 76.5 77.6 **90.5** 70.2 72.5
RTN [@CNN-DA_Long16] **93.0** 74.7 76.6 87.0 70.8 73.0
BP [@CNN-DA_Ganin15] 91.9 77.3 78.3 89.2 73.8 75.9
ATI 92.4 78.2 78.8 85.1 **77.7** **78.4**
ATI-$\lambda$ **93.0** **79.2** **79.8** 84.0 76.5 77.6
ATI-$\lambda$-N1 91.9 78.3 78.9 84.6 74.2 75.6
---------------------- ---------- ------------- ---------- ---------- ------------- ----------
: Open set domain adaptation on the unsupervised Office dataset with 10 shared classes (OS) using all samples per class [@DA_Gong13]. For comparison, results for closed set domain adaptation (CS) and modified open set (OS$^*$) are reported.
---------------------- ---------- ------------- ---------- ---------- ------------- ----------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10)
LSVM 79.4 40.0 45.1 97.9 87.5 88.5
DAN [@CNN-DA_Long15] 83.4 53.5 57.0 96.1 87.5 88.4
RTN [@CNN-DA_Long16] 82.8 53.8 57.2 97.9 88.1 89.0
BP [@CNN-DA_Ganin15] 84.3 54.1 57.6 97.5 88.9 89.8
ATI 93.4 **70.0** 71.1 **98.5** 92.2 92.6
ATI-$\lambda$ **93.8** **70.0** **71.3** **98.5** 93.2 93.5
ATI-$\lambda$-N1 93.3 65.6 67.8 97.9 **94.0** **94.4**
---------------------- ---------- ------------- ---------- ---------- ------------- ----------
: Open set domain adaptation on the unsupervised Office dataset with 10 shared classes (OS) using all samples per class [@DA_Gong13]. For comparison, results for closed set domain adaptation (CS) and modified open set (OS$^*$) are reported.
---------------------- ---------- ------------- ---------- --------- ------------- ---------- ---------- ---------- ----------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10) CS OS$^*$ OS
LSVM 80.0 44.9 49.2 **100** 96.5 96.6 87.0 65.6 68.3
DAN [@CNN-DA_Long15] 84.9 58.5 60.8 **100** 97.5 98.3 90.5 74.0 75.8
RTN [@CNN-DA_Long16] 85.1 60.2 62.4 **100** 98.3 **98.8** 91.0 74.3 76.2
BP [@CNN-DA_Ganin15] 86.2 61.8 64.0 **100** 98.0 98.7 91.6 75.7 77.4
ATI 93.4 76.4 76.6 **100** 99.1 98.3 **93.8** 82.1 82.6
ATI-$\lambda$ **93.7** **76.5** **76.7** **100** 99.2 98.3 93.7 **82.4** **82.9**
ATI-$\lambda$-N1 93.4 71.6 72.4 **100** **99.6** **98.8** 93.5 80.6 81.3
---------------------- ---------- ------------- ---------- --------- ------------- ---------- ---------- ---------- ----------
: Open set domain adaptation on the unsupervised Office dataset with 10 shared classes (OS) using all samples per class [@DA_Gong13]. For comparison, results for closed set domain adaptation (CS) and modified open set (OS$^*$) are reported.
\[table:office\_uns\_ft\]
------------- ------------------ ------ ------------------ ------ ------------------ ------ ------------------ ------ ------------------ ------ ------------------ ------
assign-$\lambda$ LSVM assign-$\lambda$ LSVM assign-$\lambda$ LSVM assign-$\lambda$ LSVM assign-$\lambda$ LSVM assign-$\lambda$ LSVM
initial 72.6 57.5 45.1 88.5 49.2 96.6
iteration 1 78.4 76.8 74.5 69.8 73.6 68.1 90.4 90.3 71.9 70.0 89.6 97.8
iteration 2 77.7 79.1 80.1 77.6 80.4 71.3 91.5 93.5 77.2 75.9 84.7 98.3
iteration 3 75.3 79.8 77.8 76.7
------------- ------------------ ------ ------------------ ------ ------------------ ------ ------------------ ------ ------------------ ------ ------------------ ------
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10)
LSVM 84.4$\pm$5.9 63.7$\pm$6.7 66.6$\pm$5.9 76.5$\pm$2.9 48.2$\pm$4.8 52.5$\pm$4.2
TCA [@DA_Pan09] 85.9$\pm$6.3 75.5$\pm$6.6 75.7$\pm$5.9 80.4$\pm$6.9 67.0$\pm$5.9 67.9$\pm$5.5
gfk [@DA_Gong12] 84.8$\pm$5.1 68.6$\pm$6.7 70.4$\pm$6.0 76.7$\pm$3.1 54.1$\pm$4.8 57.4$\pm$4.2
SA [@DA_Fernando13] 84.0$\pm$3.4 71.5$\pm$5.9 72.6$\pm$5.3 76.6$\pm$2.8 57.4$\pm$4.2 60.1$\pm$3.7
CORAL [@CNN-DA_Sun15] 85.8$\pm$7.2 79.9$\pm$5.7 79.6$\pm$5.0 81.9$\pm$2.8 68.1$\pm$3.6 69.3$\pm$3.1
ATI **91.4$\pm$1.3** 80.5$\pm$2.0 81.1$\pm$2.8 **86.1$\pm$1.1** 73.4$\pm$2.0 **75.3$\pm$1.7**
ATI-$\lambda$ 91.1$\pm$2.1 **81.1$\pm$0.4** **82.2$\pm$2.0** 85.5$\pm$2.1 **73.7$\pm$2.6** **75.3$\pm$1.4**
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
: Open set domain adaptation on the unsupervised Office dataset with 10 shared classes (OS). We report the average and the standard deviation using a subset of samples per class in 5 random splits [@DA_Saenko10]. For comparison, results for closed set domain adaptation (CS) and modified open set (OS$^*$) are reported.
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10)
LSVM 75.5$\pm$2.1 36.1$\pm$3.7 42.2$\pm$3.3 96.2$\pm$1.0 81.5$\pm$1.5 83.1$\pm$1.3
TCA [@DA_Pan09] 88.2$\pm$1.5 71.8$\pm$2.5 71.8$\pm$2.0 97.8$\pm$0.5 92.0$\pm$0.9 91.5$\pm$1.0
gfk [@DA_Gong12] 79.7$\pm$1.0 45.3$\pm$3.7 49.7$\pm$3.4 96.3$\pm$0.9 85.1$\pm$2.7 86.2$\pm$2.4
SA [@DA_Fernando13] 81.7$\pm$0.7 52.5$\pm$3.0 55.8$\pm$2.7 96.3$\pm$0.8 86.8$\pm$2.5 87.7$\pm$2.3
CORAL [@CNN-DA_Sun15] 89.6$\pm$1.0 66.6$\pm$2.8 68.2$\pm$2.5 97.2$\pm$0.7 91.1$\pm$1.7 91.4$\pm$1.5
ATI 93.5$\pm$0.3 69.8$\pm$1.4 70.8$\pm$2.1 97.3$\pm$0.5 89.6$\pm$2.1 90.3$\pm$1.8
ATI-$\lambda$ **93.9$\pm$0.4** **71.1$\pm$0.9** **72.0$\pm$0.5** **97.5$\pm$1.1** **92.1$\pm$1.3** **92.5$\pm$0.7**
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
: Open set domain adaptation on the unsupervised Office dataset with 10 shared classes (OS). We report the average and the standard deviation using a subset of samples per class in 5 random splits [@DA_Saenko10]. For comparison, results for closed set domain adaptation (CS) and modified open set (OS$^*$) are reported.
--------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ---------- ---------- ----------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10) CS OS$^*$ OS
LSVM 72.5$\pm$2.7 34.3$\pm$4.9 39.9$\pm$4.4 99.1$\pm$0.5 89.8$\pm$1.5 90.5$\pm$1.3 84.1 58.9 62.5
TCA 85.5$\pm$3.3 68.1$\pm$5.1 68.6$\pm$4.6 98.8$\pm$0.9 94.1$\pm$2.9 93.6$\pm$2.6 89.5 78.1 78.2
gfk 75.0$\pm$2.9 43.2$\pm$5.1 47.6$\pm$4.6 99.0$\pm$0.5 92.0$\pm$1.5 92.2$\pm$1.4 85.2 64.7 67.3
SA 76.5$\pm$3.2 49.7$\pm$5.1 53.0$\pm$4.6 98.8$\pm$0.7 92.4$\pm$2.9 92.4$\pm$2.8 85.7 68.4 70.3
CORAL 86.9$\pm$1.9 63.9$\pm$4.9 65.6$\pm$4.3 **99.2$\pm$0.7** 96.0$\pm$2.1 95.0$\pm$2.0 90.1 77.6 78.2
ATI 92.2$\pm$1.1 75.1$\pm$1.7 76.0$\pm$2.0 98.9$\pm$1.3 95.5$\pm$2.3 95.4$\pm$2.1 **93.2** 80.7 81.5
ATI-$\lambda$ **92.4$\pm$1.1** **75.4$\pm$1.8** **76.4$\pm$1.8** 98.9$\pm$1.3 **96.5$\pm$2.1** **95.8$\pm$1.8** **93.2** **81.5** **82.3**
--------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ---------- ---------- ----------
: Open set domain adaptation on the unsupervised Office dataset with 10 shared classes (OS). We report the average and the standard deviation using a subset of samples per class in 5 random splits [@DA_Saenko10]. For comparison, results for closed set domain adaptation (CS) and modified open set (OS$^*$) are reported.
\[table:office\_uns\_it\]
[**Unsupervised domain adaptation.**]{} We firstly compare the accuracy of our method in the unsupervised set-up with state-of-the-art domain adaptation techniques embedded in the training of CNN models. DAN [@CNN-DA_Long15] retrains the AlexNet model by freezing the first 3 convolutional layers, finetuning the last 2 and learning the weights from each fully connected layer by also minimising the discrepancy between both domains. RTN [@CNN-DA_Long16] extends DAN by adding a residual transfer module that bridges the source and target classifiers. BP [@CNN-DA_Ganin15] trains a CNN for domain adaptation by a gradient reversal layer and minimises the domain loss jointly with the classification loss. For training, we use all samples per class as proposed in [@DA_Gong13], which is the standard protocol for CNNs on this dataset. As proposed in [@CNN-DA_Ganin15], we use for all methods linear SVMs for classification instead of the soft-max layer for a fair comparison.
To analyse the formulations that are discussed in Section \[sec:method\], we compare several variants: ATI (*Assign-and-Transform-Iteratively*) denotes our formulation in assigning a source class to all target samples, [i.e.]{}, $\lambda = \infty$. Then, ATI-$\lambda$ includes the outlier rejection and ATI-$\lambda$-$N_1$ is the unsupervised version of the locality constrained formulation corresponding to with 1 nearest neighbour. In addition, we denote LSVM as the linear SVMs trained on the source domain without any domain adaptation.
--------------- -------- ---------- -------- ---------- -------- ---------- -------- ---------- -------- ---------- -------- ---------- -------- ----------
OS-SVM LSVM OS-SVM LSVM OS-SVM LSVM OS-SVM LSVM OS-SVM LSVM OS-SVM LSVM OS-SVM LSVM
No Adap. 67.5 72.6 58.4 57.5 54.8 45.1 80.0 88.5 55.3 49.2 94.0 96.6 68.3 68.3
ATI-$\lambda$ 72.0 **79.8** 65.3 **77.6** 66.4 **71.3** 82.2 **93.5** 71.6 **76.7** 92.7 **98.3** 75.0 **82.9**
--------------- -------- ---------- -------- ---------- -------- ---------- -------- ---------- -------- ---------- -------- ---------- -------- ----------
-------------------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ----------
ATI-$\lambda$ ($\mathcal{C}$ w/o unknown) 79.0 77.1 70.5 93.4 75.8 98.2 82.3
ATI-$\lambda$ ($\mathcal{C}$ with unknown) **79.8** **77.6** **71.3** **93.5** **76.7** **98.3** **82.9**
-------------------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ----------
The results of these techniques using the described open set protocol are shown in Table \[table:office\_uns\_ft\]. Our approach ATI improves over the baseline without domain adaptation (LSVM) by +6.8% for CS and +14.3% for OS. The improvement is larger for the combinations that have larger domain shifts, [i.e.]{}, the combinations that include the $Amazon$ dataset. We also observe that ATI outperforms all CNN-based domain adaptation methods for the closed (+2.2%) and open setting (+5.2%). It can also be observed that the accuracy for the open set is lower than for the closed set for all methods, but that our method handles the open set protocol best. While ATI-$\lambda$ does not obtain any considerable improvement compared to ATI in CS, the outlier rejection allows for an improvement in OS. The locality constrained formulation, ATI-$\lambda$-$N_1$, which we propose only for the semi-supervised setting, decreases the accuracy in the unsupervised setting.
[ The evolution of the percentage of correct assignments and the intermediate classification accuracies are shown in Table \[table:accuracy\_assigns\]. The approach converges after two or three iterations. While the accuracy of the LSVMs that are trained on the transformed source samples increases with each iteration, the accuracy of the assignment can even decrease in some cases. ]{}
Additionally, we report accuracies of popular domain adaptation methods that are not related to deep learning. We report the results of methods that transform the data to a common low dimensional subspace, including Transfer Component Analysis (TCA) [@DA_Pan09], Geodesic Flow Kernel (GFK) [@DA_Gong12] and Subspace alignment (SA) [@DA_Fernando13]. In addition, we also include CORAL [@CNN-DA_Sun15], which whitens and recolours the source towards the target data. Following the standard protocol of [@DA_Saenko10], we take 20 samples per object class when *Amazon* is used as source domain, and 8 for *DSLR* or *Webcam*. [As in the previous comparison with the CNN-based methods, we extract feature vectors from the last convolutional layer (fc7) from the AlexNet model [@CNN_Krizhevsky12]]{}. Each evaluation is executed 5 times with random samples from the source domain. The average accuracy and standard deviation of the five runs are reported in Table \[table:office\_uns\_it\]. The results are similar to the protocol reported in Table \[table:office\_uns\_ft\]. Our approach ATI outperforms the other methods both for CS and OS and the additional outlier handling (ATI-$\lambda$) does not improve the accuracy for the closed set but for the open set.
[**Impact of unknown class.**]{} The linear SVM that we employ in the open set protocol uses the unknown classes of the transformed source domain for the training. Since unknown object samples from the source domain are from different classes than the ones from the target domain, using an SVM that does not require any negative samples might be a better choice. Therefore, we compare the performance of a standard SVM classifier with a specific open set SVM (OS-SVM) [@DATA_Scheirer14], where only the 10 known classes are used for training. OS-SVM introduces an inclusion probability and labels target instances as unknown if this inclusion is not satisfied for any class. Table \[table:office\_wsvm\] compares the classification accuracies of both classifiers in the 6 domain shifts of the Office dataset. While the performance is comparable when no domain adaptation is applied, ATI-$\lambda$ obtains significantly better accuracies when the learning includes negative instances.
As discussed in Section \[sec:uns\], the unknown class is also part of the labelling set $\mathcal{C}$ for the target samples. The labelled target samples are then used to estimate the mapping $W$ (\[eq:objMat\]). To evaluate the impact of including the unknown class, Table \[table:office\_W\] compares the accuracy when the unknown class is not included in $\mathcal{C}$. Adding the unknown class improves the accuracy slightly since it enforces that the negative mean of the source is mapped to a negative sample in the target. The impact, however, is very small.
Additionally, we also analyse the impact of increasing the amount of unknown samples in both source and target domain on the configuration *Amazon* $\rightarrow$ *DSLR+Webcam*. Since the domain shift between *DSLR* and *Webcam* is close to zero (same scenario, but different cameras), they can be merged to get more unknown samples. Following the described protocol, we take 20 samples per known category, also in this case for the target domain, and we randomly increase the number of unknown samples from 20 to 400 in both domains at the same time. As shown in Table \[table:office\_num\_unknown\], that reports the mean accuracies of 5 random splits, adding more unknown samples decreases the accuracy if domain adaptation is not used (LSVM), but also for the domain adaptation method CORAL [@CNN-DA_Sun15]. This is expected since the unknowns are from different classes and the impact of the unknowns compared to the samples from the shared classes increases. Our method handles such an increase and the accuracies remain stable between 80.3% and 82.5%.
[|l|c|c|c|c|c|c|c|c|]{}\
*number of unknowns* & 20 & 40 & 60 & 80 & 100 & 200 & 300 & 400\
*unknown / known* & 0.10 & 0.20 & 0.30 & 0.40 & 0.50 & 1.00 & 1.50 & 2.00\
LSVM & 74.2 & 70.0 & 66.2 & 63.4 & 61.4 & 53.9 & 50.4 & 48.2\
CORAL [@CNN-DA_Sun15] & 77.2 & 76.4 & 76.2 & 74.8 & 73.7 & 71.5 & 70.8 & 69.7\
ATI-$\lambda$ & 80.3 & 82.4 & 81.2 & 81.7 & 82.5 & 80.9 & 80.7 & 81.9\
[**Subsampling of target samples.**]{} In order to evaluate the robustness of our method when having a reduced amount of target samples for domain adaptation, we subsample the target data. Fig. \[fig:numtgt\] shows the results for ATI-$\lambda$ on the 6 domain shifts of the Office dataset with the standard open set protocol (OS). We vary the number of target samples from 50 to the total number of instances. For a fixed number of target samples, we randomly sample 5 times from the target data and plot the lowest, highest and average accuracy of the 5 runs. The accuracy is always measured on the whole target dataset. The results show that between 300 and 400 target instances are sufficient to achieve similar accuracies than our method with all target samples. When the domain shifts are smaller, [e.g.]{}, *D$\rightarrow$ W* and *W$\rightarrow$ D*, even less target samples are required.
[ [**Scalability analysis of target samples.**]{} The number of sampled target samples has an impact on the execution time of the assignment and the transformation steps of the iterative process. Therefore, we also test the scalability of the two steps of our method with respect to the number of target samples. The average execution times of both techniques in the domain shift *Amazon* $\rightarrow$ *DSLR+Webcam* for all the random splits and unknown sets of the previous evaluation are shown in Fig. \[fig:timings\]. We observe that the assignment problem takes less than a second to be solved for any size of target data from the evaluated settings. Most of the computation time is required for estimating the transformation $W$, which requires at least 120 seconds. The computation time of this step, however, increases only moderately with respect to the number of target samples. ]{}
[**Impact of parameter $\rho$.**]{} The cost that determines whether a target sample is considered as outlier during the assignment process is defined by $\lambda$ , which is based on the current minimum and maximum distance between the source clusters and target samples. Thus, $\lambda$ is updated at each iteration. The value of $\lambda$, however, also depends on the parameter $\rho$. For all experiments, we use $\rho = 0.5$ as default value, aiming for a moderate outlier rejection. Fig. \[fig:lambda\] shows the impact of $\rho$ on the accuracy. Using $\rho = 0.5$, which rejects around 10-20$\%$ of the target samples, achieves the best results in 5 out of the 6 domain shifts on the Office dataset. When $\rho$ gets closer to 0 the accuracy drops substantially since too many samples are discarded. [ [**Impact of constraint $\sum_{t} x_{ct} \geq 1$.**]{} Our formulation in (\[eq:assignments\]) ensures that at least one target sample is assigned to an object category. Therefore, all classes contribute to the estimation of the transformation matrix $W$. In order to measure its impact on the adaptation problem, we run experiments with $\sum_{t} x_{ct} \geq 1$ and without the constraint, i.e., when a class might not be assigned to any target sample at all. As illustrated in Fig. \[fig:lambda\], the inclusion of this constraint provides higher accuracies when $\rho < 0.3$. For greater values of $\rho$, the constraint can be omitted since it does not influence the accuracy. ]{}
[ [**Impact of wrong assignments.**]{} During the iterative process of our method, wrong assignments take part in the optimisation of $W$, introducing false associations between the source and the target domain that negatively affect the final transformation. A general assumption in our method is that the correct assignments largely compensate the wrong ones and, thus, the transformed source data allows for better classification accuracies in the target domain. Therefore, we artificially generate assignments in the first iteration by assigning a random subset of target samples to the correct class in the source domain and the remaining target samples to random classes. We then run our approach without any additional modifications until it converges. We report in Table \[table:gt\_rnd\_evolution\] the average percentage of correct assignments of 5 random splits for the domain shift *Amazon* $\rightarrow$ *DSLR+Webcam* with 400 unknown samples. While the first iteration represents the accuracy of correct and random assignments that we generate, the last row shows the accuracies after the approach has converged. As it can be observed, the approach ends in a local optimum, but the accuracies increase for all cases except if we initialise the approach with 100% correct assignments. It is expected that the assignment accuracy does not remain at 100% since the image manifolds are not perfectly linearised and even for the best estimate of $W$ wrong assignments can occur. ]{}
[|c|c|c|c|c|c|c|c|c|c|c||c|]{}\
$\%$gt (+rnd) & 10 & 20 & 30 & 40 & 50 & 60 & 70 & 80 & 90 & 100 & *std*\
iteration 1 & 18.2 & 27.0 & 36.1 & 45.2 & 54.3 & 63.5 & 72.7 & 81.7 & 90.7 & 100.0 & *85.1*\
final & 24.4 & 40.1 & 54.7 & 65.4 & 72.8 & 79.2 & 83.6 & 88.8 & 93.1 & 96.7 & *88.6*\
[**Semi-supervised domain adaptation.**]{} We also evaluate our approach for open set domain adaptation on the *Office* dataset in its semi-supervised setting. Applying again the standard protocol of [@DA_Saenko10] with the subset of source samples, we also take 3 labelled target samples per class and leave the rest unlabelled. We compare our method with the deep learning method MMD [@CNN-DA_Tzeng14]. As baselines, we report the accuracy for the linear SVMs without domain adaptation (LSVM) when they are trained only on the source samples (s), only on the annotated target samples (t) or on both (st). As expected, the baseline trained on both performs best as shown in Table \[table:office\_sup\]. Our approach ATI outperforms the baseline and the CNN approach [@CNN-DA_Tzeng14]. As in the unsupervised case, the improvement compared to the CNN approach is larger for the open set (+4.8%) than for the closed set (+2.2%). While the locality constrained formulation, ATI-$\lambda$-$N$, decreased the accuracy for the unsupervised setting, it improves the accuracy for the semi-supervised case since the formulation enforces that neighbours of the target samples are assigned to the same class. The results with one (ATI-$\lambda$-$N1$) or two neighbours (ATI-$\lambda$-$N2$) are similar.
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10)
LSVM (s) 85.8$\pm$3.2 62.1$\pm$7.9 65.9$\pm$6.2 76.4$\pm$2.1 45.7$\pm$5.0 50.4$\pm$4.5
LSVM (t) 92.3$\pm$3.9 68.2$\pm$5.2 71.1$\pm$4.7 91.5$\pm$4.9 59.6$\pm$3.7 63.2$\pm$3.4
LSVM (st) 95.7$\pm$1.3 82.5$\pm$3.0 84.0$\pm$2.6 92.4$\pm$1.8 72.5$\pm$3.7 74.8$\pm$3.4
MMD [@CNN-DA_Tzeng14] 94.1$\pm$2.3 86.1$\pm$2.3 86.8$\pm$2.2 92.4$\pm$2.8 76.4$\pm$1.5 78.3$\pm$1.3
ATI 95.4$\pm$1.3 89.0$\pm$1.4 89.7$\pm$1.3 95.9$\pm$1.3 84.0$\pm$1.7 85.1$\pm$1.5
ATI-$\lambda$ 97.1$\pm$1.1 **89.5$\pm$1.4** 90.2$\pm$1.3 96.1$\pm$2.0 84.1$\pm$1.8 85.2$\pm$1.5
ATI-$\lambda$-N1 97.6$\pm$1.0 **89.5$\pm$1.3** **90.3$\pm$1.2** **96.4$\pm$1.7** **84.4$\pm$3.6** **85.5$\pm$1.5**
ATI-$\lambda$-N2 **97.9$\pm$1.4** 89.4$\pm$1.2 90.1$\pm$1.0 92.8$\pm$1.6 84.3$\pm$2.4 85.4$\pm$1.5
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
: Open set domain adaptation on the semi-supervised Office dataset with 10 shared classes (OS). We report the average and the standard deviation using a subset of samples per class in 5 random splits [@DA_Saenko10].[]{data-label="table:office_sup"}
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10)
LSVM (s) 85.2$\pm$1.7 40.3$\pm$4.3 45.2$\pm$3.8 97.2$\pm$0.7 81.4$\pm$2.4 83.0$\pm$2.2
LSVM (t) 88.7$\pm$2.2 52.8$\pm$6.0 57.0$\pm$5.5 91.5$\pm$4.9 59.6$\pm$3.7 63.2$\pm$3.4
LSVM (st) 91.9$\pm$0.7 68.7$\pm$2.5 71.2$\pm$2.3 98.7$\pm$0.9 87.3$\pm$2.3 88.5$\pm$2.1
MMD [@CNN-DA_Tzeng14] 90.2$\pm$1.8 69.0$\pm$3.4 71.3$\pm$3.0 98.5$\pm$1.0 85.5$\pm$1.6 86.7$\pm$1.4
ATI **93.5$\pm$0.2** 74.4$\pm$2.7 76.1$\pm$2.5 98.7$\pm$0.7 91.6$\pm$1.7 92.4$\pm$1.5
ATI-$\lambda$ **93.5$\pm$0.2** 74.4$\pm$2.5 76.2$\pm$2.3 98.7$\pm$0.8 91.6$\pm$1.7 92.4$\pm$1.5
ATI-$\lambda$-N1 93.4$\pm$0.2 74.6$\pm$2.5 76.4$\pm$2.3 98.9$\pm$0.5 92.0$\pm$1.6 92.7$\pm$1.5
ATI-$\lambda$-N2 **93.5$\pm$0.1** **74.9$\pm$2.3** **76.7$\pm$2.1** **99.3$\pm$0.5** **92.2$\pm$1.9** **92.9$\pm$1.7**
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
: Open set domain adaptation on the semi-supervised Office dataset with 10 shared classes (OS). We report the average and the standard deviation using a subset of samples per class in 5 random splits [@DA_Saenko10].[]{data-label="table:office_sup"}
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ---------- ---------- ----------
CS (10) OS$^*$ (10) OS (10) CS (10) OS$^*$ (10) OS (10) CS OS$^*$ OS
LSVM (s) 78.8$\pm$2.9 32.4$\pm$3.8 38.2$\pm$3.4 99.5$\pm$0.3 88.7$\pm$2.2 89.6$\pm$1.9 87.1 58.4 62.0
LSVM (t) 88.7$\pm$2.2 52.8$\pm$6.0 57.0$\pm$5.5 92.3$\pm$3.9 68.2$\pm$5.2 71.1$\pm$4.7 90.9 60.2 63.8
LSVM (st) 90.8$\pm$1.3 66.2$\pm$4.4 69.0$\pm$4.1 99.4$\pm$0.7 93.5$\pm$2.7 94.0$\pm$2.5 94.8 78.4 80.3
MMD [@CNN-DA_Tzeng14] 89.1$\pm$3.2 65.1$\pm$3.8 67.8$\pm$3.4 98.2$\pm$1.4 93.9$\pm$2.9 94.4$\pm$2.7 93.8 79.3 80.9
ATI **93.0$\pm$0.5** 71.3$\pm$4.6 74.3$\pm$4.3 99.3$\pm$0.6 96.3$\pm$1.8 96.6$\pm$1.7 96.0 84.4 85.7
ATI-$\lambda$ **93.0$\pm$0.5** 71.5$\pm$4.8 73.6$\pm$4.4 **99.5$\pm$0.6** 96.3$\pm$1.8 96.6$\pm$1.7 96.3 84.6 85.7
ATI-$\lambda$-N1 **93.0$\pm$0.6** 72.2$\pm$4.5 74.2$\pm$4.1 99.3$\pm$0.6 **96.7$\pm$2.1** **97.0$\pm$1.9** 96.4 **84.9** **86.0**
ATI-$\lambda$-N2 **93.0$\pm$0.6** **72.8$\pm$4.2** **74.8$\pm$3.9** 99.3$\pm$0.6 95.5$\pm$2.2 95.9$\pm$2.0 **96.6** 84.8 **86.0**
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ---------- ---------- ----------
: Open set domain adaptation on the semi-supervised Office dataset with 10 shared classes (OS). We report the average and the standard deviation using a subset of samples per class in 5 random splits [@DA_Saenko10].[]{data-label="table:office_sup"}
### Dense Cross-Dataset Analysis
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10)
LSVM 82.4$\pm$2.4 66.6$\pm$4.0 75.1$\pm$0.4 59.0$\pm$2.7 43.0$\pm$2.0 24.2$\pm$3.0 53.5$\pm$2.1 40.1$\pm$1.9 76.9$\pm$4.3 62.5$\pm$1.2 46.3$\pm$2.7 28.2$\pm$1.4
TCA [@DA_Pan09] 74.9$\pm$3.0 62.8$\pm$3.8 68.4$\pm$4.0 56.6$\pm$4.5 38.3$\pm$1.7 29.6$\pm$4.2 49.2$\pm$1.1 38.9$\pm$1.9 73.1$\pm$3.6 60.2$\pm$1.4 45.9$\pm$3.6 29.7$\pm$1.6
gfk [@DA_Gong12] 82.0$\pm$2.2 66.2$\pm$4.0 74.3$\pm$1.0 58.3$\pm$3.1 42.2$\pm$1.4 23.8$\pm$2.0 53.2$\pm$2.6 40.2$\pm$1.8 77.1$\pm$3.3 62.2$\pm$1.5 46.2$\pm$3.0 28.5$\pm$1.0
SA [@DA_Fernando13] 81.1$\pm$1.8 66.0$\pm$3.4 73.9$\pm$0.9 57.8$\pm$3.2 41.9$\pm$2.4 24.3$\pm$2.6 53.4$\pm$2.5 40.3$\pm$1.7 77.3$\pm$4.2 62.5$\pm$.8 46.1$\pm$3.3 29.0$\pm$1.5
CORAL [@CNN-DA_Sun15] 80.1$\pm$3.5 68.8$\pm$3.3 73.7$\pm$2.0 60.9$\pm$2.6 42.2$\pm$2.4 27.2$\pm$3.9 53.6$\pm$2.9 40.7$\pm$1.5 78.2$\pm$5.1 64.0$\pm$2.6 48.2$\pm$3.9 31.4$\pm$0.8
ATI 86.3$\pm$1.6 **71.4$\pm$1.8** 80.1$\pm$0.7 68.0$\pm$1.9 **49.2$\pm$3.2** 36.8$\pm$1.2 53.2$\pm$3.4 45.4$\pm$3.4 81.7$\pm$3.7 66.7$\pm$4.2 52.0$\pm$3.4 35.8$\pm$1.8
ATI-$\lambda$ **86.7$\pm$1.3** **71.4$\pm$2.3** **80.6$\pm$2.4** **69.0$\pm$2.8** 48.6$\pm$2.5 **37.4$\pm$2.6** **54.2$\pm$1.9** **45.7$\pm$3.0** **82.2$\pm$3.7** **67.9$\pm$4.2** **53.1$\pm$2.8** **37.5$\pm$2.7**
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ---------- ----------
CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10) CS (10) OS (10)
LSVM **59.1$\pm$2.0** 42.7$\pm$2.0 86.2$\pm$2.6 73.3$\pm$3.9 50.1$\pm$4.0 32.1$\pm$3.2 33.1$\pm$1.7 16.4$\pm$1.1 53.1$\pm$2.6 27.9$\pm$2.9 52.3$\pm$1.8 25.2$\pm$0.5 59.3 41.5
TCA [@DA_Pan09] 56.1$\pm$3.8 40.9$\pm$2.9 83.4$\pm$3.2 68.6$\pm$1.8 49.3$\pm$2.6 34.5$\pm$3.8 30.6$\pm$1.3 19.4$\pm$2.1 47.5$\pm$3.5 32.0$\pm$3.9 45.2$\pm$1.9 31.1$\pm$4.6 55.2 42.0
gfk [@DA_Gong12] 58.7$\pm$1.9 42.6$\pm$2.4 86.1$\pm$2.7 73.3$\pm$3.6 49.5$\pm$3.6 32.7$\pm$3.6 33.3$\pm$1.4 16.9$\pm$1.5 53.1$\pm$3.0 28.6$\pm$3.8 52.5$\pm$2.0 26.4$\pm$1.1 59.0 41.6
SA [@DA_Fernando13] 58.7$\pm$1.8 43.1$\pm$1.6 85.9$\pm$2.9 72.8$\pm$3.1 50.0$\pm$3.6 32.2$\pm$3.7 34.2$\pm$1.1 17.5$\pm$1.6 52.5$\pm$3.2 29.2$\pm$4.2 52.6$\pm$2.4 27.1$\pm$1.3 59.0 41.1
CORAL [@CNN-DA_Sun15] 58.5$\pm$2.7 44.6$\pm$2.5 85.8$\pm$1.5 74.5$\pm$3.4 49.5$\pm$4.8 35.4$\pm$4.4 32.9$\pm$1.6 18.7$\pm$1.2 52.1$\pm$2.8 33.6$\pm$5.3 52.9$\pm$1.8 31.3$\pm$1.3 59.0 44.2
ATI 57.9$\pm$1.9 **48.8$\pm$2.3** 89.3$\pm$2.2 77.1$\pm$2.6 55.0$\pm$5.0 42.2$\pm$4.0 **34.9$\pm$2.6** 22.8$\pm$3.1 59.8$\pm$1.3 46.9$\pm$2.5 **60.8$\pm$3.4** 32.9$\pm$2.2 63.4 49.5
ATI-$\lambda$ 58.6$\pm$1.4 48.7$\pm$1.8 **89.7$\pm$2.3** **77.5$\pm$2.2** **55.3$\pm$4.3** **43.4$\pm$4.8** 34.1$\pm$2.4 **23.2$\pm$3.2** **60.2$\pm$2.7** **47.3$\pm$2.9** 60.3$\pm$2.4 **33.0$\pm$1.1** **63.6** **50.2**
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ---------- ----------
In order to measure the performance of our method and the open set protocol across popular datasets with more intra-class variation, we also conduct experiments on the *dense* set-up of the *Testbed for Cross-Dataset Analysis* [@DATA_Tommasi14]. This protocol provides 40 classes from 4 well known datasets, *Bing (B)*, *Caltech256 (C)*, *ImageNet (I)* and *Sun (S)*. While the samples from the first 3 datasets are mostly centred and without occlusions, *Sun* becomes more challenging due to its collection of object class instances from cluttered scenes. As for the Office dataset, we take the first 10 classes as shared classes, the classes 11-25 are used as unknowns in the source domain and the classes 26-40 as unknowns in the target domain. We use the provided DeCAF features (DeCAF7). Following the unsupervised protocol described in [@DATA_Tommasi15], we take 50 source samples per class for training and we test on 30 target images per class for all datasets, except *Sun*, where we take 20 samples per class.
The results reported in Table \[table:testbed\] are consistent with the Office dataset. ATI outperforms the baseline and the other methods by +4.1% for the closed set and by +5.3% for the open set. ATI-$\lambda$ obtains the best accuracies for the open set.
### Sparse Cross-Dataset Analysis
C$\rightarrow$O C$\rightarrow$P O$\rightarrow$C O$\rightarrow$P P$\rightarrow$C P$\rightarrow$O AVG.
----------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------
*shared classes* 8 7 8 4 7 4
*unknown / all (t)* 0.52 0.30 0.90 0.81 0.54 0.78
LSVM 46.3 36.1 60.8 29.7 78.8 70.1 53.6
TCA [@DA_Pan09] 45.2 33.8 58.1 31.1 63.4 61.1 48.8
gfk [@DA_Gong12] 46.4 36.2 61.0 29.7 79.1 **72.6** 54.2
SA [@DA_Fernando13] 46.4 36.8 61.1 30.2 79.8 71.1 54.2
CORAL [@CNN-DA_Sun15] 48.0 35.9 60.2 29.1 78.9 68.8 53.5
ATI **51.6** **52.1** 63.1 38.8 80.6 70.9 59.5
ATI-$\lambda$ 51.5 52.0 **63.4** **39.1** **81.1** 71.1 **59.7**
: Unsupervised open set domain adaptation on the sparse set-up from [@DATA_Tommasi14].[]{data-label="table:open_uns"}
C$\rightarrow$O C$\rightarrow$P O$\rightarrow$C O$\rightarrow$P P$\rightarrow$C P$\rightarrow$O AVG.
------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ----------
LSVM (s) 46.5$\pm$0.1 36.2$\pm$0.1 60.8$\pm$0.3 29.7$\pm$0.0 79.5$\pm$0.3 73.5$\pm$0.7 54.4
LSVM (t) 53.1$\pm$3.7 44.6$\pm$2.1 73.7$\pm$1.5 40.5$\pm$3.0 81.1$\pm$2.5 70.5$\pm$4.3 60.6
LSVM (st) 56.0$\pm$1.3 44.5$\pm$1.2 68.9$\pm$1.1 40.9$\pm$2.2 80.9$\pm$0.6 76.7$\pm$0.3 61.3
ATI 59.6$\pm$1.2 55.2$\pm$1.3 75.8$\pm$1.2 45.2$\pm$1.4 81.6$\pm$0.2 **77.1$\pm$0.8** 65.8
ATI-$\lambda$ 60.3$\pm$1.2 56.0$\pm$1.2 75.8$\pm$1.1 **45.8$\pm$1.2** 81.8$\pm$0.2 76.9$\pm$1.3 66.1
ATI-$\lambda$-N1 **60.7$\pm$1.2** **56.3$\pm$1.2** **76.7$\pm$1.6** **45.8$\pm$1.4** **82.0$\pm$0.4** 76.7$\pm$1.1 **66.4**
: Semi-supervised open set domain adaptation on the sparse set-up from [@DATA_Tommasi14] with 3 labelled target samples per shared class.[]{data-label="table:open_supOffice"}
We also introduce an open set evaluation using the *sparse* set-up from [@DATA_Tommasi14] with the datasets *Caltech101 (C)*, *Pascal07 (P)* and *Office (O)*. These datasets are quite unbalanced and offer distinctive characteristics: *Office* contains centred class instances with barely any background (17 classes, 2300 samples in total, 68-283 samples per class), *Caltech101* allows for more class variety (35 classes, 5545 samples in total, 35-870 samples per class) and *Pascal07* gathers more realistic scenes with partially occluded objects in various image locations (16 classes, 12219 samples in total, 193-4015 samples per class). For each domain shift, we take all samples of the shared classes and consider all other samples as unknowns. Table \[table:open\_uns\] summarises the amount of shared classes for each shift and the percentage of unknown target samples, which varies from 30% to 90%.
[**Unsupervised domain adaptation.**]{} For the unsupervised experiment, we conduct a single run for each domain shift using all source and unlabelled target samples. The results are reported in Table \[table:open\_uns\]. ATI outperforms the baseline and the other methods by +5.3% for this highly unbalanced open set protocol. ATI-$\lambda$ improves the accuracy of ATI slightly.
[**Semi-supervised domain adaptation.**]{} In order to evaluate the semi-supervised setting, we take all source samples and 3 annotated target samples per shared class as it is done in the semi-supervised setting for the Office dataset [@DA_Saenko10]. The average and standard deviation over 5 random splits are reported in Table \[table:open\_supOffice\]. While ATI improves over the baseline trained on the source and target samples together (st) by +4.5%, ATI-$\lambda$ and the locality constraints with one neighbour boost the performance further. ATI-$\lambda$-$N_1$ improves the accuracy of the baseline by +5.1%.
### Action recognition
We extend the applicability of our technique to the field of action recognition in video sequences. We introduce an open set domain adaptation protocol between the *Kinetics Human Action Video Dataset* [@Data_Kay17] (Kinetics) and the *UCF101 Action Recognition Dataset* [@Data_Soomro12] (UCF101). The Kinects dataset is used as source domain and contains a total of 400 human action classes. The UCF101 dataset serves as target domain including 101 action categories, mainly of sports events. Since the labels of the same action differ between the datasets, [e.g.]{}, *massaging persons head* (Kinetics) and *head massage* (UCF101), we manually map the class labels between the datasets. Additionally, we also merge all action classes in one datasets if they correspond to a single class in the other dataset, [e.g.]{}, *dribbling basketball*, *playing basketball*, *shooting basketball* (Kinetics) are merged and associated to *basketball* (UCF101). We finally obtain an open set protocol with 66 shared action classes. The list of shared classes, as well as all unrelated categories between both datasets, are provided in the supplemental material. For action recognition, we use the features extracted from the 5c layer of the spatial and temporal stream of the I3D model [@3DCNN_Carreira17], which is pretrained on Kinetics [@Data_Kay17]. We forward the complete video sequences through the spatial and temporal stream of I3D [@3DCNN_Carreira17] and the 5c layer of each stream provides an $7 \times 7 \times 1024$ output for a temporal fragment. We then apply spatial average pooling using a $7\times 7$ kernel and average over time to obtain a $1024$-dimensional feature vector from both the spatial and temporal stream of the I3D model [@3DCNN_Carreira17]. Finally, the feature vectors from the spatial and temporal streams are concatenated to get a single $2048$-dimensional feature vector per video sequence.
[**Unsupervised domain adaptation.**]{} In the unsupervised setting, we evaluate our method by taking all source samples in a single run. Table \[table:ar\] shows that the proposed approach outperforms the baseline and other approaches. ATI-$\lambda$ achieves the highest accuracy and improves the accuracy by +12.0$\%$ compared to LSVM. The resulting confusion matrices of LSVM and ATI-$\lambda$ are shown in Fig. \[fig:ar\]. LSVM misclassifies many instances of shared classes in the target domain as unknown instances (last column of confusion matrix), which is a well-known problem for open set recognition. Although ATI-$\lambda$ does not resolve this problem completely, it reduces this effect substantially.
[|c|c|c|c|c|c|c|]{}\
LSVM & TCA [@DA_Pan09] & gkf [@DA_Gong12] & SA [@DA_Fernando13] & CORAL [@CNN-DA_Sun15] & ATI & ATI-$\lambda$\
64.9 & 71.2 & 64.9 & 65.1 & 69.4 & 76.6 & **76.9**\
[**Semi-supervised domain adaptation.**]{} We extend the unsupervised protocol to evaluate our method on a semi-supervised setting by labelling 3 target samples per shared class. We report the average accuracies of 5 random splits in Table \[table:ar\_sup\]. Like in the previous semi-supervised experiments, ATI-$\lambda$-N1 obtains the best classification accuracies, outperforming the baseline without adaptation, LSVM (st), by +11.0%.
[|c|c|c|c|]{}\
LSVM (st) & ATI & ATI-$\lambda$ & ATI-$\lambda$-N1\
73.5$\pm$0.5 & 84.1$\pm$0.7 & 84.2$\pm$0.8 & **84.5$\pm$0.6**\
### Synthetic data
We also introduce another open set protocol with a domain shift between synthetic and real data. In this case, we take 152,397 synthetic images of the VISDA’17 challenge [@DATA_Peng17] as source domain and 5970 instances of real images from the training data of the Pascal3D dataset [@Data_Xiang14] as target domain. Since both datasets contain several types of vehicles, we obtain 6 shared classes, namely, *aeroplane*, *bicycle*, *bus*, *car*, *motorbike* and *train*, within the 12 categories of each dataset. Following the protocol used in Section \[exp:office\], we extract deep features from the fully connected layer-7 (fc7) from the AlexNet model [@CNN_Krizhevsky12] with 4096 dimensions. In addition, we also extract features from the VGG-16 model [@SimonyanZ14a] to evaluate the impact of using deeper features.
The results of the classification task are shown in Table \[table:syn\]. The proposed domain adaptation method achieves the best results for both types of CNN features. When we compare the performance of the deep features from AlexNet and VGG-16, the accuracy of the baseline (LSVM) increases by +5.6$\%$ when using the deeper network VGG-16 instead of AlexNet. ATI and ATI-$\lambda$, however, benefit even more from the deeper architecture. For instance, the accuracy of ATI-$\lambda$ increases by +10.5$\%$. This coincides with the observation that deeper networks have a stronger linearisation effect on manifolds of image domains [@BengioMDR13; @UpchurchGPPSBW17] than shallow networks. Since the proposed approach learns a linear mapping from the feature space of the source domain to the feature space of the target domain, it benefits from a better linearisation. The confusion matrices of the classification task with features extracted from the VGG-16 model are shown in Fig. \[fig:syn\]. ATI-$\lambda$ improves the overall accuracy of LSVM by +18.3$\%$ since it resolves confusions between similar classes. For instance, LSVM frequently misclassifies *bicycle* as *motorbike* and *car* as instances of trucks, which are part of the unknown class.
--------- ------ ----------------- ------------------ --------------------- ----------------------- ---------- ---------------
LSVM TCA [@DA_Pan09] gkf [@DA_Gong12] SA [@DA_Fernando13] CORAL [@CNN-DA_Sun15] ATI ATI-$\lambda$
AlexNet 48.0 49.7 50.1 51.2 52.0 61.1 **61.4**
VGG-16 53.6 55.0 55.2 56.5 60.0 **72.0** 71.9
--------- ------ ----------------- ------------------ --------------------- ----------------------- ---------- ---------------
: Open set domain adaptation using synthetic images from the VISDA’17 challenge [@DATA_Peng17] as source and real data from the Pascal3D dataset [@Data_Xiang14] as target dataset. There are 6 shared classes between both datasets.[]{data-label="table:syn"}
Closed set domain adaptation
----------------------------
We also report the accuracies of our method for popular domain adaptation datasets using the standard closed set protocols, where all classes are known in both domains.
### Office dataset {#office-dataset}
For the *Office* dataset [@DA_Saenko10], we run experiments for the 6 domain shifts of the three provided datasets and use deep features extracted from the fc7 feature map from the AlexNet [@CNN_Krizhevsky12] and VGG-16 [@SimonyanZ14a] models.
[**Unsupervised domain adaptation.**]{} For unsupervised domain adaptation, we first report the results for the protocol from [@DA_Saenko10], where we run 5 experiments for each domain shift using randomised samples of the source dataset. The results are shown in Table \[table:office\_uns\], where we compare our method with generic domain adaptation methods, [i.e.]{}, TCA [@DA_Pan09], gfk [@DA_Gong12], SA [@DA_Fernando13] and CORAL [@CNN-DA_Sun15] using AlexNet features. The results are in accordance with the observations from Section \[exp:office\]. While ATI outperforms all generic domain adaptation methods in average and ATI-$\lambda$ performs slightly better than ATI, ATI-$\lambda$-$N_1$ decreases the accuracy in the unsupervised setting. [In addition, we also include the accuracies of using nearest neighbours without domain adaptation, NN, which reports significant lower accuracies than LSVM. LSVM also outperforms NN in other closed set evaluation protocols by a large margin.]{}
A$\rightarrow$D A$\rightarrow$W D$\rightarrow$A D$\rightarrow$W W$\rightarrow$A W$\rightarrow$D AVG.
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ----------
NN 51.3$\pm$1.4 45.7$\pm$2.1 26.0$\pm$0.9 65.5$\pm$1.4 28.0$\pm$0.5 69.8$\pm$1.8 47.7
LSVM 62.3$\pm$3.8 55.8$\pm$3.1 42.8$\pm$1.6 90.1$\pm$0.6 41.2$\pm$0.4 92.6$\pm$1.5 64.1
TCA [@DA_Pan09] 60.3$\pm$4.0 54.7$\pm$3.0 49.4$\pm$1.6 90.7$\pm$0.4 46.9$\pm$2.3 92.0$\pm$0.9 65.7
gfk [@DA_Gong12] 61.3$\pm$3.7 55.7$\pm$3.0 45.6$\pm$1.6 90.6$\pm$0.4 43.1$\pm$2.3 93.4$\pm$0.9 65.0
SA [@DA_Fernando13] 60.6$\pm$3.5 55.0$\pm$3.1 47.3$\pm$1.6 90.9$\pm$0.6 44.4$\pm$1.4 93.3$\pm$0.8 65.3
CORAL [@CNN-DA_Sun15] 64.4$\pm$3.9 58.9$\pm$3.3 52.1$\pm$1.2 **92.6$\pm$0.3** 50.0$\pm$1.0 **94.0$\pm$0.6** 68.7
ATI **67.6$\pm$3.0** 62.3$\pm$3.1 54.8$\pm$1.3 90.3$\pm$0.8 52.4$\pm$2.1 92.6$\pm$1.7 70.0
ATI-$\lambda$ 67.3$\pm$2.3 **62.6$\pm$2.5** **55.2$\pm$2.6** 90.1$\pm$0.6 **53.4$\pm$2.5** 92.7$\pm$2.5 **70.2**
ATI-$\lambda$-$N_1$ 64.6$\pm$2.9 60.9$\pm$1.3 51.9$\pm$1.9 90.2$\pm$0.9 48.1$\pm$1.6 93.7$\pm$2.1 68.2
: Comparison on the unsupervised Office dataset [@DA_Saenko10] with 31 shared classes and 6 domain shifts using the protocol from [@DA_Saenko10] and features from the AlexNet model (fc7 layer).[]{data-label="table:office_uns"}
------------------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------
A$\rightarrow$D A$\rightarrow$W D$\rightarrow$A D$\rightarrow$W W$\rightarrow$A W$\rightarrow$D AVG.
NN 55.9 49.7 27.4 75.3 31.5 86.2 54.3
LSVM 65.7 60.3 43.2 94.7 44.0 98.9 67.8
DAN [@CNN-DA_Long15] 66.8 68.5 50.0 96.0 49.8 99.0 71.7
DAH [@CNN-DA_Venkateswara17] 66.5 68.3 55.5 96.1 53.0 98.8 73.0
RTN [@CNN-DA_Long16] **71.0** 73.3 50.5 96.8 51.0 **99.6** 73.7
BP [@CNN-DA_Ganin15] - 73.0 - 96.4 - 99.2 -
ADDA [@CNN-DA_Tzeng17] - **75.1** - **97.0** - **99.6** -
ATI 70.3 68.7 55.3 95.0 **56.9** 98.7 **74.2**
ATI-$\lambda$ 69.0 67.0 **56.2** 95.0 **56.9** 98.7 73.8
NN 61.3 55.4 33.1 78.6 49.4 88.8 61.1
LSVM 76.1 68.6 55.3 95.9 61.5 99.6 76.2
DAN [@CNN-DA_Long15] 74.4 76.0 61.5 95.9 60.3 98.6 77.8
AutoDIAL [@CNN-DA_Carlucci17] **82.3** **84.2** 64.6 **97.9** 64.2 **99.9** **82.2**
ATI 80.6 81.4 **67.1** 96.1 66.4 99.3 81.8
ATI-$\lambda$ 80.8 81.3 66.9 96.1 **66.5** 98.9 81.8
------------------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------
: Comparison on the unsupervised Office dataset [@DA_Saenko10] with 31 shared classes and 6 domain shifts taking all source samples as in [@DA_Gong13]. []{data-label="table:office_uns_ft_CS"}
We also compare our method with current state-of-the-art CNN-based domain adaptation methods [@CNN-DA_Long15; @CNN-DA_Long16; @CNN-DA_Ganin15; @CNN-DA_Venkateswara17; @CNN-DA_Tzeng17; @CNN-DA_Carlucci17]. In this case, we report the accuracies when all source samples are used in a single run as described by [@DA_Gong13]. As shown in Table \[table:office\_uns\_ft\_CS\], our method achieves competitive results even for the standard closed set protocol.
-------------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ----------
A$\rightarrow$D A$\rightarrow$W D$\rightarrow$A D$\rightarrow$W W$\rightarrow$A W$\rightarrow$D AVG.
LSVM (st) 82.6$\pm$5.5 77.0$\pm$2.5 63.4$\pm$1.6 94.0$\pm$0.8 61.8$\pm$1.1 96.3$\pm$0.8 79.2
DDC [@CNN-DA_Tzeng14] - 84.1$\pm$0.6 - 95.4$\pm$0.4 - 96.3$\pm$0.3 -
DAN [@CNN-DA_Long15] - **85.7$\pm$0.3** - **97.2$\pm$0.2** - 96.4$\pm$0.2 -
MMC [@CNN-DA_Tzeng15] 86.1$\pm$1.2 82.7$\pm$0.8 **66.2$\pm$0.3** 95.7$\pm$0.5 65.0$\pm$0.5 **97.6$\pm$0.2** **82.2**
ATI (labels t) 85.0$\pm$2.1 78.3$\pm$2.3 63.6$\pm$1.5 94.0$\pm$0.8 62.3$\pm$0.9 96.4$\pm$0.8 79.9
ATI 85.5$\pm$2.9 82.4$\pm$1.1 65.1$\pm$1.3 93.4$\pm$0.9 65.6$\pm$1.5 95.7$\pm$1.1 81.3
ATI-$\lambda$ 85.6$\pm$2.6 82.6$\pm$0.5 65.3$\pm$1.3 93.3$\pm$1.0 65.7$\pm$1.7 95.7$\pm$1.1 81.4
ATI-$\lambda$-$N_1$ **88.1$\pm$1.7** 83.1$\pm$2.3 66.0$\pm$1.4 93.9$\pm$1.2 **65.9$\pm$1.5** 96.2$\pm$0.8 **82.2**
ATI-$\lambda$-$N_2$ 87.0$\pm$3.5 84.6$\pm$3.5 65.3$\pm$1.0 93.6$\pm$1.2 **65.9$\pm$1.8** 95.8$\pm$1.3 82.0
LSVM (st) 86.1$\pm$1.5 83.4$\pm$1.2 67.9$\pm$1.0 96.1$\pm$0.7 67.1$\pm$0.6 **96.6$\pm$1.0** 82.9
SO [@CNN-DA_Koniusz17] 84.5$\pm$1.7 86.3$\pm$0.8 65.7$\pm$1.7 97.5$\pm$0.7 66.5$\pm$1.0 95.5$\pm$0.6 82.7
CCSA [@CNN-DA_Motiian17] 88.2$\pm$1.0 **89.0$\pm$1.2** **72.1$\pm$1.0** **97.6$\pm$0.4** **71.8$\pm$0.5** 96.4$\pm$0.8 **85.8**
ATI-$\lambda$-$N_1$ **90.3$\pm$1.9** 88.0$\pm$1.4 70.8$\pm$0.9 95.1$\pm$0.7 70.3$\pm$2.0 96.3$\pm$0.9 85.1
-------------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ----------
: Comparison on the semi-supervised Office dataset [@DA_Saenko10] with 31 shared classes and 6 domain shifts, following the protocol from [@DA_Saenko10]. []{data-label="table:office_sup_cs"}
[**Semi-supervised domain adaptation.**]{} We also evaluate our approach for semi-supervised domain adaptation on the *Office* dataset. We follow the protocol from [@DA_Saenko10] and report the accuracies and standard deviations over 5 runs with random samples. In the first experiment with AlexNet features, we also include ATI-$\lambda$-$N_2$ with locality constraints using 2 nearest neighbours and compare our approach with state-of-the-art CNN-based methods [@CNN-DA_Tzeng14; @CNN-DA_Long15; @CNN-DA_Tzeng15]. As in Section \[exp:office\], we train the SVMs on the transformed source samples and labelled target samples (st). The results are reported in Table \[table:office\_sup\_cs\].
Our method achieves the same average accuracy as MMC [@CNN-DA_Tzeng15] and performs slightly worse than [@CNN-DA_Motiian17] for the VGG-16 features. In addition, we report the accuracy for AlexNet features when the mapping $W$ is estimated using only the labelled target samples without solving the individual assignments . This variant is denoted by ATI (labels t) and performs worse than ATI.
### Office+Caltech dataset
----------------------- ----------------- ----------------- ----------------- ----------------- ----------------- -----------------
A$\rightarrow$C A$\rightarrow$D A$\rightarrow$W C$\rightarrow$A C$\rightarrow$D C$\rightarrow$W
NN 78.4 78.1 71.7 90.7 84.4 80.8
LSVM 83.3 84.1 77.5 91.8 89.1 82.3
CORAL [@CNN-DA_Sun15] 83.2 86.5 79.6 91.4 86.6 82.1
BP [@CNN-DA_Ganin15] 84.6 92.3 90.2 91.9 92.8 93.2
DDC [@CNN-DA_Tzeng14] 83.5 88.4 83.1 91.9 88.8 85.4
DAN [@CNN-DA_Long15] 84.1 91.1 91.8 92.0 89.3 90.6
RTN[@CNN-DA_Long16] **88.1** **95.5** **95.2** 93.7 **94.2** **96.9**
ATI 86.5 92.8 88.7 **93.8** 89.6 93.6
ATI-$\lambda$ 87.1 90.6 90.7 93.4 85.4 93.4
NN 86.7 84.4 83.4 91.4 88.2 88.0
LSVM 87.8 88.7 87.2 93.3 91.8 91.4
ATI 91.0 92.4 95.9 94.7 93.1 97.4
ATI-$\lambda$ 90.4 92.4 91.4 94.5 93.9 96.0
----------------------- ----------------- ----------------- ----------------- ----------------- ----------------- -----------------
: Classification accuracies on the unsupervised Office+Caltech dataset [@DA_Gong12] with 10 shared classes and 12 domain shifts using deep features. We take all source samples on a single run [@DA_Gong13]. []{data-label="table:saenko_uns_ft"}
----------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------
D$\rightarrow$A D$\rightarrow$C D$\rightarrow$W W$\rightarrow$A W$\rightarrow$C W$\rightarrow$D AVG
NN 64.2 58.6 89.0 63.2 58.8 95.4 76.1
LSVM 79.4 70.2 97.9 80.0 72.7 **100.0** 84.0
CORAL [@CNN-DA_Sun15] 87.3 77.5 **99.3** 85.2 76.1 **100.0** 86.2
BP [@CNN-DA_Ganin15] 84.0 74.9 97.8 86.9 77.3 **100.0** 88.2
DDC [@CNN-DA_Tzeng14] 89.0 79.2 98.1 84.9 73.4 **100.0** 87.1
DAN [@CNN-DA_Long15] 90.0 80.3 98.5 92.1 81.2 **100.0** 90.1
RTN[@CNN-DA_Long16] **93.8** 84.6 99.2 **95.5** **86.6** **100.0** **93.4**
ATI 93.4 **85.9** 98.9 93.6 86.3 **100.0** 91.9
ATI-$\lambda$ 93.6 85.8 **99.3** 93.6 86.1 **100.0** 91.8
NN 78.9 75.0 95.2 80.9 78.5 100.0 85.6
LSVM 82.5 77.9 98.4 87.8 84.9 100.0 89.3
ATI 93.7 89.8 98.1 95.1 90.3 99.5 94.3
ATI-$\lambda$ 94.6 89.4 98.4 95.3 89.4 99.6 93.8
----------------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------------- ----------
: Classification accuracies on the unsupervised Office+Caltech dataset [@DA_Gong12] with 10 shared classes and 12 domain shifts using deep features. We take all source samples on a single run [@DA_Gong13]. []{data-label="table:saenko_uns_ft"}
We also evaluate our approach on the extended version of the Office evaluation set [@DA_Gong12], which includes the additional *Caltech (C)* dataset. This results in 12 domain shifts, but reduces the amount of shared classes to only 10. As shown in Table \[table:saenko\_uns\_ft\], our method obtains very competitive results with AlexNet features, outperforming in overall the generic domain adaptation method [@CNN-DA_Sun15] and 3 out of 4 CNN-based methods. If features from a deeper network such as VGG-16 are used, our method obtains the best overall results.
### Dense Testbed for Cross-Dataset Analysis
B$\rightarrow$C B$\rightarrow$I B$\rightarrow$S C$\rightarrow$B C$\rightarrow$I C$\rightarrow$S
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------
LSVM 63.8$\pm$2.2 57.4$\pm$0.7 20.2$\pm$1.0 38.3$\pm$0.8 62.9$\pm$0.9 21.7$\pm$1.6
TCA [@DA_Pan09] 53.8$\pm$1.3 49.1$\pm$1.1 17.1$\pm$1.1 35.6$\pm$1.8 59.2$\pm$0.8 18.9$\pm$1.2
gfk [@DA_Gong12] 63.4$\pm$1.8 57.2$\pm$1.1 20.6$\pm$1.3 38.3$\pm$0.9 62.9$\pm$1.2 21.7$\pm$1.4
SA [@DA_Fernando13] 63.0$\pm$1.9 57.1$\pm$1.4 20.2$\pm$1.4 38.3$\pm$0.9 62.8$\pm$1.0 21.5$\pm$1.2
CORAL [@CNN-DA_Sun15] 63.9$\pm$2.1 57.8$\pm$0.8 20.4$\pm$2.0 38.3$\pm$0.8 63.4$\pm$0.9 22.5$\pm$1.2
ATI 69.1$\pm$1.3 62.4$\pm$1.9 23.4$\pm$1.1 **39.0$\pm$1.4** **66.9$\pm$1.2** 25.2$\pm$0.9
ATI-$\lambda$ **69.4$\pm$1.4** **62.9$\pm$1.3** **23.6$\pm$1.0** **39.0$\pm$1.4** **66.9$\pm$1.1** **25.3$\pm$0.9**
: *Testbed* dataset [@DATA_Tommasi14] with 40 common classes and 12 domain shifts.[]{data-label="table:testbed_dense"}
I$\rightarrow$B I$\rightarrow$C I$\rightarrow$S S$\rightarrow$B S$\rightarrow$C S$\rightarrow$I AVG
----------------------- ------------------ ------------------ ------------------ ------------------ ------------------ ------------------ ----------
LSVM 39.3$\pm$1.4 70.8$\pm$1.5 24.6$\pm$1.8 16.6$\pm$1.0 26.1$\pm$2.0 26.3$\pm$0.7 39.0
TCA [@DA_Pan09] 36.4$\pm$1.2 66.3$\pm$2.3 22.2$\pm$1.4 13.8$\pm$1.4 23.2$\pm$1.5 23.2$\pm$1.5 34.9
gfk [@DA_Gong12] 38.8$\pm$1.3 70.9$\pm$1.1 24.4$\pm$1.4 16.3$\pm$0.9 26.7$\pm$1.8 26.1$\pm$1.0 38.9
SA [@DA_Fernando13] 39.0$\pm$1.3 71.1$\pm$1.3 24.2$\pm$1.4 16.0$\pm$0.9 26.8$\pm$1.9 26.4$\pm$1.1 38.9
CORAL [@CNN-DA_Sun15] 39.0$\pm$1.2 71.2$\pm$1.3 24.9$\pm$1.6 16.8$\pm$1.0 27.4$\pm$2.2 27.7$\pm$0.5 39.4
ATI 39.7$\pm$1.8 74.4$\pm$1.6 **25.9$\pm$2.1** 18.3$\pm$1.1 37.1$\pm$3.2 **35.0$\pm$1.0** 42.8
ATI-$\lambda$ **39.8$\pm$1.8** **74.8$\pm$1.5** 25.8$\pm$2.0 **18.7$\pm$0.7** **37.4$\pm$2.9** 34.8$\pm$0.8 **43.2**
: *Testbed* dataset [@DATA_Tommasi14] with 40 common classes and 12 domain shifts.[]{data-label="table:testbed_dense"}
We also present an evaluation on the Dense dataset of the Testbed for Cross-Dataset Analysis [@DATA_Tommasi15] using the provided DeCAF features. This protocol comprises 12 domain shifts between the 4 datasets *Bing (B)*, *Caltech (C)*, *ImageNet (I)* and *Sun (S)*, which share 40 classes. Following the protocol described in [@DATA_Tommasi15], we take 50 source samples per class for training and we test on 30 target images per class for all datasets, except *Sun*, where we take 20 samples per class. The results reported in Table \[table:testbed\_dense\] show that ATI-$\lambda$ outperforms other generic domain adaptation methods.
### Sentiment Analysis
B$\rightarrow$E D$\rightarrow$B E$\rightarrow$K K$\rightarrow$D AVG.
----------------------- ------------------ ------------------ ------------------ ------------------ ----------
LSVM 75.5$\pm$1.6 78.2$\pm$2.5 83.1$\pm$1.8 73.3$\pm$1.8 77.5
TCA [@DA_Pan09] 76.6$\pm$2.2 78.5$\pm$1.6 **83.8$\pm$1.5** 75.0$\pm$1.4 78.5
gfk [@DA_Gong12] 77.0$\pm$2.0 **79.2$\pm$1.8** 83.7$\pm$1.7 73.7$\pm$1.9 78.4
SA [@DA_Fernando13] 75.9$\pm$1.9 78.4$\pm$2.1 83.0$\pm$1.7 72.1$\pm$1.9 77.4
CORAL [@CNN-DA_Sun15] 76.2$\pm$1.7 78.4$\pm$2.0 83.1$\pm$2.0 74.2$\pm$3.0 78.0
ATI **79.9$\pm$2.0** **79.2$\pm$1.9** 83.7$\pm$2.1 **75.6$\pm$1.9** **79.6**
ATI-$\lambda$ 79.6$\pm$1.4 79.0$\pm$1.8 83.6$\pm$2.1 74.4$\pm$1.7 79.2
: Accuracies of 4 domain shifts on the Sentiment dataset [@DA_Blitzer07] using the bag-of-words features and the protocol from [@DA_Gong13ICML].[]{data-label="table:sentiment"}
To show the behaviour of our method with a different type of feature descriptor, we also present an evaluation on the *Sentiment analysis* dataset [@DA_Blitzer07]. This dataset gathers reviews from Amazon for four products: *books (B)*, *DVDs (D)*, *electronics (E)* and *kitchen appliances (K)*. Each domain contains 1000 reviews labelled as *positive* and another set of 1000 reviews as *negative*. We use the data provided by [@DA_Gong13ICML], which extracts bag-of-words features from the 400 words with the largest mutual information across domains. We report the mean accuracy over 20 splits, where for each run 1600 samples are randomly selected for training and the other 400 for testing. The results in Table \[table:sentiment\] show that our approach not only works very well for image and video data, but it can also be applied to other types of data. This demonstrates the versatility of the proposed approach.
Conclusions
===========
We have introduced the concept of open set domain adaptation in the context of image classification and action recognition. In contrast to closed set domain adaptation, we do not assume that all instances in the source and target domain belong to the same set of classes, but allow that each domain contains instances of classes that are not present in the other domain. We furthermore proposed an approach for unsupervised and semi-supervised domain adaptation that achieves state-of-the-art results for open sets and competitive results for closed sets. In particular, the flexibility of the approach, which can be used for images, videos and other types of data, makes the approach a versatile tool for real-world applications.
Acknowledgments {#acknowledgments .unnumbered}
===============
Acknowledgment {#acknowledgment .unnumbered}
==============
The work has been supported by the ERC Starting Grant ARCA (677650) and the DFG projects GA 1927/2-2 (DFG Research Unit FOR 1505 Mapping on Demand) and GA 1927/4-1 (DFG Research Unit FOR 2535 Anticipating Human Behavior).
[Pau Panareda Busto]{} received his B.Sc. and Masters degrees in computer engineering from the Technical University of Catalonia (2010) and his Masters degree in media informatics from the RWTH Aachen Universtiy (2013). Since September 2013, he has been a Ph.D. candidate at the University of Bonn in collaboration with Airbus Group. His research interests include computer vision, computer graphics and machine learning.
\[[{width="1in" height="1.25in"}]{}\][Ahsan Iqbal]{} obtained his master degree in Computer Science from the University of Bonn in 2017. He did an internship in Amazon Berlin from December 2016 to March 2017. Since April 2017, he is a Ph.D. student at the University of Bonn. His research interests are action recognition and action detection.
\[[{width="1in" height="1.25in"}]{}\][Juergen Gall]{} obtained his B.Sc. and his Masters degree in mathematics from the University of Wales Swansea (2004) and from the University of Mannheim (2005). In 2009, he obtained a Ph.D. in computer science from the Saarland University and the Max Planck Institut f[ü]{}r Informatik. He was a postdoctoral researcher at the Computer Vision Laboratory, ETH Zurich, from 2009 until 2012 and senior research scientist at the Max Planck Institute for Intelligent Systems in T[ü]{}bingen from 2012 until 2013. Since 2013, he is professor at the University of Bonn and head of the Computer Vision Group.
[^1]: Manuscript received January 29, 2018.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
In cold dark matter cosmologies, small mass halos outnumber larger mass halos at any redshift. However, the lower bound for the mass of a galaxy is unknown, as are the typical luminosity of the smallest galaxies and their numbers in the universe. The answers depend on the extent to which star formation in the first population of small mass halos may be suppressed by radiative feedback loops operating over cosmological distance scales. If early populations of dwarf galaxies did form in significant number, their relics should be found today in the Local Group. These relics have been termed “fossils of the first galaxies”. This paper is a review that summarizes our ongoing efforts to simulate and identify these fossils around the Milky Way and Andromeda.
It is widely believed that reionization of the intergalactic medium would have stopped star formation in the fossils of the first galaxies. Thus, they should be among the oldest objects in the Universe. However, here we dispute this idea and discuss a physical mechanism whereby relatively recent episodes of gas accretion and star formation would be produced in some fossils of the first galaxies. We argue that fossils may be characterized either by a single old population of stars or by a bimodal star formation history. We also propose that the same mechanism could turn small mass dark halos formed before reionization into gas-rich but starless “dark galaxies”.
We believe that current observational data supports the thesis that a fraction of the new ultra-faint dwarfs recently discovered in the Local Group are in fact fossils of the first galaxies.
author:
- Massimo Ricotti
bibliography:
- './2ndyear.bib'
- '/Users/ricotti/Latex/TeX/archive.bib'
title: The First Galaxies and the likely discovery of their Fossils in the Local Group
---
Introduction
============
There are many questions that remain open in cosmology with regard to the mass, number and properties of the smallest galaxies in the universe. Have we already discovered the smallest galaxies in the universe or we are still missing an elusive but large population of ultra-faint dwarf galaxies?
In cold dark matter (CDM) cosmologies most of the dark halos that formed before reionization had masses smaller than $10^8-10^9$ M$_{\odot}$ [, @GnedinO:97]. The small mass halos that survived tidal destruction to the modern epoch, were they able to form stars, would constitute a sub-population of dwarf satellites orbiting larger halos. Small mass dark halos significantly outnumber more massive galaxies like the Milky Way and can be located in the voids between luminous galaxies [, @Hoeft:06; @Ricotti:09].
However, until recently (, before 2005) observations did not show a large number of satellites around massive galaxies like the Milky Way and Andromeda. This became known as the “missing galactic satellite problem, [@Klypin:99; @Moore:99]. The voids between bright galaxies appear to be devoid of dwarf galaxies [@Karachentsev:04; @Karachentsevetal06; @Tullyetal06]. While the abundance of dwarfs in large voids may not pose a problem to CDM cosmology, as shown by [@TinkerC:09], it is unclear whether the predictions of the number of faint dwarfs in the Local Group is consistent with both the number of observed Milky Way dwarf satellites and the number of relatively isolated dwarfs in the local voids.
Historically, the discrepancy between observation and theory on the number of dwarf galaxies has been interpreted in two ways: 1) as a problem with the CDM paradigm that could be solved by a modification of the dark matter properties – for instance by introducing warm dark matter [, @BodeO:01] – or 2) as an indicator of feedback processes that are exceptionally efficient in preventing star formation in small mass halos, which remain mostly dark [, @HaimanAR:00].
The recently discovered population of ultra-faint dwarfs [@Belokurovetal06a; @Belokurovetal07; @Irwinetal07; @Walshetal07; @Willmanetal05ApJ; @Willmanetal05AJ; @Zuckeretal06a; @Zuckeretal06b; @Ibataetal07; @Majewskietal07; @Martinetal06] in combination with a proper treatment of observational incompleteness [@SimonGeha07; @Koposov:08; @Tollerudetal08; @Gehaetal08] has increased the estimated number of Milky Way satellites to a level that can be more easily reconciled with theoretical expectations. For instance, the suppression of dwarf galaxy formation due to intergalactic medium (IGM) reheating during reionization [@Babul:92; @Efstathiou:92b; @ShapiroG:94; @HaimanRL:96; @Thoul:96; @Quinn:96; @Weinberg:97; @NavarroS:97; @Bullock:00; @Gnedin:00; @Somerville:02; @Dijkstra:04; @Shapiro:04; @Hoeft:06; @OkamotoGT08; @Ricotti:09], in conjunction with a strong suppression of star formation in small mass pre-reionization dwarfs, may be sufficient to explain the observed number of Milky Way satellites. In the near future we can hope to answer perhaps a more interesting question: what is the minimum mass that a galaxy can have? This is a non trivial and fundamental question in cosmology. Answering it requires a better understanding of the feedback mechanisms that regulate the formation of the first galaxies before reionization and the details of the process of reionization feedback itself.
The formation of the first dwarf galaxies - before reionization - is self-regulated on cosmological distance scales. This means that the fate of small mass halos (, whether they remain dark or form stars) depends on local and global feedback effects. This type of galaxy feedback differs from the more familiar model operating in normal galaxies (, SN explosions, AGN feedback, etc), where the feedback is responsible for regulating the star formation rate within the galaxy itself but does not impact star formation in other distant galaxies. Rather, before reionization, each proto-dwarf galaxy reacts to the existence of all the others. Different theoretical assumptions and models for the cosmological self-regulation mechanisms will, of course, produce different predictions for the number and luminosity of the first dwarf galaxies [@HaimanAR:00; @RicottiGnedinShull02b; @WiseA:08; @OShea:08; @RicottiGnedinShull08].
We now introduce the basic concepts on how feedback-regulated galaxy formation operates in the early universe (, before reionization):\
A cooling mechanism for the gas is required in order to initiate star formation in dark halos. In proto-galaxies that form after reionization this is initially provided by hydrogen Lyman-alpha emission. This cooling is efficient at gas temperatures of 20,000 K but becomes negligible below $T \sim 10,000$ K. Later, as the temperature drops below 10,000 K, the cooling is typically provided by metal line cooling. In the first galaxies, however, both these cooling mechanisms may be absent. This is because the first dwarf galaxies differ when compared to present-day galaxies in two respects: 1) they lack important coolants – such as carbon and oxygen – because the gas is nearly primordial in composition, and 2) due to the smaller typical masses of the first dark halos, the gas initially has a temperature that is too low to cool by Lyman-alpha emission.
The gas in small mass halos with circular velocity $v_{\rm
vir}=(GM_{\rm tot}/r_{\rm vir})^{1/2} \simlt 20$ km s$^{-1}$, where $r_{\rm vir}$ is the virial radius – roughly corresponding to a mass $M_{\rm tot} \simlt 10^8$ M$_{\odot}$ at the typical redshift of virialization – has a temperature at virialization $T \simlt 10,000$ K. Hence, if the gas has primordial composition it is unable to cool by Lyman-alpha emission and initiate star formation unless it can form a sufficient amount of primordial H$_2$ (an abundance $x_{H_2} \simgt
10^{-4}$ is required). Because molecular hydrogen is easily destroyed by far-ultraviolet (FUV) radiation in the Lyman-Werner bands ($11.3<h\nu<13.6$ eV) emitted by the first stars, it is widely believed that the majority of galaxies with $v_{\rm
vir}<20$ km s$^{-1}$ remain dark [, @HaimanAR:00]. However, several studies show that even if the FUV radiation background is strong, a small amount of H$_2$ can always form, particularly in relatively massive halos with virial temperature of several thousands of degrees [@WiseA:07; @OShea:08]. Thus, negative feedback from FUV radiation may only delay star formation in the most massive pre-reionization dwarfs rather than fully suppress it [@Machacek:00; @Machacek:03]. We will argue later that hydrogen ionizing radiation ($h\nu > 13.6$ eV) in the extreme-ultraviolet (EUV) plays a far more important role in regulating the formation of the first galaxies than FUV radiation. Thus, in our opinion, models that do not include 3D radiative transfer of H and He ionizing radiation cannot capture the most relevant feedback mechanism that regulates galaxy formation in the early universe [@RicottiGnedinShull02a; @RicottiGnedinShull02b].
After reionization, the formation of dwarf galaxies with $v_{\rm
vir}<20$ km s$^{-1}$ is strongly inhibited by the increase in the Jeans mass in the IGM. Thus, according to this model, reionization feedback and negative feedback due to H$_2$ photo-dissociation by the FUV background (important before reionization) determine the mass of the smallest galactic building blocks. The resultant circular velocity of the smallest galactic building blocks is $v_{\rm vir} \sim
20$ km s$^{-1}$, roughly corresponding to masses $M_{\rm tot} \sim
10^8-10^9$ M$_\odot$. If this is what really happens in the early universe, the “missing Galactic satellite problem” can be considered qualitatively solved because the predicted number of Milky Way satellites with $v_{\rm vir}>20$ km s$^{-1}$ is already comparable to the estimated number of observed satellites after applying completeness corrections (although this model may still have problems reproducing the observations in detail).
However, as briefly mentioned above, we have argued for some time that most simulations of the first galaxies cannot capture the main feedback mechanism operating in the early universe because they do not include a key physical ingredient: radiative transfer of H and He ionizing radiation. Our simulations of the formation of the first galaxies are to date the only simulations of a cosmologically representative volume of the universe (at $z \sim 10$) that include 3D radiative transfer of H and He ionizing radiation [@RicottiGnedinShull02a; @RicottiGnedinShull02b; @RicottiGS:08]. Figure \[fig:pfr2\] shows the evolution of ionized bubbles around the first galaxies in a cubic volume of 1.5 Mpc in size at redshifts $z=21.2, 17.2, 15.7, 13.3$ from one of our simulations. The results suggest that negative feedback from the FUV background is not the dominant feedback mechanism that regulates galaxy formation before reionization. Rather, “positive feedback” on H$_2$ formation from ionizing radiation [@HaimanRL:96; @RicottiGS:01] dominates over the negative feedback of H$_2$ dissociating radiation. Hence, a strong suppression of galaxy formation in halos with $v_{\rm vir}<20$ km s$^{-1}$ does not take place. In this latter case, some galactic satellites would be the fossil remnants of the first galaxies. Comparisons of simulated pre-reionization fossils to dwarf spheroidals in the Local Group show remarkable agreement in properties [@RicottiGnedin05 hereafter RG05]. Based on the results of the simulations, we also suggested the existence of the ultra-faint population before it was discovered about a year later [see RG05, @BovillR:09].
Definition of “pre-reionization fossils”
-----------------------------------------
Throughout this paper we define “pre-reionization fossils” as the dwarfs hosted in halos with a maximum circular velocity remaining below 20 km s$^{-1}$ at all times during their evolution: $v_{\rm
max}(t)<20$ km s$^{-1}$. It will become clear in this paper that this definition is [*not*]{} directly related to the ability of fossils to retain gas and form stars after reionization: in § \[sec:infall\] we describe a mechanism in which small mass halos with $v_{\rm max}(t)< 20$ km s$^{-1}$ are able to have a late phase of gas accretion and possibly star formation.
Our definition of fossil reflects the [*special cooling mechanisms and feedback processes*]{} that regulate star formation and the number of luminous halos with $v_{\rm max}(t)<20$ km s$^{-1}$, before and [*after*]{} reionization. In proto-fossil galaxies – even adopting the most conservative assumption of maximum efficiency of shock heating of the gas during virialization – the gas is heated to a temperature below $T \sim 10,000$ K. Thus the gas cannot cool by Lyman-alpha emission, a very efficient coolant. The cooling of the gas is dominated either by H$_2$ roto-vibrational line emission or by metal cooling, important if the metallicity exceeds $Z \sim
10^{-3}$ Z$_\odot$ [, @BrommF:01; @SantoroS:06; @RicottiGS:08; @Smithetal:09]. These coolants are much less efficient than Lyman-alpha emission. Moreover, H$_2$ abundance and cooling is modulated and often suppressed by the FUV and EUV radiation fields. The FUV radiation in the H$_2$ Lyman-Werner bands and hard ultraviolet radiation have large mean free paths with respect to the typical distances between galaxies, thus their feedback is global in nature. Qualitatively, this explains why the first galaxies have low luminosities and low surface-brightness, similar to dwarf spheroidal (dSph) galaxies in the Local Group [RG05, @BovillR:09; @SalvadoriF:09].
Simulations also show that stars in the the first galaxies do not form in a disk but in a spheroid [RG05, @RicottiGS:08]. A thin galactic disk is not formed because of the high merger rates and the low masses of dark halos in the early universe. Roughly, pre-reionization fossils have a mass at virialization $M_{\rm
tot}<10^8$ M$_\odot$, assuming they form at $z_{\rm vir} \sim 10$, but their mass may increase by up to one order of magnitude by $z=0$ due to secondary infall [RGS02a, b, @RicottiGnedinShull08]. Secondary infall does not affect $v_{\rm
max}$, which remains roughly constant after virialization.
Pre-reionization fossils and reionization
-----------------------------------------
The critical value of $v_{\rm max, crit}$ for which dwarf galaxy formation is suppressed by reionization feedback is close to the 20 km s$^{-1}$ value that defines a fossil, but it is [*not necessarily the same value*]{}. Indeed, it can be significantly larger than 20 km s$^{-1}$ if the IGM is heated to $T \gg 10,000$ K [@RicottiGS:00]. Thus, we expect that the virialization of new “pre-reionization fossils” is strongly suppressed after reionization due to IGM reheating (, they mostly form before reionization). However, pre-reionization fossils and dark halos with $v_{\rm max}<20$ km s$^{-1}$ that virialized before reionization may accrete gas and, in certain cases, form new stars after reionization at redshifts $z<1-2$ [@Ricotti:09].
Unfortunately, the value of $v_{\rm max, crit}$ is uncertain due to our poor understanding of the thermal history of the IGM [@RicottiGS:00]. The uncertainty surrounding the IGM equation of state may partially explain the differences found in literature on the values for $v_{\rm max, crit}$ and the different levels of suppression of star formation as a function of the halo mass after reionization [, @Weinberg:97; @Gnedin:00; @Hoeft:06; @OkamotoGT08]. Regardless of assumptions for the reionization feedback model, one should bear in mind that no halo with $v_{\rm max}
<20$ km s$^{-1}$ can form stars after reionization unless the gas in those halos has been significantly pre-enriched with metals. For instance, the model by [@Koposov:09] assumes star formation after reionization in halos as small as $v_{\rm max}\sim
10$ km s$^{-1}$. With this assumption they find that their model is consistent with observations of ultra-faint dwarfs but claim that fossils are not needed to explain the data. However, star formation in such small halos can only take place in a gas that was pre-enriched with metals, suggesting the existence of older populations of stars in those halos. Indeed, according to our definition, the smallest post-reionization dwarfs with $10~{\rm km~s}^{-1}< v_{\rm max} <
20$ km s$^{-1}$ in the [@Koposov:09] model are “fossils”. As stated above, fossils may also be able to form stars after reionization due to a late phase of cold gas accretion from the IGM [@Ricotti:09].
Identification of pre-reionization fossils in observations
----------------------------------------------------------
Pre-reionization fossils are not easily identifiable because $v_{\rm
max}$ cannot be measured directly from observations. Understanding the star formation history of dwarf galaxies may help in this respect, as fossils likely show some degree of suppression of their star formation rate occurring about 12.5 Gyr ago due to reionization. However, their identification based on their star formation history may be complicated because some pre-reionization fossils in the last 10 Gyr may have had a late phase of gas accretion and star formation. The caveat is that star formation histories cannot be measured with accuracy better than to within 1-2 Gyr and the accuracy becomes increasingly poorer for old stellar populations. Thus, it is impossible to prove whether an old population of stars formed before reionization (which happened about 1 Gyr after the Big Bang) or at $z \sim 3$, when the Milky Way was assembled. Nevertheless, ultra-faint dwarfs that show some degree of bimodality in their star formation history are candidates for being pre-reionization fossils.
According to results by RG05 and [@BovillR:09], Willman 1, Bootes II, Segue 1 and Segue 2 do not lie on the luminosity-surface brightness relationship of simulated pre-reionization fossils. This result is based on the assumption that fossil properties are not modified by tides. Their surface brightness is larger than the model predictions for objects with such low-luminosity. An as yet undiscovered population of ultra-faints with lower surface brightness is instead predicted by our simulations. It is likely that the properties of the lowest luminosity ultra-faints may have been modified by tidal forces due to their proximity to the Milky Way disk.
Although it is difficult to identify individual fossils, statistical arguments suggest that at least some ultra-faint dwarf galaxies are pre-reionization fossils. This is because the number of satellites from N-body simulations with $v_{\rm max}(t)>20$ km s$^{-1}$ is substantially smaller than the estimated number of observed satellites after completeness corrections. Admittedly the current theoretical and observational uncertainties on the number of satellites are still large. However, if the estimated number (after completeness corrections) of ultra-faint dwarfs increases further, the existence of pre-reionization fossils will be inescapably proven. This is especially the case if a population of ultra-faint dwarfs with luminosities similar to Willman 1, Bootes II, Segue 1 and Segue 2 but surface brightness below the current sensitivity limit of the SDSS – as predicted by our simulations – is discovered.
The possibility of identifying the fossils of the first galaxies in our own backyard is very exciting. It would greatly improve our understanding of the physics involved in self-regulating the formation of the first galaxies before reionization. Clearly, even the launch of the James Webb Space Telescope ([*JWST*]{}), would not yield the wealth of observational data on the formation of the first galaxies that could be obtained by studying ultra-faint galaxies in the Local Group.
The rest of the paper is organized as follows. In § \[sec:obs\] we briefly review and discuss observational data on Galactic satellites, in § \[sec:theo\] we summarize the results of simulations of the formation of the first galaxies in a cosmological volume and the effect of reionization feedback on galaxy formation. In § \[sec:infall\] we discuss a recently proposed model for “late gas accretion” from the IGM onto small mass halos. In § \[sec:comp\] we compare the theoretical properties of simulated pre-reionization fossils to observations. In § \[sec:disc\] we compare different ideas for the origin of classical and ultra-faint dwarf spheroidals. We present our conclusions in § \[sec:conc\].
Observations {#sec:obs}
============
The ultra-faint satellites of the Milky Way and Andromeda
---------------------------------------------------------
Prior to 2005, the number of observed dwarf satellites of the Milky Way and Andromeda was about 30 [@Mateo98]. One of the most evident properties of the dwarfs in the Local Group is a type segregation, with “gas free” dwarf spheroidal (dSph) galaxies distributed near the center of their host galaxy and gas rich dwarf Irregulars (dIrr) at larger distances from the galactic centers. Notable exceptions are the Magellanic Clouds that are dIrr less than 100 kpc from the center of the Milky Way and a few isolated dSphs like Tucana and Antlia. One popular explanation for this segregation is the transformation of dIrr into dSph due to tidal and ram pressure stripping as dwarfs fall toward the Milky Way center [@Mayeretal07; @PenarrubiaNM08]. In addition, simulations showed that the number of dark matter satellites of the Milky Way with mass $>10^8$ M$_\odot$ (, with mass sufficiently large to expect star formation in them) was an order of magnitude larger than the number of known dwarf satellites [@Klypin:99; @Moore:99]. This posed a problem for CDM cosmologies.
Since 2005-2006 the number of known Local Group satellites has begun to increase dramatically, with the discovery of a new population of ultra-faint dwarfs. The new galaxies have been discovered by data mining the SDSS and other surveys of the halo around M31, resulting in the discovery of 14 new ultra-faint Milky Way satellites [@Belokurovetal06a; @Belokurovetal07; @Irwinetal07; @Walshetal07; @Willmanetal05ApJ; @Willmanetal05AJ; @Zuckeretal06a; @Zuckeretal06b; @Koposov:08; @Walsh:09] and 11 new companions for M31 [@Ibataetal07; @Majewskietal07; @Martinetal06; @Martinetal09]. Unofficial reports from members of the SDSS collaboration state that there are actually at least 17 new ultra-faint Milky Way dwarfs, but several of them are as yet unpublished (anonymous referee’s private communication).
The new Milky Way satellites have been slowly discovered since SDSS Data Release 2, with the most recent, Segue 2 discovered in Data release 7 [@Adelman-McCarthyetal06; @Adelman-McCarthyetal07]. [@Koposov:08] and [@Walsh:09] systematically searched Data Releases 5 and 6, respectively. Due to the partial sky coverage of the SDSS, and assuming isotropic distribution of satellites [but see @Kroupa:05; @Zentner:05], the total number of ultra-faint dwarfs in the Milky Way should be at least 5.15 times larger than the observed number [@Tollerudetal08]. With this simple but conservative correction, the number of Milky Way satellites within $400$ kpc is about $12 + 5.15 \times (14 \pm 3.7) \sim 84 \pm
19$. The quoted uncertainty is simply Poisson error due to the relatively small number of known ultra-faint dwarfs.
In estimating the completeness correction for the number of Milky Way dwarfs, one should account for selection effects inherent in the method used to find the ultra-faints in the SDSS data. In addition to completeness corrections for the survey’s coverage of the sky, the most important selection effect is the total number of stars from the object seen in the survey, , horizontal branch stars or main sequence and/or red giant stars for the lowest luminosity ultra-faints like Coma or Segue 1. This sets a limiting surface brightness cutoff at roughly 30 ${\rm mag~arcsec}^{-2}$ for the SDSS [@Koposov:08] (but see [@Martin:08] that find a limiting surface brightness about 6.4 times larger). There is also a distance-dependent absolute magnitude cutoff. The efficiency of finding ultra-faint dwarfs by data mining the SDSS typically drops rapidly at Galactocentric distances beyond $50-150$ kpc for the ultra-faints (depending on their luminosity) [@SimonGeha07; @Koposov:08; @Tollerudetal08; @Walsh:09]. Of the new Milky Way dwarfs, only Leo T is well beyond this distance threshold[^1], and 11 of the 14 new Milky Way satellites are within $200$ kpc.
The luminous satellites can be radially biased. So, the abundance of the faintest satellites within 50 kpc, that is the most complete sample, may not be used to correct for incompleteness at larger distances from the Galactic center without prior knowledge of this bias. And, of course, satellites of different luminosity and surface brightness will have different completeness limits. These selection biases have been considered in a paper by [@Tollerudetal08]. Their study finds that there may be between 300 to 600 luminous satellites within $400$ kpc. Their estimate for the number of luminous satellites within a Galactocentric distance of $200$ kpc is between 176 to 330.
Recent surveys of M31 [@Martinetal06; @Ibataetal07; @Martinetal09] have covered approximately a quarter of the space around the M31 spiral. The surveys have found 11 new M31 satellites, bringing the total number to 19. If we make a simple correction for the covered area of the survey, the estimated number of M31 satellites, including the new dwarfs, increases from $8$ to $52 \pm 13$.
The sensitivity limits of the surveys for Andromeda do not allow the detection of ultra-faint dwarfs that would correspond with those with the lowest luminosity found in the Milky Way. However, despite the fact that Andromeda and the Milky Way are thought to have approximately the same mass (within a factor of two), their satellite systems show interesting differences for even the satellites at the brighter end of the luminosity function. For instance, there are small differences in the galactocentric distance distribution of satellites and in the morphology of the satellites (, number of dIrr, dE and dSphs).
Peculiar ultra-faint dwarfs {#sec:leoT}
---------------------------
Many of the newly discovered dwarfs are dSphs with a dominant old population of stars and virtually no gas, which makes them candidates for being pre-reionization fossils. However, there are notable exceptions that we discuss below that may not perfectly fit the properties of simulated “fossils”. For instance, the dwarf galaxy Leo T resembles all the other ultra-faints but has gas and recent star formation [@Irwinetal07; @SimonGeha07]. We argue that Leo T could be a true “fossil” with $v_{\rm max}(t)<20$ km s$^{-1}$, but may have experienced a late phase of gas accretion and star formation due to the mechanism discussed in § \[sec:infall\].
[*Leo T*]{} has a stellar velocity dispersion of $\sigma_{Leo T} = 7.5
\pm 1.6 {\rm km~s}^{-1}$ [@SimonGeha07], or an estimated dynamical mass of $10^7$ M$_\odot$ within the stellar spheroid (although its total halo mass may be much larger). Leo T shows no sign of recent tidal destruction by either the Milky Way or M31 [@deJongetal08] and is located in the outskirts of the Milky Way at a Galactocentric distance of $400$ kpc. Leo T photometric properties are consistent with those of pre-reionization fossils. On the other hand, the halo of Leo T could be sufficiently massive ($v_{\rm max}(t) >20$ km s$^{-1}$) to retain or accrete mass after reionization and not be a pre-reionization fossil. However, as discussed in § \[sec:infall\] it is also possible that Leo T is a pre-reionization fossil that has been able to accrete gas from the IGM at late times due to an increase in the concentration of its dark halo and a decrease in the IGM temperature [@Ricotti:09]. Under this scenario, Leo T stopped forming stars after reionization, but was able to start accreting gas again from the IGM at $z \simlt 1-2$ and therefore would have a bimodal stellar population. [@deJongetal08] have found evidence for bimodal star formation in Leo T. Our model would explain why Leo T does not resemble other dIrr and is similar to dSphs and pre-reionization fossils, while not suffering significant tidal stripping.
[*Willman 1, Bootes II, Segue1 and Segue2*]{} are among the lowest luminosity ultra-faint dwarfs discovered so far; however, they do not fit the typical properties of pre-reionization fossils (see § \[sec:comp\]). For instance Willman 1 has a dynamical mass within the largest stellar orbit ($r \sim 100$ pc) of $5 \times 10^{5}
M_{\odot}$ and a mass-to-light ratio $\sim 470$, similar to other ultra-faint dwarfs [@Willmanetal05AJ]. However, given its low luminosity, Willman 1 has central surface brightness that is too large when compared to simulated “fossils”. Simulated fossils with luminosities $L_V<10^3$ L$_\odot$ should have a typical surface brightness that falls below the detection limit of $\sim$30 ${\rm
mag~arcsec}^{-2}$ estimated for the SDSS [@Koposov:08]. Hence, the lowest luminosity fossils may still be undiscovered. Although the nature of the lowest luminosity ultra-faints is unknown, due to selection effects they can only be found within $\sim 50$ kpc of the Galactic center. Thus it is possible, and perhaps to be expected, that their properties have been affected by tidal forces [@PenarrubiaNM08; @MartinJR:08].
Formation of first galaxies in CDM {#sec:theo}
==================================
The first episodes of star formation in the universe are thought to take place at redshift $z \sim 30-50$, in the center of dark matter halos with typical mass $M_{\rm tot} \sim 10^5-10^6$ M$_\odot$. The gas in these halos is metal free and simulations show that a single or binary massive star per halo is formed [@Bromm:02; @Abel:02; @Saigo:04; @Gao:05; @OShea:07; @Yoshida:08; @Turk:09; @StacyGB:09]. Such stars are called Pop III and their mass, although not well constrained by the simulations, is quite large: in the range between $20$ M$_\odot$ to a few $100$ M$_\odot$. Whether or not we can refer to minihalos containing a single massive star (or a binary star) in their center as the “first galaxies” is debatable. However, the crucial point to be made here is that there is a gap of 2 to 3 orders of magnitude between the typical halo mass in which Pop III stars are born ($10^5-10^6$ M$_\odot$) and the typical mass of the population of dwarf galaxies that are not strongly influenced by radiative and reionization feedback ($10^8-10^9$ M$_\odot$). The primordial dwarfs that fill the gap are those that we refer to as [ *pre-reionization fossils*]{}.
If the formation of pre-reionization fossils is not drastically suppressed by radiative feedback, their number may be several orders of magnitude larger than the number of more massive dwarfs. This is because in CDM cosmologies the number of dark halos per unit comoving volume roughly scales with the mass as $N \propto M_{\rm dm}^{-2}$.
It is widely believed that nearly all halos with mass $M_{\rm
tot}>10^8-10^9$ M$_\odot$ host luminous galaxies, although there can be substantial disagreement among theorists on their luminosity. However, most of the theoretical controversy rests in understanding the fate of the halos with mass between $10^6-10^8$ M$_\odot$ and the dominant feedback that determines whether they become luminous or remain dark. We will elaborate on this statement in the next sections.
Radiative feedback {#sec:sims}
------------------
Simulating the formation of the first stars is a relatively well defined initial condition problem, given the cosmological parameters. However, these simple initial conditions must soon be modified to take into account the effects of other newborn stars, whose properties are still quite uncertain. The physics becomes more complex as competing feedback effects determine the fate of the first galaxies: radiative feedback regulates the formation and destruction of H$_2$ and metals are injected into the IGM and into protogalaxies.
### Negative feedback from H$_2$ photo-dissociating radiation
The net effect of radiative feedback on the global star formation history of the universe before the redshift of reionization is uncertain. An FUV background (at energies between $11.34$ eV and $13.6$ eV) destroys H$_2$, the primary coolant at the start of galaxy formation. The FUV radiation emitted by the first few Pop III stars is sufficient to [*suppress or delay*]{} galaxy formation in halos with circular velocities $v_{\rm vir}< 20$ km s$^{-1}$ that are too small to cool by Lyman-alpha emission [@Haimanetal00; @Ciardietal00; @Machaceketal00; @OShea:08]. Hence, according to this scenario, most halos with masses $<10^8-10^9$ M$_\odot$ remain dark. More work is needed to quantify the level of suppression of galaxy formation and examine how these models compare to observations of Milky Way satellites.
Figure \[fig:pfr0\] illustrates the effect of H$_2$ dissociating radiation on the IGM. The two panels show slices through a simulation in [@RicottiGnedinShull02b] at $z=19.44$ and $z=18.5$. The top-right tiles in the two panels show H$_2$ abundance. At $z=19.4$, the H$_2$ has its relic abundance everywhere in the IGM except inside the dissociation spheres around the first galaxies, where it is destroyed. At $z=18.5$, the dissociation spheres are still visible, but the UV background starts to dissociate H$_2$ everywhere in the IGM except the denser filaments.
### Positive Feedback Regions
Our main criticism for the “negative feedback” model is that it does not take into account the effect of hydrogen ionizing radiation [@RicottiGnedinShull01; @Whalenetal07] that, according to simulations, may indeed play a dominant role in regulating galaxy formation before reionization [@RicottiGnedinShull02a; @RicottiGnedinShull02b]. Simulations including 3D radiation transfer show that star formation in the first small mass halos is inefficient, partially due to winds produced by internal UV sources. This produces galaxies that are extremely faint and have very low surface brightnesses. However, our simulations show that a large number of ultra-faint dwarfs (a few hundred galaxies per comoving Mpc$^3$) form before reionization at $z \sim 7-10$. Hence, according to this model, the Local Group may contain thousands of ultra-faint dwarf galaxies.
Ionizing radiation from the first stars enhances the production of H$_2$ (we refer to this as “positive feedback”) by creating free electrons and promoting the formation of H$^-$, the main catalyst for the formation of H$_2$ in a low metallicity gas [@ShapiroKang87; @HaimanReesLoeb96; @RicottiGnedinShull01; @AlvarezBrommShapiro06; @Ciardietal06]. [@RicottiGnedinShull01] found that shells of H$_2$ can be continuously created in precursors around the Strömgren spheres produced by ionizing sources and, for a bursting mode of star formation, inside recombining regions. We refer to these shells as “positive feedback regions”. This is because the catalyst H$^-$, and hence H$_2$, is formed most efficiently in regions where the gas ionization fraction is about $50\%$. This local “positive feedback” is difficult to incorporate into cosmological simulations because the implementation of spatially inhomogeneous, time-dependent radiative transfer is computationally expensive.
Fig. \[fig:pfr1\] shows “positive feedback regions” in one of our simulations. The figure shows a slice through a simulation at 4 different times (at $z= 17.3, 12.2, 11.3$, and 10.2). We recognize the two main processes that create H$_2$ in the filaments. In the top-left frame at $z=17.3$ we can see a “positive feedback region” as an irregular shell of H$_2$ surrounding the region that is barely intersected by the slice. In the bottom-left frame ($z=11.3$) two regions are clearly visible. Inside the regions, the H$_2$ is destroyed. In the bottom-right frame ($z=10.2$) the regions are recombining (demonstrating that the star formation is bursting) and new H$_2$ is being reformed inside the relic regions. A finer inspection[^2] of the time evolution of this slice shows that at least five regions form and recombine between $z = 20$ and $z=10$ in this small region of the simulation.
There are two reasons why our results are still controversial. First, our simulations do not yet have sufficient resolution to ensure their convergence. Second, there are no other published simulations to compare our results with. Only recently have some groups started to include the effect of 3D radiative transfer on hydrodynamics [, @WiseA:07; @WiseA:08]. However, currently there are no other simulations of the formation of the first galaxies in a cosmological volume suited for comparison with observations of dwarfs in the Local Group other than our own [@RicottiGnedinShull02a; @RicottiGnedinShull02b; @RicottiGnedin05; @RicottiGnedinShull08; @BovillR:09]. Hence, our results may differ from other numerical studies because of the inclusion in the code of the effects of “positive feedback regions” and galactic winds from ionizing radiation.
Simulations by [@WiseA:08] include a self-consistent treatment of hydro and 3D radiative transfer that is more accurate than our approximate, but faster method. However, because the authors use ray-tracing for the radiative transfer, only a few sources of radiation can be simulated at the same time. This limits the volume and number of galaxies that can be simulated. Due to these limitations the simulations are not suited for comparison between the primordial dwarf populations and the ultra-faint dwarfs. In addition, at the moment, the aforementioned simulations do not include metal cooling and the formation of normal stars (other than Pop III).
### The simulations
The simulation used for comparison to observations of ultra-faint dwarfs has been thoroughly described in Ricotti, Gnedin & Shull(2002a, 2002b) as run “256L1p3”. Here we remind the reader that the simulation includes $256^3$ dark matter particles, an equal number of baryonic cells, and more than 700,000 stellar particles in a box of size $\sim 1.5$ comoving Mpc. The mass of the dark matter particles in our simulation is 4930 M$_\odot$, and real comoving spatial resolution (twice the Plummer softening length) is $150~h^{-1}$ pc (which corresponds to a physical scale of 24 parsecs at $z=8.3$). This resolution allows us to resolve cores of all simulated galaxies that would correspond to the observed Local Group dwarfs. The stellar masses are always smaller than the initial baryon mass in each cell but can vary from $\sim 0.6$ h$^{-1}$ M$_\odot$ to $600$ h$^{-1}$ M$_\odot$ with a mean of $6$ h$^{-1}$ M$_\odot$. Stellar particles do not represent individual stars but, in general, a collection of stars (, OB associations).
The simulation includes most of the relevant physics, including time-dependent spatially-variable radiative transfer using the OTVET approximation [@GnedinA:01], detailed radiative transfer in Lyman-Werner bands, non-equilibrium ionization balance, etc. In addition to primordial chemistry and 3D radiative transfer, the simulations include a sub-grid recipe for star formation, metal production by SNe and metal cooling. The code also includes mechanical feedback by SN explosions. However, we found that for a Salpeter IMF, the effect of SNe is not dominant when compared to feedback produced by ionizing radiation from massive stars [@RicottiGnedinShull08 hereafter RGS08]. The effect of SN explosions is somewhat model dependent and uncertain because it is treated using a sub-grid recipe. Hence, the simulation analyzed in this work includes metal pollution but not mechanical feedback by SNe.
In RG05, we included the effect of reionization in the simulation 256L1p3. Because the size of the simulation box has been fixed at $\sim 1.5$ comoving Mpc, the simulation volume is too small to model the process of cosmological reionization with sufficient accuracy. We therefore assume that the simulation volume is located inside an region of a bright galaxy at a higher redshift. Specifically, we introduce a source of ionizing radiation within the computational box, properly biased, which corresponds to a star-forming galaxy with the constant star formation rate of 1 solar mass per year (similar to star formation rates of observed Lyman Break Galaxies at $z \sim 4$, Steidel et al. 1999). The source is switched on at $z=9.0$, and by $z=8.3$ the whole simulation box is completely ionized.
### Summary of main results
The main results of the simulations are the following [see @RicottiGnedinShull08 for details]:
1. [*Negative feedback*]{}. H$_2$ photo-dissociation from FUV radiation, the main negative feedback thought to suppress the formation of the first galaxies, is not the dominant feedback. If we modify the spectrum of the sources of radiation to artificially increase or decrease the FUV flux emitted by up to one order of magnitude, we do not find any appreciable effect on the global star formation history.
2. [*Self-regulation*]{}. Feedback by hydrogen ionizing radiation (EUV) plays the key role. Galactic outflows, produced by UV photo-heating from massive stars, and H$_2$ formation/photo-dissociation induces a bursting star formation mode in the first galaxies that acts as the catalyst for H$_2$ re-formation inside relic (recombining) regions and in the “precursors” of cosmological Strömgren spheres ([*i.e.*]{}, positive feedback regions). As a result, the formation of the first galaxies is self-regulated on a cosmological distance scale. It is significantly reduced by radiative feedback but it is not completely suppressed, even in halos with $v_{\rm max} \sim
5-10$ km s$^{-1}$. Note that our sub-grid recipe for star formation assumes a Schmidt law with a tunable efficiency parameter $\epsilon_*$ (the fraction of gas converted into stars per crossing time). We find that the global star formation history and the fraction of baryons converted into stars in each galaxy, $f_*=M_*/M_{\rm bar}$, is nearly independent on the assumed value of $\epsilon_*$. This is typical for feedback regulated star formation. Hence, the star formation efficiency, $f_*$, is not an assumed parameter but it is extracted from the simulations. Thus, the derived star formation efficiency $f_*$ is a very generic prediction of our feedback model because it is nearly independent of the assumed value of $\epsilon_*$, that is instead quite uncertain.
3. [*Contribution to Reionization*]{}. Due to the feedback-induced bursting mode of star formation in pre-reionization dwarfs, the cosmological regions that they produce remain confined in size and never reach the overlap phase (, Fig. \[fig:pfr2\] above). Reionization is completed by more massive galaxies.
4. [*Gas photo-evaporation and metallicity.*]{} Star-forming dwarf galaxies show large variations in their gas content because of the combined effects of stellar feedback from internal sources and IGM reheating. Ratios of gas to dark matter lie below the cosmic mean in halos with masses $M_{\rm dm}<10^8$ M$_\odot$. Fig. \[fig:feed\](left) shows the fraction of baryons retained by dark and luminous halos. It is clear that small mass luminous halos lose most of their gas before reionization due to internal radiation sources. Dark halos instead are able to retain gas for a longer time [see also @Hoeft:06]. Gas depletion increases with decreasing redshift: the lower-mass halos lose all their gas first but, as the universe evolves, larger halos with $M_{\rm dm} \sim 10^8$ M$_\odot$ also lose a large fraction of their gas. Gas photoevaporation reduces the metallicity of pre-reionization dwarfs to levels consistent with observations of dSph galaxies.
5. [*Number of luminous galaxies*]{}. Only about 1% of dwarf dark matter halos with mass $M_{\rm dm}\sim 5 \times 10^6$ M$_\odot$, assembled prior to reionization, are able to form stars. The fraction of luminous halos scales with the halo mass as $F_{\rm lum} \propto
M_{\rm dm}^{2}$. Thus, most halos with mass $\simgt 5 \times
10^7$ M$_\odot$ are luminous. Fig. \[fig:feed\](right) shows $F_{lum}$ as a function of the halo mass at redshifts $z=12.5$ and $z=10.2$. The figure also illustrated that $f_*$ depends on the environment. Namely, it depends on the proximity of the pre-reionization dwarfs to other luminous galaxies.
We find $\sim 450$ dwarf galaxies per Mpc$^{3}$ with bolometric luminosity between $10^4$ and $10^8$ L$_\odot$. The luminosity function is rather flat at low luminosities, with about $10$ galaxies per Mpc$^{3}$ in the range $10^7<L<10^8$ L$_\odot$, and about $220$ Mpc$^{-3}$ in the range $10^4<L<10^5$ L$_\odot$ and $10^5<L<10^6$ L$_\odot$.
6. [*Star formation efficiency and mass-to-light*]{}. The mean star formation efficiency , where $M_{\rm bar}^{\rm max} \simeq
M_{\rm dm}/7$, is found to be [*nearly independent of redshift*]{} and to depend on total mass as $\langle f_* \rangle \propto M_{ \rm
dm}^{\alpha}$ with $\alpha=2$ if the radiative feedback is strong (, top heavy IMF and/or large $\langle f_{esc}\rangle$) and $\alpha=1.5$ if the feedback is weak. This is shown in Fig. \[fig:sfe\](left), where we plot the mean star formation efficiency, $\langle f_*\rangle $, and the mean gas fraction $\langle f_g\rangle$ in halos of mass $M_{\rm dm}$.
7. [*Scatter of the mass-to-light ratio*]{}. A tight relationship between the star formation efficiency $f_*$ and the total mass of halos holds only for galaxies with $M_{\rm dm}>5 \times 10^7$ M$_\odot$. In lower-mass halos, the scatter around the mean $\langle f_* \rangle$ is increasingly large (see Fig. \[fig:sfe\](right)). For a given halo mass, the galaxy can be without stars (dark galaxy) or have star formation efficiency $f_*
\sim 0.1$. However, only a few dark galaxies with mass at formation $M_{\rm dm} \sim 1-5 \times 10^{7}$ M$_\odot$ should exist in the Local Group.
8. [*Size and morphology of stellar component*]{}. Galaxies with masses $M_{\rm dm}<10^8$ M$_\odot$ have a low surface brightness and extended stellar spheroid. At $z \sim 10$, the outer edges of the stellar spheroid nearly reaches the virial radius. In more massive galaxies that cool more efficiently by Lyman-alpha radiation, the stars and light are more centrally concentrated. Fig. \[fig:prof\] shows the structural properties of the dark matter and stellar halo in three of the most massive galaxies in our simulation. These dwarf galaxies have properties similar to Draco and Umi dSphs. The figure also shows that the velocity dispersion of the stars in these dwarfs is about a factor of two smaller than $v_{\rm max}$.
### Photo-evaporation and Reionization feedback {#sec:reio}
The small total mass of the first galaxies has two other implications. First, the ionizing radiation emitted by massive stars can blow out most of the gas before SN-driven winds become important, further reducing star formation rates (see RGS08). Second, the increase in temperature of the IGM to $10,000-20,000$ K due to reionization prevents the gas from condensing into newly virialized halos with circular velocities smaller than $10-20$ km s$^{-1}$ [, @Babul:92; @Efstathiou:92b; @Bullock:00; @Gnedin-filteringmass00; @Hoeft:06; @OkamotoGT08]. It follows that dwarf galaxies with $v_{\rm max} < 10-20$ km s$^{-1}$ lose most of their gas before reionization and stop accreting new gas and forming stars after reionization.
The value $v_{\rm max} \sim 20$ km s$^{-1}$ that we use to define a fossil is motivated by the fundamental differences in cooling and feedback processes discussed above that regulate star formation in the early universe. It is not the critical value for suppression of gas accretion due to reionization. Indeed, we discuss in § \[sec:infall\] that pre-reionization fossils may have a late phase of gas accretion and star formation well after reionization, at redshift $z<1-2$. Thus, a complete suppression of star formation after reionization (about 12 Gyr ago) is not the defining property of a fossil dwarf.
Late time cold accretion from the IGM {#sec:infall}
=====================================
The ability of the IGM gas to condense at the center of dark halos depends on the ratio, $\Gamma=v_{\rm vir}/c_{\rm s, igm}$, of the circular velocity to the IGM sound speed, and also on the dark halo concentration, $c$ [@Ricotti:09]. Typically, the concentration of a halo is $c_{\rm vir} \sim 4$ at the redshift of virialization [@Bullock:01; @WechslerB:02] but, as the halo evolves in the expanding universe, its concentration increases $\propto (1+z_{\rm
vir})/(1+z)$. The evolution of the halo concentration with redshift can be understood in the context of the theory of cosmological secondary infall of dark matter [@Bertschinger:85] and has been quantified using N-body simulations [@Bullock:01; @WechslerB:02]. Thus, primordial halos with $v_{\rm vir}<10-20$ km s$^{-1}$ stop accreting gas immediately after reionization, but, in virtue of their increasing concentration and the decreasing temperature of the IGM at $z<3$ (after reionization), they may start accreting gas from the IGM at later times [see, @Ricotti:09]. As a result, we expect that if the fossils of the first galaxies exist in the Local Group (RG05), they may have a more complex star formation history than previously envisioned. A signature of this model is a bimodal star formation history with an old ($\sim 13$ Gyr) and a younger ($\simlt
5-10$ Gyr, depending on the halo mass) population of stars. Leo T properties can be reproduced by this simple model for late gas accretion [@Ricotti:09]. In addition, Leo T seems to show a bimodal star formation history [@deJongetal08] as expected in our model. Still, other models may also explain the observed star formation history of Leo T [@Stinsonetal07].
Perhaps more important is the possible existence of dark galaxies: small mass halos containing only gas but no stars. Dark galaxies are most likely to exist if pre-reionization fossils do not form efficiently due to dominant negative feedback in the early universe (, H$_2$ photo-dissociation by the FUV background).
The late gas accretion from the IGM is shown in Figure \[fig:infall\] for dark halos with circular velocity at virialization $v_{\rm vir}=18, 15, 12, 9$ and $6$ km s$^{-1}$ (lines from the top to the bottom). The lines show the evolution of the gas density in the core of the halo as a function of redshift. The core radius is typically $100$ pc and the labels show the circular velocity at the core radius ($v_{\rm cir}(r_{\rm core}) \approx 0.66 v_{vir}
\approx 0.624 v_{\rm max}$, where $v_{\rm vir}$ is circular velocities at the virial radius and $v_{\rm max}$ is the maximum circular velocity) . We show the evolution of the gas density only for halos that are affected by reionization feedback. More massive halos will also accrete gas from the IGM as they evolve in isolation after virialization, but the gas accretion is continuous and not affected by reionization.
Under the scenario in which halos with masses smaller than the critical value of $10^8-10^9$ M$_\odot$ remain dark due to feedback effects, the increase in their dark matter concentration and the temperature evolution of the IGM will produce a late phase of gas accretion at redshift $z<1-2$. If the gas has very low metallicity or is metal free, it is unlikely that the accreted gas will be able to form stars in the smallest mass halos. This is why we envisioned the possible existence of dark galaxies. However, their mass would be smaller than the mass of any luminous galaxy and their discovery would be challenging.
The level of metal pre-enrichment necessary to initiate star formation in minihalos that experience a late phase of gas accretion can be roughly estimated from the cooling function from hyperfine transitions of oxygen and carbon: $\Lambda_{23} \sim 10^{-3}~(Z/Z_\odot)$, where $\Lambda_{23}=10^{-23}$ erg s$^{-1}$ cm$^{3}$ and $Z$ is the gas metallicity. A necessary condition for star formation is $t_{\rm cool}
\approx (0.7~{\rm yr})~T/(n_{\rm g, core}\Lambda_{-23}) < t_{\rm H}$, that can be written as $n_{\rm g, core} > 0.03~{\rm
cm}^{-3}(Z/10^{-2}~Z_\odot)^{-1}$. The left panel in Figure \[fig:core2\] shows $n_{\rm g, core}$ and $N_{\rm H}$ in minihalos that evolve isothermally at $T \sim 10^4$ K but that do not form stars (, candidates for extragalactic CHVCs and dark galaxies). The horizontal lines show the requirement for metal cooling and star formation assuming gas metallicity $Z=0.1$ and $0.01$ Z$_\odot$. The right panel in Figure \[fig:core2\] shows $n_{\rm g, core}$, $N_{\rm H}$ and $M_{\rm dyn}/M_{\rm gas}$ (the dynamical mass to gas mass ratio) in the core of minihalos that are able to cool to $T_{\rm g, core}=5000$ K (roughly the temperature of the ISM in Leo T), and thus form stars. The symbols show the observed value for Leo T.
Comparison of theory and observations {#sec:comp}
=====================================
Number of fossils and non-fossil satellites in the Milky Way {#ssec:count}
------------------------------------------------------------
N-body simulations can be used to estimate the number of dark halos in the Milky Way with maximum circular velocity $v_{\rm
max}>20$ km s$^{-1}$. However, there is a complication to this naive calculation. A significant fraction of dark halos that today have $v_{\rm max} < 20$ km s$^{-1}$ were once more massive, due to tidal stripping [@KravtsovGnedinKlypin04]. According to our definition, dwarf galaxies formed in these dark halos would not be pre-reionization fossils if they had at any time during their evolution $v_{\rm max}(t)>20$ km s$^{-1}$ (see § \[sec:reio\]). If the number of observed Milky Way satellites exceeds the estimated number of these massive halos we must conclude that at least a fraction of the observed Milky Way satellites are pre-reionization fossils.
However, there is an assumption in this scenario. One must assume that the stars in these halos survive tidal stripping for as long as the dark matter. In this case tidally stripped halos may indeed account for a fraction or all of the newly discovered ultra-faint dwarfs. However, [@PenarrubiaNM08] find that tidally stripped dark halos lose their stars more rapidly than they lose their dark matter. Thus, they may become dark halos even though they were initially luminous satellites. These dark halos should not be counted as ultra-faint dwarfs.
Using results of published N-body simulations of the Milky Way, [@BovillR:09] have estimated the number of dark halos that have or had in the past $v_{\rm max}(t)>20$ km s$^{-1}$ (, non pre-reionization fossils). In Table \[tab:count\] we summarize the results of the counts for dark matter and luminous satellites for two large N-body simulations of a Milky Way type halo: the “Aquarius’ simulation [@Springeletal08] and the Via Lactea I simulation [@Diemandetal07b].
---------------------- ------------ ------- ------------- ------- --------------
Distance Luminous
from dwarfs
center today any time today any time
\[0.5ex\] $<200$ kpc 176 to 330 14 $36 \pm 8$ 34 $91 \pm 20$
$<417$ kpc 304 to 576 28 $73 \pm 16$ 69 $182 \pm 40$
\[1ex\]
---------------------- ------------ ------- ------------- ------- --------------
: Number of observed satellites versus number of dark halos with $v_{\rm max}(t)>20$ km s$^{-1}$ (, non pre-reionization fossils) for the Milky Way
\[tab:count\]
The number of luminous satellites that exist within the Milky Way is highly uncertain beyond a distance from the Galactic center of $200$ kpc. [@Tollerudetal08], after applying incompleteness corrections, estimated satellites within $417$ kpc and about within $200$ kpc (the numbers are from their Table 3). As shown in Table \[tab:count\], the existence of some pre-reionization fossils among the ultra-faint dwarfs appears to be favored by the data. However, the current uncertainties on the completeness corrections of observations and on the simulations are too large to deem the existence of fossils as necessary.
The error bars on the theoretical estimate of the number of fossils in the Milky Way shown in Table \[tab:count\] come from uncertainties in the fraction of halos that were more massive in the past. This fraction was derived from simulations by [@KravtsovGnedinKlypin04]. Another uncertainty in the simulation results can be attributed to the different predictions for the number of Milky Way satellites in the Via Lactea I and II and Aquarius simulations. The discrepancy can be partially attributed to different cosmology in the simulations but mostly because the Via Lactea I simulation likely used erroneous initial conditions. Finally, [@Tollerudetal08] corrections on the number of observed satellites also rely on the radial distribution of dark matter sub halos extracted from Via Lactea I simulations that may be erroneous. Once the discrepancies among different simulations are better understood the number of simulated satellites of the Milky Way may be known with greater certainty.
Using comparisons between the predicted and observed Galactocentric distributions of dwarf satellites around the Milky Way, [@GnedinKravtsov06], hereafter GK06, have estimated that pre-reionization fossils may constitute about $1/3$ of Milky Way dwarfs. GK06 estimated the number of fossils in the Milky Way using data from the simulations of the first galaxies in RG05. GK06 defined a fossil as a simulated halo which survives at $z = 0$ and remains below the critical circular velocity of $20$ km s$^{-1}$ with no appreciable tidal stripping (the usual definition of fossil adopted in this paper as well). They calculate the probability, $P_S(v_{\rm max},
r)$, of a luminous halo with a given maximum circular velocity $v_{\rm
max}$ to survive from $z = 8$ (the final redshift of the RG05 simulation) to $z = 0$. The surviving halos are assigned a luminosity based on the $L_{\rm V}$ versus $v_{\rm max}$ relationship from RG05. At $z = 0$, GK06 has a population of dwarf galaxies with a resolution limit of $v_{\rm max} = 13$ km s$^{-1}$. This limit corresponds to a lower luminosity limit of $L_{\rm V} \sim 10^5$ L$_{\odot}$, which includes Leo T and Canes Venatici I, but excludes all the other new ultra-faint Milky Way satellites.
In Figure \[LF\], we show the cumulative luminosity function from GK06 for the Milky Way and M31 satellites. The lower panel shows satellites with distance from their host $d<100$ kpc, the middle panel $d<300$ kpc and the upper panel $d<1$ Mpc. The gray lines show the GK06 predictions, and the shaded region encompasses the error bars. The resolution limits in GK06 cause halos with $v_{ \rm
max}<17$ km s$^{-1}$ to be preferentially destroyed by tidal effects. The dashed line show the predicted luminosity function corrected for the resolution effects. Both the uncorrected (solid lines) and corrected (dashed lines) luminosity functions are plotted in the lower panel. The points with error bars show the observed luminosity function of dSph galaxies around the Milky Way and M31 corrected only for limits in sky coverage of the SDSS survey. The plot is from GK06 but has been updated to include the new ultra-faint dwarfs with $L_{\rm V} \simgt 10^5$ L$_\odot$.
The results of this model are consistent with the observations. The model reproduces the Galactocentric distribution of the most luminous dSphs, even though in this model dSphs are not tidally stripped dIrrs. It also shows a good agreement with observations for luminosities that can be considered nearly complete within a given Galactocentric distance.
Statistical Properties of pre-reionization Fossils {#sec:prop}
--------------------------------------------------
In this section, we compare the properties of the new dwarf galaxies discovered in the Local Group to the theoretical predictions of simulations of primordial galaxies formed before reionization. The argument that justifies this comparison is that star formation stops or is greatly reduced after reionization (but see § \[sec:infall\]). We do not expect two perfectly distinct populations of fossil galaxies with $v_{\rm max}<20$ km s$^{-1}$ and non-fossils with $v_{\rm max}\ge 20$ km s$^{-1}$, but a gradual transition of properties from one population to the other. Some fossils may become more massive than $v_{\rm max} \sim 20$ km s$^{-1}$ after reionization, accrete some gas from the IGM, and form a younger stellar population. These dwarfs are no longer defined as “fossils”. However, if the dark halo circular velocity remains close to $20$ km s$^{-1}$ the young stellar population is likely to be small with respect to the old one. In RG05 we call these galaxies “polluted fossils” because they have the same basic properties of “fossils” with a sub-dominant young stellar population. A similar argument can be made regarding the late phase of gas accretion that may produce objects similar to Leo T.
In Figs. \[Kor\]-\[ZL\] we compare the RG05 predictions for the fossils of primordial galaxies to the observed properties of the new Milky Way and M31 dwarfs. The symbols and lines in Figs. \[Kor\]-\[ZL\] have the following meanings. All known Milky Way dSphs are shown by circles; Andromeda’s dSphs satellites are shown by triangles; simulated fossils are shown by the small solid squares. The solid and open symbols refer to previously known and new dSphs, respectively. The transition between fossils and non-fossil galaxies is gradual. In order to illustrate the different statistical trends of “non-fossil” galaxies we show dwarf irregulars (dIrr) as asterisks and the dwarf ellipticals (dEs) as crosses, and we show the statistical trends for more luminous galaxies as thick dashed lines on the right side of each panel.
Figure \[Kor\](left) shows how the surface brightness (top panel) and half light radius (bottom panel) of all known Milky Way and Andromeda satellites as a function of V-band luminosity compares to the simulated fossils. The surface brightness limit of the SDSS is shown by the thin solid lines in both panels of the figure. The new dwarfs agree with the predictions up to this threshold, suggesting the possible existence of an undetected population of dwarfs with $\Sigma_{\rm V}$ below the SDSS sensitivity limit. The new M31 satellites have properties similar to their previously known Milky Way counterparts (, Ursa Minor and Draco). Given the similar host masses and environments, further assuming similar formation histories for the halos of M31 and the Milky Way, we may be tempted to speculate on the existence of an undiscovered population of dwarfs orbiting M31 equivalent to the new SDSS dwarfs.
The large mass outflows due to photo-heating by massive stars and the subsequent suppression of star formation after an initial burst make reionization fossils among the most dark matter dominated objects in the universe, with predicted mass-to-light ratios as high as $10^4$ and $L_{\rm V} \sim 10^3-10^4$ L$_{\odot}$. Figure \[Kor\](right) shows the velocity dispersion (bottom panel) and mass-to-light ratios, $M_\sigma/L_{\rm V}$ (top panel), as a function of V-band luminosity of the new and old dwarfs from observations in comparison to simulated fossils. The symbols are the same as in the previous figures. Theoretical and observed dynamical masses are calculated from the velocity dispersions of stars (, $M_\sigma=2r_{1/2}\sigma^2/G$), and do not necessarily reflect the total mass of the dark halo at virialization.
Observations show that the value of the dynamical mass within the stellar spheroid, $M \sim (1 \pm 5) \times 10^7$ M$_\odot$, remains relatively constant as a function of $L_{\rm V}$ [@Mateo98]. Recent work by [@Strigarietal08] shows analogous results to the one found by [@Mateo98]. The dynamical mass of dwarf spheroidals within a radius of 300 pc is relatively constant: $M
\sim 10^7$ M$_\odot$. The radii of the stellar spheroids in these dwarf galaxies may be either larger or smaller than 300 pc. In the later case, the determination of the mass of the dwarfs is uncertain.
Our simulation provides some insight into the reason why the dynamical mass remains roughly constant in dSphs. The simulations show that in pre-reionization dwarfs, the ratio of the radius of the stellar spheroid to the virial radius of the dark halo decreases with increasing dark halo mass (, the stellar profile becomes more concentrated for more luminous dwarfs). Thus, as the halo mass and virial radius increases, the stellar spheroid becomes increasingly concentrated in the deepest part of the potential well. If follows that the ratio, $f_\sigma \equiv M_\sigma/M_{\rm dm}$, of the dynamical mass within the largest stellar orbits to total dark matter mass is also reduced. Thus, the decrease of $f_\sigma$ for increasing dark matter mass of halos maintains the value of the dynamical mass $M_\sigma = f_\sigma M_{\rm dm}$ (measured by the velocity dispersion of the stars) almost constant, even though the total mass of the halo increases. The extent of the stellar spheroids in the lowest mass dwarfs is comparable in size to their virial radii at formation (see § \[sec:sims\]).
The metallicity-luminosity relation of the observed and simulated dwarfs is shown in the left panel of Figure \[ZL\]. \[Fe/H\] is plotted against V-band luminosity in solar units. Symbols for the previously known dwarfs, the new, ultra-faint dwarfs, and simulated fossils are the same as in Figure \[Kor\]. In this plot we color code simulated fossils according to their star formation efficiency, $f_*$. Red symbols show simulated dwarfs with $f_*<0.003$, blue $0.003
\le f_* \le 0.03$ and green $f_*>0.03$.
Using the data for the metallicity collected in [@BovillR:09], the new ultra-faint dwarfs do not appear to follow the tight luminosity-metallicity relationship observed in more luminous galaxies (although error bars are large). Note that here, as well as in [@BovillR:09] (although Table 3 in that paper was erroneously not updated), we have plotted data from [@Kirby:08] for the 6 ultra-faint Milky Way satellites provided in that paper. There are several physical mechanisms that may produce the observed scatter in metallicities of dwarfs at a given constant luminosity. The large spread of star formation efficiencies producing a dwarf of a given luminosity in our simulations is responsible for at least part of the large spread of the luminosity-metallicity relation. At this point it is unclear whether our simulations can reproduce the scatter of metallicities of simulated fossils, or if perhaps the luminosity of the lowest luminosity ultra-faint dwarfs has been reduced due to tidal interactions. As mentioned before we have suggested that the lowest luminosity ultra-faint dwarfs have not yet been discovered because their surface brightness lies below the SDSS detection limits.
Figure \[ZL\](right) shows the scatter of the metallicity of the stars, $\sigma_{\rm [Fe/H]}$, plotted against V-band luminosity and \[Fe/H\] respectively. The various point types and colors are the same used in the left panel of figure \[ZL\]. The large spread in the metallicity of the stars is a natural consequence of the hierarchical assembly of dwarf galaxies in cosmological simulations. It is not necessarily an indication that star formation in dwarf satellites was protracted for longer than $1$ Gyr, as argued in [@Grebel:04] to prove that star formation in dwarf spheroidals is not stopped by reionization feedback.
Discussion: The tidal scenario vs fossil hypothesis {#sec:disc}
===================================================
According to the results summarized above in Table \[tab:count\], the number of dark matter satellites of the Milky Way that have or had in the past $v_{\rm max}>20$ km s$^{-1}$ is smaller than the number of observed luminous satellites (after applying completeness corrections). This implies that non-fossil galaxies cannot account for all the observed satellites. Thus, pre-reionization fossils should exist.
However, we have already discussed the several uncertainties in estimating the numbers summarized in Table \[tab:count\]. Additional uncertainties that render the identification of fossils uncertain are the following. The mass and circular velocity of the dark matter halo of the Milky Way are not known precisely. Simulations should take into account the effect of baryons in modifying the density profile and the radial distribution of satellites. The effect of tidal stripping on the properties of the stars in the satellites is not well understood, thus we do not know if the tidal scenario is consistent with observations of ultra-faint dwarfs. The luminosity and stellar properties of non-fossil dwarf satellites are not known.
Non-fossil galaxies with $v_{\rm max}>20$ km s$^{-1}$ may lose a substantial fraction of their mass due to tidal interactions. If they survive the interaction, their properties, such as surface brightness and half light radius, may be modified. [@KravtsovGnedinKlypin04] estimate that about 10$\%$ of Milky Way dark matter satellites were at least ten times more massive at their formation than they are today. Although their simulation does not include stars, they favor the idea that the stellar properties of these halos would remain unchanged (, dSphs are not tidally stripped dIrr). In their model the majority of brighter dwarf satellites have been considerably more massive in the past and could have formed their stars undisturbed after reionization. More precisely, the redshift of reionization does not affect the results of their model for classic dwarfs because the probability of these to be fossils is low.
This version of the tidal model may be hard to distinguish observationally from the model we propose for the fossils because in both models the properties of the dwarfs are not modified by tidal forces (, their properties are those at formation). In addition, fossil dwarfs may stop forming stars for only about 2 to 4 Gyr after reionization, before starting to accrete gas again from the IGM. Thus, reionization may imprint a bimodal star formation history in some fossils, but this signature is not a robust discriminant because the star formation history of dwarfs cannot be determined with sufficient accuracy.
Observations seem to suggest that star formation in dwarf galaxies slightly more massive than $10^8-10^9$ M$_\odot$ may be similar to star formation in fossils and thus fit the observed properties of classical dSphs without requiring significant tidal stripping of stars. If star formation was included in [@KravtsovGnedinKlypin04], their model may have reproduced the properties of observed dwarf satellites that our simulations of pre-reionization dwarf galaxies already does. The differences between the two models will depend on whether fossil galaxies are allowed to form and on their properties. In other words, the two models may differ on the assumed mass of the smallest dark halo that can host luminous satellites. This critical mass cannot be directly observed in dwarf galaxies but, in principle, can be constrained by comparing the observed number of luminous satellites to the model predictions. Determining the minimum mass for a dark halo to become luminous is of great importance in understanding galaxy formation in the early universe.
To summarize, there are a few observational tests that can be used to distinguish true fossils from dSphs or dEs that form in more massive halos and form stars unaffected by IGM reionization. True fossils should have either a single old stellar population or have a bimodal star formation history produced by a temporary suppression of star formation after reionization and late gas accretion. In addition, if the number of observed Milky Way satellites (or the number of isolated dwarfs) exceeds some critical value determined using N-body simulations (, see Table \[tab:count\]), we may conclude that some pre-reionization fossils do exist in the Local Group.
It is likely that these tests will prove inconclusive for some time to come, unless the number (after corrections for completeness) of new ultra-faint galaxies surges in the coming years. The weakness of the star formation history test is that it requires measurements with precision of 1-2 Gyr of the stellar populations in order to be really discriminating between models that are quite similar to each other. This is hard to achieve especially for ultra-faint dwarfs with few stars. If the number of ultra-faint dwarfs remains about the same as today, the number argument may also remain controversial until more detailed theoretical modeling can reduce the current uncertainties surrounding the expected number of dark halos in the Milky Way and the completeness corrections of the observations. Ultimately, the case for the origin of ultra-faint dwarfs must be made on the basis of the model that does the best job of reproducing available observations.
Finally, even if pre-reionization fossils do not exist (, halos with $v_{\rm max}<20$ km s$^{-1}$ are all dark), a fraction of them should be able to accrete some gas at redshift $z<1-2$ and might be discoverable in the outer parts of the Local Group using H$\alpha$ or 21 cm surveys [, ALFALFA survey, @Giovanelli:05; @Giovanelli:07]. Of course, one should prove that the gas clouds are embedded in dark halos. Measurements of the gas cloud size, column density, and velocity broadening of the emission/absorption lines can be used to discriminate between “dark galaxies” and tidal debris. This is because the gas in dark galaxies is confined by the gravitational potential of the dark matter halo, while tidal debris or clouds formed via thermal instability are confined by the external gas pressure [@Ricotti:09]. This is another promising direction for determining the minimum mass of luminous galaxies in the universe.
Another variant of the tidal hypothesis for the origin of dSphs is a scenario in which dIrr galaxies transform into dSphs as they fall into the Milky Way and Andromeda, due to tidal and ram pressure stripping [@Mayeretal07; @Mayeretal06]. A work by [@PenarrubiaNM08] explores the idea that ultra-faint dSphs are tidally stripped dIrrs. They achieve some success in reproducing observed properties of ultra-faint dwarfs. While this type of tidal stripping can reproduce properties of an individual galaxy, it is unable to completely reproduce all the trends in the ultra-faint population. This is primarily seen in the kinematics of the ultra-faint dwarfs. Tidal stripping predicts a steeper than observed drop in the velocity dispersion of the stars with decreasing $L_V$ [@PenarrubiaNM08]. In addition several dSph do not show signs of strong tidal stripping. And XII and And XIV may be on their first approach to the Local Group [@Martinetal06; @Chapmanetal07]. Other examples of dSphs that are found distant from the center of their host galaxies are And XVIII, Cetus and Tucana [@McConnachieetal08].
Finally, another interesting case study is Leo T, that we have discussed extensively above in § \[sec:leoT\] and § \[sec:infall\]. Leo T properties can be explained in some detail as being a fossil that experienced a late phase of gas accretion [@Ricotti:09]. However, another possibility that should be explored quantitatively with simulations is that Leo T is more massive than a fossil but less massive than dIrr galaxies.
Conclusions and Future work {#sec:conc}
===========================
We have summarized our work on the formation of the first galaxies before reionization (, pre-reionization dwarfs) and the quest to identify the fossils of these first galaxies in the Local Group. The definition of a pre-reionization fossil is not directly related to the suppression of star formation experienced by these galaxies due to reionization feedback. Indeed, we discussed how pre-reionization fossils may experience a late phase of gas accretion and possibly star formation at redshift $z<1-2$. Most importantly, fossils are a population of dwarf galaxies whose formation (, the fraction of halos that are luminous) is self-regulated on cosmological distance scales by radiative processes. Their existence is not certain due to a possible strong negative feedback that may prevent the majority of these halos from ever forming stars. In addition, if negative feedback heavily suppresses the number and luminosity of these first galaxies, more massive halos with $v_{\rm max}>20$ km s$^{-1}$ will evolve differently because of the lower level of metal pre-enrichment of the IGM. To summarize, the critical circular velocity $v_{\rm max} \sim
20$ km s$^{-1}$ that we adopt to define a fossil is primarily motivated by fundamental differences in cooling and feedback processes that regulate star formation in these halos in the early universe. However, it is also close to the critical value for continued gas accretion after IGM reionization [@Gnedin-filteringmass00; @Hoeft:06; @OkamotoGT08].
The number of Milky Way and M31 satellites provides an indirect test of galaxy formation and the importance of positive and negative feedback in the early universe. This test, although the uncertainties are large, supports the idea that a fraction of the new ultra-faint dwarfs are fossils. The good agreement of the SDSS and new M31 ultra-faint dwarf properties with predictions of our simulations (RG05, GK06, Bovill & Ricotti 2009) does not prove the primordial origin of the new ultra-faint dwarfs, but it supports this possibility.
More theoretical work and more observational data are needed to prove that some dwarfs in the Local Group are true fossils of the first galaxies. Future theoretical work should focus on improving the accuracy of predictions on the properties of dwarf galaxies formed before reionization and their evolution to the present day. Modeling the evolution of the baryonic component after reionization in dwarf satellites and in the Milky Way – Andromeda system may be necessary to make robust predictions. More observational data will certainly be available in the near future. A large number of surveys, both at optical and radio wavelengths will be online in the near future (, [*Pan-STARRS, LSST, ALMA, EVLA, JWST, SKA*]{} to mention a few). Different survey strategies may be used to find and characterize fossil dwarf galaxies. A deep pencil beam survey would be useful to find the faintest dwarf satellites of the Milky Way and determine more precisely their Galactocentric distribution. A shallower all sky survey could be used to quantify the degree of anisotropy in the distribution of satellites around the Milky Way.
The star formation history of the dwarf galaxies is not strongly discriminatory because fossil galaxies may have a late phase of gas accretion and star formation during the last $9-10$ Gyrs [@Ricotti:09]. The distinction between fossils and non-fossil galaxies may be quite elusive but it is nevertheless important to understand galaxy formation and feedback in the early universe. Arguments based on counting the number of dwarfs in the Local universe are among the more solid arguments that could be used to prove the existence of fossil galaxies (see Table \[tab:count\]).
Future tests may be provided by deep surveys looking for ultra-faint galaxies in the local voids or looking for gas in dark galaxies (, dark halos that have been able to accrete gas from the IGM at $z<1-2$). Ultra-faint dwarfs should be present in the voids if dwarf galaxies formed in large numbers before reionization (Bovill & Ricotti, in preparation). If pre-reionization dwarfs never formed due to dominant negative feedback in the early universe, it is possible that a faint (in H$\alpha$ and 21 cm emission) population of dark galaxies exists in the outer parts of the Local Group. Hence, another way to detect fossil galaxies in the outer parts of the Milky Way or outside the super-galactic plane would be to search for neutral or ionized gas that they may have accreted from the IGM. Future radio telescopes (, [*ALMA, EVLA, SKA*]{}) may be able to detect neutral hydrogen in dark galaxies or in ultra-faint dwarfs. Ionized gas in the outer parts of dark halos may be observed in absorption along the line of sight of distant quasars (for instance in or with [*COS*]{} on the [*HST*]{}). However, the probability that a line of sight toward a quasar intersects the ionized gas collected from the IGM by dark or fossil galaxies might be small. Additional theoretical work is required to address these issues.
[^1]: Leo T was found because it contains a young stellar population and gas. Otherwise, it would not have been identified as an ultra-faint dwarf due to its large Galactocentric distance.
[^2]: Movies of 2D slices and 3D rendering of the simulations are publicly available on the web at the URL: http://astro.umd.edu/$\sim$ricotti/movies.html
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
Kopitzki et al (preceeding comment) claim that the relationship between Renormalized and Kullback-Leibler entropies has already been given in their previous papers. Moreover, they argue that the first can give more useful information for e.g. localizing the seizure-generating area in epilepsy patients.
In our reply we stress that if the relationship between both entropies would have been known by them, they should have noticed that the condition on the effective temperature is unnecessary. Indeed, this condition led them to choose different reference segments for different channels, even if this was physiologically unplausible. Therefore, we still argue that it is very unlikely that renormalized entropy will give more information than the conventional Kullback-Leibler entropy.
---
\
$^1$[*Sloan-Swartz Center for Theoretical Neurobiology.*]{}\
[*California Institute of Technology, MC 216-76. 91125 Pasadena, CA. USA*]{}\
$^2$[*John von Neumann Institute for Computing,*]{}\
[*Forschungszentrum Jülich GmbH, D - 52425 Jülich, Germany*]{}\
$^3$[*Department of Epileptology, University of Bonn,*]{}\
[*Sigmund-Freud Str. 25, D - 53105 Bonn, Germany*]{}\
We thank the authors of the preceding comment for pointing out a misprint in [@quian] (in the line following Eq. (8) it should read $p\equiv \tilde{q}$ instead of $p\equiv q$), and a numerical inconsistency in Figs. 2-4 of ref. [@quian]. The latter resulted from an error in our code. The corrected data is shown in Fig. 1 below, where we added the renormalized entropy values calculated with a pre-seizure reference for completeness. Note that Eq. (10) of [@quian], i.e. $| \Delta H | \leq K(p|q)$, is now verified in all cases [@comment1]. Despite this correction, the data are qualitatively similar to the ones presented in [@quian], and we still conclude that renormalized entropy does not give more information than standard Kullback-Leibler (KL) entropy.
Apart from this, we do not agree with any of the other claims raised in the preceding comment (and we still have some discrepancy in details with the numerical results shown in [@kopitzki] whose origin is unclear to us).
The first main point of ref. [@quian] was to show that the “renormalized entropy" (RE) proposed in [@saparin] and applied to EEG data in [@kopitzki] was indeed a KL entropy, but taking an unusual “renormalized" reference. We maintain, in contrast to claims made in the comment, that this relation (Eq. (9) in [@quian]) was not mentioned in [@saparin; @klimo-physica], and not in [@kopitzki] either. Indeed, due to it, the condition $T_{eff} \geq 1$ postulated in [@saparin; @klimo-physica; @kopitzki] is not needed to obtain the inequality $\Delta H
\leq 0$. The fact that the latter was claimed in [@saparin; @klimo-physica; @kopitzki] to hold only for $T_{eff} \geq 1$ indicates that the authors were not aware of the relation to KL (or “relative") entropy. Apart from this, we also wanted to give a simple treatment free of all allusions to statistical thermodynamics, the latter making the treatments in [@saparin; @kopitzki] hard to understand.
Our second point was that RE is very unlikely to be more useful than the usual (un-renormalized) KL entropy for the analysis of EEGs from epileptic patients, as claimed in [@kopitzki]. On the one hand this was based on the numerical similarity between RE and standard KL entropies, which is enforced by several inequalities and which makes it unlikely a priori that either is superior. On the other hand, we verified this explicitly by detailed numerical calculations which indeed showed that both behaved very similar. It is clear from Fig. 1 that major differences are due to the choice of the reference window. In contrast to what is suggested in the comment, we did not base this conclusion entirely on theoretical arguments.
Finally, we also stressed that the condition $T_{eff} \geq 1$ – which is not needed at all – has led the authors of [@kopitzki] to choose reference points which are physiologically very unfortunate. Again, we remark that it is very unreliable to compare a relative entropy measure obtained from EEG recordings at different electrodes by using different references (from the pre-seizure stage, during the seizure, or from the post-seizure stage) for each electrode in order to localize an epileptic focus. Thus the failure of realizing that RE is a sort of KL entropy – or at least of drawing the obvious consequences from this observation – has hampered its application to EEGs.
[10]{}
R. Quian Quiroga, J. Arnhold, K. Lehnertz, and P. Grassberger, Phys. Rev. E [**62**]{}, 8382 (2000).
K. Kopitzki, P.C. Warnke, and J. Timmer, Phys. Rev. E [**58**]{}, 4859 (1998).
P. Saparin, A. Witt, J. Kurths, and V. Anishchenko, Chaos, Solitons and Fractals [**4**]{}, 1907 (1994).
Yu. L. Klimontovich, Physica A [**142**]{}, 390 (1987).
We do not understand how this relationship proved in ref. [@quian] can be invalidated by “basic mathematical operations" as claimed by the authors in their reference \[9\].
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
In this paper, we investigate an electrodynamics in which the physical modes are coupled to a Lorentz-violating (LV) background by means of a higher-derivative term. We analyze the modes associated with the dispersion relations (DRs) obtained from the poles of the propagator. More specifically, we study Maxwell’s electrodynamics modified by a LV operator of mass dimension 6. The modification has the form ${D_{\beta\alpha}}\partial_{\sigma}F^{\sigma\beta}\partial_{\lambda}
F^{\lambda\alpha}$, i.e., it possesses two additional derivatives coupled to a *CPT*-even tensor $D_{\beta\alpha}$ that plays the role of the fixed background. We first evaluate the propagator and obtain the dispersion relations of the theory. By doing so, we analyze some configurations of the fixed background and search for sectors where the energy is well-defined and causality is assured. A brief analysis of unitarity is included for particular configurations. Afterwards, we perform the same kind of analysis for a more general dimension-6 model. We conclude that the modes of both Lagrange densities are possibly plagued by physical problems, including causality and unitarity violation, and that signal propagation may become physically meaningful only in the high-momentum regime.
author:
- 'Rodolfo Casana$^{a}$'
- 'Manoel M. Ferreira, Jr.$^{a}$'
- 'Letícia Lisboa-Santos$^{a}$'
- 'Frederico E.P. dos Santos$^{b}$'
- 'Marco Schreck$^{a}$'
title: |
Maxwell electrodynamics modified by *CPT*-even and Lorentz-violating\
dimension-6 higher-derivative terms
---
Introduction
============
Physics beyond the Standard Model has been under extensive development in the latest years, encompassing Lorentz-violating (LV) theories as one branch of investigation. The minimal Standard-Model Extension (SME) [@Colladay; @Samuel] is a general gauge-invariant and power-counting renormalizable framework that incorporates terms of Lorentz invariance violation by means of tensor-valued background fields fixed under particle Lorentz transformations. These background fields can be interpreted as vacuum expectation values that are generated by spontaneous symmetry breaking taking place in a more fundamental theory. Studies in the SME have been pursued to look for LV effects and to develop a precision programme that allows us to scrutinize the limits of Lorentz symmetry in several physical interactions. In this sense, many investigations were performed in the context of the SME fermion sector [@Fermion1; @Fermion2], *CPT*-violating contributions [@CPT], the *CPT*-odd electromagnetic sector [@Adam1; @Cherenkov1], the *CPT*-even electromagnetic sector [@Cherenkov2; @KM], fermion-photon interactions [@KFint; @Interac; @Schreck1], and radiative corrections [@Radio]. Lorentz-violating theories are also connected to higher-dimensional operators. In this sense, nonminimal extensions of the SME were developed, both in the photon [@Kostelec1] and fermion sector [@Kostelec2], composed by *CPT*-even and *CPT*-odd higher-derivative operators. Other higher-dimensional LV theories [@Myers1; @Marat] were also proposed and examined. Nonminimal higher-dimensional couplings that do not involve higher derivatives have been proposed and constrained, as well [@NModd1; @NM3; @NMkost].
Models of a higher-derivative electrodynamics have been investigated since the advent of Podolsky’s theory [@Podolsky1], characterized by the Lorentz- and gauge-invariant term $\partial_{\alpha} F^{\alpha\beta} \partial_{\lambda} F^{\lambda}{}_{\beta}$, one of the simplest dimension-6 structures that can be built with the electromagnetic field. Podolsky’s Lagrangian is $$\label{eq:lagrangian-podolsky}
\mathcal{L}=-\frac{1}{4}F_{\mu\nu}F^{\mu}{}^{\nu}+\frac{\theta^{2}}{2}\partial_{\alpha}F^{\alpha}{}^{\beta}\partial_{\lambda}F_{\phantom{\lambda}\beta}^{\lambda}-j_{\mu}A^{\mu}\,,$$ where $F_{\mu\nu}=\partial_{\mu}A_{\nu}-\partial_{\nu}A_{\mu}$ is the field strength tensor assigned to the vector field $A_{\mu}$ and $\theta$ is Podolsky’s parameter having mass dimension $-1$. The vector field is coupled to a conserved four-current $j_{\mu}$. One of the most remarkable characteristics of Podolsky’s electrodynamics is the generation of a massive mode without the loss of gauge symmetry, being in this aspect different from Proca’s theory. The Podolsky propagator contains two poles, one corresponding to the massless photon and the other one associated with the massive photon. At the classical level, the massive mode of the model has the advantage of curing divergences connected to the pointlike self-energy, but at the quantum level it is associated with the occurrence of ghosts [@Accioly1]. The suitable gauge condition to address Podolsky’s electrodynamics is not the usual Lorenz gauge, but a modified gauge relation [@GalvaoP], compatible with the existence of five degrees of freedom (two related to a massless photon and three related to the massive mode). Other aspects of quantum field theories in the presence of the Podolsky term, such as path integral quantization and finite-temperature effects [@Podolsky3; @Podolsky4], renormalization [@Podolsky5], as well as multipole expansion and classical solutions [@Podolsky6] were also examined.
Another dimension-6 term, proposed in the late sixties, $F_{\mu\nu}\partial_{\alpha}\partial^{\alpha}F^{\mu\nu}$, defines Lee-Wick electrodynamics [@LeeWick], which leads to a finite self-energy for a pointlike charge in (1+3) spacetime dimensions. Furthermore, it produces a bilinear contribution to the Lagrangian similar to that of Podolsky’s term but with opposite sign. This wrong sign yields energy instabilities at the classical level, while it leads to negative-norm states in the Hilbert space at the quantum level. Lee and Wick also proposed a mechanism to preserve unitarity, which removes all states containing Lee-Wick photons from the Hilbert space. This theory regained attention after the proposal of the Lee-Wick standard model [@LW], based on a non-Abelian gauge structure free of quadratic divergences. Such a model had a broad repercussion, with many contributions in both the theoretical and phenomenological sense [@LW2]. In the Lee-Wick scenario, studies of ghost states [@LWghost], constructions endowed with higher derivatives [@LWhd], renormalization aspects [@LWren], and finite-temperature investigations [@LWft] have been reported, as well. The study of higher-derivative terms in quantum field theories was also motivated by their role as ultraviolet regulators [@A.A]. Some works were dedicated to investigating interactions between stationary sources in the context of Abelian Lee-Wick models, with emphasis on sources distributed along parallel branes with arbitrary dimension and the Dirac string in such a context [@Barone1]. Lee-Wick electrodynamics was also studied for the case of the self-energy of pointlike charges in arbitrary dimensions, exhibiting a finite ultraviolet result for $d=1$ and $d=3$ (spatial dimensions) [@Barone2], for the case of perfectly conducting plates [@Barone3], and in the evaluation of the interaction between two pointlike charges [@Accioly2]. Recently, a LV higher-derivative and dimension-6 term, $d_{\beta}d_{\sigma}\partial_{\alpha}F^{\alpha}{}^{\beta}\partial_{\lambda}
F^{\lambda\sigma}$ with an observer four-vector $d_{\mu}$, radiatively generated in Ref. [@Petrov], was considered in the context of Maxwell’s Lagrangian [@Barone4]. The latter study focused on interactions between external sources with the modified electromagnetic field, as performed in Refs. [@Barone1; @Barone2].
For almost 20 years now, Lorentz-violating contributions of mass dimensions 3 and 4 have been investigated extensively from both a phenomenological and a theoretical point of view. Many of the corresponding controlling coefficients are tightly constrained, especially in the photon and lepton sector . Since a point in time not long ago, a significant interest was aroused in field theories endowed with higher-order derivatives. In the *CPT*-even photon sector, the leading-order contributions in an expansion in terms of additional derivatives are the dimension-6 ones. Hence, these are also the most prominent ones that could play a role in nature, if higher-derivative Lorentz violation existed. Note that Lorentz-violating terms of mass dimension 6 are known to emerge from theories of noncommutative spacetimes, as the noncommutativity tensor has a mass dimension of $-2$.
In the present work, we investigate basic features of a higher-derivative electrodynamics, in which the physical fields are coupled to a *CPT*-even and LV background by means of a dimension-6 term. More specifically, we study Maxwell’s theory modified by a LV operator of mass dimension 6, which possesses two additional derivatives coupled to a *CPT*-even fixed background, $D_{\beta\alpha}$, in a structure of the form ${D_{\beta\alpha}}
\partial_{\sigma}F^{\sigma\beta}\partial _{\lambda} F^{\lambda\alpha}$. The latter is a kind of anisotropic Podolsky term, i.e., it is a natural Lorentz-violating extension of Podolsky’s theory. Initially, the condition is discussed (nonzero trace) for which a LV structure of this kind comprises the Podolsky term. It is interesting to note that in the recent article [@Bonetti:2016vrq], Lorentz violation is considered in a scenario with intact supersymmetry. The Lorentz-violating background fields are assumed to be linked to a supersymmetric multiplet and their effect on photon propagation is studied. After integrating out the contributions from the photino, the effective Lagrangian given by their Eq. (9) incorporates the type of modified Podolsky term that we are studying.
So far, not much is known about the properties of Lorentz-violating theories including higher derivatives. It is a well-established fact, though, that higher-dimensional operators lead to a rich plethora of new effects as well as additional issues. For example, the existence of additional time derivatives may produce exotic modes that cannot be considered as perturbations of the standard ones. These modes can lead to an indefinite metric in Hilbert space, which is connected to the occurrence of states with negative norm. The procedure developed by Lee and Wick [@LeeWick] makes it possible to deal with such modes in a quantized theory such that a breakdown of unitarity is prevented. Before delving into these possibly very profound problems of Lorentz-violating theories including higher derivatives, the classical properties of these frameworks should be well understood first.
Describing classical aspects of LV theories with higher derivatives is the main motivation of the current paper. Hence, we are interested in obtaining the Green’s function of the field equations and the dispersion relations as well as developing an understanding of classical causality. Using a technique already employed in some previous LV models [@Propagator], which consists of finding a closed projector algebra, and using the prescription $D_{\mu\nu}=({B_{\mu}}C_{\nu}+{B_{\nu}}C_{\mu})/2$ with two observer four-vectors $B_{\mu}$ and $C_{\mu}$, the propagator is derived and the dispersion relations are determined from its poles. The goal of this work is to examine signal propagation within a Podolsky electrodynamics modified by the term ${D_{\beta\alpha}}\partial_{\sigma}F^{\sigma\beta}\partial_{\lambda}F^{\lambda\alpha}$. Thus, the modes described by the corresponding dispersion relations are analyzed for several configurations of the fixed background and we search for sectors where the energy is well-defined and causality is assured. Furthermore, we will perform a brief analysis of unitarity of the theory for a vanishing Podolsky parameter. After doing so, we present a more general dimension-6 higher-derivative Lagrangian that can be proposed in the presence of the rank-2 background $D_{\beta\alpha}$. The latter also involves a kind of anisotropic Lee-Wick term ${D^{\mu\nu}}\partial_{\sigma}F^{\sigma\lambda} \partial_{\mu}F_{\nu\lambda}$. The corresponding propagator and the dispersion relations are derived again. Mode propagation is examined for several configurations of the background, revealing that the dispersion relations of these LV dimension-6 model may exhibit a physical behavior in the limit of large momenta only. We finally show that the dimension-6 terms considered here are contained in the nonminimal SME developed by Kostelecký and Mewes [@Kostelec1]. Throughout the paper, natural units will be employed with $\hbar=c=1$.
Maxwell electrodynamics modified by higher-derivative terms: some possibilities
===============================================================================
As a first step, we propose a Maxwell electrodynamics modified by a higher-derivative, *CPT*-even term of mass dimension 6 including two additional derivatives coupled to a fixed tensor $D_{\beta\alpha},$ that is, $${D_{\beta\alpha}}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda}{}^{\alpha}\,, \label{TADLV1}$$ which represents a kind of anisotropic and generalized Podolsky term. Without a restriction of generality, $D_{\beta\alpha}$ can be taken to be symmetric, as its antisymmetric part does not contribute, anyhow. Hence, a possible Lagrangian to be considered is $$\mathcal{L}=-\frac{1}{4}F_{\mu\nu}F^{\mu}{}^{\nu}+\eta^{2}{D_{\beta\alpha}}
\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda}{}^{\alpha}\,, \label{LVSLAD1}$$ where the parameter $\eta$ has dimension of (mass)$^{\text{--1}}$. A property to check is if the anisotropic piece (\[TADLV1\]) contains a sector that is equivalent to the Podolsky term of Eq. (\[eq:lagrangian-podolsky\]). Such a term is generated from a nonvanishing trace of $D_{\mu\nu}$, what can be shown quickly. If $D_{\mu\nu}$ contains nonzero diagonal components of the form $D_{11}=D_{22}=D_{33}=-D_{00}$, the tensor involves a trace that is given by $${D^{\lambda}_{\phantom{\lambda}\lambda}}=D_{00}-D_{ii}=4D_{00}\,.$$ We define a new traceless tensor $\tilde{D}_{\beta\alpha}$ by subtracting the trace from the latter: $${\tilde{D}_{\beta\alpha}\equiv D_{\beta\alpha}-\frac{1}{4}g_{\beta\alpha}D^{\rho}_{\phantom{\rho}\rho}=D_{\beta\alpha}-g_{\beta\alpha}D_{00}\,,} \label{Dtraceless}$$ which fulfills $\tilde{D}^{\kappa}_{\phantom{\kappa}\kappa}=0$ and leads to $${\tilde{D}_{\beta\alpha}}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda\alpha}={D_{\beta\alpha}}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda\alpha}
-{D_{00}(\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F_{\phantom{\lambda}\beta}^{\lambda})}\,. \label{HDLVT4}$$ The second term on the right-hand side corresponds to a Podolsky term, i.e., such a term appears in connection to the trace of $D_{\mu\nu}$, indeed. In this sense, there are two possibilities that can be pursued, i.e., we can assess a dimension-6 electrodynamics that either contains or does not contain the Podolsky term. One option that exhibits, in principle, the same physical content of Lagrangian (\[LVSLAD1\]) is to consider Podolsky’s electrodynamics modified by the traceless LV dimension-6 term of Eq. (\[Dtraceless\]), that is, $$\mathcal{L}=-\frac{1}{4}F_{\mu\nu}F^{\mu}{}^{\nu}+\frac{\theta^{2}}{2}\partial_{\alpha}F^{\alpha}{}^{\beta}\partial_{\lambda}F_{\phantom{\lambda}\beta}^{\lambda}+\eta^{2}{\tilde{D}_{\beta\alpha}}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda}{}^{\alpha}\,.
\label{LVSLAD2}$$ A cleaner option including only the LV dimension-6 contribution in the context of Maxwell’s electrodynamics, is defined when the Podolsky sector is zero, $D^{\kappa}_{\phantom{\kappa}\kappa}=0$, so that the Lagrangian to be addressed is $$\mathcal{L}=-\frac{1}{4}F_{\mu\nu}F^{\mu}{}^{\nu}+\eta^{2}{\tilde{D}_{\beta\alpha}}\partial_{\sigma
}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda}{}^{\alpha}\,. \label{a2}$$
\[c\][lllll]{}& $C$ & $P$ & $T$ & *CPT*\
$D_{00}$ & $+$ & $+$ & $+$ & $\ +$\
$D_{0i}$ & $+$ & $-$ & $-$ & $\ +$\
$D_{ij}$ & $+$ & $+$ & $+$ & $\ +$\
& & & &
The components of the tensor $D_{\beta\alpha}$ can be classified in accordance with the behavior under the discrete *C*, *P*, and *T* operations. To do so, we decompose the sum over the contracted indices in the term (\[TADLV1\]) into components of the electric and magnetic fields $\mathbf{E}$, $\mathbf{B}$: $$\begin{aligned}
{D_{\beta\alpha}}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda}{}^{\alpha
} & ={D_{00}}\partial_{a}E^{a}{}\partial_{b}E^{b}{}-2\epsilon^{bil}{D_{0i}}\partial_{a}E^{a}{}\partial_{b}B^{l}\nonumber\\
& \phantom{{}={}}-2{D_{0i}}\partial_{a}E^{a}{}\partial_{0}E{}^{i}+{D_{ij}}\partial
_{0}E{}^{i}\partial_{0}E{}^{j}\nonumber\\
& \phantom{{}={}}+2\epsilon^{bjl}{D_{ij}}\partial_{0}E{}^{i}\partial_{b}B^{l}
+\epsilon^{bjl}\epsilon^{aim}{D_{ij}}\partial_{a}B^{m}\partial_{b}B^{l}\,,
\label{Dcomp}$$ remembering that $F^{a0}{}=E^{a},F^{b}{}^{j}=-\epsilon^{bjl}B^{l}$ with the three-dimensional Levi-Civita symbol $\epsilon^{ijk}$. Under charge conjugation (*C*), the electric and magnetic fields behave according to $\mathbf{E}\rightarrow-\mathbf{E}$ and $\mathbf{B}\rightarrow-\mathbf{B},$ while $\partial_{\mu}\rightarrow\partial_{\mu}.$ In this way, we notice that the coefficients $D_{00},D_{0i},D_{ij}$ are *C*-even. Under parity (*P*), $\mathbf{E}\rightarrow-\mathbf{E}$ and $\mathbf{B}\rightarrow\mathbf{B}$, $\partial_{a}\rightarrow-\partial_{a},\partial
_{0}\rightarrow\partial_{0},$ so that $D_{0i}$ is parity-odd, and $D_{00}$ and $D_{ij}$ are parity-even. Under time reversal ($T),$ $\mathbf{B}\rightarrow-\mathbf{B}$, $\mathbf{E}\rightarrow\mathbf{E}$, with $\partial
_{a}\rightarrow\partial_{a},\partial_{0}\rightarrow-\partial_{0}.$ This implies that $D_{0i}$ is $T$-odd, while $D_{00}$ and $D_{ij}$ are *T*-even. A summary of these properties can be found in Tab. \[tab:transformation-properties\].
Propagator of the dimension-6 generalized Podolsky theory
=========================================================
In this section, we consider the Maxwell Lagrangian modified by the Podolsky and the dimension-6 anisotropic higher-derivative term (\[TADLV1\]), given as $$\mathcal{L}=-\frac{1}{4}F_{\mu\nu}F^{\mu}{}^{\nu}
+\frac{\theta^2}{2}\partial_{\alpha}F^{\alpha}{}^{\beta}\partial_{\lambda}F_{\phantom{\lambda}\beta}^{\lambda}
+\eta^{2}D_{\beta\alpha}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda\alpha}{}
+\frac{1}{2\xi}\left(\partial_{\mu}A^{\mu}\right)^2\,, \label{LPVL1a}$$ where the last contribution is introduced to fix the gauge.[^1] The parameters $\theta$, $\eta$ have dimension of (mass)$^{\text{--1}}$ and $D_{\beta\alpha}$ is the dimensionless *CPT*-even tensor introduced before. The coefficients $\theta^2$ and $\eta^2$ are here considered as positive in analogy with Podolsky’s theory, where $\theta^2>0$ is a necessary condition for obtaining a physical dispersion relation. In a broad context, there exists the possibility of considering $\theta^2$ and $\eta^2$ as negative. These choices have the potential for altering dispersion relations and other physical properties of the theory such as unitarity. However, an investigation of this sector is beyond the scope of the paper, which is why we will assume that both $\theta^2\geq 0$ and $\eta^2\geq 0$. The Lagrangian (\[LPVL1a\]) can be written in the bilinear form, $$\label{eq:lagrange-density-bilinear}
\mathcal{L}=\frac{1}{2}A^{\nu}O_{\mu\nu}A^{\mu}\,,$$ where $O_{\mu\nu}$ is$$\begin{aligned}
O_{\mu\nu} & ={(1+\theta^2\square)\square\Theta_{\mu\nu}}-\frac{1}{\xi}\square\Omega_{\mu\nu}\nonumber\\
&\phantom{{}={}}+2\eta^2[{D_{\nu\mu}}\square^2
-{D_{\nu\alpha}}\square\partial_{\mu}\partial^{\alpha}
-{D_{\sigma\mu}}\square\partial_{\nu}\partial^{\sigma}
+{D_{\sigma\alpha}}\partial_{\nu}\partial_{\mu}\partial^{\sigma}\partial^{\alpha}]\,. \label{OLV1}$$
$$\begin{array}
[c]{cccccc}\toprule & \Theta_{\phantom{\nu}\alpha}^{\nu} & \Omega_{\phantom{\nu}\alpha
}^{\nu} & B_{\alpha}\partial^{\nu} & B_{\alpha}C^{\nu} & C^{\nu}\partial_{\alpha}\\
\colrule \Theta_{\mu\nu} & \Theta_{\mu\alpha} & 0 & 0 & B_{\alpha}C_{\mu}
-\frac{\rho}{\square}B_{\alpha}\partial_{\mu} & C_{\mu}\partial_{\alpha}-\rho\Omega_{\mu\alpha}\\
\colrule \Omega_{\mu\nu} & 0 & \Omega_{\mu\alpha} & B_{\alpha}\partial_{\mu} &
\frac{\rho}{\square}B_{\alpha}\partial_{\mu} & \rho\Omega_{\mu\alpha}\\
\colrule B_{\nu}\partial_{\mu} & B_{\alpha}\partial_{\mu}-\kappa\Omega
_{\mu\alpha} & \kappa\Omega_{\mu\alpha} & \kappa B_{\alpha}\partial_{\mu} &
(B\cdot C) B_{\alpha}\partial_{\mu} & (B\cdot
C)\square\Omega_{\mu\alpha}\\
\colrule B_{\nu}C_{\mu} & B_{\alpha}C_{\mu}-\frac{\kappa}{\square}C_{\mu
}\partial_{\alpha} & \frac{\kappa}{\square}C_{\mu}\partial_{\alpha} & \kappa
B_{\alpha}C_{\mu} & (B\cdot C)B_{\alpha}C_{\mu} & (
B\cdot C) C_{\mu}\partial_{\alpha}\\
\colrule C_{\mu}\partial_{\nu} & 0 & C_{\mu}\partial_{\alpha} & \square
B_{\alpha}C_{\mu} & \rho B_{\alpha}C_{\mu} & \rho C_{\mu}\partial_{\alpha}\\
\colrule B_{\mu}\partial_{\nu} & 0 & B_{\mu}\partial_{\alpha} & \square
B_{\mu}B_{\alpha} & \rho B_{\mu}B_{\alpha} & \rho B_{\mu}\partial_{\alpha}\\
\colrule B_{\mu}C_{\nu} & B_{\mu}C_{\alpha}-\frac{\rho}{\square}B_{\mu
}\partial_{\alpha} & \frac{\rho}{\square}B_{\mu}\partial_{\alpha} & \rho
B_{\mu}B_{\alpha} & C^2B_{\mu}B_{\alpha} & C^2B_{\mu}\partial_{\alpha}\\
\colrule C_{\nu}\partial_{\mu} & C_{\alpha}\partial_{\mu}-\rho\Omega
_{\mu\alpha} & \rho\Omega_{\mu\alpha} & \rho B_{\alpha}\partial_{\mu} &
C^{2}B_{\alpha}\partial_{\mu} & C^2\square\Omega_{\mu\alpha}\\
\colrule B_{\mu}B_{\nu} & B_{\mu}B_{\alpha}-\frac{\kappa}{\square}B_{\mu
}\partial_{\alpha} & \frac{\kappa}{\square}B_{\mu}\partial_{\alpha} & \kappa
B_{\mu}B_{\alpha} & (B\cdot C) B_{\mu}B_{\alpha} & (
B\cdot C) B_{\mu}\partial_{\alpha}\\
\colrule C_{\mu}C_{\nu} & C_{\mu}C_{\alpha}-\frac{\rho}{\square}C_{\mu
}\partial_{\alpha} & \frac{\rho}{\square}C_{\mu}\partial_{\alpha} & \rho
B_{\alpha}C_{\mu} & C^2B_{\alpha}C_{\mu} & C^2C_{\mu}\partial_{\alpha}\\
\botrule
\end{array}$$
Here, we have used the longitudinal and transverse projectors, $$\label{eq:projectors}
\Omega_{\beta\lambda}=\frac{\partial_{\beta}\partial_{\lambda}}{\square}\,,\quad
\Theta_{\beta\lambda}=g_{\beta\lambda}-\Omega_{\beta\lambda}\,,$$ respectively, where $g_{\mu\nu}$ is the Minkowski metric with signature $(+,-,-,-)$. To derive the propagator, we propose the following parameterization:$$D_{\mu\nu}=\frac{1}{2}({B_{\mu}}C_{\nu}+{B_{\nu}}C_{\mu})\,,
\label{Presc1}$$ where $C_{\mu},B_{\nu}$ are two independent observer four-vectors giving rise to preferred spacetime directions. This parameterization is nearly general and describes most of the configurations of the symmetric $D_{\beta\alpha}$ tensor. It is used for technical reasons, mainly connected to the construction of the propagator of this theory. Furthermore, it allows us to classify different sectors of the theory by geometric properties related to the two vectors, e.g., orthogonality of their spatial parts. With the latter choice, the operator of Eq. (\[OLV1\]) becomes $$\begin{aligned}
O_{\mu\nu} & =(1+\theta^2\square)\square\Theta_{\mu\nu}+\left(2\eta^2{\kappa\rho}-\frac{1}{\xi}\right)\square\Omega_{\mu\nu}
+\eta^2{(B_{\mu}C_{\nu}+B_{\nu}C_{\mu})}\square^2 \nonumber\\
&\phantom{{}={}}-\eta^2\square({C_{\nu}}\kappa \partial_{\mu}
+{C_{\mu}}\kappa\partial_{\nu}+{B_{\nu}}\rho\partial_{\mu}
+{B_{\mu}}\rho\partial_{\nu})\,,
\label{OLV2}$$ with$$\kappa=B_{\alpha}\partial^{\alpha}\,,\quad\rho=C_{\alpha}\partial^{\alpha}\,.
\label{karo}$$
$$\begin{array}
[c]{cccccc}\toprule & B^{\nu}\partial_{\alpha} & B^{\nu}C_{\alpha} & C_{\alpha}\partial^{\nu} & B^{\nu}B_{\alpha} & C^{\nu}C_{\alpha}\\
\colrule \Theta_{\mu\nu} & B_{\mu}\partial_{\alpha}-\kappa\Omega_{\mu\alpha} &
B_{\mu}C_{\alpha}-\frac{\kappa}{\square}C_{\alpha}\partial_{\mu} & 0 & B_{\mu
}B_{\alpha}-\frac{\kappa}{\square}B_{\alpha}\partial_{\mu} & C_{\mu}C_{\alpha
}-\frac{\rho}{\square}C_{\alpha}\partial_{\mu}\\
\colrule \Omega_{\mu\nu} & \kappa\Omega_{\mu\alpha} & \frac{\kappa}{\square
}C_{\alpha}\partial_{\mu} & C_{\alpha}\partial_{\mu} & \frac{\kappa}{\square
}B_{\alpha}\partial_{\mu} & \frac{\rho}{\square}C_{\alpha}\partial_{\mu}\\
\colrule B_{\nu}\partial_{\mu} & \square B^{2}\Omega_{\mu\alpha} &
B^2C_{\alpha}\partial_{\mu} & \kappa C_{\alpha}\partial_{\mu} &
B^2B_{\alpha}\partial_{\mu} & (B\cdot C) C_{\alpha}\partial_{\mu}\\
\colrule B_{\nu}C_{\mu} & B^2C_{\mu}\partial_{\alpha} & B^2C_{\mu
}C_{\alpha} & \kappa C_{\mu}C_{\alpha} & B^2B_{\alpha}C_{\mu} & (
B\cdot C) C_{\mu}C_{\alpha}\\
\colrule C_{\mu}\partial_{\nu} & \kappa C_{\mu}\partial_{\alpha} & \kappa
C_{\mu}C_{\alpha} & \square C_{\mu}C_{\alpha} & \kappa B_{\alpha}C_{\mu} &
\rho C_{\mu}C_{\alpha}\\
\colrule B_{\mu}\partial_{\nu} & \kappa B_{\mu}\partial_{\alpha} & \kappa
B_{\mu}C_{\alpha} & \square B_{\mu}C_{\alpha} & \kappa B_{\mu}B_{\alpha} &
\rho B_{\mu}C_{\alpha}\\
\colrule B_{\mu}C_{\nu} & (B\cdot C)B_{\mu}\partial_{\alpha} &
(B\cdot C)B_{\mu}C_{\alpha} & \rho B_{\mu}C_{\alpha} & (
B\cdot C) B_{\mu}B_{\alpha} & C^2B_{\mu}C_{\alpha}\\
\colrule C_{\nu}\partial_{\mu} & \square(B\cdot C) \Omega
_{\mu\alpha} & (B\cdot C)C_{\alpha}\partial_{\mu} & \rho
C_{\alpha}\partial_{\mu} & (B\cdot C)B_{\alpha}\partial_{\mu} &
C^2C_{\alpha}\partial_{\mu}\\
\colrule B_{\mu}B_{\nu} & B^2B_{\mu}\partial_{\alpha} & B^2B_{\mu}C_{\alpha} &
\kappa B_{\mu}C_{\alpha} & B^2B_{\mu}B_{\alpha} & (B\cdot
C)B_{\mu}C_{\alpha}\\
\colrule C_{\mu}C_{\nu} & (B\cdot C)C_{\mu}\partial_{\alpha} &
(B\cdot C)C_{\mu}C_{\alpha} & \rho C_{\mu}C_{\alpha} &
(B\cdot C)B_{\alpha}C_{\mu} & C^2C_{\mu}C_{\alpha}\\
\botrule
\end{array}$$
To derive the propagator, we need to invert the operator $O_{\mu\nu},$ composed of the projectors $\Theta_{\mu\nu}$, $\Omega_{\mu\nu}$, $B_{\nu}\partial_{\mu}$, $B_{\mu}\partial_{\nu}$, $C_{\mu}B_{\nu}$, $C_{\nu}B_{\mu}$, $C_{\mu}\partial_{\nu}$, $C_{\nu}\partial_{\mu}$. In this sense, the *Ansatz* $$\begin{aligned}
\Delta_{\phantom{\nu}\alpha}^{\nu} & =a\Theta_{\phantom{\nu}\alpha}^{\nu}
+b\Omega_{\phantom{\nu}\alpha}^{\nu}+cB_{\alpha}\partial^{\nu}
+dC^{\nu}B_{\alpha}+eC^{\nu}\partial_{\alpha}\nonumber\\
&\phantom{{}={}}+fB^{\nu}\partial_{\alpha}+gC_{\alpha}B^{\nu}+hC_{\alpha}\partial^{\nu}
+iB^{\nu}B_{\alpha}+jC^{\nu}C_{\alpha}\,, \label{PropLV3}$$ for the Green’s function $\Delta_{\nu\alpha}=\Delta_{\nu\alpha}(x-y)$ is proposed obeying the condition, $O_{\mu\nu}\Delta_{\text{ }}^{\nu\alpha}=\delta
_{\mu}^{\phantom{\mu}\alpha},$ or$$O_{\mu\nu}\Delta_{\phantom{\nu}\alpha}^{\nu}=\Theta_{\mu\alpha}+\Omega_{\mu\alpha}\,. \label{Prop1a}$$ The tensor projectors contained in Eq. (\[PropLV3\]) fulfill a closed algebra, as shown in Tabs. \[tab:closed-algebra-1\], \[tab:closed-algebra-2\]. By inserting Eq. (\[PropLV3\]) into Eq. (\[Prop1a\]), we obtain a system of ten equations for the ten coefficients to be determined, whose solution is given by
$$\begin{aligned}
a & =\frac{1}{\square(1+\theta^2\square)}\,,\quad
c=f=\frac{\eta^4({C^2}\square-\rho^2){\kappa}-\eta^2\rho\Pi}{\Gamma\square(1+\theta^2\square)}\,,\displaybreak[0]\\[2ex]
b & =-\frac{\xi}{\square}+\frac{\eta^2\left[-\eta^2\rho^2({B^2}\square-\kappa^2)-\eta^2\kappa^2({C^2}\square-\rho^2)
+2{\kappa\rho}\Pi\right]}{\Gamma\square(1+\theta^2\square)}\,,\displaybreak[0]\\[2ex]
d & =g=\frac{\eta^2\Pi}{(1+\theta^2\square)\Gamma}\,,\quad e=h=\eta^2\frac{[\eta^2\rho({B^2}\square-\kappa^2)
-{\kappa}\Pi]}{\Gamma\square(1+\theta^2\square)}\,,\displaybreak[0]\\[2ex]
i & =-\frac{\eta^4({C^2}\square-\rho^2)}{\Gamma(1+\theta^2\square)}\,,\quad
j=-\frac{\eta^4({B^2}\square-\kappa^2)}{\Gamma(1+\theta^2\square)}\,,\end{aligned}$$
where
$$\begin{aligned}
\Gamma & =\eta^4({B^2}\square
-\kappa^2)({C^2}\square-\rho^2)-\Pi^2,\displaybreak[0]\\[2ex]
\Pi & =1+\theta^2\square+\eta^2{(B\cdot C)}\square-\eta^2{\kappa\rho}\,.\end{aligned}$$
In momentum space, the propagator is
$$\begin{aligned}
\Delta_{\nu\alpha}&=-\frac{\mathrm{i}}{p^2(1-\theta^2p^2)\Gamma(p)}\Big\{\Gamma(p)\Theta_{\nu\alpha}+\left[b^{\prime}-\xi(1-\theta^2p^2)\Gamma(p)\right]\Omega_{\nu\alpha}\nonumber\\
&\phantom{{}={}-p^2(1-\theta^2p^2)\Gamma(p)\big\{}-\mathrm{i}F(p)(B_{\nu}p_{\alpha}+B_{\alpha}p_{\nu})-2\eta^{2}{D_{\nu\alpha}}p^{2}\Pi(p)-\mathrm{i}H(p)(C_{\nu}p_{\alpha}+C_{\alpha}p_{\nu})\nonumber\\[1ex]
&\phantom{{}={}-p^2(1-\theta^2p^2)\Gamma(p)\big\{}\left.
+\,\eta^4p^2{B_{\nu}B_{\alpha}}\left[(C\cdot p)^2-{C^2}p^2\right]
+\eta^4p^2{C_{\nu}C_{\alpha}}\left[(B\cdot p)^2-{B^2}p^2\right]\right\}\,, \label{PropLV5}$$
where $p_{\mu}$ is the four-momentum and
\[eq:definitions-bprime-f-h\]$$\begin{aligned}
b^{\prime}&=\eta^2\left\{\eta^2(C\cdot p)^2\left[(B\cdot p)^2-{B^2}p^2\right]\right.
\displaybreak[0]\label{blinha}\\
&\phantom{{}={}\eta^2\big\{}\left.+\,\eta^{2}(B\cdot p)^2\left[(C\cdot p)^2-{C^2}p^2\right]
-2{(B\cdot p)}(C\cdot p)\Pi(p)\right\}\,,\displaybreak[0]\nonumber\\[2ex]
F(p)&=\mathrm{i}\eta^{2}{(C\cdot p)}\Pi(p)-\mathrm{i}\eta^{4}(B\cdot p)\left[(C\cdot p)^2
-{C^2}p^2\right] \,,\label{Fp}\displaybreak[0]\\[2ex]
H(p)&=\mathrm{i}\eta^2(B\cdot p)\Pi(p)-\mathrm{i}\eta^4(C\cdot p)\left[(B\cdot p)^2-{B^2}p^2\right]\,, \label{Hp}$$ with
\[eq:definitions-gamma-pi\]$$\begin{aligned}
\Gamma(p)& =\eta^4{\left[(B\cdot p)^2-B^2p^2\right]}\left[(C\cdot p)^2-{C^2}p^2\right]
-\Pi^2(p)\,,\displaybreak[0]\label{lambp}\\[2ex]
\Pi(p)&=1-\theta^2p^2-\eta^2{(B\cdot C)}p^2+\eta^2{(B\cdot p)}(C\cdot p)\,.\end{aligned}$$ Note that $F$ and $H$ have dimensions of (mass)$^{-1}$, while $\Gamma$ and $\Pi$ are dimensionless. In the absence of the LV term, $\eta=0$, and the propagator (\[PropLV5\]) takes the form,
$$\Delta_{\nu\alpha}=-\mathrm{i}\left(\frac{\Theta_{\nu\alpha}}{p^2(1-\theta^2p^2)}-\frac{\xi}{p^2}\Omega_{\nu\alpha}\right)\,,
\label{PPodolsky}$$
recovering Podolsky’s propagator, as expected. Setting $\theta^{2}=0$ in the propagator (\[PropLV5\]), the result is $$\begin{aligned}
\Delta_{\nu\alpha}&=-\frac{\mathrm{i}}{p^2\Gamma(p)}\Big\{\Gamma(p)\Theta_{\nu\alpha}+(b^{\prime}-\xi\Gamma)\Omega_{\nu\alpha}\nonumber\\
&\phantom{{}={}-p^{2}\Gamma(p)}\,-\mathrm{i}F(p)(B_{\nu}p_{\alpha}+B_{\alpha}p_{\nu})-2\eta^2{D_{\nu\alpha}}p^{2}\Pi(p)
-\mathrm{i}H(p)(C_{\nu}p_{\alpha}+C_{\alpha}p_{\nu})\nonumber\\[1ex]
&\phantom{{}={}-p^2\Gamma(p)}\left.\,+\,\eta^4{B_{\nu}B_{\alpha}}\left[(C\cdot p)^2-{C^2}p^2\right]{p^2}
+\eta^4{C_{\nu}C_{\alpha}}\left[(B\cdot p)^2-{B^2}p^2\right]{p^2}\right\}\,, \label{PropLV6}$$ where $b^{\prime},F\left( p\right) ,H\left( p\right) ,\Gamma\left(
p\right) $ are given by the same expressions of Eqs. (\[eq:definitions-bprime-f-h\]), (\[eq:definitions-gamma-pi\]), with $$\Pi_{\theta=0}(p)=1-\eta^{2}{(B\cdot C)}p^{2} +\eta^{2}{(B\cdot p)}(C\cdot p)\,. \label{P3}$$ In this situation there are still two poles, $p^{2}=0,$ $\Gamma\left(
p\right) =0,$ associated with the Maxwell modes and those related to the LV higher-derivative term. Thus, in principle, the Lagrangian (\[PropLV6\]) has degrees of freedom linked to both modes, yielding a counting of modes analogue to that in Podolsky’s electrodynamics (see Ref. [@GalvaoP]).
Dispersion relations
--------------------
The dispersion equations of the modified electrodynamics defined by Eq. (\[LPVL1a\]) can be red off the poles of the propagator (\[PropLV5\]) in momentum space,
\[eq:n3\] $$\begin{aligned}
p^2(1-\theta^2p^2)&=0\,,\label{n3a}\\[2ex]
\Gamma(p)&=0\,. \label{n3b}$$
Since these poles also appear in terms that are not connected to the gauge fixing parameter $\xi$, they must be physical. As before, the dispersion equation $p^{2}=0$ represents the well-known Maxwell modes, while $1-\theta^{2}p^{2}=0$ stands for the Podolsky modes. Both are not modified by the higher-derivative term, whose effect is fully encoded in the dispersion equation (\[n3b\]). In the absence of the Podolsky term, $\theta^{2}=0,$ the modified Eq. (\[n3b\]) yields:$$\eta^{4}{\left[(B\cdot p)^2-B^2p^2\right]\left[(C\cdot p)^2-C^2p^2\right]}-\left[ 1-\eta^2{(B\cdot C)}p^2+\eta^2{(B\cdot p)}(C\cdot p)\right]^2=0\,. \label{n4}$$ With the help of *FORM*[^2] [@Vermaseren:2000nd], the dispersion equations above can be generalized to an arbitrary choice of $D_{\mu\nu}$. For a general choice of this tensor, the dispersion equation $p^2=0$ remains, but $1-\theta^2p^2=0$ and Eq. (\[n4\]) merge into a single equation. Although, in principle, there are many possibilities of forming observer Lorentz scalars from a general two-tensor and the momentum four-vector, the latter result collapses when taking into account that $D_{\mu\nu}$ is symmetric. The equation can then be conveniently written as follows: $$\begin{aligned}
\label{4n}
0&=-\,(1-\theta^2p^2)^3-2(1-\theta^2p^2)^2\eta^2(D_{\mu\nu}p^{\mu}p^{\nu}-p^2D^{\mu}_{\phantom{\mu}\mu}) \notag \\
&\phantom{{}={}}+2(1-\theta^2p^2)p^2\eta^4\left[p^2D_{\mu\nu}D^{\mu\nu}-2D_{\mu\nu}D^{\mu\varrho}p^{\nu}p_{\varrho}+D^{\mu}_{\phantom{\mu}\mu}\left(2D_{\nu\varrho}p^{\nu}p^{\varrho}-p^2D^{\nu}_{\phantom{\nu}\nu}\right)\right] \notag \\
&\phantom{{}={}}+\frac{4}{3}p^4\eta^6\left[3D_{\mu\nu}D^{\mu\nu}\left(D_{\varrho\sigma}p^{\varrho}p^{\sigma}-p^2D^{\varrho}_{\phantom{\varrho}\varrho}\right)+(D^{\mu}_{\phantom{\mu}\mu})^2\left(p^2D^{\nu}_{\phantom{\nu}\nu}-3D_{\nu\varrho}p^{\nu}p^{\varrho}\right)\right. \notag \\
&\phantom{{}={}+\frac{4}{3}p^4\eta^6\Big[}\left.+\,2p^2D^{\mu\nu}D_{\mu}^{\phantom{\mu}\varrho}D_{\nu\varrho}+6\left(D_{\mu\nu}D^{\mu\varrho}p^{\nu}p_{\varrho}D^{\sigma}_{\phantom{\sigma}\sigma}-D_{\mu\nu}D^{\mu\varrho}D^{\nu\sigma}p_{\varrho}p_{\sigma}\right)\right]\,.\end{aligned}$$ Note that it is not possible to factor out $1-\theta^2p^2$, as there is a contribution proportional to $\eta^6$ that does not contain the latter term. This contribution vanishes when $D_{\mu\nu}$ is decomposed into two four-vectors according to Eq. (\[Presc1\]), whereupon $(1-\theta^2p^2)\Gamma(p)=0$ is reproduced.
Analysis of some sectors of the theory {#sec:analysis-sectors}
--------------------------------------
In this section, we analyze some sectors of Eqs. (\[eq:n3\]) – (\[4n\]) to search for LV configurations that exhibit a consistent physical behavior. We are especially interested in causality. The behavior of the group and front velocity [@Brillouin:1960] allows for conclusions to be drawn on it where $$\mathbf{u}_{\mathrm{gr}}\equiv\frac{\partial p_{0}}{\partial\mathbf{p}}\,,\quad u_{\mathrm{fr}}\equiv\lim_{|\mathbf{p}|\mapsto\infty}\frac{p_{0}}{|\mathbf{p}|}\,.$$ We use a notion of classical causality that requires $u_{\mathrm{gr}}\equiv|\mathbf{u}_{\mathrm{gr}}|\leq1$ and $u_{\mathrm{fr}}\leq1$. Dispersion laws that do not describe standard photons for vanishing Lorentz-violation will be referred to as “exotic,” which includes Podolsky-type dispersion relations, but not necessarily unphysical ones. Dispersion relations that exhibit divergent group/front velocities or velocities greater than 1 will be called “spurious.” For all the investigated background configurations, there are three distinct poles, including $p^{2}=0$, which describes the usual photon.[^3] The first and the second of the isotropic cases examined below cannot be parameterized with two four-vectors as proposed in Eq. (\[Presc1\]). Therefore, they must be studied by using the more general Eq. (\[4n\]).
### Isotropic trace sector {#sec:isotropic-case-sector}
As an initial cross check of Eq. (\[4n\]), we study the case of a diagonal tensor $D_{\beta\alpha}$ with $D_{00}=-D_{11}= -D_{22}=-D_{33}$, $D_{ij}=0$ for $i\neq j$, i.e., $D_{\beta\alpha}= D_{00}g_{\beta\alpha}$, which gives $D^{\kappa}_{\phantom{\kappa}\kappa}=4D_{00}$. It is a pure trace configuration that is Lorentz-invariant, though. The dispersion equation for this configuration of $D_{\mu\nu}$ reads: $$\left[1-\Theta^2(p_0^2-\mathbf{p}^2)\right]^3=0\,,\quad\Theta^2=\theta^2+2\eta^2D_{00}\,,$$ from which we deduce $$\label{n5}p_{0}=\sqrt{\mathbf{p}^{2}+\frac{1}{\Theta^2}}\,.$$ This result is Podolsky’s dispersion relation, obviously with a redefined Podolsky parameter $\Theta$ that involves the standard Podolsky parameter $\theta$ and the nonvanishing controlling coefficient. If a Lorentz-violating background field $D_{\mu\nu}$ existed in nature, its trace would mimic the Podolsky term. Finding such a contribution experimentally, could even hint towards the existence of a Lorentz-violating background. The dispersion relation found is causal and compatible with a well-behaved propagation of signals as long as $\Theta^2>0$, i.e., for $D_{00}>-\theta^2/(2\eta^2)$.
### Complete isotropic sector
The traceless isotropic sector can be expressed in the form $D_{\mu\nu}=-D_{00}\times\mathrm{diag}(3,1,1,1)_{\mu\nu}$ where the global minus sign is chosen to be consistent with the definitions in Sec. \[sec:isotropic-case-sector\]. This choice of the background tensor is traceless and can be covariantly expressed in terms of the Minkowski metric tensor and the purely timelike preferred direction $\lambda^{\mu}=(1,0,0,0)^{\mu}$ as follows: $$D_{\mu\nu}=D_{00}\left(g_{\mu\nu}-4\lambda_{\mu}\lambda_{\nu}\right)\,.$$ Hence, the isotropic case under consideration contains the Lorentz-invariant part considered in Sec. \[sec:isotropic-case-sector\] and a Lorentz-violating contribution dependent on the preferred direction $\lambda_{\mu}$, which is similar to that of Sec. \[sec:timelike-isotropic-sector\]. This particular framework exhibits two distinct dispersion relations that can be obtained from Eq. (\[4n\]):
$$\begin{aligned}
p_0^{(1)}&=\sqrt{\mathbf{p}^2+\frac{1}{\Theta^2}}\,,\quad \Theta^2=\theta^2+2\eta^2D_{00}\,, \\[2ex]
p_0^{(2)}&=\sqrt{\Lambda\mathbf{p}^2+\frac{1}{\Theta^2}}\,,\quad \Lambda=1-\frac{8\eta^2D_{00}}{\Theta^2}\,.\end{aligned}$$
The first dispersion law is a perturbation of the usual Podolsky dispersion relation, which corresponds to Eq. (\[n5\]) found before. We already know that this mode has a well-defined energy and is causal for $D_{00}\geq -\theta^2/(2\eta^2)$. The second dispersion law is a perturbation, as well, but it is at bit more involved, as the momentum-dependent terms are also modified. The front velocity of the latter is given by $u_{\mathrm{fr}}^{(2)}=\sqrt{\Lambda}$. Thus, the front velocity is real for $D_{00}\leq \theta^2/(6\eta^2)$ and it is $\leq 1$ as long as $D_{00}>0$. The modulus of the group velocity can be cast into the form $u_{\mathrm{gr}}^{(2)}=\Lambda|\mathbf{p}|/p_0^{(2)}$. This function monotonically increases from 0 to a constant value given by $$\lim_{|\mathbf{p}|\mapsto\infty} u_{\mathrm{gr}}^{(2)}=u_{\mathrm{fr}}^{(2)}=\sqrt{\frac{\theta^2-6\eta^2D_{00}}{\theta^2+2\eta^2D_{00}}}\,.$$ The latter quantity is $\leq 1$ for $D_{00}>0$, i.e., the framework is causal and well-behaved for a controlling coefficient in the range $D_{00}\in [0,\theta^2/(6\eta^2)]$.
### Timelike isotropic sector {#sec:timelike-isotropic-sector}
The simplest isotropic LV configuration is based on the decomposition of Eq. (\[Presc1\]) with the purely timelike directions $B_{\mu}=(B_0,\mathbf{0})_{\mu}$, $C_{\mu}=(C_0,\mathbf{0})_{\mu}$. This sector corresponds to $D_{00}=C_0B_0$ and $D_{0i}=D_{i0}=D_{ij}=0$. For the latter, Eq. (\[eq:n3\]) can be solved for $p_0$ providing
$$\begin{aligned}
p_0^{(1)}&=\sqrt{\mathbf{p}^2+\frac{1}{\theta^2}}\,, \\[2ex]
p_0^{(2)}&=\sqrt{\Psi\mathbf{p}^2+\frac{1}{\theta^2}}\,,\quad \Psi=1+\frac{2\eta^2D_{00}}{\theta^2}\,,\end{aligned}$$
where $p_0^{(1)}$ is the original Podolsky dispersion relation and $p_0^{(2)}$ is a perturbation of the latter. The first is known to describe a proper propagation of signals. The front velocity of the second is $u_{\mathrm{fr}}^{(2)}=\sqrt{\Psi}$, which requires $D_{00}$ to be nonpositive such that $u_{\mathrm{fr}}^{(2)}\leq 1$. Furthermore, the modulus of the group velocity amounts to $u_{\mathrm{gr}}^{(2)}=\Psi|\mathbf{p}|/p_0^{(2)}$. It rises monotonically from 0 to a constant value that corresponds to the front velocity of this mode: $$\lim_{|\mathbf{p}|\mapsto\infty} u_{\mathrm{gr}}^{(2)}=u_{\mathrm{fr}}^{(2)}\,.$$ Therefore, $D_{00}\leq 0$ has to hold to grant causal signal propagation.
### Parity-even anisotropic sector (with $\theta=0)$ {#sec:parity-even-anisotropic}
For the configuration described by the two purely spacelike directions $B_{\mu}=(0,\mathbf{B})_{\mu}$ and $C_{\mu}=(0,\mathbf{C})_{\mu}$, the components of the two-tensor background field are $D_{00}=0$, $D_{0i}=D_{i0}=0$, and $D_{ij}=({B_i}C_j+{B_j}C_i)/2$. Furthermore, the preferred directions obey $B^2=-\mathbf{B}^2$, $C^2=-\mathbf{C}^2$, $B\cdot p=-\mathbf{B\cdot
p}$, and $C\cdot p=-\mathbf{C\cdot p}$. Inserting this information into Eq. (\[n4\]) delivers: $$\begin{aligned}
0&=\eta^{4}\left[{\mathbf{B}^2}\mathbf{C}^2p^4+{\mathbf{B}^2}(\mathbf{C\cdot p})^2p^2
+\mathbf{C}^2(\mathbf{B\cdot p})^2p^2+{(\mathbf{B\cdot p})^2}(\mathbf{C\cdot p})^2\right] \nonumber\\
&\phantom{{}={}}-\left[1+\eta^2{(\mathbf{B\cdot C})p^2}+\eta^2{(\mathbf{B\cdot p})}(\mathbf{C\cdot p})\right]^2\,.\end{aligned}$$ Employing the additional restriction of parallel or antiparallel vectors, $\mathbf{C}=\alpha\mathbf{B}$ with $\alpha\in\mathbb{R}\setminus \{0\}$ providing $(\mathbf{B}\cdot\mathbf{C})^2=\mathbf{B}^2\mathbf{C}^2$ and $\mathbf{B}(\mathbf{C}\cdot\mathbf{p})=\mathbf{C}(\mathbf{B}\cdot\mathbf{p})$, produces a nonbirefringent dispersion relation $$p_0^2=\mathbf{p}^2-(\mathbf{\hat{B}}\cdot\mathbf{p})^2-\frac{1}{2\eta^2\alpha\mathbf{B}^2}\,,\label{n6b}$$ with the unit vector $\mathbf{\hat{B}=B/}\vert\mathbf{B}\vert$. Here, $\alpha>0$ stands for parallel vectors, whereas $\alpha<0$ describes antiparallel ones. Introducing an angle via the relationship $\mathbf{B\cdot p}=\vert\mathbf{B}\vert\vert\mathbf{p}\vert\cos\vartheta$, yields: $$p_{0}=\sqrt{\mathbf{p}^2\sin^2\vartheta-\frac{1}{2\eta^{2}\alpha\mathbf{B}^2}}\,.
\label{n6b2}$$This dispersion relation is similar to that of Podolsky’s theory, but the momentum-dependent part is modified by the factor $\sin^2\vartheta$. Moreover, the “mass” term, $1/(2\eta
^{2}\alpha\mathbf{B}^2)$, appears with a negative sign for $\alpha>0$, which jeopardizes the energy definition. Relation (\[n6b\]) can be investigated with respect to causality.
First, the front velocity is given by $u_{\mathrm{fr}}=|\sin\vartheta|\leq1$ independently from $\alpha$ and does not show any vicious behavior. Second, the group velocity reads $${\mathbf{u}_{\mathrm{gr}}}=\frac{\mathbf{p}-\mathbf{\hat{\mathbf{B}}}(\mathbf{\hat{\mathbf{B}}}\cdot\mathbf{p})}{p_{0}}\,,$$ having a modulus given by $$\label{ug1}
u_{\mathrm{gr}}=\frac{1}{\sqrt{1-\mathrm{sgn}(\alpha)[2x^2\sin^2\vartheta]^{-1}}}\,,\quad x=\eta|\mathbf{p}|\sqrt{|\alpha|}|\mathbf{B}|\,.$$ For $\alpha>0$, the velocity (\[ug1\]) is only well-defined for $\mathbf{p}^{2}>{(2\eta^{2}\alpha\vert \mathbf{B}\vert ^{2}\sin
^{2}{\vartheta})^{-1}}.$ One problem is that this minimum cut-off depends on the angle $\vartheta$, which is why it does not really work as a true cut-off. Besides, the mode does not behave like a standard photon for vanishing $\eta$. It is easy to notice that the group velocity is always larger than 1 and even singular for $\eta\vert \mathbf{p}\vert\vert \mathbf{B}\vert=\csc{\vartheta}/\sqrt{2\alpha}$, i.e., the corresponding mode violates causality and is spurious. Its behavior is depicted in Fig. \[caseNzero-positive-alpha\] as a function of $x$. In contrast, the group velocity for $\alpha<0$ does not exhibit any singularities. This mode is exotic, as its group velocity increases from zero and approaches the limit $u_{\mathrm{gr}}=1$ from below. A plot of its modulus is presented in Fig. \[caseNzero-negative-alpha\].
Another possibility worthwhile to consider is that of orthogonal $\mathbf{B}$ and $\mathbf{C}$, $\mathbf{B\cdot C}=0$. This configuration is especially interesting, because it implies $D_{ii}=0$ and Tr($D_{\mu\nu})=0$, whereupon it represents a higher-derivative electrodynamics that does not contain Podolsky’s sector at all. The dispersion equation (\[n4\]) is$$\label{n7}
p^4(\eta^4{\mathbf{B}^2\mathbf{C}^2})+\eta^4p^2\left[{\mathbf{B}^2(\mathbf{C\cdot p})^2}+\mathbf{C}^2(\mathbf{B\cdot p})^2\right]
-2\eta^2{(\mathbf{B\cdot p})}(\mathbf{C\cdot p})-1=0\,.$$ To analyze this situation, we adopt a coordinate system in which the vectors $\mathbf{C}$, $\mathbf{B}$ point along the $x$ and $y$ axis, respectively: $\mathbf{C}={|\mathbf{C}|}\mathbf{\hat{e}}_{x}$, $\mathbf{B}={|\mathbf{B}|}\mathbf{\hat{e}}_{y}$. The momentum $\mathbf{p}$ shall enclose an angle $\vartheta$ with the $z$ axis. In this system, it holds that $$\label{basis}
\mathbf{p}=\left\vert \mathbf{p}\right\vert \left( \sin\vartheta\cos\phi
,\sin\vartheta\sin\phi,\cos\vartheta\right) ,$$ which implies ${\mathbf{B\cdot p}}=\left\vert \mathbf{B}\right\vert \left\vert
\mathbf{p}\right\vert \sin\vartheta\cos\phi$ and ${\mathbf{C\cdot p}}=\left\vert \mathbf{C}\right\vert \left\vert \mathbf{p}\right\vert
\sin\vartheta\sin\phi.$ The relation (\[n7\]) is rewritten as
$${(p_0^{(\pm)})^2}=\alpha_{\vartheta}\vert\mathbf{p}\vert^2\pm{\frac{1}{\eta^2{\vert\mathbf{B}\vert}\vert\mathbf{C}\vert}
\sqrt{\frac{1}{4}\nu_1^2\vert\mathbf{p}\vert^4+\nu_1\vert \mathbf{p}\vert^2\sin2\phi+1}}\,,$$
with$$\alpha_{\vartheta}={1-\frac{1}{2}\sin^2\vartheta}\,,\quad
\nu_1=\eta^2{\vert\mathbf{B}\vert}\vert\mathbf{C}\vert\sin^2\vartheta\,.$$
In what follows, modes denoted by the glyph $\oplus$ will be identified with the upper sign choice of a dispersion relation, whereas modes called $\ominus$ will correspond to the lower one. The front velocity reads $u_{\mathrm{fr}}=1$ independently of the angles $\vartheta$ and $\phi$, which is a result compatible with causality. The next step is to evaluate the group velocity, whose modulus can be expressed in terms of the basis $\mathbf{B}={|\mathbf{B}|}\mathbf{\hat{e}}_{y},\mathbf{C}={|\mathbf{C}|}\mathbf{\hat{e}}_{x}$ by using the representation of the momentum of Eq. [(\[basis\])]{}:
\[eq:group-velocity-parity-even-orthogonal\]$$\label{ug3}
{u_{\mathrm{gr}}^{(\pm)}}=\sqrt{\frac{4x^{2}\widetilde{\mathcal{S}}+2x^{4}\widetilde{\mathcal{T}}+x^{6}\widetilde{\mathcal{U}}}{Y^{2}\{x^{2}[3+\cos(2\vartheta)]\pm2Y\}}}\,,$$ with $$\begin{aligned}
\widetilde{\mathcal{S}} &
=3+\cos(2\vartheta)\pm Y\sin^{2}\vartheta\sin(2\phi)\,,\displaybreak[0]\\[2ex]
\widetilde{\mathcal{T}} & =2[3+\cos(2\vartheta)]\sin^{2}\vartheta\sin
(2\phi)\pm Y\sin^{4}\vartheta\,,\displaybreak[0]\\[2ex]
\widetilde{\mathcal{U}} & =[3+\cos(2\vartheta)]\sin^{4}\vartheta
\,,\displaybreak[0]\\[2ex]
Y & =\sqrt{4+x^{2}\sin^{2}\vartheta\lbrack x^{2}\sin^{2}\vartheta+4\sin
(2\phi)]}\,,\quad x=\eta|\mathbf{p}|\sqrt{|\mathbf{B}||\mathbf{C}|}\,.\end{aligned}$$ The graph of Fig. \[caseN\] displays the [modulus of the group velocity]{} of the two modes $\oplus$, $\ominus$ considered in Eq. (\[ug3\]) for different choices of the angles. Some comments are [worthwhile]{}. First, we point out that the group velocity of the mode $\ominus$ diverges for the following choice of the dimensionless variable $x$:
$$x_{\mathrm{sing}}=\frac{|\sec\vartheta|}{2}\sqrt{\sqrt{[3+\cos(2\vartheta
)]^{2}-4\sin^{4}\vartheta\cos^{2}(2\phi)}+2\sin^{2}\vartheta\sin(2\phi)}\,.$$
This value is indicated by the dashed, vertical lines in Fig. \[caseN\]. Because of the singularity, the behavior of the latter mode at momenta lying in this regime is unphysical. In contrast, the group velocity of the mode $\oplus$ (continuous line) does not have any singularities. For small momenta, the group velocities can be expanded as follows: $$\label{eq:group-velocity-parity-even-small-momentum}
{u_{\mathrm{gr}}^{(\pm)}}=\widetilde{\Theta}^{(\pm)}|\mathbf{p}|\,,\quad \widetilde{\Theta}^{(\pm)}=\eta\sqrt{|\mathbf{B}||\mathbf{C}|}\sqrt{\frac{1}{2}[\sin(2\phi)\mp 1]\sin^2\vartheta\pm 1}\,.$$ Each behaves like a Podolsky mode whose mass depends on the momentum direction. Note that such expansions can also be interpreted as expansions valid for a small controlling coefficient $\eta$. Furthermore, Fig. \[spacelike-case-orthogonal-vectors-1\] shows that the group velocity of the mode $\oplus$ has a maximum, approaches 1 from above, and becomes larger than 1 (breaking causality) for a certain range of parameters. For the same values of the angles, the group velocity of the mode $\ominus
$ decreases from its initial singularity to reach a minimum and finally approaches 1 from below. The graph \[spacelike-case-orthogonal-vectors-2\], generated for other values of the angles does not reveal any maxima or minima, with the group velocity of the mode $\ominus$ approaching 1 from above, and the mode $\oplus$ approaching 1 from below. For this parameter choice, the mode $\oplus$ is exotic. It was also verified that the mode $\ominus$ does not propagate for certain angles. For example, the group velocity vanishes identically for $\vartheta=\pi/2$ and $\phi=\pi/4$ where for $\vartheta=\pi/2$ and $\phi<[\pi+\arcsin(1/x^2)]/2$ it even takes complex values.
So, we conclude that the spacelike configuration, $B_{\mu}=(0,\mathbf{B})_{\mu}$ and $C_{\mu}=(0,\mathbf{C})_{\mu}$, with parallel or orthogonal $\mathbf{B}$ and $\mathbf{C}$ can yield both exotic and spurious dispersion relations whose group velocities diverge or break causality. In general, such relations suggest meaningful signal propagation only in a regime of large momenta that is compatible with [$u_{\mathrm{fr}}=1$]{}. However, this interpretation is not capable of recovering a certain sector, once the dispersion relations associated cannot be considered as physical for a given momentum range.
### Parity-odd anisotropic sector (with $\theta=0)$ {#sec:parity-odd-anisotropic}
A parity-odd anisotropic configuration is characterized by a purely timelike direction $C_{\mu}=(C_0,0)_{\mu}$ and a purely spacelike one $B_{\mu}=(0,\mathbf{B})_{\mu}$, which leads to
$$p_{0}=\frac{\mathrm{sgn}(C_0)\left\vert \mathbf{p}\right\vert \sqrt
{\eta^4C_0^2|\mathbf{B}\times\mathbf{p}|^4+\mathbf{B}^2}
-\mathbf{B\cdot p}}{\eta^{2}C_0|\mathbf{B}\times\mathbf{p}|^2}\,,$$
or $$\label{eq:dispersion-relation-parity-odd}
p_0=\frac{\mathrm{sgn}(C_0)\sqrt{1+\eta^4C_0^2\mathbf{B}^2\mathbf{p}^4\sin^4\vartheta_B}-\cos\vartheta_B}{\eta^2C_0\vert\mathbf{B}\vert\vert\mathbf{p}\vert\sin^2\vartheta_B}\,,$$
with ${\mathbf{B\cdot p}}=\vert\mathbf{B}\vert\vert\mathbf{p}\vert\cos\vartheta_B$, $|\mathbf{B}\times\mathbf{p}|=|\mathbf{B}||\mathbf{p}|\sin\vartheta_B$. The dispersion relation is positive, real-valued, and not defined in the limit $C_0\mapsto0$. The result for the front velocity is simply $u_{\mathrm{fr}}=1,$ not revealing any problems with causality in the large-momentum regime. The associated group velocity, however, is
$$\begin{aligned}
{u_{\mathrm{gr}}} & =\left[(\eta|\mathbf{p}|)^{-4}C_0^{-2}\Upsilon^{-3}
\left\{(\mathbf{B}\cdot\widehat{\mathbf{p}})\left[{\mathrm{sgn}(C_0)}2\Xi(\Upsilon+4\mathbf{B}^2)-4\mathbf{B}^2(\mathbf{B}\cdot\widehat{\mathbf{p}})\right] +2\mathbf{B}^2(\Upsilon-2\mathbf{B}^2)
-4\Xi^2\Upsilon\right\}\right. \nonumber\\
& \left.\phantom{{}={}}+5-4\Upsilon^{-1}\mathbf{B}^2\left[\mathrm{sgn}(C_{0})2\Xi^{-1}(\mathbf{B}\cdot\widehat{\mathbf{p}})-1\right]
-4(\eta|\mathbf{p}|)^4\mathbf{B}^2C_0^2\Upsilon\Xi^{-2}\right]^{1/2}\,,\end{aligned}$$
with $$\Upsilon=(\mathbf{B}\cdot\widehat{\mathbf{p}})^2-\mathbf{B}^2\,,\quad
\Xi=\sqrt{\mathbf{B}^2+(\eta|\mathbf{p}|)^{4}C_0^2\Upsilon^2}\,.$$
![Group velocity of Eq. for $\vartheta_B=3\pi/10$, $\mathrm{sgn}(C_0)=1$ (blue, plain) and $\mathrm{sgn}(C_0)=-1$ (red, dashed). The horizontal asymptote is illustrated as a black, dotted line.[]{data-label="caseN2"}](parity-odd-case-new.pdf)
Note that there is, in principle, a single mode only, i.e., the framework considered describes a nonbirefringent vacuum for photons. However, both the dispersion law and the group velocity depend on the sign of the component $C_{0}$ explicitly. In terms of the angle $\vartheta_{B}$ between $\mathbf{B}$ and $\mathbf{p}$, the result is expressed as
$$\begin{aligned}
\label{eq:group-velocity-parity-odd-case}
u_{\mathrm{gr}} & =\left[1-4\csc^2\vartheta_B+{\mathrm{sgn}(C_0)}
\frac{8}{Y}\cot\vartheta_B\csc\vartheta_B+\frac{4x^4}{Y^2}\sin^2\vartheta_B\right. \nonumber\\
&\phantom{{}={}}\left.\,+\,\frac{1}{8{x^4}}\left\{\csc^6\left(\frac{\vartheta_B}{2}\right)(1-{\mathrm{sgn}(C_0)}Y)
+\sec^6\left(\frac{\vartheta_B}{2}\right)(1+{\mathrm{sgn}(C_0)}Y)\right\}\right]^{1/2}\,,\end{aligned}$$
where $Y=\sqrt{1+{(x\sin\vartheta)^{4}}}\,,$ $x={\eta|\mathbf{p}|\sqrt{|\mathbf{B}||C_{0}|}}\,.$ The group velocity becomes singular for a vanishing momentum only, see Fig. \[caseN2\], which characterizes an unphysical regime for each sign of $C_{0}$, therefore. An expansion for a small momentum and controlling coefficient $\eta$, respectively, provides $$u_{\mathrm{gr}}=\frac{1}{2\sqrt{2}x^2}\sqrt{[1+\mathrm{sgn}(C_0)]\sec^6\left(\frac{\vartheta_B}{2}\right)+[1-\mathrm{sgn}(C_0)]\csc^6\left(\frac{\vartheta_B}{2}\right)}\,,$$ explicitly revealing the singularities for both possible signs of $C_0$. Note that the group velocity of one of the modes (for $C_0<0$) has a minimum and approaches 1 from below for increasing $x$, whereas the group velocity of the second mode (for $C_0>0$) has both a minimum and a maximum and approaches 1 from above. The modulus of the group velocity becomes larger than 1 for both modes at some values of $x,$ which is a behavior that corresponds to causality violation. We also point out that the two modes merge into a single mode for $\vartheta=\pi/2,$ and that they interchange their role for $\vartheta\in(\pi/2,\pi]$. Both modes do not behave like standard photons when $\eta$ vanishes. As their group velocities exhibit singularities, they are spurious.
The compilation of the results obtained seems to reveal that this higher-derivative electrodynamics only exhibits well-behaved signal propagation in the limit of large momenta. But, the dispersion relations cannot be valid only for a given momentum range, unless a physically reasonable cutoff can be imposed on the theory. So, this interpretation will not be considered at this stage. For small momenta, some of the modes behave like a Podolsky mode where the Podolsky parameter can depend on the direction of the momentum due to anisotropy. We encountered exotic modes such as the dispersion relation $\oplus$ for the parameter choice used in Fig. \[spacelike-case-orthogonal-vectors-2\]. The others that exhibit nonphysical features must be considered as spurious.
Unitarity analysis {#sec:unitarity-analysis}
------------------
At this point, we present some discussion on unitarity of this higher-derivative model. The analysis of unitarity at tree-level can be performed by means of a contraction of the propagator with external currents [@Veltman], which leads to a Lorentz scalar that is often referred to as the saturated propagator ($\mathit{SP}$):$$\mathit{SP}\equiv J^{\mu}\Delta_{\mu\nu}J^{\nu}\,.\label{Sat2}$$ The gauge current $J^{\mu}$ is taken as real and satisfies the conservation law $\partial_{\mu}J^{\mu}=0$, which in momentum space reads $p_{\mu
}J^{\mu}=0$. Therefore, a contraction with this current eliminates all contributions whose tensor structure contains at least a single four-momentum. The latter are associated with the gauge choice, i.e., these terms do not describe the physics of the system under consideration. In accordance with this method, unitarity is assured whenever the imaginary part of the residue of the saturation $SP$ evaluated at a vanishing denominator is positive.
We first assess the usual Podolsky theory, whose propagator is given by Eq. (\[PPodolsky\]). Contractions with the external four-current provide $$\mathit{SP}=-\mathrm{i}\left( \frac{J^{\nu}\Theta_{\nu\alpha}J^{\alpha}}{p^{2}\left(
1-\theta^{2}p^{2}\right) }-\frac{\xi}{p^{2}}J^{\nu}\Omega_{\nu\alpha}J^{\alpha}\right)\,,$$ which simplifies to $$\mathit{SP}=\frac{\mathrm{i}}{\theta^{2}}\left( \frac{J^{2}}{p^{2}\left( p^{2}-1/\theta
^{2}\right) }\right)\,,$$ due to current conservation. As properly noted in Ref. [@Turcati], the latter can be written as$$\mathit{SP}=\mathrm{i}\left(-\frac{J^{2}}{p^{2}}+\frac{J^{2}}{p^{2}-1/\theta
^{2}}\right)\,.$$ For a physical spacelike current, $J^{2}<0,$ the pole $p^{2}=0$ yields a positive imaginary part of the residue of $\mathit{SP}$, while the pole $p^{2}=1/\theta^{2}$ provides a negative one, so that this theory reveals a nonunitary behavior associated with the massive mode. The Lee-Wick theories are plagued by the same problem, but mechanisms to recover unitarity by suppressing the negative-norm states are stated in the literature [@LWghost]. Within this scenario, we point out that working with a Poldosky Lagrangian, whose term proportional to $\theta^{2}$ has a reversed sign, does not solve the unitarity problem at all (it only moves the problem from one pole to the other). Instead, it leads to a dispersion relation that is not well-defined for small enough momenta.
Considering the propagator of Eq. (\[PropLV6\]), defined in the absence of the Podolsky term ($\theta^{2}=0$), the saturation reads$$\begin{aligned}
\label{eq:saturated-propagator}
\mathit{SP} & =-\frac{\mathrm{i}}{p^{2}\Gamma(p) }\Big\{\Gamma(
p)J^{2}-\,2\eta^{2}p^{2}(B\cdot J)(C\cdot
J)\Pi_{\theta=0}(p)\nonumber\\
& \phantom{{}={}}\hspace{1.6cm}\left. +\,\eta^{4}p^{2}(B\cdot J) ^{2}\left[
(C\cdot p) ^{2}-C^{2}p^{2}\right] +\eta^{4}p^{2}(C\cdot J)^{2}\left[ ( B\cdot
p) ^{2}-B^{2}p^{2}\right] \right\} \,,\end{aligned}$$ with $\Pi_{\theta=0}(p)$ given by Eq. (\[P3\]). The latter should be analysed with respect to the residue of each denominator.
- First pole: The residue of the saturation evaluated at $p^2=0$ is simply given by $$\mathrm{Res}(\textit{SP})\vert_{p^2=0}=-\mathrm{i}J^{2}\,.$$ A spacelike $J^{2}<0$ yields $\mathrm{Im}[\mathrm{Res}(\textit{SP})\vert_{p^2=0}]>0$, which is why unitarity is assured. Therefore, we deduce that a quantization of this mode, which is the standard one of electrodynamics, corresponds to the photon.
- Second pole: Evaluating the residue at $\Gamma (p)=0$ works in the same manner, in principle: $$\begin{aligned}
\mathrm{Res}(\mathit{SP})|_{\Gamma (p)=0}& =-\mathrm{i}\eta^2\Big\{\eta
^{2}(B\cdot J)^{2}\left[ (C\cdot p)^{2}-C^{2}p^{2}\right] +\eta ^{2}(C\cdot
J)^{2}\left[ (B\cdot p)^{2}-B^{2}p^{2}\right] \notag \\
& \phantom{{}={}}\hspace{1.0cm}-2(B\cdot J)(C\cdot J)\Pi _{\theta =0}(p)\Big\}\Big|_{\Gamma (p)=0}\,.\end{aligned}$$The situation now is more involved, and requires a careful analysis to find the global sign of the expression in curly brackets. Considering Eq. ([lambp]{}), an evaluation of $\Gamma (p)=0$ provides the useful relationship$$\mathrm{sgn}(\Pi_{\theta=0})\Pi (p)|_{\substack{ \theta =0 \\ \Gamma (p)=0}}=\eta ^{2}\sqrt{\left[ (B\cdot p)^{2}-B^{2}p^{2}\right] \left[ (C\cdot
p)^{2}-C^{2}p^{2}\right] }\big|_{\Gamma (p)=0}\,.
\label{eq:definitions-gamma-pi}$$The sign function on the left-hand side must be taken into account, as $\Pi
_{\theta =0}(p)$ can be a negative quantity onshell, whereas the right-hand side is manifestly positive. Furthermore, we again have to distinguish between two cases. After investigating a large number of Lorentz-violating sectors numerically, we deduced that either $p^{\mu }\in M_{1}$ or $p^{\mu
}\in M_{2}$ with the two sets $M_{1}\equiv \{(B\cdot
p)^{2}-B^{2}p^{2}>0,(C\cdot p)^{2}-C^{2}p^{2}>0\}$ or $M_{2}\equiv \{(B\cdot
p)^{2}-B^{2}p^{2}<0,(C\cdot p)^{2}-C^{2}p^{2}<0\}$. Both options assure a real $\Pi(p).$ For the first option, the residue can be expressed in terms of the square of a sum of two terms. For the second option, a global minus sign must be pulled out of the whole expression, reversing the global saturation sign. The residues can then neatly be written as follows: $$\mathrm{Res}(\mathit{SP})|_{\Gamma (p)=0}=\mathrm{i}\eta^4\left\{
\begin{array}{lcl}
-\Big\{(B\cdot J)\sqrt{(C\cdot p)^{2}-C^{2}p^{2}} & \text{for} & p^{\mu }\in
M_{1}\,, \\
\phantom{{}}-\mathrm{sgn}(\Pi_{\theta=0})(C\cdot J)\sqrt{(B\cdot p)^{2}-B^{2}p^{2}}\Big\}^{2}\Big|_{\Gamma (p)=0} & & \\
\Big\{(B\cdot J)\sqrt{|(C\cdot p)^{2}-C^{2}p^{2}|} & \text{for} & p^{\mu
}\in M_{2}\,. \\
\phantom{{}}+\mathrm{sgn}(\Pi_{\theta=0})(C\cdot J)\sqrt{|(B\cdot p)^{2}-B^{2}p^{2}|}\Big\}^{2}\Big|_{\Gamma (p)=0} & & \\
\end{array}\right.$$
Hence, $\mathrm{Im}[\mathrm{Res}(\mathit{SP})|_{\Gamma (p)=0}]<0$ for the first case, which is a behavior indicating a breakdown of unitarity. The second case behaves in the opposite way, whereupon unitarity is preserved. The parity-odd configuration with $(C_{\mu})=(C_0,\mathbf{0})$, $(B_{\mu})=(0,\mathbf{B})$ analysed in Sec. \[sec:parity-even-anisotropic\] is covered by the latter case. We find $$(B\cdot p)^{2}-B^{2}p^{2}=-\frac{|\mathbf{B}|}{2\eta ^{2}|\mathbf{C}|}<0\,,\quad (C\cdot p)^{2}-C^{2}p^{2}=-\frac{|\mathbf{C}|}{2\eta ^{2}|\mathbf{B}|}<0\,,$$which is why $p^{\mu}\in M_{2}$ in this sector. Thus, unitarity is preserved for this configuration.
For the parity-odd case introduced in Sec. [sec:parity-odd-anisotropic]{} it holds that $(C\cdot
p)^{2}-C^{2}p^{2}=C_{0}^{2}\mathbf{p}^{2}>0$. However, it is a bit more involved to evaluate the second condition. The inequality $(B\cdot
p)^{2}-B^{2}p^{2}\geq 0$ to be checked can be cast into a more transparent form as follows:
$$2x^4\sin ^{4}\vartheta _{B}\cos ^{2}\vartheta _{B}+\left\{ 3+\cos
(2\vartheta _{B})-4\,\mathrm{sgn}(C_{0})\cos (\vartheta _{B})\sqrt{1+x^4\sin ^{4}\vartheta _{B}}\right\} \geq 0\,,$$
which is equivalent to $$2\left\{1-\mathrm{sgn}(C_{0})\cos \vartheta _{B}\sqrt{1+x^4\sin
^{4}\vartheta _{B}}\right\} ^{2}\geq 0\,.$$
The latter involves the angle $\vartheta _{B}$ defined directly below Eq. (\[eq:dispersion-relation-parity-odd\]) and the dimensionless quantity $x$ defined under Eq. (\[eq:group-velocity-parity-odd-case\]). For $C_{0}>0$, the equality sign is valid for $\vartheta _{B}=0$ only, whereas for $C_{0}<0$ it is valid for $\vartheta _{B}=\pi $. Apart from these very special values at the boundaries of the interval for $\vartheta _{B}$, the condition $(B\cdot
p)^{2}-B^{2}p^{2}>0$ is fulfilled manifestly, whereby $p^{\mu }\in M_{1}$. Therefore, according to our criterion, there are issues with unitarity for this sector.
There is an alternative possibility of calculating the saturated propagator. Using the transverse (*T*) and longitudinal (*L*) projection operators $\Theta_{\mu\nu}$ and $\Omega_{\mu\nu}$ of Eq. (\[eq:projectors\]) transformed to momentum space, the propagator can be decomposed into the four parts $\Delta^{TT}_{\mu\nu}\equiv \Theta_{\mu\alpha}\Delta^{\alpha\beta}\Theta_{\beta\nu}$, $\Delta^{TL}_{\mu\nu}\equiv \Theta_{\mu\alpha}\Delta^{\alpha\beta}\Omega_{\beta\nu}$, $\Delta^{LT}_{\mu\nu}\equiv \Omega_{\mu\alpha}\Delta^{\alpha\beta}\Theta_{\beta\nu}$, and $\Delta^{LL}_{\mu\nu}\equiv \Omega_{\mu\alpha}\Delta^{\alpha\beta}\Omega_{\beta\nu}$. After performing contractions with the conserved current, only $\Delta^{TT}_{\mu\nu}$ will survive. Following this procedure, the saturated propagator can be cast into the form
$$\begin{aligned}
\textit{SP}&=J^{\mu}\Delta^{TT}_{\mu\nu}J^{\nu}\,,\quad \Delta^{TT}_{\mu\nu}=-\mathrm{i}\frac{(M^{-1})_{\mu\nu}}{p^2}\,, \\[2ex]
M^{\mu\nu}&=\left[(1-\theta^2p^2)g-2\eta^2D^{TT}p^2\right]^{\mu\nu}\,,\quad (D^{TT})_{\mu\nu}\equiv \Theta_{\mu\alpha}D^{\alpha\beta}\Theta_{\beta\nu}\,.\end{aligned}$$
Hence, the behavior of unitarity is completely controlled by the totally transverse part of the matrix $D_{\mu\nu}$. Performing a formal partial-fraction decomposition, the standard pole $p^2=0$ can be separated from the remaining expressions involving the Podolsky parameter and the Lorentz-violating coefficient: $$\label{eq:propagator-transverse-transverse}
\Delta_{\mu\nu}^{TT}=-\mathrm{i}\left\{\frac{g_{\mu\nu}}{p^2}+(\theta^2\delta+2\eta^2D^{TT})_{\mu}^{\phantom{\mu}\alpha}(M^{-1})_{\alpha\nu}\right\}\,.$$ Using the explicit decomposition of $D_{\mu\nu}$ in terms of two four-vectors proposed in Eq. (\[Presc1\]) and setting $\theta=0$, we can deduce that $\det(M)=\Gamma(p)$ with $\Gamma(p)$ from Eq. (\[eq:definitions-gamma-pi\]) for $\theta=0$. The result of Eq. (\[eq:propagator-transverse-transverse\]) is suitable to investigate unitarity for the case of a nonzero $\theta$ and a general $D_{\mu\nu}$ whose form does not rely on a decomposition into two four-vectors. The analysis of such cases is an interesting open problem.
To conclude this section, unitarity violation seems to be connected to the appearance of additional time derivatives in the Lorentz-violating contributions of the Lagrange density, cf. Eqs. (\[eq:lagrange-density-bilinear\]), (\[OLV1\]). A behavior of this kind is expected. For a Lorentz-invariant, higher-derivative theory it was observed in [@LeeWick], amongst other works. The authors of [@Colladay:2001wk] pointed out that additional time derivatives in the minimal SME fermion sector lead to issues with the time evolution of asymptotic states. Related problems in the nonminimal SME were found in the third and second paper of [@Kostelec1] and [@Kostelec2], respectively.
Generalized model involving anisotropic Podolsky and Lee-Wick terms
===================================================================
There are two other *CPT*-even dimension-6 terms endowed with two additional derivatives, besides the LV modification (\[TADLV1\]), that can be expressed in terms of the electromagnetic field strength tensor and the tensor $D^{\mu\nu},$ namely:$${D^{\alpha\beta}}F_{\mu\nu}\partial_{\alpha}\partial_{\beta}F^{\mu}{}^{\nu}\,,\quad
{D^{\mu\nu}}\partial_{\sigma}F^{\sigma\lambda}{}\partial_{\mu}F_{\nu\lambda}\,.
\label{LVSLAD2B}$$ The first one is a kind of anisotropic Lee-Wick term, while the second yields a bilinear contribution similar to it, but with the opposite sign, so that only one of these terms will be considered. The very same correspondence observed for the case of the anisotropic Podolsky term holds for the anisotropic Lee-Wick term, too: for the configuration of a diagonal tensor $D_{\beta\alpha}$ of the form $D_{\beta\alpha}=D_{00}g_{\beta\alpha}$, it becomes proportional to the usual Lee-Wick term, that is, $${D^{\alpha\beta}}F_{\mu\nu}\partial_{\alpha}\partial_{\beta}F^{\mu\nu}=D_{00}(F_{\mu\nu}\square F^{\mu\nu})\,.$$ The behavior of the coefficients of the tensor $D_{\beta\alpha}$ under discrete *C*, *P*, and *T* operations, described in Tab. \[tab:transformation-properties\], does not depend on the way how the tensor is coupled to the electromagnetic field, being equally valid for the anisotropic Lee-Wick structures of Eq. (\[LVSLAD2B\]).
In principle, the most general LV dimension-6 electrodynamics, modified by a rank-2 symmetric tensor, $D_{\beta\alpha},$ also includes the second of the anisotropic Lee-Wick contributions of Eq. (\[LVSLAD2B\]), and is represented by the following Lagrangian:$$\mathcal{L}=-\frac{1}{4}F_{\mu\nu}F^{\mu}{}^{\nu}+\frac{\theta^{2}}{2}\partial_{\alpha}F^{\alpha}{}^{\beta}\partial_{\lambda}F_{\phantom{\lambda}\beta}^{\lambda}
+\eta_{1}^{2}D_{\beta\alpha}\partial_{\sigma}F^{\sigma}{}^{\beta}\partial_{\lambda}F^{\lambda\alpha}
+\eta_2^2D^{\beta\alpha}\partial_{\sigma}F^{\sigma}{}^{\lambda}\partial_{\beta}F_{\alpha\lambda}
+\frac{1}{2\xi}(\partial_{\mu}A^{\mu})^2\,,\label{L3}$$ with $\theta\geq 0$, $\eta_1\geq 0$, and $\eta_2\geq 0$. The latter incorporates the standard Podolsky term, its LV modification considered earlier, and the LV anisotropic Lee-Wick contribution. Such a Lagrangian can be written in the form $\mathcal{L}=(1/2)A^{\mu}\bar{O}_{\mu\nu}A^{\nu}$ where$$\begin{aligned}
\bar{O}_{\mu\nu} & =(1+\theta^2\square+2\eta_2^2{D^{\beta\alpha}}\partial_{\beta}\partial_{\alpha})\square
\Theta_{\mu\nu}+\left(2\eta_1^2D^{\beta\alpha}\partial_{\beta}\partial_{\alpha}-\frac{1}{\xi}\right)\square\Omega_{\mu\nu}+2\eta_1^2
{D_{\nu\mu}}\square^2\nonumber\\
&\phantom{{}={}}-2\eta_1^2{D_{\nu\alpha}}\square\partial_{\mu}\partial^{\alpha}
-2\eta_1^2{D_{\sigma\mu}}\square\partial_{\nu}\partial^{\sigma}\,.\end{aligned}$$ Based on the prescription (\[Presc1\]) for the symmetric tensor $D_{\beta\alpha}$ we obtain $$\begin{aligned}
\bar{O}_{\mu\nu}&=(1+\theta^2\square+2\eta_2^2{\kappa\rho})\square\Theta_{\mu\nu}
+\left(2\eta_1^2{\kappa\rho}-\frac{1}{\xi}\right)\square\Omega_{\mu\nu}
+\eta_1^2{(B_{\mu}C_{\nu}+B_{\nu}C_{\mu})}\square^2\nonumber\\
&\phantom{{}={}}-\eta_1^2{(C_{\nu}\partial_{\mu}+C_{\mu}\partial_{\nu})\kappa}\square
-\eta_1^2{(\partial_{\nu}B_{\mu}+\partial_{\mu}B_{\nu})\rho}\square \,,\label{prop1}$$ with $\kappa$ and $\rho$ of Eq. (\[karo\]). Using the same tensor algebra as that of the first case, the following propagator can be derived:
$$\begin{aligned}
\Delta_{\nu\alpha}& =-\frac{\mathrm{i}}{p^2\Delta(p)}\left\{\tilde{\Gamma}(p)\Theta_{\nu\alpha}
+\left[b^{\prime}-\xi\Delta(p)\right]\Omega_{\nu\alpha}\right. \nonumber\\
&\phantom{{}={}-\mathrm{i}p^2\Delta(p)}\left.-\,\mathrm{i}\tilde{F}(p)(B_{\nu}p_{\alpha}+B_{\alpha}p_{\nu})
-2\eta_1^2D_{\nu\alpha}p^2\tilde{\Pi}(p)-\mathrm{i}\tilde{H}(p)(C_{\nu}p_{\alpha}+C_{\alpha}p_{\nu})\right. \nonumber\\[1ex]
&\phantom{{}={}-\mathrm{i}p^2\Delta(p)}
\left.+\,\eta_1^4B_{\nu}B_{\alpha}\left[(C\cdot p)^2-C^2p^2\right]p^2+\eta_1^4C_{\nu}C_{\alpha}\left[(B\cdot p)^2-B^2p^2\right]p^2\right\}\,,\end{aligned}$$
where $$\begin{aligned}
\Delta(p) &= [1-\theta^2p^2-2\eta_2^2(B\cdot p)(C\cdot p)]\tilde{\Gamma}(p)\,, \displaybreak[0]\\[2ex]
\tilde{\Gamma}(p)& =\eta_1^4\left[(B\cdot p)^2-B^2p^2\right]\left[(C\cdot p)^2-C^2p^2\right]-\tilde{\Pi}^2(p)\,,\displaybreak[0]\\[2ex]
\tilde{\Pi}(p) & =1-\theta^2p^2-\eta_1^2p^2(B\cdot C)+[\eta_1^2-2\eta_2^2](B\cdot p)(C\cdot p)\,, \displaybreak[0]\\[2ex]
\tilde{F}(p)& = F(p)|_{\eta=\eta_1}\,,\quad \tilde{H}(p)=H(p)|_{\eta=\eta_1}\,,\end{aligned}$$
with $F(p)$ and $H(p)$ of Eqs. (\[Fp\]) and (\[Hp\]), respectively. The new dispersion equations read
$$\begin{aligned}
0&=p^2[1-\theta^2p^2-2\eta_2^2(B\cdot p)(C\cdot p)]\,,\label{PDR6A}\\[2ex]
0&=\tilde{\Gamma}(p)\,.\label{PDR6B}$$
We now observe that in contrast to Eq. (\[n3a\]), Podolsky’s dispersion relation is also modified, as shown in Eq. (\[PDR6A\]).
[Dispersion relations]{}
------------------------
To analyse the dispersion equations (\[PDR6A\]) and (\[PDR6B\]), we consider the main two background configurations – timelike and spacelike – discussed in the first model. The Podolsky parameter $\theta$ will not be discarded.
### [Timelike isotropic sector (with $\theta\neq0)$]{}
The timelike isotropic configuration is characterized by $B_{\mu}=(B_0,\mathbf{0})_{\mu}$, $C_{\mu}=(C_0,\mathbf{0})_{\mu}$. As already mentioned, the dispersion equation (\[PDR6A\]) shows that the usual Podolsky dispersion relation is now modified by a contribution resulting from the anisotropic Lee-Wick term. In this case, the dispersion relation obtained from Eq. (\[PDR6A\]) takes the simple form, $$p_0=\sqrt{\frac{1}{\varepsilon}\left(\mathbf{p}^2+\frac{1}{\theta^2}\right)}\,,
\label{DR51}$$ with $\varepsilon=1+2\eta_2^2B_0C_0/\theta^2$. The latter is a Podolsky-like dispersion relation, modified by a kind of dielectric constant, $\varepsilon$. For $B_0C_0>0$, the front velocity is less than one, $u_{\mathrm{fr}}<1,$ and the group velocity, $$u_{\mathrm{gr}}=\frac{\vert\mathbf{p}\vert}{\sqrt{\varepsilon}\sqrt{\mathbf{p}^2+1/\theta^2}}\,,$$ is always less than 1, as well, ensuring the validity of causality for this configuration. The DR (\[DR51\]) only makes sense in the presence of the Podolsky term ($\theta^2\neq0)$. Note that $\theta^2$ appears inside $\varepsilon$, as well. Now, we analyse the dispersion equation (\[PDR6B\]) for [the]{} isotropic configuration with $B_{\mu}=(B_0,\mathbf{0})_{\mu}$, $C_{\mu}=(C_0,\mathbf{0})_{\mu}$, which can be written as
$$E^{(\pm)}=\psi^{(\pm)}|\mathbf{p}|\,,\quad\psi^{(\pm)}=\psi^{(\pm)}(|\mathbf{p}|)\,,\label{eq:group-velocity-purely-timelike-case}$$
where $$\begin{aligned}
\psi^{(+)} & =\sqrt{\frac{1+\theta^2|\mathbf{p}|^2}{2y^2+\theta^2|\mathbf{p}|^2}}\,,\\[2ex]
\psi^{(-)} & =\sqrt{\frac{1+2x^2+\theta^2|\mathbf{p}|^2}{2y^2+\theta^2|\mathbf{p}|^2}}\,,\end{aligned}$$ and $x=\sqrt{B_0C_0}\eta_1|\mathbf{p}|$, $y=\sqrt{B_{0}C_{0}}\eta_{2}|\mathbf{p}|$ are dimensionless coefficients. Note that both $x$ and $y$ are linear functions of the three-momentum magnitude where the ratios $x/(\theta|\mathbf{p}|)=\sqrt{B_0C_0}\eta_1/\theta\equiv\alpha$ and $y/(\theta|\mathbf{p}|)=\sqrt{|\mathbf{B}||\mathbf{C}|}\eta_2/\theta
\equiv\beta$ are dimensionless. Hence, it is reasonable to investigate the expressions for the characteristic velocities after replacing $x$ by $\alpha\theta|\mathbf{p}|$ and $y$ by $\beta\theta|\mathbf{p}|$, respectively. Now, the front velocities for both modes are simply given by
$$\begin{aligned}
u_{\mathrm{fr}}^{(+)} & =\frac{1}{\sqrt{1+2\beta^2}}\,,\displaybreak[0]\\[2ex]
u_{\mathrm{fr}}^{(-)} & =\sqrt{\frac{1+2\alpha^2}{1+2\beta^2}}\,.\end{aligned}$$
The first of these expressions is always smaller than 1, whereas the second can be larger than 1 for $\alpha>\beta$. As the case under consideration is isotropic, it is not too involved to obtain the group velocities:
\[eq:group-velocities-timelike-isotropic-theta\]$$\begin{aligned}
u_{\mathrm{gr}}^{(+)} & =\frac{\psi^{(+)}}{1+(\theta|\mathbf{p}|)^{-2}}\,,\\[2ex]
u_{\mathrm{gr}}^{(-)} & =\psi^{(-)}\left\{1-\frac{1}{(\psi^{(-)})^2\left[2y^2+(\theta|\mathbf{p}|)^2\right]}\right\}\,.\end{aligned}$$
Both expressions do not exhibit any singularities. The graph of Fig. \[caseN3\] reveals decreasing group velocities for declining momenta where each mode behaves as a Podolsky mode in this regime:
$$\begin{aligned}
u_{\mathrm{gr}}^{(\pm)}&=\widehat{\Theta}^{(\pm)}|\mathbf{p}|\,, \\[2ex]
\widehat{\Theta}^{(+)}&=\frac{\theta}{\sqrt{1+2\beta^2}}\,,\quad\widehat{\Theta}^{(-)}=\frac{(1+2\alpha^2)\theta}{\sqrt{1+2\beta^2}}\,.\end{aligned}$$
As the case under consideration is isotropic, the Podolsky parameters do not depend on the momentum direction, but they involve the Lorentz-violating coefficients. Causality violation can occur for the mode $\ominus$ dependent on the chosen parameters. The group velocity for each mode rises from 0 to a finite value that is given by the previously obtained expressions for the front velocity: $$\lim_{z\mapsto\infty}u_{\mathrm{gr}}^{(+)}=u_{\mathrm{fr}}^{(+)}\,,\quad
\lim_{z\mapsto\infty}u_{\mathrm{gr}}^{(-)}=u_{\mathrm{fr}}^{(-)}\,.$$ Hence, the group velocity of the mode $\ominus$ can reach values larger than 1, breaking causality, which characterizes a spurious mode. On the other hand, the mode $\oplus$ mode does not violate causality, although it does not behave like a standard photon for vanishing $\eta_1$ or $\eta_2$. So it is an exotic mode. Note that vanishing Lorentz violation translates into the limits $\alpha\mapsto 0$ and $\beta\mapsto 0$, which provides the Podolsky-type dispersion law.
### [Parity-even anisotropic sector (with $\theta\neq0)$]{}
For the spatially anisotropic configuration, $B_{\mu}=(0,\mathbf{B})_{\mu}$, $C_{\mu}=(0,\mathbf{C})_{\mu}$, we can express the dispersion relation following from Eq. (\[PDR6A\]) as follows:$$p_0^2=\frac{1}{\theta^2}+\mathbf{p}^2-{\frac{2\eta_2^2}{\theta^2}}
(\mathbf{B\cdot p})(\mathbf{C\cdot p})\,,
\label{DR52A}$$ which for $\eta_2\mapsto 0$ recovers Podolsky’s dispersion relation. To analyze the dispersion relation for $\eta _{2}\neq 0$, we first consider the situation in which the vectors $\mathbf{C}$, $\mathbf{B}$ are orthogonal. Using the coordinate system employed in Sec. \[sec:parity-even-anisotropic\] with the momentum of Eq. (\[basis\]), implies $$p_0=\sqrt{\frac{1}{\theta^2}+\mathbf{p}^2\left(1-\frac{\eta_2^2}{\theta^2}
\vert\mathbf{B}\vert\vert\mathbf{C}\vert\sin^2\vartheta\sin2\phi\right)}\,.
\label{DR52B}$$
![Group velocity of Eq. for $\beta=1$, $\vartheta=\phi=0$ (blue, plain), for $\beta=3/2$, $\vartheta=\pi/5$, $\phi=\pi/3$ (green, dash-dotted), and for $\beta=2$, $\vartheta=\pi/2$, $\phi=\pi/3$ (red, dashed). The asymptotes are indicated by black, dotted lines.[]{data-label="caseparityodd"}](full-theory-parity-odd-case-new.pdf)
Here, the front velocity is$$u_{\mathrm{fr}}=\sqrt{1-{\frac{\eta_2^2}{\theta^2}\vert\mathbf{B}\vert\vert\mathbf{C}\vert}
\sin^2\vartheta\sin(2\phi)}\,,$$ which can be larger than 1 for some values of $\phi.$ The modulus of the group velocity is given by $${u_{\mathrm{gr}}=\sqrt{2+\frac{y^4\sin^2\vartheta-z^2(2+z^2)}{z^2\left[1+z^2-y^2\sin^2\vartheta\sin(2\phi)\right]}}}\,,
\label{eq:group-velocity-full-parity-pdd}$$ with $y=\sqrt{|\mathbf{B}||\mathbf{C}|}\eta_{2}|\mathbf{p}|$, $z=\theta|\mathbf{p}|$. Here we introduce the constant ratio $y/z=\sqrt
{|\mathbf{B}||\mathbf{C}|}\eta_{2}/\theta\equiv\beta$. The graph of Fig. \[caseparityodd\] shows the behavior of the group velocity for distinct values of the parameter $\beta$ and the angles. The latter $\beta$ can be chosen such that $u_{\mathrm{gr}}$ rises steadily from 0 and approaches a certain constant from below for increasing $z$. Explicitly, this constant exceeds 1 at $$z|_{u_{\mathrm{gr}}=1}=\frac{\csc\vartheta}{|\beta|\sqrt{\beta^{2}-\sin(2\phi)}}\,,$$ whereupon there will be issues with causality. This value of $z$ is real for $\beta>\sqrt{\sin(2\phi)}$ only, which demonstrates that causality problems do not arise necessarily. Besides, $u_{\mathrm{gr}}$ also exhibits a singularity for a suitable choice of the parameters that lies at the value $$z_{\mathrm{sing}}=\frac{1}{\beta^{2}\sin^{2}\vartheta\sin(2\phi)-1}\,,$$ revealing an unphysical regime. For small momenta, the group velocity of Eq. ([eq:group-velocity-full-parity-pdd]{}) behaves as $$u_{\mathrm{gr}}=\breve{\Theta}|\mathbf{p}|\,,\quad\breve{\Theta}=\theta\sqrt{1+\beta^2[\beta^2-2\sin(2\phi)]\sin^2\vartheta}\,,$$so the corresponding mode propagates as a Podolsky-type mode with a modified mass depending on the momentum direction. Note the similarities to Eq. . The limit $\beta\mapsto 0$ for vanishing Lorentz violation reproduces the behavior of the conventional Podolsky mode.
For this spacelike configuration, we now investigate the dispersion relation derived from Eq. (\[PDR6B\]) by initially studying the case of orthogonal vectors $\mathbf{B}$, $\mathbf{C}$. This configuration yields two different additional dispersion relations, written as
$$(p_0^{(\pm)})^2=\frac{\mathbf{p}^2}{z^4-x^4}\left\{\mathcal{V}/2
\pm\sqrt{(z^4-x^4){\left[x^4\cos^2\vartheta-\mathcal{W}\right]}
+(\mathcal{V}/2)^2}\right\}\,,
\label{DR44}$$
where $$\begin{aligned}
\mathcal{V} & =2z^2(1+z^2)-2x^4+\sin^2\vartheta\left[x^2+(x^2-2y^2)z^2\sin(2\phi)\right]\,, \\[2ex]
\mathcal{W} & =\left[1+z^2-y^2\sin^{2}\vartheta\sin(2\phi)\right]\left[1+z^2+(x^2-y^2)\sin^2\vartheta\sin(2\phi)\right]\,,\end{aligned}$$ and$$x=\sqrt{|\mathbf{B}||\mathbf{C}|}\eta_{1}|\mathbf{p}|\,,\quad
y=\sqrt{|\mathbf{B}||\mathbf{C}|}\eta_{2}|\mathbf{p}|\,,\quad z=\theta|\mathbf{p}|\,.
\label{DEFIN}$$ The modulus squared of the group velocity is given by a highly involved expression
\[eq:group-velocity-full-theory-complicated-dispersion-equation-spacelike-orthogonal\]$$(u_{\mathrm{gr}}^{(\pm)})^2=\frac{|\mathbf{p}|^2}{(p_0^{(\pm)})^2}\left(
\cos^2\vartheta+\frac{{f_{\pm}^2}(\vartheta,\phi)+{f_{\pm}^2}(\vartheta,\pi/2-\phi)}{{g_{\pm}^2}(\vartheta,\phi)}\right)\,,$$ with $$\begin{aligned}
f_{\pm}(\vartheta,\phi) & =2y^2(y^2-x^2)\sin^3\vartheta\sin(2\phi)\sin\phi\nonumber\\
&\phantom{{}={}}+(x^2-2y^2)\left\{1+z^2\left[2\sin^2\vartheta\cos^2\phi+1-(e^{(\pm)})^{2}\right]\right\}\sin\vartheta\sin\phi\nonumber\\
&\phantom{{}={}}+\left(x^4\left[(e^{(\pm)})^2-\cos^2\vartheta\right]
+2z^2\left\{1+\left[1-(e^{(\pm)})^{2}\right]z^2\right\}\right)\sin\vartheta\cos\phi\,,\displaybreak[0]\\[2ex]
g_{\pm}(\vartheta,\phi) & =\left[2(e^{(\pm)})^2-1-\cos^2\vartheta\right]x^4+\left[2+(x^2-2y^2)\sin^2\vartheta\sin(2\phi)\right]z^2\nonumber\\
&\phantom{{}={}}+2\left[1-(e^{(\pm)})^{2}\right]z^4\,,\end{aligned}$$
where $e^{(\pm)}\equiv p_0^{(\pm)}/|\mathbf{p}|$. Fig. \[caseOrtog2\] shows the group velocity for the modes $\oplus$ and $\ominus$ for different angles and parameters $\sqrt{|\mathbf{B}||\mathbf{C}|}\eta_1/\theta\equiv\alpha$ and $\sqrt{|\mathbf{B}||\mathbf{C}|}\eta_2/\theta\equiv\beta$. The group velocity of the mode $\oplus$ is badly behaved for a broad range of parameters and angles, as it exhibits one or even two singularities that lie at the following values of $z$:
$$z_{\mathrm{sing}}^{(\pm)}=\frac{\sqrt{2}}{\sqrt{(2\beta^2-\alpha^2)
\sin^2\vartheta\sin(2\phi)-2\pm\alpha^2\sqrt{4\cos^2\vartheta+\sin^4
\vartheta\sin^2(2\phi)}}}\,.$$
Besides, when there are singularities, the group velocity even becomes complex for large domains of the momentum. For small momenta, the group velocity for both modes exhibits the following asymptotic behavior:
$$\begin{aligned}
u_{\mathrm{gr}}^{(\pm)}&=\bar{\Theta}^{(\pm)}|\mathbf{p}|\,, \\[2ex]
\bar{\Theta}^{(\pm)}&=\theta\sqrt{1\mp\alpha^2+\frac{\sin^2\vartheta\left[\alpha^2(\pm\alpha^2-2)\pm 2\beta^2(\alpha^2-\beta^2)+(\alpha^2\mp 2)(\alpha^2-2\beta^2)\sin(2\phi)\right]}{2(\alpha^2\mp1)}}\,.\end{aligned}$$
Hence, these modes are modifications of the Podolsky mode with anisotropic Podolsky parameters that can become complex when the parameters are chosen suitably. The Podolsky mode is reproduced in the combined limit $\alpha\mapsto 0$ and $\beta\mapsto 0$. For certain choices, the group velocity is well-behaved in the sense that it increases monotonically from $0$ until it approaches a value smaller than 1 from below, cf. the green curve in Fig. \[fig:full-spacelike-orthogonal-case-1\]. It then behaves like an exotic mode. The other curves in this figure correspond to parameter choices that provide spurious behaviors. In contrast, the mode $\ominus$ in Fig. \[fig:full-spacelike-orthogonal-case-2\] does not exhibit any singularities. It rises from 0 and converges to a value smaller then 1 from below or larger than 1 from above. The latter behavior, though, is interpreted as a violation of causality, which occurs at least for some parameter choices. Therefore, the mode represented by the blue curve is exotic, whereas the mode illustrated by the red curve is spurious.
We can also present expansions for the front velocities following from DR (\[DR44\]) in the limit of a large Podolsky parameter $\theta$ in comparison to the LV coefficients $\eta_{1}$, $\eta_{2}$:
$$\begin{aligned}
u_{\mathrm{fr}}^{(+)} & \approx1+\frac{1}{2}\left[\left(\frac{x}{z}\right)^2
-\left(\frac{y}{z}\right)^2\right]\sin^2\vartheta\sin(2\phi)\,,\\[2ex]
u_{\mathrm{fr}}^{(-)} & \approx1-\frac{1}{2}\left(\frac{y}{z}\right)^2\sin^2\vartheta\sin(2\phi)\,.\end{aligned}$$
Both can be larger than 1, depending on the relative sizes of $x$, $y$ and on the angle $\phi,$ which spoils the premises of causality. Thus, while the mode $\oplus$ is unphysical in several senses, the mode $\ominus$ breaks causality for certain choices of parameters.
The other configuration to be examined is the case of parallel vectors $\mathbf{B}$, $\mathbf{C}$ where the dispersion equation (\[PDR6B\]) yields
$$\begin{aligned}
E^{(\pm)} & =\zeta^{(\pm)}|\mathbf{p}|\,,\quad{\zeta^{(\pm)}=\zeta^{(\pm)}(|\mathbf{p}|)}\,,\displaybreak[0]\\[2ex]
\zeta^{(+)} & =\frac{\sqrt{1+z^2-2y^2\cos^2\vartheta}}{z}\,,\displaybreak[0]\\[2ex]
\zeta^{(-)} & =\sqrt{\frac{1+z^2-2(x^2\sin^2\vartheta+y^2\cos^2\vartheta)}{z^2-2x^2}}\,.\end{aligned}$$
The front velocities of both modes are given by
$$\begin{aligned}
u_{\mathrm{fr}}^{(+)} & =\sqrt{1-2\beta^2\cos^2\vartheta}\,,\displaybreak[0]\\[2ex]
u_{\mathrm{fr}}^{(-)} & =\sqrt{\frac{1-2(\alpha^2\sin^2\vartheta
+\beta^2\cos^2\vartheta)}{1-2\alpha^2}}\,,\end{aligned}$$
where these results are expressed in terms of the constant ratios $x/z\equiv\alpha$ and $y/z\equiv\beta$. The front velocity of the mode $\oplus$ cannot exceed 1, which does not hold for the mode $\ominus$, though, as $u_{\mathrm{fr}}^{(-)}$ can be larger than 1 for $\beta<\alpha$. The next step is to obtain the magnitudes of the group velocities:
\[eq:group-velocity-spacelike-case-parallel-vectors\]$$\begin{aligned}
u_{\mathrm{gr}}^{(+)} & =\sqrt{\frac{4y^2(y^2-z^2)\cos^2\vartheta+z^4}{z^2(1+z^2-2y^2\cos^2\vartheta)}}\,,\displaybreak[0]\\[2ex]
u_{\mathrm{gr}}^{(-)} & =\sqrt{\frac{4(x^2-y^2)(x^2+y^2-z^2)\cos^2\vartheta
-(2x^2-z^2)^{2}}{(2x^2-z^2)\left[1+z^2-2(x^2\sin^2\vartheta+y^2\cos^2\vartheta)\right]}}\,,\end{aligned}$$
with the definitions (\[DEFIN\]). By introducing the parameters $\alpha$, $\beta$ into the previous results, the asymptotic behavior of the group velocity for small momenta is
$$\begin{aligned}
u_{\mathrm{gr}}^{(\pm)}&=\bar{\bar{\Theta}}^{(\pm)}|\mathbf{p}|\,, \\[2ex]
\bar{\bar{\Theta}}^{(+)}&=\theta\sqrt{1+4\beta^2(\beta^2-1)\cos^2\vartheta}\,,\quad\bar{\bar{\Theta}}^{(-)}=\theta\sqrt{1-2\alpha^2+\frac{4(\alpha^2-\beta^2)[1-(\alpha^2+\beta^2)]\cos^2\vartheta}{1-2\alpha^2}}\,.\end{aligned}$$
Each mode again resembles an anisotropic Podolsky mode. The limit of vanishing Lorentz violation leads to the standard Podolsky dispersion relation. Also, the modes can exhibit spurious behavior, since singularities are present for $$z_{\mathrm{sing}}^{(+)}=\frac{1}{\sqrt{2\beta^2\cos^2\vartheta-1}}\,,\quad
z_{\mathrm{sing}}^{(-)}=\frac{1}{\sqrt{2(\alpha^2\sin^2\vartheta+\beta^2\cos^2\vartheta)-1}}\,,$$ again pointing to the existence of unphysical regimes (see the red curves of Fig. \[caseN6\]). The parameters $\alpha$ and $\beta$ can be chosen such that these values are complex, whereupon there are no singularities. In this case, the group velocity rises from 0 to a finite value. This behavior is characteristic of an exotic mode if this finite value is smaller than 1. The blue curve in Fig. \[fig:full-theory-complicated-dispersion-equation-spacelike-parallel-case-1\] represents an example for this case. If the finite value is larger than 1 the mode is spurious whereby an example is given by the blue curve in Fig. \[fig:full-theory-complicated-dispersion-equation-spacelike-parallel-case-2\]. Each expression for the group velocity exceeds 1 for $$z|_{u_{\mathrm{gr}}^{(+)}=1}=\frac{1}{\sqrt{4\beta^2-2}|\beta\cos\vartheta|}\,,\quad
z|_{u_{\mathrm{gr}}^{(-)}=1}=\sqrt{\frac{1-2\alpha^2}{(1-2\beta^2)(\alpha^2-\beta^2)[1+\cos(2\vartheta)]}}\,.$$ The first value becomes complex for $\beta<1/\sqrt{2}$ and the second for $\alpha>1/\sqrt{2}$, $\beta>1/\sqrt{2}$, and $\beta>\alpha$ at the same time or for $\alpha>1/\sqrt{2}>\beta$ or when $1/\sqrt{2}>\beta>\alpha$. Under these conditions, the group velocity stays below 1. The profiles of $u_{\mathrm{gr}}$ for these two modes are depicted in Fig. \[caseN6\].
Connection to nonminimal SME
============================
An interesting question to ask is whether the LV contribution of Lagrangian (\[LVSLAD1\]) is contained in the nonminimal SME by Kostelecký and Mewes [@Kostelec1]. As we are addressing LV and *CPT*-even dimension-6 terms, the first higher-order LV coefficient to be considered is that of dimension 6, namely$$(\hat{k}_{F})^{\kappa\lambda\mu\nu}=
(k_{F}^{(6)})^{\kappa\lambda\mu\nu\alpha_1\alpha_2}\partial_{\alpha_1}\partial_{\alpha_2}\,,
\label{k6AM}$$ where we have omitted the usual dimension-4 *CPT*-even coefficients of the photon sector of the SME, $(k_F^{(4)})^{\kappa\lambda\mu\nu}$. It is possible to show that the parameterization
\[eq:preliminary-mappings\] $$(k_{F}^{(6)})^{\mu\nu\varrho\sigma\alpha_1\alpha_2}=\eta^{2}D^{\nu\sigma}
g^{\mu\alpha_1}g^{\varrho\alpha_2}\,,$$ successfully reproduces the LV term present in Lagrangian (\[LVSLAD1\]). Indeed,$$(k_{F}^{(6)})^{\mu\nu\varrho\sigma\alpha_1\alpha_2}\partial_{\alpha_1}
\partial_{\alpha_2}F_{\mu\nu}F_{\rho\sigma}=\eta^2D^{\nu\sigma}\partial^{\mu}\,F_{\mu\nu}\partial^{\varrho}F_{\rho\sigma}\,,$$
which demonstrates that this LV piece is contained in the nonminimal SME [@Kostelec1]. In an analogue way, the correspondence $$(k_F^{(6)})^{\mu\nu\varrho\sigma\alpha_1\alpha_2}=
\eta_1^2D^{\nu\sigma}g^{\mu\alpha_1}g^{\varrho\alpha_2}
-\eta_2^2D^{\alpha_2\sigma}g^{\mu\alpha_1}g^{\nu\varrho}\,,
\label{Eq-88}$$ yields the LV terms of the generalized model of Lagrangian (\[L3\]). Finally, the symmetries of $(\hat{k}_F)^{\kappa\lambda\mu\nu}$ have to be taken into account. The latter fourth-rank observer tensor is antisymmetric in the first and second pair of indices and symmetric under an exchange of both index pairs, which is evidently not the case for the choices in Eqs. (\[eq:preliminary-mappings\]), (\[Eq-88\]). Performing the (anti)symmetrization explicitly and considering the factor 1/4 before $(\hat{k}_F)^{\kappa\lambda\mu\nu}$ in the definition of the $CPT$-even term of the SME photon sector, the final mapping of our terms to the SME is given by $$\begin{aligned}
(\hat{k}_F)^{\mu\nu\varrho\sigma}&=&\frac{1}{2}\left[(k_F^{(6)})^{\mu\nu\varrho\sigma\alpha_1 \alpha_2}-(k_F^{(6)})^{\nu\mu\varrho\sigma\alpha_1\alpha_2}-(k_F^{(6)})^{\mu\nu\sigma\varrho \alpha_1\alpha_2}+(k_F^{(6)})^{\nu\mu\sigma\varrho\alpha_1\alpha_2}\right. \notag \\
&&\left.\phantom{\frac{1}{2}\Big[}+\,(\mu\leftrightarrow\varrho,\nu\leftrightarrow \sigma)\right]\partial_{\alpha_1}\partial_{\alpha_2}\,.\end{aligned}$$ It can be checked that the tensor constructed in this way automatically satisfies the Bianchi identity for an arbitrary triple of indices chosen from the first four indices. In addition, antisymmetrization on any three indices must be imposed via Eq. (55) of the first paper in [@Kostelec1]. The Podolsky term is Lorentz-invariant and cannot be mapped to SME coefficients. However, it is introduced into the field equations of the nonminimal $CPT$-even photon sector as follows:
$$\begin{aligned}
M^{\mu\nu}A_{\nu}&=0\,, \\[2ex]
M^{\mu\nu}&=\left\{(1-\theta^2p^2)\left[g^{\mu\nu}g^{\alpha\beta}-g^{\mu\alpha} g^{\nu\beta}\right]+2(\hat{k}_F)^{\mu\alpha\nu\beta}\right\}p_{\alpha}p_{\beta}\,.\end{aligned}$$
Conclusion and remarks
======================
In this work we have addressed an electrodynamics endowed with LV terms of mass dimension 6. We considered aspects not yet analysed in the literature, such as an evaluation of the propagator, the dispersion relations, and the propagation modes. First, a Maxwell electrodynamics modified by an anisotropic Podolsky-type term of mass dimension 6 was examined. For certain parameter choices, all the modes, which are affected by the higher-dimensional term, were found to exhibit a nonphysical behavior. The magnitude of the group velocity was demonstrated to become larger than 1 (breaking of causality), to have singularities or to vanish for small momenta (absence of propagation). The first two behaviors described are interpreted as unphysical and we referred to dispersion relations having characteristics of this kind as spurious. Features such as decreasing group or front velocities for small momenta only indicate that these modes decouple from the low-energy regime. According to the criteria used in the paper, a mode endowed with such properties was called exotic as long as it neither involved singularities nor superluminal group or front velocities.
The results obtained indicate that an electrodynamics modified by this kind of higher-derivative LV term can only provide a physical propagation of signals in the limit of large momenta. Such an interpretation could recover a meaningful behavior of the theory in the case that it is possible to state a suitable cutoff without generating bad side effects. For small momenta, some of the modes correspond to Lorentz-violating modifications of the Podolsky-type dispersion relation, which are not necessarily unphysical. We also found dispersion relations providing group velocities that are larger than 1 or divergent at some points, which are behaviors characterizing spurious propagation modes. A brief discussion of unitarity for this modification was delivered as well. The result of our analysis was that additional time derivatives in the Lagrange density are likely to spoil unitarity. This behavior is expected and was already observed in Lorentz-invariant higher-derivative theories as well as alternative LV higher-derivative modifications of the photon and fermion sector.
At a second stage, we analyzed a more general extension of the modified electrodynamics in the presence of an additional anisotropic Lee-Wick term of mass dimension 6. Again, the propagator was derived and the corresponding dispersion relations were obtained from it. Subsequently, these dispersion relations were examined with respect to signal propagation. The structure of this second framework is more involved in comparison to the first theory under consideration. The modified dispersion relations are associated with exotic and spurious modes, as already observed in the first model. There exist parameter choices that are connected to issues with causality. Another observation was that some of the dispersion relations can exhibit singularities, i.e., they become unphysical for a certain range of momenta, which is enough to classify these modes as spurious. However, we also found sectors for which the dispersion laws do not suffer from any of the problems just mentioned. Therefore, those can be considered as well-behaved according to the criteria that the current paper is based on.
The results of the paper suggest which sectors of the framework should be subject to quantization and which ones should be discarded. These sectors are classified according to the geometrical properties of the two four-vectors that the background tensor is composed of. Both spurious and exotic modes were found dependent on choices of the parameters. There is no need of introducing lower cutoffs of the four-momentum, which would be difficult to realize from a physical point of view, anyhow. Instead, the regions of parameter space associated with a vicious classical behavior should be removed from the theory. If the theories proposed are subject to studies at the quantum level in the future, the results of the paper demonstrate, which sectors could be quantized properly. Physical dispersion relations must be clearly free of singularities. Modes without violations of classical causality may be the preferred ones to be studied. However, it ought to be taken into account, as well, that superluminal velocities do not necessarily cause problems at the quantum level, e.g., for microcausality [@Klinkhamer:2010zs].
One of the most significant questions to be answered in future works would be related to gauge fixing. It was mentioned in the paper that gauge fixing in vector field theories including higher derivatives is even more subtle than in the standard case. Whether or not the gauge fixing conditions proposed in [@GalvaoP] for Podolsky’s theory are still applicable in the Lorentz-violating case, would be an interesting aspect to investigate. Furthermore, to quantize the theory, proper field operators will have to be constructed and a treatment of the spurious dispersion laws at the quantum level would be mandatory. A related question is whether the prescription of how to preserve (perturbative) unitarity, which was introduced by Lee and Wick [@LeeWick], still works. Finally, the free theory should be coupled to fermions to be able to study physical processes. Such questions have been considered recently for a particular fermionic framework including higher derivatives [@Reyes:2016pus], but how to proceed in the photon sector is still an open problem.
The authors are grateful to FAPEMA, CAPES, and CNPq (Brazilian research agencies) for invaluable financial support.
[99]{}
D. Colladay and V.A. Kostelecký, Phys. Rev. D **55**, 6760 (1997); D. Colladay and V.A. Kostelecký, Phys. Rev. D **58**, 116002 (1998); S.R. Coleman and S.L. Glashow, Phys. Rev. D **59**, 116008 (1999).
V.A. Kostelecký and S. Samuel, Phys. Rev. Lett. **63**, 224 (1989); Phys. Rev. D **39**, 683 (1989); Phys. Rev. D **40**, 1886 (1989); Phys. Rev. Lett. **66**, 1811 (1991); V.A. Kostelecký and R. Potting, Nucl. Phys. B **359**, 545 (1991); Phys. Rev. D **51**, 3923 (1995); Phys. Lett. B **381**, 89 (1996).
V.A. Kostelecký and C.D. Lane, J. Math. Phys. **40**, 6245 (1999); V.A. Kostelecký and R. Lehnert, Phys. Rev. D **63**, 065008 (2001); D. Colladay and V.A. Kostelecký, Phys. Lett. B **511**, 209 (2001); R. Lehnert, Phys. Rev. D **68**, 085003 (2003); R. Lehnert, J. Math. Phys. **45**, 3399 (2004); B. Altschul, Phys. Rev. D **70**, 056005 (2004); G.M. Shore, Nucl. Phys. B **717**, 86 (2005).
W.F. Chen and G. Kunstatter, Phys. Rev. D **62**, 105029 (2000); O.G. Kharlanov and V.Ch. Zhukovsky, J. Math. Phys. **48**, 092302 (2007); B. Gonçalves, Y.N. Obukhov, and I.L. Shapiro, Phys. Rev. D **80**, 125034 (2009); S.I. Kruglov, Phys. Lett. B **718**, 228 (2012); T.J. Yoder and G.S. Adkins, Phys. Rev. D **86**, 116005 (2012); S. Aghababaei, M. Haghighat, I. Motie, Phys. Rev. D **96**, 115028 (2017).
R. Bluhm, V.A. Kostelecký, and N. Russell, Phys. Rev. Lett. **79**, 1432 (1997); R. Bluhm, V.A. Kostelecký, and N. Russell, Phys. Rev. D **57**, 3932 (1998); R. Bluhm, V.A. Kostelecký, and N. Russell, Phys. Rev. Lett. **82**, 2254 (1999); V.A. Kostelecký and C.D. Lane, Phys. Rev. D **60**, 116010 (1999); R. Bluhm and V.A. Kostelecký, Phys. Rev. Lett. **84**, 1381 (2000); R. Bluhm, V.A. Kostelecký, and C.D. Lane, Phys. Rev. Lett. **84**, 1098 (2000); R. Bluhm, V.A. Kostelecký, C.D. Lane, and N. Russell, Phys. Rev. Lett. **88**, 090801 (2002).
S.M. Carroll, G.B. Field, and R. Jackiw, Phys. Rev. D **41**, 1231 (1990); A.A. Andrianov and R. Soldati, Phys. Rev. D **51**, 5961 (1995); A.A. Andrianov and R. Soldati, Phys. Lett. B **435**, 449 (1998); A.A. Andrianov, R. Soldati, and L. Sorbo, Phys. Rev. D **59**, 025002 (1998); C. Adam and F.R. Klinkhamer, Nucl. Phys. B **607**, 247 (2001); C. Adam and F.R. Klinkhamer, Nucl. Phys. B **657**, 214 (2003); V.Ch. Zhukovsky, A.E. Lobanov, and E.M. Murchikova, Phys. Rev. D **73**, 065016 (2006); A.A. Andrianov, D. Espriu, P. Giacconi, and R. Soldati, J. High Energy Phys. **0909**, 057 (2009); J. Alfaro, A.A. Andrianov, M. Cambiaso, P. Giacconi, and R. Soldati, Int. J. Mod. Phys. A **25**, 3271 (2010); Y.M.P. Gomes, P.C. Malta, Phys. Rev. D **94**, 025031 (2016); A. Martín-Ruiz, C.A. Escobar, Phys. Rev. D **95**, 036011 (2017).
A.P. Baêta Scarpelli, H. Belich, J.L. Boldo, and J.A. Helayël-Neto, Phys. Rev. D **67**, 085021 (2003); R. Lehnert and R. Potting, Phys. Rev. Lett. **93**, 110402 (2004); R. Lehnert and R. Potting, Phys. Rev. D **70**, 125010 (2004); C. Kaufhold and F.R. Klinkhamer, Nucl. Phys. B **734**, 1 (2006); B. Altschul, Phys. Rev. D **75**, 105003 (2007); H. Belich, L.D. Bernald, P. Gaete, and J.A. Helayël-Neto, Eur. Phys. J. C **73**, 2632 (2013).
B. Altschul, Phys. Rev. Lett. **98**, 041603 (2007); C. Kaufhold and F.R. Klinkhamer, Phys. Rev. D **76**, 025024 (2007); B. Altschul, Nucl. Phys. B **796**, 262 (2008).
V.A. Kostelecký and M. Mewes, Phys. Rev. Lett. **87**, 251304 (2001); V.A. Kostelecký and M. Mewes, Phys. Rev. D **66**, 056005 (2002); V.A. Kostelecký and M. Mewes, Phys. Rev. Lett. **97**, 140401 (2006); C.A. Escobar and M.A.G. Garcia, Phys. Rev. D **92**, 025034 (2015); A. Martín-Ruiz and C.A. Escobar, Phys. Rev. D **94**, 076010 (2016).
B. Altschul, Phys. Rev. Lett. **98**, 041603 (2007); C. Kaufhold and F.R. Klinkhamer, Phys. Rev. D **76**, 025024 (2007); F.R. Klinkhamer and M. Risse, Phys. Rev. D **77**, 016002 (2008); F.R. Klinkhamer and M. Risse, Phys. Rev. D **77**, 117901 (2008); F.R. Klinkhamer and M. Schreck, Phys. Rev. D **78**, 085026 (2008).
A. Moyotl, H. Novales-Sánchez, J.J. Toscano, and E.S. Tututi, Int. J. Mod. Phys. A [**29**]{}, 1450039 (2014); Int. J. Mod. Phys. A [**29**]{}, 1450107 (2014).
M. Schreck, Phys. Rev. D **86**, 065038 (2012); G. Gazzola, H.G. Fargnoli, A.P. Baêta Scarpelli, M. Sampaio, and M.C. Nemes, J. Phys. G **39**, 035002 (2012); A.P. Baêta Scarpelli, J. Phys. G **39**, 125001 (2012); B. Agostini, F.A. Barone, F.E. Barone, P. Gaete, and J.A. Helayël-Neto, Phys. Lett. B **708**, 212 (2012); L.C.T. Brito, H.G. Fargnoli, and A.P. Baêta Scarpelli, Phys. Rev. D **87**, 125023 (2013).
R. Jackiw and V.A. Kostelecký, Phys. Rev. Lett. **82**, 3572 (1999); J.-M. Chung, Phys. Rev. D **60**, 127901 (1999); M. Pérez-Victoria, Phys. Rev. Lett. **83**, 2518 (1999); J.-M. Chung and B.K. Chung, Phys. Rev. D **63**, 105015 (2001); G. Bonneau, Nucl. Phys. B **593**, 398 (2001); M. Pérez-Victoria, J. High Energy Phys. **0104**, 032 (2001); O.A. Battistel and G. Dallabona, J. Phys. G **27**, L53 (2001); Nucl. Phys. B **610**, 316 (2001); A.P.B. Scarpelli, M. Sampaio, M.C. Nemes, and B. Hiller, Phys. Rev. D **64**, 046013 (2001); O.A. Battistel and G. Dallabona, J. Phys. G **28**, L23 (2002); B. Altschul, Phys. Rev. D **70**, 101701(R) (2004); T. Mariz, J.R. Nascimento, E. Passos, R.F. Ribeiro, and F.A. Brito, J. High Energy Phys. **0510**, 019 (2005); J.R. Nascimento, E. Passos, A.Yu. Petrov, and F.A. Brito, J. High Energy Phys. **0706**, 016 (2007); A.P.B. Scarpelli, M. Sampaio, M.C. Nemes, and B. Hiller, Eur. Phys. J. C **56**, 571 (2008); F.A. Brito, J.R. Nascimento, E. Passos, and A.Yu. Petrov, Phys. Lett. B [**664**]{}, 112 (2008); F.A. Brito, L.S. Grigorio, M.S. Guimaraes, E. Passos, and C. Wotzasek, Phys. Rev. D **78**, 125023 (2008); O.M. Del Cima, J.M. Fonseca, D.H.T. Franco, and O. Piguet, Phys. Lett. B [**688**]{}, 258 (2010).
V.A. Kostelecký and M. Mewes, Phys. Rev. D **80**, 015020 (2009); M. Mewes, Phys. Rev. D [**85**]{}, 116012 (2012); M. Schreck, Phys. Rev. D [**89**]{}, 105019 (2014).
V.A. Kostelecký and M. Mewes, Phys. Rev. D [**88**]{}, 096006 (2013); M. Schreck, Phys. Rev. D [**90**]{}, 085025 (2014).
R.C. Myers and M. Pospelov, Phys. Rev. Lett. **90**, 211601 (2003); C.M. Reyes, L.F. Urrutia, and J.D. Vergara, Phys. Rev. D **78**, 125011 (2008); J. Lopez-Sarrion and C.M. Reyes, Eur. Phys. J. C **72**, 2150 (2012).
C.M. Reyes, L.F. Urrutia, and J.D. Vergara, Phys. Lett. B **675**, 336 (2009); C.M. Reyes, Phys. Rev. D **82**, 125036 (2010); C.M. Reyes, Phys. Rev. D **87**, 125028 (2013).
H. Belich, T. Costa-Soares, M.M. Ferreira, Jr., and J.A. Helayël-Neto, Eur. Phys. J. C **41**, 421 (2005); H. Belich, L.P. Colatto, T. Costa-Soares, J.A. Helayël-Neto, and M.T.D. Orlando, Eur. Phys. J. C **62**, 425 (2009); B. Charneski, M. Gomes, R.V. Maluf, and A.J. da Silva, Phys. Rev. D **86**, 045003 (2012); A.F. Santos, and Faqir C. Khanna, Phys. Rev. D **95**, 125012 (2017).
R. Casana, M.M. Ferreira, Jr., E. Passos, F.E.P. dos Santos, and E.O. Silva, Phys. Rev. D **87**, 047701 (2013); J.B. Araujo, R. Casana, M.M. Ferreira, Jr., Phys. Lett. B **760**, 302 (2016).
Y. Ding and V.A. Kostelecký, Phys. Rev. D [**94**]{}, 056008 (2016).
B. Podolsky, Phys. Rev. [**62**]{}, 68 (1942); B. Podolsky and C. Kikuchi, Phys. Rev. [**65**]{}, 228 (1944).
A. Accioly and E. Scatena, Mod. Phys. Lett. A [**25**]{}, 269 (2010).
C.A.P. Galvão and B.M. Pimentel, Can. J. Phys. [**66**]{}, 460 (1988).
R. Bufalo, B.M. Pimentel, and G.E.R. Zambrano, Phys. Rev. D [**83**]{}, 045007 (2011).
C.A. Bonin, R. Bufalo, B.M. Pimentel, and G.E.R. Zambrano, Phys. Rev. D [**81**]{}, 025003 (2010).
R. Bufalo, B.M. Pimentel, and G.E.R. Zambrano, Phys. Rev. D [**86**]{}, 125023 (2012).
C.A. Bonin, B.M. Pimentel, and P.H. Ortega, “Multipole expansion in generalized electrodynamics,” arXiv:1608.00902.
T.D. Lee and G.C. Wick, Nucl. Phys. B [**9**]{}, 209 (1969); T.D. Lee and G.C. Wick, Phys. Rev D [**2**]{}, 1033 (1970).
B. Grinstein and D. O’Connell, Phys. Rev. D [**78**]{}, 105005 (2008); B. Grinstein, D. O’Connell, and M.B. Wise, Phys. Rev. D [**77**]{}, 025012 (2008); J.R. Espinosa, B. Grinstein, D. O’Connell, and M.B. Wise, Phys. Rev. D [**77**]{}, 085002 (2008); T.E.J. Underwood and R. Zwicky, Phys. Rev. D [**79**]{}, 035016 (2009); R.S. Chivukula, A. Farzinnia, R. Foadi, and E.H. Simmons, Phys. Rev. D [**82**]{}, 035015 (2010).
T.G. Rizzo, J. High Energy Phys. [**0706**]{}, 070 (2007); T.G. Rizzo, J. High Energy Phys. [**0801**]{}, 042 (2008); E. Álvarez, C. Schat, L. Da Rold, and A. Szynkman, J. High Energy Phys. [**0804**]{}, 026 (2008); C.D. Carone and R. Primulando, Phys. Rev. D [**80**]{}, 055020 (2009).
A.M. Shalaby, Phys. Rev. D [**80**]{}, 025006 (2009).
C.D. Carone, Phys. Lett. B [**677**]{}, 306 (2009); C.D. Carone and R.F. Lebed, J. High Energy Phys. [**0901**]{}, 043 (2009).
B. Grinstein and D. O’Connell, Phys. Rev. D [**78**]{}, 105005 (2008); R.S. Chivukula, A. Farzinnia, R. Foadi, and E.H. Simmons, Phys. Rev. D [**82**]{}, 035015 (2010).
B. Fornal, B. Grinstein, and M.B. Wise, Phys. Lett. B [**674**]{}, 330 (2009); C.A. Bonin, R. Bufalo, B.M. Pimentel, and G.E.R. Zambrano, Phys. Rev. D [**81**]{}, 025003 (2010); C.A. Bonin and B.M. Pimentel, Phys. Rev. D [**84**]{}, 065023 (2011).
A.A. Slavnov, Teor. Mat. Fiz [**13**]{}, 174 (1972).
F.A. Barone, G. Flores-Hidalgo, and A.A. Nogueira, Phys. Rev. D [**88**]{}, 105031 (2013).
F.A. Barone, G. Flores-Hidalgo, and A.A. Nogueira, Phys. Rev. D [**91**]{}, 027701 (2015).
F.A. Barone and A.A. Nogueira, Eur. Phys. J. C [**75**]{}, 339 (2015).
A. Accioly, J. Helayël-Neto, F.E. Barone, F.A. Barone, and P. Gaete, Phys. Rev. D [**90**]{}, 105029 (2014).
L.H.C. Borges, A.G. Dias, A.F. Ferrari, J.R. Nascimento, and A.Yu. Petrov, Phys. Lett. B [**756**]{}, 332 (2016).
L.H.C. Borges, A.F. Ferrari, and F.A. Barone, Eur. Phys. J. C [**76**]{}, 599 (2016).
V.A. Kostelecký and N. Russell, Rev. Mod. Phys. [**83**]{}, 11 (2011).
L. Bonetti, L.R. dos Santos Filho, J.A. Helayël-Neto, and A.D.A.M. Spallicci, Phys. Lett. B [**764**]{}, 203 (2017).
A.P. Baêta Scarpelli, H. Belich, J.L. Boldo, and J.A. Helayël-Neto, Phys. Rev. D [**67**]{}, 085021 (2003); R. Casana, M.M. Ferreira, Jr., A.R. Gomes, and F.E.P. dos Santos, Phys. Rev. D [**82**]{}, 125006 (2010); R. Casana, M.M. Ferreira, Jr., and R.P.M. Moreira, Eur. Phys. J. C [**72**]{}, 2070 (2012); A. Moyotl, H. Novales-Sánchez, J.J. Toscano, and E.S. Tututi, Int. J. Mod. Phys. A [**29**]{}, 1450107 (2014); T.R.S. Santos, R.F. Sobreiro, and A.A. Tomaz, Phys. Rev. D [**94**]{}, 085027 (2016).
J.A.M. Vermaseren, “New features of *FORM*,” math-ph/0010025.
L. Brillouin, *Wave propagation and group velocity* (Academic Press, New York and London, 1960).
F.R. Klinkhamer and M. Schreck, Nucl. Phys. B [**848**]{}, 90 (2011).
C.M. Reyes and L.F. Urrutia, Phys. Rev. D [**95**]{}, 015024 (2017); M. Maniatis and C.M. Reyes, Phys. Rev. D [**89**]{}, 056009 (2014); C.M. Reyes, Phys. Rev. D [**87**]{}, 125028 (2013).
R. Turcati and M.J. Neves, Adv. High Energy Phys. [**153953**]{}, 1 (2014).
M. Veltman, in *Methods in Field Theory*, edited by R. Balian and J. Zinn-Justin (World Scientific, Singapore, 1981).
D. Colladay and V.A. Kostelecký, Phys. Lett. B [**511**]{}, 209 (2001).
[^1]: We mentioned in the introduction that gauge conditions used in Maxwell’s electrodynamics may cause problems in Podolsky’s extension [@GalvaoP]. For example, as the vector field component $A^0$ is nondynamical, Lorenz gauge requires that the solutions of the field equations be transverse, which is not the case for massive modes. It is paramount to consider alternative gauge conditions when quantizing the theory. However, we neither obtain the solutions of the equations of motion nor do we quantize the framework under investigation. The focus is on studying the modified dispersion relations, which is a classical analysis. Also, the gauge condition will not modify the dispersion relations. Therefore, to reduce technical complications in computing the propagator, we will still employ the usual set of gauges used in Maxwell’s electrodynamics.
[^2]: *FORM* is a programming language that allows for symbolic manipulations of mathematical expressions to be performed. It is widely used for evaluating lengthy algebraic expressions that occur in computations of quantum corrections in high-energy physics.
[^3]: This fact will be demonstrated in Sec. \[sec:unitarity-analysis\] explicitly.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We have experimentally observed re-equilibration of a magnetically trapped cloud of metastable neon atoms after it was put in a non-equilibrium state. Using numerical simulations we show that anharmonic mixing, equilibration due to the collisionless dynamics of atoms in a magnetic trap, is the dominant process in this equilibration. We determine the dependence of its time on trap parameters and atom temperature. Furthermore we observe in the simulations a resonant energy exchange between the radial and axial trap dimensions at a ratio of trap frequencies $\omega_r/\omega_z=3/2$. This resonance is explained by a simple oscillator model.'
address:
- ' Physics Department, Eindhoven University of Technology, Postal Box 513, Eindhoven, The Netherlands'
- ' Physics and Astronomy Department, University College London, Gower Street, London WC1E 6BT, United Kingdom'
- '§ School of Physics, University of Melbourne, Parkville, Vic. 3010, Australia'
author:
- 'R Gommers, B J Claessens, C J Hawthorn§, H C W Beijerinck and E J D Vredenbregt'
title: Anharmonic mixing in a magnetic trap
---
Introduction
============
Magnetic traps have become a standard and inexpensive tool in cold atom physics in recent years [@Esslinger]. Together with the application of cooling techniques such as evaporative cooling and sympathetic cooling, they have enabled the formation of Bose-Einstein condensates [@Anderson] and, more recently, degenerate Fermi gases [@Regal]. The most commonly used species for cold atom experiments (Rb, Cs, Na) are easily cooled to the micro-Kelvin regime, where the atomic dynamics in a magnetic trap is usually well described by regular harmonic motion. For other species such as the noble gases [@Robert; @Tempelaars] and the group II elements [@Simien; @Daily] however, the lowest attainable temperatures can be in the milli-Kelvin regime. In this regime the atomic dynamics is more complicated, with coupling between the different dimensions of the trap becoming important.
Anharmonic mixing is a process that couples the motion of atoms in different dimensions of a magnetic trap. This coupling enables redistribution of energy over the coupled dimensions. Therefore, anharmonic mixing has to be considered when applying techniques that rely on the motion of atoms in a trap, such as evaporative cooling [@Luiten], Doppler cooling in a magnetic trap [@Helmerson] and rethermalization experiments aimed at measuring a scattering length [@Monroe]. In order to investigate on what timescale anharmonic mixing plays a role, we perform a Monte Carlo simulation of a cloud of atoms in a Ioffe-Quadrupole magnetic trap. We determine the dependence of this timescale on trap parameters and on atom temperature. Surkov *et. al.* [@Surkov] have investigated anharmonic mixing in the limit of low temperature where it can be treated as a perturbation of regular harmonic motion, but to the best of our knowledge this is the first paper discussing anharmonic mixing at higher temperatures. We will also present here an experiment showing that anharmonic mixing does play an important role in the atomic dynamics, and can even be the dominant mechanism of energy redistribution under certain conditions. To conclude we will look at possible applications of anharmonic mixing.
Monte Carlo calculations
========================
A Ioffe-Quadrupole magnetic trap (MT) is one of the most common magnetic traps. Its potential is given by $$\begin{aligned}
\label{TotalPot}
U(x,y,z)&=\mu[B_{0}^2+(\alpha^2-B_0\beta)(x^2+y^2)+2B_0\beta
z^2\nonumber\\&+\frac{1}{4}\beta^2(x^2+y^2)^2
+\beta^2z^4+2\alpha\beta(x^2-y^2)z]^{1/2}-\mu
B_0,
\end{aligned}$$ where $x$ and $y$ are the radial trap dimensions, $z$ is the axial trap dimension, $\mu$ is the magnetic moment of the atom, $\alpha$ the gradient of the magnetic field, $\beta$ the curvature of the magnetic field and $B_0$ the magnetic bias field. In this expression for the trap potential terms of order higher than four have been neglected. The MT has trap frequencies $\omega_r$ and $\omega_z$ in the limit $3k_BT\ll\mu B_0$, where the trap shape is harmonic: $$\label{HarmPot}
U(r,z)=\frac{m}{2}(\omega_r^2r^2+\omega_z^2z^2).$$ Here $m$ is the atomic mass, $r=(x^2+y^2)^{1/2}$ the radial coordinate, $\omega_r=[\frac{\mu}{m}(\alpha^2/B_0-\beta)]^{1/2}$ the radial trap frequency, and $\omega_z=(\mu\beta/m)^{1/2}$ the axial trap frequency. The higher order terms in (\[TotalPot\]) are in this limit negligible compared to the harmonic terms, the term $2\alpha\beta(x^2-y^2)z$ that couples the motion in the axial and radial directions is therefore absent from (\[HarmPot\]).
The starting point of a simulation is an atom cloud in equilibrium in a harmonic trap. The initial positions and velocities of the atoms are determined by a Monte Carlo method. For the properties of the atom we use the values for metastable neon in the $^3P_2$ state, as this is the atom we use in our experiment. Two kinds of clouds have been used, namely thermal clouds and clouds in which all the atoms have the same energy. The former is used to compare the results of the simulations with experiments while the latter is used to gain a better understanding of the dynamics of atoms at a certain energy. During the simulation the position and velocity of each atom is determined as a function of time by integrating the equations of motion. At the start of the simulation the magnetic bias field $B_0$ is adiabatically ramped down, thereby increasing the energy of the atoms in the radial direction and changing the effective shape of the potential from harmonic to almost linear. The ramping is adiabatic if the condition $\frac{1}{\omega_r^2}\frac{d\omega_r}{dt}<1$ is fulfilled [@Surkov]. The simulation yields the kinetic and potential energies of the cloud in every direction as a function of time. From the kinetic energies in different directions the transfer of energy from one direction to the other can be determined, and a comparison with the experiment can be made.
We choose the trap parameters at the start of the simulation as $\alpha=1\cdot10^4$ G/m, $\beta=47.5\cdot10^4$ G/m$^2$ and $B_0=99.6$ G. The bias field is then ramped down to 1.5 G. These parameters are chosen because they are easily accessible in our experiment. A typical calculated result of the temperature evolution, or the average kinetic energy, in the radial direction of a thermal cloud of 1 mK after ramping the bias field is shown in Figure \[Thermal\]. The first few tens of milliseconds show a linear decrease of temperature, after that the decrease is approximately exponential with a mixing time of 112 ms for this particular cloud with a temperature of 1 mK and a relative change in temperature of 8%. The linear decrease of temperature in the first tens of ms is caused by high-energy atoms that mix on a timescale of $\frac{1}{2}\omega_z^{-1}=11$ ms. This is the average time an atom needs to reach a singular point in the potential [@Surkov], where almost instantaneous mixing can occur.
The dependence of the behavior of a thermal cloud on initial trap parameters and cloud temperature is relatively weak. When $\alpha$ is varied over the range $0.9-1.2\cdot10^4$ G/m and $B_{0,end}$ over the range $1.5-20$ G, the mixing time stays within the range $100-130$ ms. Only when $\beta$ is decreased, a significant change in mixing time occurs: $\tau_{mix}\approx400$ ms when $\beta=37.5\cdot10^4$ G/m$^2$. The reason for this weak dependence on trap parameters is that a significant part of the energy of the cloud is carried by atoms that are in a regime where mixing does not occur or where it occurs on a timescale of $\frac{1}{2}\omega_z^{-1}$.
To show that indeed the timescale of energy transfer between dimensions of an atom depends strongly on the energy of that atom, we now take clouds of atoms that all have the *same energy* $E_{atom}$ in order to determine the mixing timescale as a function of energy. This gives us the average behavior of an atom with a certain energy. The result is shown in Figure \[TauTemp\]. Below $E_{atom}/3k_B=375$ $\mu$K no mixing occurs on the longest timescale we considered, i.e. the lifetime of the metastable state Ne$^{\ast}(^3P_2)$ of 14.7 s [@Zinner]. For temperatures above 1.5 mK the mixing time is determined by $\frac{1}{2}\omega_z^{-1}$ and becomes independent of energy.
Next, we investigate the dependence of $\tau_{mix}$ on the trap parameters $\alpha$ and $\beta$ while keeping the energy constant at $E_{atom}/3k_B=700$ $\mu$K. Figure \[TauAlfaBeta\] shows that $\tau_{mix}$ increases with $\alpha$ and decreases with $\beta$. This behavior can be understood qualitatively by comparing the strength of the harmonic $(\sim\alpha^2-\beta)$ and coupling $(\sim\alpha\beta)$ terms in (\[TotalPot\]).
Furthermore we determined the mixing time as a function of the value of the bias field at the end of the ramp. When the final value of the bias field is higher, less energy is added to the atoms and the final temperature of the cloud is lower. Therefore we expect that the mixing time becomes larger for higher final values of the bias field. As can be seen from the result in Figure \[TauBnul\] however, there is a broad resonance (meaning small mixing time) in the mixing time around 14 G. This occurs exactly at a ratio of radial and axial trap frequencies of $\omega_r/\omega_z=3/2$. The resonance is so broad because the radial trap frequency changes only very slowly with bias field. The reason for this slow change is that at higher bias field the final energy of the cloud is lower, resulting in a higher oscillation frequency in the MT because the trap is not harmonic, but at the same time the trap is less tightly confining, resulting in a lower oscillation frequency. These two effects almost completely cancel.
Resonance model
===============
In order to explain this resonance the dynamics of an atom in the MT is now described as two coupled one-dimensional harmonic oscillators, whose energies represent the radial and axial energies of the atom. The coupling term is chosen to be $W=x^2z$, as shown in Figure \[Coupling\]. Now we examine the forces $F_z$ and $F_x$ that the two oscillators exert on one another, assuming they oscillate as $x\sim cos(\omega_xt)$ and $z\sim cos(\omega_zt)$: $$\begin{aligned}
\label{Force}
\fl F_z=-\frac{\partial}{\partial z}W=x^2\sim\cos^2(\omega_xt)\sim\cos(2\omega_xt)+1,\nonumber\\
\fl F_x=-\frac{\partial}{\partial x}W=xz\sim\cos(\omega_xt)\cos(\omega_zt)\sim\cos((\omega_x+\omega_z)t)+\cos((\omega_x-\omega_z)t).
\end{aligned}$$
A resonance occurs if an oscillator is driven by a periodic force with its eigenfrequency, $F_i\sim cos(\omega_it)$. From (\[Force\]) it can be seen that this happens if $\omega_z=2\omega_x$. In general, a coupling term $x^az^b$ yields a resonance at $a\omega_x=b\omega_z$. When potential (\[TotalPot\]) is linearized a term $x^2z^3$ appears, explaining the resonance in Figure \[TauBnul\]. A more formal treatment explaining the resonance is possible by transforming the Hamiltonian of an atom in the potential to Birkhoff-Gustavson normal form [@Tuwankotta; @Quispel], but that is beyond the scope of this paper.
Experiment
==========
To check our simulation results, we now compare with experiment. The only difference between simulation and experiment is that in the experiment the atoms can also collide, which we did not include in the simulation. Therefore it is necessary to consider both anharmonic mixing and collisional equilibration of energy. Our setup has been described previously in [@Tempelaars; @Kuppens]. Briefly, a discharge source is used to create metastable $^3P_2$ neon atoms, and after passing through several laser-cooling sections the atomic beam flux is $6\cdot10^{10}$ s$^{-1}$ [@Tempelaars]. From this atomic beam about $10^9$ atoms are loaded into a magneto-optical trap (MOT) at a temperature of 1 mK. These atoms can be transferred to a Ioffe-Quadrupole magnetic trap with almost unity efficiency. To match the size and temperature of our MOT the trap parameters are chosen as $\alpha=1\cdot10^4$ G/m, $\beta=47.5\cdot10^4$ G/m$^2$ and $B_0=99.6$ G. The MT has trap frequencies of $\omega_r=(2\pi)376$ Hz and $\omega_z=(2\pi)44.7$ Hz.
For this experiment the MOT was operated at a detuning of $-2\Gamma$ and at an intensity of 0.5 mW/cm$^2$ for the three MOT beams together. After turning off the MOT, a $50$ $\mu$s long $\sigma^+$ spin polarization pulse of 2 mW/cm$^2$ was given to put all the atoms in the $|m_J=+2>$ state, and then the MT was turned on. The atom cloud now has a temperature of approximately 1 mK and contains approximately $1\cdot10^9$ atoms. After turning on the MT the atoms are held there long enough to allow the atom cloud to reach its equilibrium state. Then the bias field is ramped from 99.6 G to 1.5 G, adding energy to the atoms in the radial direction and compressing them spatially. The cloud radius in the radial direction as a function of time after the ramping is determined by absorption imaging. The radius is directly proportional to the potential energy and therefore the temperature because the trap shape is linear. A typical result of a series of measurements is shown in Figure \[SqueezeExp\].
The exponential fit gives a characteristic equilibration time of $130\pm25$ ms and a relative change in width of $5\pm3\%$. This result agrees with the timescale of the simulation as found from the data in Figure \[Thermal\]. If the equilibration would be due solely to collisions, this would give a collisional cross-section of $\sigma=8\pi\cdot$(200a$_0$)$^2$, with $a_0$ the Bohr radius. However, the cross-section at 1 mK can not be larger than the unitary limit $\sigma=8\pi(\lambda_{dB}/2\pi)^2=8\pi\cdot$(80a$_0$)$^2$. Here $\lambda_{dB}$ is the thermal de Broglie wavelength of the atom. Also, we did not observe a dependence of the equilibration time on the atomic density, indicating that anharmonic mixing is the dominant mechanism here.
Applications
============
An interesting application for anharmonic mixing is to use it to enhance the efficiency of Doppler cooling in a magnetic trap. This cooling technique is one-dimensional because only along the axial direction of an MT all atoms are polarized the same way. Cooling in the radial directions can be achieved by reabsorption of scattered photons [@Pfau] if the cloud is optically dense. If that is not the case anharmonic mixing can cool the radial directions, as shown in an early experiment by Helmerson *et. al.* [@Helmerson]. By making the mixing time as short as possible the cooling can be fast and the atom losses as a result of the cooling can be limited.
Another technique that could benefit from anharmonic mixing is evaporative cooling. One of the problems that can occur when an atom cloud is cooled evaporatively is gravitational sag, in which gravity shifts the equipotential surfaces of atoms in a trap in such a way that they do not coincide with the surfaces of constant magnetic field. This causes the cooling process to become one or two-dimensional, reducing its efficiency [@Thomas]. If the dimensions are coupled by anharmonic mixing the evaporation remains three-dimensional even when gravity plays a role.
A situation where anharmonic mixing is undesirable and needs to be suppressed is in rethermalization experiments to measure the scattering length [@Monroe; @Schmidt]. The scattering length is one of the properties of an atom that determine the feasibility of evaporative cooling and reaching the transition point of Bose-Einstein condensation for a given number of atoms at a certain density and temperature [@Luiten]. It can be measured by observing collisional equilibration after ramping the bias field as we did in our simulations and experiment. We can conclude that this method will only yield reliable results if the timescale on which anharmonic mixing occurs is long compared to the collision time.
Conclusion
==========
We determined the timescale on which anharmonic mixing occurs and the dependence of that timescale on trap parameters and atom temperature. We observed a resonance in the mixing time as a function of the end bias field, and explained this with a simple oscillator model. We verified experimentally that anharmonic mixing does indeed occur, and that its timescale can be short compared to the timescale needed for collisional equilibration. The application of anharmonic mixing to improve Doppler cooling and lower-dimensional evaporative cooling is possible.
References {#references .unnumbered}
==========
[18]{}
T. Esslinger. I. Bloch and T.W. Hänsch, Phys. Rev. A **58**, R2664 (1998). M. Anderson. J.R. Ensher. M.R. Matthews. C.E. Wieman. and E.A. Cornell, Science **269**, 168 (1995). C.A. Regal. C. Ticknor. J.L. Bohn and D.S. Jin, Nature **424**, 47 (2003). A. Robert. O. Sirjean. A. Browaeys. J. Poupard. S. Nowak. D. Boiron. C.I. Westbrook. and A. Aspect, Science **292**, 461 (2001). J.G.C. Tempelaars. R.J.W. Stas. P.G.M. Sebel. H.C.W. Beijerinck. and E.J.D. Vredenbregt, Eur. Phys. J. D, **18**, 113 (2002). C.E. Simien. Y.C. Chen. P. Gupta. S. Laha. Y.N. Martinez. P.G. Mickelson. S.B. Nagel and T.C. Killian, Phys. Rev. Lett. **92**, 143001 (2004). J.E. Daily. R. Gommers. E.A. Cummings. D.S. Durfee. and S.D. Bergeson, physics 0411004. O.J. Luiten. M.W. Reynolds. and J.T.M. Walraven, Phys. Rev. A **53**, 381 (1996). K. Helmerson. A. Martin. and D.E. Pritchard, J. Opt. Soc. Am. B **9**, 1988 (1992). C.R. Monroe. E.A. Cornell. C.A. Sackett. C.J. Myatt. and C.E. Wieman, Phys. Rev. Lett. **70**, 414 (1993). E.L. Surkov. J.T.M. Walraven. and G.V. Schnlyapnikov, Phys. Rev. Lett. **49**, 4778 (1994). M. Zinner. P. Spöden. T. Kraemer. G. Birkl. and W. Ertmer, Phys. Rev. A **67**, 010501 (2003). J.M. Tuwankotta and F. Verhulst, J. Appl. Math. **61**, 1369 (2000). J.M. Tuwankotta. and G.R.W. Quispel, J. Comp. Appl. Math. **154**, 229 (2003). S.J.M. Kuppens. J.G.C. Tempelaars. V.P. Mogendorff. B.J. Claessens. H.C.W. Beijerinck. and E.J.D. Vredenbregt, Phys. Rev. A **65**, 023410 (2002). P.O. Schmidt. S. Hensler. J. Werner. T. Binhammer. A. Görlitz. and T. Pfau, J. Opt. Soc. Am. B **20**, 960 (2003). A.M. Thomas. S. Hopkins. S.L. Cornish. and C.J. Foot, J. Opt. B **5**, 107 (2003). P.O. Schmidt. S. Hensler. J. Werner. A. Griesmaier. A. Görlitz. T. Pfau and A. Simoni, Phys. Rev. Lett. **91**, 193201 (2003).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'We have demonstrated the effect of hydrostatic pressure on magnetic and transport properties, and thermal transport properties in electron-doped manganites CaMn$_{1-x}$Sb$_{x}$O$_{3}$. The substitution of Sb$^{5+}$ ion for Mn $^{4+}$site of the parent matrix causes one-electron doping with the chemical formula CaMn$^{4+}_{1-2x}$Mn$^{3+}_{x}$Sb$^{5+}_{x}$O$_{3}$ accompanied by a monotonous increase in unit cell volume as a function of $x$. Upon increasing the doping level of Sb, the magnitudes of both electrical resistivity and negative Seebeck coefficient are suppressed at high temperatures, indicating the electron doping. Anomalous diamagnetic behaviors at $x=0.05$ and 0.08 are clearly observed in field cooled dc magnetization. The effect of hydrostatic pressure on dc magnetization is in contrast to the chemical pressure effect due to Sb doping. The dynamical effect of ac magnetic susceptibility measurement points to the formation of the magnetically frustrated clusters such as FM clusters embedded in canted AFM matrix.'
author:
- Yuh Murano
- Michiaki Matsukawa
- Syuya Ohuchi
- Satoru Kobayashi
- Shigeki Nimori
- Ramanathan Suryanarayanan
- Keiichi Koyama
- Norio Kobayashi
title: 'Effect of pressure on the magnetic, transport, and thermal-transport properties of the electron-doped manganite CaMn$_{1-x}$Sb$_{x}$O$_{3}$'
---
INTRODUCTION
============
Manganese oxides with perovskite structure have been extensively investigated since the discovery of colossal magnetoresistance (CMR) effect.[@TO00] The spontaneous insulator to metal transition and its associated CMR effect are well explained on the basis of the double exchange (DE) model between Mn$^{3+}$ and Mn$^{4+}$ ions.[@ZE51] Furthermore, the phase separation model, where the ferromagnetic (FM) metallic and antiferromagnetic (AFM) insulating clusters of competing electronic phases coexist, strongly supports experimental studies of manganites.[@DA01] The dynamic Jahn-Teller (JT) effect of Mn$^{3+}$ ions plays a crucial role in the physics of manganites.[@MI95]
CaMnO$_{3}$, the end member of Ca$_{1-x}$La$_{x}$MnO$_{3}$ system, undergoes a G-type antiferromagnetic transition around $T_N\sim 120$ K accompanied by a weak ferromagnetic component,[@MAC67] where each spin of Mn ions is antiparallel to its nearest neighbors of Mn. In recent years, the electron doped manganite system ($x<0.5$)[@CH96] has attracted much attentions because of the possibility of magnetoresistance effect. Neutron measurements on Ca$_{1-x}$La$_{x}$MnO$_{3}$ [@GRA03] have revealed the formation of the nanometric-scale FM clusters isolated within a G-type AFM matrix for sufficiently low doping ($x<0.03$). Furthermore, for intermediate doping (0.03$\leq x\leq $0.14), a canting of the G-type AFM structure occurs with the nanometric FM clusters. In addition to a large number of researches on the A-site substituted electron doped manganites, it is shown that substituting of the Mn site of CaMnO$_{3}$ with higher valence ions than 4+ such as CaMn$^{4+}_{1-2x}$Mn$^{3+}_{x}$M$^{5+}_{x}$O$_{3}$, with M = Nb, Ta, V, Ru,[@RA00; @PI03; @SHA04; @GU05; @ANG06; @ZHO09]and CaMn$^{4+}_{1-3x}$Mn$^{3+}_{2x}$M$^{6+}_{x}$O$_{3}$, with M = Mo, W,[@RA00; @MAR01; @PI03; @MAI02; @MIC07; @ROZ08; @MAR09] creates Mn$^{3+}$ ions, i.e., electrons leading to the CMR effect. The crystallographic and magnetic properties of CaMn$_{1-x}$Ru$_{x}$O$_{3}$ system ($x\geq 0.1$) allow us to apply the phase separated FM + AFM model to the ground state of this system.[@SHA04] For CaMn$_{1-x}$Mo$_{x}$O$_{3}$ system with low doping of Mo ($x=0.04$), the low temperature magnetic ground state is better described by the canted AFM magnetic structure than by the phase separated state.[@ROZ08; @MAR09] At higher doping levels of Mo, the charge ordered state is established within the parent matrix CaMnO$_{3}$.[@MAR01] It thus is interesting to examine the physical properties of the Mn-site substituted compositions for our understanding of electronic phase diagram of electron-doped manganites.
In this paper, we demonstrate the effect of hydrostatic pressure on magnetic and transport properties, and thermal transport properties in electron-doped manganites CaMn$_{1-x}$Sb$_{x}$O$_{3}$, in order to examine a relationship between lattice and spin. The physical pressure effect is a powerful probe to investigate the electronic states of manganese oxides varying the one-electron band width at doping level fixed because the application of external pressure gives rise to a shrinkage of Mn-O bond length and/or straightening of a Mn-O-Mn bond angle.[@KU97] There has been several studies on the effect of pressure on magnetism of calcium based electron doped manganites so far.[@MAR04] For CaMn$_{1-x}$Ru$_{x}$O$_{3}$ with $x=0.1$, the applied pressure dramatically suppresses the ferromagnetic phase accompanied by a rise of the magnetic transition temperature up to $\sim $14K.
In previous works of slightly electron doped CaMnO$_{3}$ with B-site substitution, negative magnetization properties have been demonstrated.[@ANG06; @MU10] The negative magnetization phenomena in manganites have been originally reported in compounds with two sublattices of Mn ions and rare-earth ions (Nd, Gd, Dy), such as NdMnO$_{3}$[@BAR05; @TRO06], (La,Gd)MnO$_{3}$,[@HEM04] (Nd,Ca)MnO$_{3}$,[@TRO03] (Gd,Ca)MnO$_{3}$,[@PEN02] and (Dy,Ca)MnO$_{3}$.[@NO96; @MO02] Some of these studies were discussed on the basis of ferrimagnetic scenario leading to negative magnetization below a compensation temperature, where Mn and some rare-earths sublattices are antiferrmagnetically coupled. Moreover, a phase separation model between the ferromagnetic clusters and the canted AFM matrix is proposed, in order to account for a possible origin of the negative magnetization in manganites.[@BAR05; @TRO03] In addition to the B site substituted manganite, the negative magnetization was observed in some manganites without magnetic rare-earth ions, such as LaMnO$_{3}$ nano particles. [@MAR08] Accordingly, the nature of the negative magnetization in manganites is one of crucial issues to be unveiled, which may be close to a phase segregated state.
Furthermore, we carry out the ac magnetic susceptibility measurements for CaMn$_{1-x}$Sb$_{x}$O$_{3}$, in order to examine the dynamic effect linked to magnetically frustrated properties. For spin glass or cluster glass system, a visible anomaly in the ac susceptibility appears upon lowering $T$ across freezing temperature of spins or clusters when the magnetic relaxation time becomes longer than the measuring time.[@MU81] In particular, we believe that a phase separated state realized in manganites is not consistent with a typical spin glass phase at low field.[@DE01] It thus is very intriguing to demonstrate a close relationship between static and dynamic responses of magnetization to the applied field, for our further understanding of complicated magnetic behavior of the present samples.
EXPERIMENT
==========
![(Color online) (a) X ray diffraction patterns of CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x$=0.02, 0.05, 0.08 and 0.1) taken at room temperature. Cross marks denote intensity peaks of orthorhombic CaMnO$_{3}$. (b) The lattice parameters and unit cell volume as a function of Sb contents from 0.02 up to 0.1. []{data-label="lattice"}](FigLattice.ps){width="10cm"}
Polycrystalline samples of CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x$=0.02, 0.05, 0.08 and 0.1) were prepared with a solid-state reaction method. The stoichiometric mixtures of CaCO$_{3}$, Mn$_{3}$O$_{4}$, and Sb$_{2}$O$_{3}$ high purity powders were calcined in air at 1000 $^{\circ}$C for 24 h. The products were then ground and pressed into cylindrical pellets. The pellets were finally sintered at $1400\sim 1450$ $^{\circ}$C for 12 h. X-ray diffraction data revealed that all samples are almost single phase with orthorhombic structures ($Pnma$)(Fig.\[lattice\]). The lattice parameters of the $x$=0.08 sample are $a=5.326$ Å, $b=7.512$ Å, and $c=5.310$ Å, which is in fair agreement with a previous work.[@PO04] The cell parameters and unit cell volume increase with increasing the Sb doping because the ion radius of Sb$^{5+}$ (0.61Å) is greater than the value of Mn$^{4+}$ (0.54Å). In addition, the Mn$^{4+}$ ions are replaced by Mn$^{3+}$ ions (0.645Å) with one extra electron, which contributes to the increased cell volume. The electrical resistivity was measured with a four-probe method. Seebeck coefficient was determined from both measurements of a thermoelectric voltage and temperature difference along the longitudinal direction of the measured sample. The thermal conductivity was collected with a conventional heat flow method. The dc and ac magnetization measurement was carried out using commercial superconducting quantum interference device (SQUID) magnetometers both at Iwate Univ. and National Institute for Materials Science. The ac magnetic susceptibility measurement for $x=0.02$, 0.05, and 0.08 samples was measured as a function of frequency and dc magnetic field at the ac magnetic field of 0.5 mT. In particular, to remove the influence of remanent magnetic field, the SQUID magnetometer with the option of magnet reset mode was used in low field measurements[@REM; @MU10]. Hydrostatic pressures in magnetization and electrical resistivity measurements were applied by using a clamp-type CuBe cell up to 1 GPa. Fluorinert was used as a pressure transmitting medium. The magnitude of pressure was calibrated by the pressure dependence of the critical temperature of lead. Magnetoresistance effect was measured by using a superconducting magnet at the High Field Laboratory for Superconducting Materials, Institute for Materials Research, Tohoku University.
Results and discussion
======================
![(Color online) (a) Temperature variation of electrical resistivity $\rho$ for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02, 0.05, 0.08, and 0.1). For comparison, the data of the parent compound are presented. (b) The $\rho-T$ plot magnified between 180 K and room temperature. The solid lines denote the fits of the high-$T$ data by using a small polaron model. The activation energy $W$ is listed as a function of the Sb content in Table \[lattice\]. (c) Magnetoresistance effect, $MR=[R(H)-R(0)]/R(0)\times100\%$ at 20 K for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02, 0.05, and 0.08). []{data-label="RT"}](FigRT.ps){width="10cm"}
![(Color online) (a) Temperature variation of electrical resistance $R$ for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02 and 0.08) at an applied pressure of 0.8 GPa. For comparison, the data at ambient pressure are presented. (b) $\Delta R/R =[R(T,0.8$GPa$)-R(T,0$GPa$)]/R(T,0$GPa) $\times100\%$ for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02, 0.05, and 0.08). The inset displays $\Delta R/R$ as a function of Sb content at 100 K and 300 K. []{data-label="RTP"}](FigRTP.ps){width="8cm"}
--------- --------------- ------- ------------ --------------------- ------- -- -- --
Content $\rho_{300K}$ $W$ $MR_{20K}$ $\Delta R/R_{100K}$ $T_m$
$x$ ($\Omega $cm) (meV) ($\%$) ($\%$) (K)
0 1.5 120
0.02 0.057 17.4 -60 -17 110
0.05 0.051 24.2 -85 -22 99
0.08 0.049 33.9 -64 -47 88
0.1 0.044 45.2 -43 38
--------- --------------- ------- ------------ --------------------- ------- -- -- --
: \[table1\] Electrical transport characteristics of the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ series. $\rho_{300K}$ represents the resistivity at 300 K. $W$ is calculated from the best fits of the high-$T$ transport data using the small polaron model. The $MR$ under an applied field of 7 T and $\Delta R/R$ under an applied pressure of 0.8 GPa are estimated from the effect of magnetic field and pressure on the resistivity, respectively.(In detail, see the captions of Fig. \[RT\] and Fig. \[RTP\]). Magnetic transition temperature $T_m$ is estimated from the inflection point of the field cooled $M-T$ curve of Fig. \[MT\] (a), (c), and (e).
Effect of magnetic field and external pressure on electrical transport
-----------------------------------------------------------------------
The temperature variation of electrical resistivity $\rho$ for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system is shown in Fig. \[RT\] as a function of the Sb content. The Sb substitution for Mn site up to $x=0.1$ gives rise to a substantial decrease in high temperature $\rho$(300 K) from 1.5 $\Omega$cm at the parent $x=0$ sample through 0.057 $\Omega$cm at the lightly doped $x=0.02$ sample down to 0.044 $\Omega$cm at the intermediately doped $x=0.1$ sample, indicating the carrier doping into Mn-site. All samples exhibit semiconducting behaviors upon decreasing temperatures except for a metallic conduction at high temperatures for the Sb light doping. The $\rho-T$ curve of the $x=0.02$ and 0.05 samples remains a metallic like character in the paramagnetic region down to $\sim$ 200 K and 250 K, respectively. On the other hand, the magnitude of the resistivity of the $x$=0.08 and 0.1 is enhanced at lower temperatures since the further doping of Sb ion breaks some of conduction paths along the Mn-O-Mn network, resulting in reinforcing carrier localization. In previous works,[@JA96; @MAI98] it is reported that the high temperature transport in doped manganites is described by a small polaron hopping model. Now, we try to fit the resistivity data at high temperatures by using such an expression as $\rho=AT$exp($W/kT$), where the fitting parameter $W$ represents the activation energy. As listed in Table \[table1\], with increasing the Sb concentration, the value of the activation energy shows a gradual increase from 17.4 meV at $x=0.02$ up to 45.2 meV at $x=0.1$ in spite of a monotonous suppression in the resistivity vs Sb content at high temperatures. In the preceding section of Thermal transport, we will discuss the transport mechanism in the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system. We notice that the resistivity data of parent CaMnO$_{3}$ observed here are, both in temperature dependence and magnitude, similar to the $\rho-T$ curve of the stoichiometric composition without oxygen defects, CaMnO$_{3-\delta }$ ($ \delta=0 $).[@ZE99]
Next, let us examine the negative magnetoresistance effect on the electron doped manganite samples with $x=0.02$, 0.05, and 0.08 as shown in Fig. \[RT\](c). At the $x$=0.05 sample, the gigantic magnetoresistance attaining $-85\%$ at 20 K under the applied field of 7 T is observed. For the other samples, we obtain the MR ratio of $\sim -60\%$. The MR data are comparable to those reported in a previous work[@RA00] on CaMn$_{1-x}$M$_{x}$O$_{3}$ (M=Nb,Ta) system. Theses findings strongly indicate that the spin polarized ferromagnetic metal clusters are established by the applied field and the field-induced delocalized state is realized within the samples studied.
Finally, let us show in Fig. \[RTP\] the effect of pressure on the electrical resistance $R$ as a function of temperature between 77 K and 300 K for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x$=0.02 and 0.08). For comparison, the data at ambient pressure are presented. The applied pressure of 0.8 GPa on the $x=0.08$ sample suppresses the magnitude of $\rho$ from 25$\%$ at 300 K down to about 50 $\%$ at lower temperatures around 80 K upon decreasing temperature. The effect of pressure on the electrical transport observed here seems to be more enhanced below near the magnetic transition temperatures $T_{m}$=110 K at $x=0.02$ and 88 K at $x=0.08$, where $T_{m}$ is determined from the inflection point of the magnetization data of Fig. \[MT\]. The pressure dependence of the resistance $\Delta R/R$ (the inset of Fig. \[RTP\]) exhibits a maximum at the heavily doped sample of $x=0.08$, which is consistent with the effect of strong pressure on the magnitude of magnetization below $T_{m}$.(see Figs. \[MT\](e) and \[MT\](f))
Thermal transport (Seebeck coefficient and thermal conductivity)
-----------------------------------------------------------------
![(Color online) (a) Temperature variation of Seebeck coefficient $S$ for the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02, 0.05, 0.08 and 0.1). The solid lines denote the $T$ linear fits to the $S(T)$ data in the temperature region above 130 K. The arrows denote the magnetic transition temperatures determined from the magnetization data. (b) Thermal conductivity $ \kappa $ of CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x$=0, 0.02, 0.05, 0.08, and 0.1). For comparison, the $\kappa $ data of the parent compound CaMnO$_{3}$ are presented. The inset also represents the $S$ data of parent CaMnO$_{3}$. []{data-label="SK"}](FigSK.ps){width="9cm"}
Seebeck coefficient $S$ is suppressed upon increasing Sb doping as shown in Fig. \[SK\](a). For all samples studied, $S$ shows a negative value over a wide range of temperatures, strongly suggesting electron doping into the parent matrix. In the case of the substitution of Sb$^{5+}$ ion for Mn$^{4+}$ site, the Mn$^{4+}$ ions are replaced by Mn$^{3+}$ ions with one extra electron. For lower doped samples ($x$=0.02 and 0.05), the local maximum observed in $S$ is located near the magnetic transition temperature as mentioned above. With further doping of Sb, the local maximum suppressed in $S$ of $x$=0.08 is still located around 130 K, but it has little relationship with the weakly magnetic transition near 90 K. We notice a common feature in the $T$-linear dependence of Seebeck coefficient for three samples ($x$=0.02, 0.05 and 0.08) between 130 K and room temperature, although their $T$ dependences of the resistivity are quite different over the same temperature range as shown in Fig. \[RT\](b). In our previous section, we demonstrated that the high-$T$ resistivity is well fitted by using a small polaron model. The corresponding expression for Seebeck coefficient is given in the form of $$S(T)=\frac{k_{B}}{e}\left\{\frac{E_{S}}{k_{B}T}\right\}+S_{\infty },$$ where $E_{S}$ and $S_{\infty }$ represent the thermal activation energy and the spin entropy in the high temperature limit, respectively.[@JA96] In Fig. \[SK\], we note that the $1/T$ law is violated at high temperatures. The electrical resistivity of the present samples is greatly influenced by grain boundaries in comparison to Seebeck measurement, giving rise to little reliable information about the electronic states.[@SE] In fact, such disagreements between these transports of polycrystalline oxide materials have been pointed out as far.[@FI00; @MAE87]
In previous works[@MAI98; @FI00] of electron doped manganites, the $T$-linear dependence of $S(T)$ has been discussed on the basis of the conventional metal model, or Culter and Mott (CM) model. The former is given in the form of $$S(T)=\pi^{2} \frac{k_{B}}{3e}\left\{k_{B}T\frac{N(E)}{n}+const.\right\}_{E=E_{F}},$$ where $N(E)$ is the density of states and $n$ is the carrier density.[@MAI98] The latter formula for the CM model is expressed as $$S(T)=-\pi^{2} \frac{k_{B}}{3e} \left\{ k_{B}T\frac{\partial \ln (\mu_{0} N(E))}{\partial E} -\frac{\partial W}{\partial E} \right\} _{E=E_{F}},$$ where $\mu =\mu_{0} $exp$(-W(E)/k_{B}T )$ is an activated mobility and $ W(E)$ is an activation energy.[@CU69] The CM model for the random hopping system well describes a random distribution of localized states of electrons around Fermi level. The typical fitting parameters $N(E)/n$ and $\{\partial $ln$(\mu_{0} N)/\partial E \}^{-1}$, for the conventional and CM models are listed in Table \[TableS\]. The fitted result shows that $N(E)/n$ ratio is almost independent of the nominal composition. The magnitudes of $\{\partial $ln$(\mu_{0} N)/\partial E \}^{-1}$ are similar to those of the A-site substituted manganite Ca$_{1-x}$La$_{x}$MnO$_{3}$ ($x$=0.017 and 0.033).[@FI00] For the insulating Li$_{1+x}$Ti$_{2-x}$O$_{4}$ sintered samples, it has been reported that $S(T)$ is proportional to temperature and the $S(T)$ behavior is analyzed using the CM model.[@MAE87] The Sb substitution induced lattice disorder influences the electronic states of the parent matrix, causing the validity of the CM model. Concerning the B-site substituted manganites, there is a common trend in the temperature dependence of $S(T)$ between CaMn$_{1-x}$W$_{x}$O$_{3}$($x=0.02$ and 0.04) and the present samples with $x=0.02$, 0.05, and 0.08.[@MIC07]
--------- ------------- ------------------ ---------------------------- ---------------------------------------------------------------------- -- -- -- --
Content $S_{300K}$ $\frac{N(E)}{n}$ $\frac{n(Mn)}{n(Mn^{3+})}$ $\left( \frac{\partial \ln (\mu_{0} N(E))}{\partial E} \right)^{-1}$
$x$ ($\mu $V/K) (1/eV) (eV)
0 -340
0.02 -107 8.2 50 0.12
0.05 -84 9.3 20 0.11
0.08 -63 8 12.5 0.13
0.1 -51
--------- ------------- ------------------ ---------------------------- ---------------------------------------------------------------------- -- -- -- --
: \[TableS\] Thermal transport characteristics of the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ series. $S_{300K}$ represents Seebeck coefficient at 300 K. $N(E)/n$ is calculated from the best fits of the high-$T$ data using the conventional model. $n(Mn)/n(Mn^{3+})$ is the ratio of the number of total Mn ions to that of Mn$^{3+}$ ion. We obtain from the inverse slopes of $S$,$\{\partial $ln$(\mu_{0} N)/\partial E \}^{-1}$, by using the CM model.
For $x$=0.1, the $S(T)$ data remain a nearly constant from room temperature down to near 100 K, then increase rapidly and finally exhibit a positive value across the horizontal axis at lower temperatures. Such an unusual behavior in $S(T)$ has been reported both in A-site substituted system Ca$_{1-x}$Sm$_{x}$MnO$_{3}$($x$=0.2)[@HE99] and B-site system CaMn$_{1-x}$W$_{x}$O$_{3}$($x\geq 0.07$)[@MIC07], indicating a holelike character of charge carriers.
Finally, let us show in Fig. \[SK\] the temperature variation of thermal conductivity in CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x$=0, 0.02, 0.05, 0.08, and 0.1). The thermal conduction for all samples studied here is carried by acoustic phonons because the electron contribution is negligible by an estimation from the electric resistivity data using the Wiedemann-Franz law. Upon decreasing temperature crossing the antiferromagnetic temperature T$_{N}$, a sharp increase in $ \kappa $ of parent CaMnO$_{3}$ is observed as reported in previous works.[@HE99; @CO02] We believe that the thermal conductivity in the paramagnetic phase is strongly suppressed by phonon scattering due to nanoscale strains generated by spin correlations.[@CO02] This finding is thus explained by a rapid reduction of the phonon scattering when the AFM long range order is established below T$_{N}$. The light doping of Sb$^{5+}$ strongly suppresses the magnitude of thermal conductivity from 120 mW/cmK (50K) at the pure x=0 sample down to 14 mW/cmK at $x$=0.02 (Fig.\[SK\]). The Sb$^{5+}$ doping removes the Mn$^{4+}$ ion and instead produces the Mn$^{3+}$ ion for Mn sites which is the Jahn-Teller active ion with one e$_{g}$-electron. It is believed that the local lattice distortion due to JT effect causes phonon scattering, which is close to the depressed thermal conduction.[@MA03] In addition, the lattice deformation due to the Sb doping with its larger ion radius affects the neighboring Mn$^{3+}$O$_{6}$ octahedron, resulting in some variation of the orbital-state of e$_{g}$-electron through the local JT effect.
Effect of pressure and magnetic field on dc magnetization
---------------------------------------------------------
![(Color online) Temperature dependence of zero field cooled and field cooled magnetization of the Sb substituted CaMn$_{1-x}$Sb$_{x}$O$_{3}$ recorded at the magnetic field of 10 mT. In the case of ambient pressure, (a) $x=0.02$, (c) $x=0.05$, and (e) $x=0.08$. In the case of 0.8 GPa, (b) $x=0.02$, (d) $x=0.05$, and (f) $x=0.08$. For comparison, the ZFC data of the parent sample are given in (a). In the inset of (e), the $M-T$ curve with the $x=0.1$ sample is presented at 0 GPa. []{data-label="MT"}](FigMTL.ps){width="9cm"}
![(Color online) ZFC and FC magnetization curves of the Sb substituted CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x=0.02$, 0.05, and 0.08) collected at 50 mT. For ambient pressure, we get (a) $x=0.02$, (c) $x=0.05$, and (e) $x=0.08$. For applied pressure of 0.8 GPa, (b) $x=0.02$, (d) $x=0.05$, and (f) $x=0.08$. []{data-label="MT50"}](FigMTH.ps){width="9cm"}
Let us show in Figs. \[MT\](a), \[MT\](c), and \[MT\](e) the zero-field cooled (ZFC) and field cooled (FC) magnetization data of the Sb substituted CaMn$_{1-x}$Sb$_{x}$O$_{3}$ measured under an applied magnetic field of 10 mT at 0 GPa. First of all, temperature dependent magnetization reversal appears in the FC curves of both $x=0.05$ and 0.08, as it has been reported in the V doped CaMn$_{1-x}$V$_{x}$O$_{3}$ manganites.[@ANG06] However, lowering temperature down to 2 K, the ZFC data of both samples show no negative value. Upon increasing the Sb content up to $x=0.1$, a large difference between ZFC and FC curves is observed but such negative magnetization phenomenon vanishes. The magnetic transition temperature is suppressed from 110 K at $x=0.02$, through 99 K at $x=0.05$, down to 38 K at$x=0.1$ due to the Sb doping because the Sb$^{5+}$ ion is non magnetic one with the closed shell of $4d^{10}$. The substitution of non magnetic Sb ions for Mn sites introduces electron doping, forming ferromagnetic clusters, but destroys magnetic ordering between Mn ions within the parent matrix.
The FC data of $x=0.08$ indicates that the antiparallel component of the canted AFM spins to the applied field is stabilized by the Sb substitution. The spontaneously canted magnetic moment contributing to the negative magnetization is close to a variation of the local easy axis which is caused by the local lattice distortion of Sb substitution in comparison to the Sb free sample. If the number of canted spin clusters contributing to the antiparallel component is dominant over that of these clusters contributing to the parallel one along the applied field, we then obtain the negative magnetization. Furthermore, we need to point out the significant roles of the Dzyaloshinsky-Moriya interaction which prefers canted spin arrangements. If the antisymmetric exchange interaction between nearest neighbor Mn ions is dominant, such mechanism then stabilizes non collinear spin configuration, leading to complicated magnetic behaviors. In order to account for the negative magnetization observed in the lightly doped Nd$_{1-x}$Ca$_{x}$MnO$_{3}$ series, the low-temperature spin reorientation is explained in the framework of a two-phase model, where the samples consist of exchange coupled ferromagnetic and weak ferromagnetic phases.[@TRO03] The reorientation of magnetic moment of Nd ions in both phases plays a crucial role on the magnetization reversal through the f-d exchange interaction between the Nd and manganese sublattices. It seems that the proposal model is ruled out because the magnetic rare earth ion is absent in the present system. Next, we attempt to measure the effect of pressure on the magnetization for CaMn$_{1-x}$Sb$_{x}$O$_{3}$, to examine a relation between lattice distortion and spin arrangement (Figs. \[MT\](b), \[MT\](d), and \[MT\](f)). The application of hydrostatic pressure up to 0.8 GPa enhances $T_{m}$ by 8$\sim $9 K at $x=0.02$ and 0.05, and a stable increase in $T_{m}$ reaches about 5 K at the $x=0.08$ sample. We expect that a shrinkage of lattice parameters makes stronger a super-exchange interaction between nearest neighbors $t_{2g}$ spins of Mn$^{4+}$ ions, resulting in stabilization of the G-type AFM magnetic structure. The magnitude of magnetization is not largely changed at $x=0.02$ and 0.05, but the FC magnetization of the $x=0.08$ sample exhibits a rapid rise in $M_{max}$ from 0.18 emu/g at 0 GPa up to 0.36 emu/g at 0.8 GPa, which is responsible for a sharp increase of FM clusters induced by pressure with a stable rise of $T_{m}$. Both the ZFC and FC curves at $x=0.02$ under 0.8 GPa are similar to those under ambient pressure but the magnitude of magnetization is suppressed by the application of pressure. The application of external pressure on the $x=0.05$ sample changes the temperature induced magnetization reversal observed at 0 GPa and results in positive FC curve, approaching the ambient magnetization curves of the low doped sample ($x=0.02$). At further Sb content of $x=0.08$, the negative magnetization of the FC curve at 0.8 GPa are limited at lower temperatures below $\sim $30 K. However, it seems that the ZFC curves of all samples remain qualitatively unchanged even under the applied pressure. The orthorhombic distortion due to Sb doping is suppressed by application of pressure, resulting in a similar magnetic behavior to the lower doped case under ambient pressure. At the applied field of 50 mT, the negative magnetization disappears except for the low temperature region of $x=0.08$ at ambient pressure as shown in Fig. \[MT50\].
Here, we make remarks about a notable difference between the magnetic properties of the $x=0.1$ sample and other ones ($x=0.02$, 0.05, and 0.08). As mentioned above, the magnetic transition temperature $T_m$ is monotonously decreased from 110 K at $x=0.02$ through 99 K at $x=0.05$ to 88 K at $x=0.08$ upon further Sb doping. At a maximum content of $x=0.1$, $T_m$ is strongly depressed down to 40 K. These findings indicate that the magnetic interaction working between manganese ions is considerably suppressed beyond $x=0.08$. For the $x=0.1$ sample, the nominal content of $e_g$ electron is taken as a maximum in the all samples studied. In fact, the room temperature resistivity and corresponding Seebeck coefficient of the $x=0.1$ sample exhibit the lowest values. However, upon lowering temperatures, the suppression of magnetic coupling working among the Mn ions around each Sb ion probably becomes dominant over the carrier doping effect and causes a strongly reduced magnetization accompanied by a remarkable decrease of $T_m$. In addition to it, we point out the giant pressure effect on the magnitude of magnetization of the $x=0.08$ sample in applied fields of 10 mT and 50 mT as depicted in Fig. \[MT\](d) and Fig. \[MT50\](d). In particular, the weaken magnetic interaction due to Sb doping under ambient pressure is strengthened under applied pressure of 0.8 GPa, giving the remarkable increase in the magnetization by a factor of about 2. The critical content is located near $x=0.08$, separating negative and normal magnetic tendencies.
------ ------- ----------- ----------- --------------- ----------- ---------------- ---------------- --
$T_m$ $T_m^{p}$ $M_{max}$ $M_{max}^{p}$ $\Theta $ $\mu_{exp}$ $\mu_{cal}$
$x$ (K) (K) (emu/g) (emu/g) (K) ($\mu_{B}$/Mn) ($\mu_{B}$/Mn)
0 120 3.87
0.02 110 119 0.95 0.71 -119 3.51 3.85
0.05 99 107 1.23 1.70 -1 3.41 3.83
0.08 88 93 0.18 0.36 28 3.70 3.81
0.1 38 3.80
------ ------- ----------- ----------- --------------- ----------- ---------------- ---------------- --
: \[table3\] Effect of pressure on magnetic transition temperature $T_m$ of the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02, 0.05, and 0.08 ). $T_m$ and $T_m^{p}$ are determined from the inflection of $M-T$ curves of Fig.\[MT\] under 0 GPa and 0.8 GPa, respectively. The $M_{max}$ and $M_{max}^{p}$ denote a maximum in FC magnetization curves below the magnetic transition temperatures under 0 GPa and 0.8 GPa, respectively. $\Theta $ and $\mu_{exp}$ represent the Curie-Weiss temperature and the effective magnetic moment per one manganese ion estimated from a high temperature linear fit of $1/M-T$ under ambient pressure. For comparison, the effective magnetic moment $\mu_{cal}$is given. (In detail,see the text)
![(Color online) (a) Temperature dependence of ZFC and FC magnetization curves of the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ ($x=0.02$ and $0.05$) samples under an applied field of 1.5 T. (b) Magnetization curves $M$ vs $H$ of $x=0.02$, 0.05, and 0.08 at 5 K. []{data-label="MH"}](FigMH.ps){width="8cm"}
Let us determine from the $1/M-T$ data taken at 0.1 T (not shown here), the Curie-Weiss temperature and the effective magnetic moment, for CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system ($x$=0.02, 0.05, and 0.08 ) as listed in Table \[table3\]. The Curie-Weiss temperature $\Theta $ is estimated from a high temperature linear fit of $1/M-T$ by using the relation $ \chi=C/(T-\Theta)$. The Curie constant $C$ gives the effective magnetic moment $\mu_{exp}$ per one manganese ion. For comparison, the effective magnetic moment $\mu_{cal}$ is calculated as a function of nominal Sb content from $\mu_{eff}^2=(1-2x)\mu^2$(Mn$^{4+}$)+$x\mu^2$(Mn$^{3+}$). For free ions in the high spin configuration, we have, $\mu $(Mn$^{4+}$) and $\mu $(Mn$^{3+}$) , 3.87$\mu_{B}$ and 4.90$\mu_{B}$, respectively. Here, we assume that $\mu =2\sqrt{S(S+1)}$ where the spin quantum number $S$ = 3/2 for Mn$^{4+}$ and $S$ = 2 for Mn$^{3+}$. In the paramagnetic regime, we obtain $\Theta $ from the negative value of -119 K at $x$=0.02 to the positive ones of 28 K at $x$=0.08, which is indicative of a qualitative variation in the magnetic interaction from antiferromagnetic to ferromagnetic character. Furthermore, the effective magnetic moment shows a slight increase from 3.51 $\mu_{B}$ at $x$=0.02 up to 3.70 $\mu_{B}$ at $x$=0.08. This finding seems to point to, the existence of ferromagnetic fluctuation or the formation of FM small clusters, in the paramagnetic phase of the Sb doped sample above the magnetic transition temperature. However, the resultant magnetic parameters are not always consistent with the low temperature magnetization suppressed at $x$=0.08. The suppression of magnetic interaction due to Sb doping as discussed above and its related magnetic frustration probably prevent establishment of a large FM cluster or a long range FM order at low temperatures.
In a relatively high field of 1.5 T, the temperature variation of magnetization for $x=0.02$ and 0.05 shows no magnetic hysteresis between ZFC and ZC curves as displayed in Fig. \[MH\](a). The $M$($H$) data with a small hysteresis do not saturate even at 5 T and rises linearly with increasing $H$ as shown in Fig. \[MH\](b), indicating the AFM matrix with small FM phase. We understand from both Figs. \[RT\](c) and \[MH\](b) that the larger magnetization at high fields corresponds the stronger magnetoresistance effect. A linear extrapolation of $M$($H$) curve at 5 K to $H=0$ gives a spontaneous magnetization $M_{s}$ from 0.97 emu/g at $x=0.02$ through 1.2 emu/g at $x=0.05$ down 0.42 emu/g at $x=0.08$, revealing the existence of small residual magnetization for all samples studied.
Finally, we suppose the valence fluctuation of Sb ion, to account for discrepancies in the effective moment between the experimental and calculated values as listed in Table \[table3\]. In a recent study of CaMn$_{1-x}$Ru$_{x}$O$_{3}$[@ZHO09], X-ray absorption measurements reveal the presence of the mixed valence states of Ru ion. In the intermediate samples (0.1$\leq x\leq $0.9), it has been reported that the valence states of Ru$^{5+}$ and Ru$^{4+}$ coexist with the mixed valence of Mn ion. However, X-ray photoelectron spectroscopy analysis of La$_{0.9}$Sb$_{0.1}$MnO$_{3}$[@DU04] indicates that the valence state of Sb ion is +5. Thus, we believe that Sb$^{5+}$ exists predominantly in the low doping region of Sb.
Frequency and dc magnetic field dependences of ac magnetic susceptibility
-------------------------------------------------------------------------
![(Color online) The real part $\chi'$ of ac magnetic susceptibility of CaMn$_{1-x}$Sb$_{x}$O$_{3}$ collected at zero dc magnetic field with frequency ranging 1 Hz up to 1 kHz. (a) $x=0.02$, (b) $x=0.05$, and (c)$x=0.08$. The amplitude of the ac magnetic field $H_{ac}$ was set to be 0.5 mT. The arrows point to the direction of increasing frequencies. []{data-label="ACRe"}](FigacRe.ps){width="8cm"}
![(Color online) The imaginary part $\chi''$ of ac magnetic susceptibility of CaMn$_{1-x}$Sb$_{x}$O$_{3}$ as a function of frequency from 1 Hz to 1 kHz at $H_{dc}$ = 0 T.($H_{ac}$ = 0.5 mT) (a) $x=0.02$, (b) $x=0.05$, and (c)$x=0.08$. []{data-label="ACIm"}](FigacIm.ps){width="8cm"}
![(Color online) The real and imaginary parts of ac magnetic susceptibility of CaMn$_{1-x}$Sb$_{x}$O$_{3}$ measured at 10 Hz under a superimposed dc field ($H_{dc}$=10 mT). For $x=0.02$, (a) $\chi'$ and (b)$\chi''$. For $x=0.05$, (c) $\chi'$ and (d)$\chi''$. For $x=0.08$, (e) $\chi'$ and (f)$\chi''$. The ac magnetization data are recorded as a function of temperature under ZFC and FC conditions. []{data-label="ACFC"}](FigReImFC.ps){width="9cm"}
![(Color online) The real and imaginary parts of ac magnetic susceptibility of CaMn$_{0.9}$Sb$_{0.1}$O$_{3}$ a function of frequency from 1 Hz to 1 kHz at $H_{dc}$ = 0 T.($H_{ac}$ = 0.5 mT). For $x=0.1$,(a) $\chi'$ and (b)$\chi''$. For comparison, the ac magnetization data recorded as a function of temperature under ZFC and FC conditions are given in (c) ($H_{dc}$=10 mT). []{data-label="AC0.1"}](FigacH.ps){width="8cm"}
Furthermore, we carry out the ac magnetic susceptibility measurements for $x=0.02$, 0.05, and 0.08 samples, in order to examine the dynamic effect linked to magnetically frustrated properties (Figs. \[ACRe\] and \[ACIm\]). The real and imaginary parts of the ac susceptibility, $\chi'$ and $\chi''$, are registered at zero dc magnetic field with frequency $f$ ranging 1 Hz up to 1 kHz. For $x=0.02$, a sharp transition is noticed around $T_{m}=110$ K, which well agrees with the dc magnetic measurements. A maximum peak of $\chi'$ curves shows no shift in temperature with increasing $f$ but a second peak at lower temperature exhibits a small shift towards high temperatures, indicating the presence of a magnetic frustration. The magnitude of $\chi'$(T) curves is suppressed with frequency over a wide range of temperatures. In addition to the two peaks observed in $\chi'$, a third peak in $\chi''$ appears at low frequencies, its peak then shifts towards higher temperatures with increasing $f$, and finally it is merged in a maximum peak at high frequencies. The new peak around 80 K at 1 Hz strongly depends on frequency. The imaginary component representing magnetic energy dissipation is close to the low field dc magnetic hysteresis between ZFC and FC. For $x=0.05$, the maximum peak is located at $T_{m}=100$ K, while the second one observed at higher temperature of 90 K is accompanied by a slight frequency shift towards low temperatures. It seems that the frequency shift of the second peak in $\chi'$ and $\chi''$ curves for both $x=0.02$ and 0.05 is indicative of the signature of a spin-glass like magnetic property. However, a substantial decrease of peak in $\chi''$ with increasing frequency is qualitatively in contrast to the behavior of conventional spin glass system,[@MU81] as previously reported in phase separated Pr$_{0.7}$Ca$_{0.3}$MnO$_{3}$.[@DE01] The dynamical effect of $\chi'$ has some relationship with the formation of the magnetically frustrated clusters such as FM clusters embedded in canted AFM matrix. On the other hand, in the case of $x=0.08$, the second peak in $\chi'$ and $\chi''$ curves is located near 55K and it shows a weak frequency dependence as shown in Figs.\[ACRe\] and \[ACIm\]. We notice that the second peaks of $x=0.02$ and 0.05 are located around 47 K and 90 K, respectively, corresponding to the maximum of the ZFC dc magnetization. However, the second peak of $x=0.08$ has no corresponding maximum in the ZFC dc curve (Fig. \[MT\](c)). We note that the ac susceptibility of the parent CaMnO$_{3}$ exhibits no magnetic peak except for its maximum peak associated with long range magnetic ordering.[@MAI98] Moreover, for the B site substituted CaMn$_{1-x}$Mo$_{x}$O$_{3}$($x=0.04$), the imaginary part $\chi''$ shows only a rapid peak at 105 K around the magnetic transition temperature and no signal at low temperatures.[@MAR09] For the low doped CaMn$_{1-x}$Ru$_{x}$O$_{3}$($x=0.06$), both $\chi'$ and $\chi''$ exhibit neither frequency dependence nor enhanced peak below $T_{m}$.[@SHA04] This finding predicts that the low temperature ground state of the light Ru substituted CaMnO$_{3}$ is better explained by the complex canted magnetic structure than by the phase separated model. For A site substituted Ca$_{0.9}$La$_{0.1}$MnO$_{3}$, a much larger hysteresis in the dc magnetization between ZFC and FC curves, and the frequency effect on the ac susceptibility demonstrate that its low temperature state is considered as a cluster glass one without long range ferromagnetism.[@MAI98]
Now, under zero field cooled and field cooled conditions, we display in Fig. \[ACFC\] the ac susceptibility measurements of $x=0.02$, 0.05, and 0.08 with a superimposed dc field ($H_{dc}$=10 mT). First of all, the steep peaks in both $\chi'$ and $\chi''$ around $T_{m}$ in the absence of dc field are strongly suppressed under the application of dc field but the second peaks of $x=0.02$ and 0.05 still remain stable. The collapse of a maximum peak at zero dc field is caused by the application of low dc field since the applied low field strongly suppresses the magnetic fluctuation associated with the magnetic transition.[@MUK96] Next, a magnetic divergence in the ac magnetization between ZFC and FC curves is visible at $x=0.02$, but upon increasing the Sb content, both ZFC and FC ac curves at $x=0.08$ become reversible and exhibit no clear differences except for the high temperature region near $T_{m}$. Finally, for the $x=0.08$ sample the maximum peak in FC $\chi''$ of Fig. \[ACFC\](f) is observed around 85 K, at the temperature where the FC dc curve reaches a maximum as shown in Fig. \[MT\](e). If we assume that the maximum peak in FC $\chi''$ is related to the formation of magnetically frustrated clusters associated with a phase separated state, we can not identify the second peak with corresponding signature in dc magnetization curve. However, the anomalous peak is located near the characteristic temperature pointing to the dc magnetization reversal. Accordingly, we suppose that the existence of the anomalous peak in $\chi''$ of $x=0.08$ has some relationship with the negative magnetization phenomena observed here. For comparison, the ac susceptibility data of the $x=0.1$ sample are presented in Fig. \[AC0.1\]. We notice frequency dependence of a maximum peak in both $\chi'$ and $\chi''$ around 40 K corresponding to the ZFC dc peak in the inset of Fig.\[MT\](e), indicating the strong evidence for magnetically frustrated state.
SUMMARY
=======
We have demonstrated the effect of hydrostatic pressure on magnetic and transport properties, in electron-doped manganites CaMn$_{1-x}$Sb$_{x}$O$_{3}$. In addition, thermal transport properties (Seebeck coefficient and thermal conductivity) of the CaMn$_{1-x}$Sb$_{x}$O$_{3}$ system have been examined as a function of $T$. Furthermore, the ac magnetic susceptibility measurements for $x=0.02$, 0.05, and 0.08 samples have been performed, in order to examine a close relation between the dynamic effect linked to magnetically frustrated properties and the static dc magnetization.
The substitution of Sb$^{5+}$ ion for Mn $^{4+}$site of the parent matrix causes one-electron doping with the chemical formula CaMn$^{4+}_{1-2x}$Mn$^{3+}_{x}$Sb$^{5+}_{x}$O$_{3}$ accompanied by a monotonous increase in unit cell volume as a function of $x$.
Upon increasing the doping level of Sb, the magnitudes of both electrical resistivity and Seebeck coefficient are suppressed at high temperatures, indicating the electron doping. The CM model applied to the random hopping system gives a better fit to Seebeck coefficient at higher temperatures. The light doping of Sb$^{5+}$ strongly suppresses the high thermal conductivity of the parent sample through the local lattice distortions.
The anomalously diamagnetic behaviors at $x=0.05$ and 0.08 are clearly observed in the field cooled spontaneous magnetization. The magnetization curves under the applied pressure of 0.8 GPa at $x=0.05$ and 0.08 exhibit similar behaviors to those of the $x=0.02$ and 0.05 samples without pressure, respectively. We expect that these findings are close to some change of the local easy axis of magnetization due to the local lattice distortion induced by the Sb doping, in comparison to the case of the A site substitution. A notable difference in the magnetization curves between the $x=0.1$ sample and other ones is attributed to the weaken magnetic interaction working among the Mn ions around Sb ion. The dynamical effect of ac magnetic susceptibility measurement has some relationship with formation of the magnetically frustrated clusters such as FM clusters embedded in canted AFM matrix. In particular, we suppose that the existence of the anomalous peak in FC $\chi''$ of $x=0.08$ is related to the negative magnetization observed here.
This work was partially supported by a Grant-in-Aid for Scientific Research from Japan Society of the Promotion of Science.
[30]{} *Colossal Magnetoresistive Oxides*, edited by Y. Tokura (Gordon and Breach, New York, 2000). C. Zener, Phys. Rev. 82, 403 (1951); P. G. deGennes, $ibid$. 118, 141 (1960). *Nanoscale Phase Separation and Colossal Magnetoresistance*, by E. Dagotto (Springer, 2003). A. J. Millis, P. B. Littlewood, and B. I. Shraiman, Phys. Rev. Lett. 74, 5144 (1995); A. J. Millis, B. I. Shraiman, and R. Mueller, $ibid$. 77, 175 (1996). J. B. MacChesney, H. J. Williams, J. F. Potter, and R. C. Sherwood, Phys. Rev. 164, 779 (1967). H. Chiba, M. Kikuchi, K. Kasuba, Y. Muraoka, and Y. Syono, Solid State Commun. 99, 499 (1996). E. Granado, C. D. Ling, J. J. Neumeier, J. W. Lynn, and D. N. Argyriou, Phys. Rev. B68, 134440 (2003). B. Raveau, Y. M. Zhao, C. Martin, M. Hervieu, and A. Maignan, J. Solid State Chem. 149, 203 (2000). L. Pi. S. Hebert, C. Martin, A. Maignan, and B. Raveau, Phys. Rev. B67, 024430 (2003). A. I. Shames, E. Rozenberg, C. Martin, A. Maignan, B. Raveau, G. Andre, and G. Gorodetsky, Phys. Rev. B70, 134433 (2004). Y. Guo, W. Li, S. Roy, and N. Ali, Chem. Mater. 17, 2735 (2005). R. Ang, Y. P. Sun, Y. Q. Ma, B. C. Zhao, X. B. Zhu and W. H. Song, J. Appl. Phys. 100, 063902 (2006). Q. Zhou, B. J. Kennedy, Z. Zhang, L. Y. Jang, and J. B. Aitken, Chem. Mater. 21, 4203 (2009). C. Martin, A. Maignan, M. Hervieu, and B. Raveau, Phys. Rev. B63, 100406 (2001). A. Maignan, C. Martin, C. Autret, M. Hervieu, B. Raveau, and J. Hejtmanek, J. Mater. Chem. 12, 1806 (2002). M. Miclau, J. Hejtmanek, R. Retoux, K. Knizek, Z. Zirak, R. Fresard, A. Maignan, S. Hebert, M. Hervieu, and C. Martin, Chem. Mater. 19, 4243 (2007). E. Rozenberg, M. Auslender, A. I. Shames, C. Martin, and S. Hebert, J. Appl. Phys. 103, 07F720 (2008). C. Martin, M. Miclau, S. Hebert, M. Giot, A. Maignan, G. Andre, and F. Bouree-Vigneron, J. Magn. Magn. Mater. 321, 3938 (2009). H. Kuwahara, Y. Moritomo, Y. Tomioka, A. Asamitsu, R. Kumai, and Y. Tokura, Phys. Rev. B56, 9386 (1997). V. Markovich, I. Fita, R. Puzniak, E. Rozenberg, C. Martin, A. Wisniewski, A. Maignan, B. Raveau, Y. Yuzhelevskii and G. Gorodetsky, Phys. Rev. B70, 024403 (2004): V. Markovich, I. Fita, R. Puzniak, E. Rozenberg, C. Martin, A. Wisniewski, Y. Yuzhelevskii and G. Gorodetsky, Phys. Rev. B71, 134427 (2005). Y. Murano, M. Matsukawa, S. Kobayashi, S. Nimori, R. Suryanarayanan,Journal of Physics. Conf. Series 200, 012114 (2010). F. Bartolome, J. Herrero-Albillos, L. M. Garcia, J. Bartolome, N. Jaouen, and A. Rogalev, J. Appl. Phys. 97, 10A503 (2005). I. O. Troyanchuk, V. A. Khomchenko, S. N. Pastushonok. O. A. Novitsky, V. I. Pavlov, and H. Szymczak, J. Magn. Magn. Mater. 303, 111 (2006). J. Hemberger, S. Lobina, H. A. Krug von Nidda, N. Tristan, V. Yu. Ivanov, A. A. Mukhin, A. M. Balbashov, and A. Loidl, Phys. Rev. B70, 024414 (2004). I. O. Troyanchuk, V. A. Khomchenko, G. M. Chobot, A. I. Kurbakov, A. N. Vasilev, V. V. Eremenko, V. A. Sirenko, M. Yu. Shvedun, H. Szymczak, and R. Szymczak, J. Phys. Condens. Mater. 15, 8865 (2003). O. Pena, M. Bahout, K. Ghanimi, P. Duran, D. Gutierrez, and C. Moure, J. Mater. Chem. 12, 2480 (2002) : O. Pena, C. Moure, P. Barahona, M. Baibich, and G. Martinez, Physica B 384, 57 (2006). C. A. Nordman, V. S. Achutharaman, V. A. Vasko, P. A. Kraus, A. R. Ruosi, A. M. Kadin, and A. M. Goldman, Phys. Rev. B54, 9023 (1996). M. Mouallem-Bahout, O. Pena, D. Gutierrez, P. Duran, and C. Moure, Solid State Commun. 122, 561 (2002). V. Markovich, I. Fita, A. Wisniewski, R. Puzniak, D. Mogilyansky, L. Titelman, L. Vradman, M. Herskowitz, and G. Gorodetsky, Phys. Rev. B77, 014423 (2008) C. A. M. Mulder, A. J. van Duyneveldt, and J. A. Mydosh, Phys. Rev. B23, 1384 (1981). I. G. Deac, J. F. Mitchell, and P. Schiffer, Phys. Rev. B63, 172408 (2001).
V. Poltavets, K. Vidyasagar and M.Jansen, J. Solid State Chem. 177, 1285 (2004). In our previous work[@MU10], we presented the low field magnetization measurements by using the SQUID magnetometer. Recently, we noticed that the remanent magnetic field in the superconducting magnet affects the temperature variation of the magnetization in the zero field cooled runs. In fact, we checked it by using the SQUID magnetometer with the function of magnet reset mode, to reduce the remanent field below about 2G. Separately, we tried to perform the magnetization measurements with the SQUID by using the fluxgate device for further reduction of the remanent value below 1 mG. The former result was in good agreement with the latter. As a result, we expect that the apparent $''$ZFC$''$ scan in the previous measurements is a field cooled one under the negative remanent field in the case of switching off the magnet reset.
M. Jaime, M. B. Salamon, M. Rubinstein, R. E. Treece, J. S. Horwitz, and D. B. Chrisey, Phys. Rev. B54, 11914 (1996). A. Maignan, C. Martin, F. Damay, B. Raveau, and J. Hejtmanek, Phys. Rev. B58, 2758 (1998). Z. Zeng, M. Greenblatt, and M. Croft, Phys. Rev. B59, 8784 (1999). Seebeck coefficient $S$ is a sensitive probe to examine the electronic state of disordered crystals or highly inhomogeneous materials. $S$ is less affected by grain boundaries of polycrystalline samples than the electrical conductivity measurement since the thermal current applied to the sample is hard to be disturbed by the grain boundaries in comparison to the electric current. B. Fisher, L. Patlagan, G. M. Reisner, and A. Knizhnik, Phys. Rev. B61, 470 (2000). I. Maekawa, F. Takagi, Y. Sakai, and N. Tsuda, J. Phys. Soc. Jpn. 56, 2119 (1987). M. Culter and N. F. Mott, Phys.Rev.181,1336(1969). J. Hejtmanek, Z. Jirak, M. Marysko, C. Martin, A. Maignan, M. Hervieu, and B. Raveau, Phys. Rev. B60, 14057 (1999).
J. L. Cohn and J. J. Neumeier, Phys. Rev. B66, 100404(R) (2002). M. Matsukawa, M. Narita, T. Nishimura, M. Yoshizawa, M. Apostu, R. Suryanarayanan, A. Revcolevschi, K. Itoh, and N. Kobayashi, Phys. Rev. B67, 104433 (2003). P. Duan, S. Y. Dai, G. T. Tan, H. B. Lu, Y. L. Zhou, B. L. Cheng, and Z. H. Chen, J. Appl. Phys. 95, 5666 (2004). S. Mukherjee, R. Ranganathan, P. S. Anikumar, and P. A. Joy, Phys. Rev. B54, 9267 (1996).
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: 'Let ${{\mathcal}{M} = (M_i \colon i<\lambda)}$ be a family consisting of finitary and cofinitary matroids on a common ground set $E$, where $\lambda$ is a finite or infinite cardinal. We prove the following Cantor-Bernstein-type result: if $E$ can be covered by sets ${(B_i \colon i<\lambda)}$ which are bases in the corresponding matroids and there are also pairwise disjoint bases of the matroids $M_i$ then $E$ can be partitioned into bases with respect to ${\mathcal}{M}$. We also show that the failure of this Cantor-Bernstein-type statement for arbitrary matroid families is consistent relative to ZFC.'
address:
- 'Joshua Erde, University of Hamburg, Department of Mathematics, Bundesstra[ß]{}e 55 (Geomatikum), 20146 Hamburg, Germany'
- 'Pascal Gollin, Discrete Mathematics Group, Institute for Basic Science (IBS), 55, Expo-ro, Yuseong-gu, Daejeon, Republic of Korea, 34126'
- 'Atilla Joó, University of Hamburg, Department of Mathematics, Bundesstra[ß]{}e 55 (Geomatikum), 20146 Hamburg, Germany'
- 'Paul Knappe, University of Hamburg, Department of Mathematics, Bundesstra[ß]{}e 55 (Geomatikum), 20146 Hamburg, Germany'
- 'Max Pitz, University of Hamburg, Department of Mathematics, Bundesstra[ß]{}e 55 (Geomatikum), 20146 Hamburg, Germany'
author:
- Joshua Erde
- Pascal Gollin
- Atilla Joó
- Paul Knappe
- Max Pitz
title: 'Base partition for finitary-cofinatary matroid families'
---
Introduction {#sec:introduction}
============
The main result
---------------
Our starting point and main motivation was the following problem in infinite graph theory. Let ${G = (V,E)}$ be a connected graph. Given a cardinal $\lambda$, a family $(T_i \colon i < \lambda)$ of spanning trees of $G$ is called
- a *$\lambda$-covering* of $G$ if the union of the $E(T_i)$ is $E$,
- a *$\lambda$-packing* of $G$ if the $T_i$ are pairwise edge-disjoint, and
- a *$\lambda$-partitioning* of $G$ if $(E(T_i) \colon i < \lambda)$ is a partition of $E$.
Do the existence of a $\lambda$-covering and a $\lambda$-packing imply the existence of a $\lambda$-partitioning? Recent, we have proved the positive answer for the case when $\lambda$ is infinite [@EGJKP19]:
\[thm:intro-graphs-infinite\] Let $\lambda$ be an infinite cardinal. Then a graph admits a $\lambda$-partitioing if and only if it admits both a $\lambda$-packing and a $\lambda$-covering.
The point of departure for this paper is the natural question whether the corresponding result also holds for finite $\lambda$ (and infinite $G$). A special case of the main result of this article tells that the answer is yes. The proof comes most naturally in the language of infinite matroids. Indeed, recall that for every connected graph ${G=(V,E)}$, finite or infinite, there is a matroid ${M(G)}$ with ground set $E$ whose circuits are the subsets of $E$ given finite cycles of $G$, and whose bases are precisely the subsets of $E$ given by spanning trees of $G$. Every matroid of the form ${M(G)}$ is *finitary*: it has the property that all of its circuits are finite. Recently, there has been a renewed interest in the theory of infinite matroids, after Bruhn, Diestel, Kriesell, Pendavingh and Wollan [@BDKW13] gave a set of cryptomorphic axioms for infinite matroids which encompasses duality, generalising the usual independent set-, bases-, circuit-, closure- and rank-axioms for finite matroids. All results about infinite matroids needed for this paper will be given in Section \[s:matroids\]. Our earlier notions about packings, coverings and decompositions of graphs into spanning trees then take the following matroidal form. Let $\lambda$ be a cardinal (finite or infinite) and ${{\mathcal}{M} = (M_i \colon i < \lambda)}$ be a family of matroids on the same ground set $E$. A family ${(B_i \colon i < \lambda)}$ where each $B_i$ is a base of $M_i$ is called
- a *covering* if $\bigcup_{i < \lambda} B_i = E$,
- a *packing* if the $B_i$ are pairwise disjoint,
- a *partitioning* if $\{B_i \colon i < \lambda\}$ is a partition of $E$.
Before stating our main result let us remind that a matroid is called *cofinitary* if its dual is finitary.
\[thm:main-thm-intro\] Let ${\mathcal}{M}$ be a family of matroids on a common ground set $ E $ consisting of finitary and cofinatory matroids. Then ${\mathcal}{M}$ admits a partitioning if and only if it admits a covering and a packing.
By considering the special case of Theorem \[thm:main-thm-intro\] where each matroid is $M(G)$ for the same connected graph $G$, we obtain the promised generalisation of Theorem \[thm:intro-graphs-infinite\] where $\lambda$ is allowed to be finite.
To introduce our relative consisteny result, we recall that a matroid is called *uniform* if whenever $I$ is independent with ${e \in I}$ and ${f \in E \setminus I}$, then ${I-e+f}$ is independent. We also remind that the *reaping number* $\mathfrak{r}$ (also called the *refinement number*) is the least cardinal such that there exists a family ${(A_i \colon i<\mathfrak{r})}$ of infinite subsets of $\mathbb{N}$ for which there is no bipartition of $\mathbb{N}$ that splits each $A_i$ into two infinite pieces, see [@vaughan]. Clearly ${\aleph_0<\mathfrak{r} \leq 2^{\aleph_0}}$, and hence under the Continuum Hypothesis we have ${\mathfrak{r} = 2^{\aleph_0}}$. However, the same conclusion also holds for example under Martin’s Axiom. On the other hand, also ${\mathfrak{r} < 2^{\aleph_0}}$ is consistently true, see [@van1984integers; @vaughan].
\[thm:unprovable-intro\] If ${\mathfrak{r} = 2^{\aleph_0}}$ then there is a uniform matroid $U$ on $\mathbb{N}$ such that the matroid family consisting of two copies of $U$ admits a packing and a covering, but not a partitioning.
Open questions
--------------
Our results give rise to a number of open questions. Undoubtedly the most pressing one is whether our main Theorem \[thm:main-thm-intro\] extends to more general matroid families than finitary-cofinitary ones: A matroid is called tame if the intersection of any circuit and cocircuit is finite. Tame matroids are the largest subclass where usually “nice behaviour” is expected. Even more special classes include the graphic matroids, infinite matroids where every finite minor is graphic, see [@bowler2018infinite] and the $\Psi$-matroids, graphic matroids that are related to the Freudenthal compactification of locally finite graphs, see [@bowler2013ubiquity].
Is the analogue of Theorem \[thm:main-thm-intro\] true for tame matroids? If not, does it hold for graphic matroids? If not, does it hold for $\Psi$-matroids?
Our second question is motivated by Theorem \[thm:unprovable-intro\].
Is there a ZFC counterexample to the analogue of Theorem \[thm:main-thm-intro\] for arbitrary matroids?
Structure of this paper
-----------------------
The paper is structured as follows. In Section \[s:matroids\] we will give a short introduction to the theory of infinite matroids. In Section \[sec\_prep\] contains some auxiliary results which are used in our proof of the main Theorem \[thm:main-thm-intro\] in Section \[sec\_mainresult\]. Finally, we end in Section \[s:consistently-false\] with the proof of Theorem \[thm:unprovable-intro\].
Infinite Matroids {#s:matroids}
=================
In this section we will gather all basic facts about infinite matroid necessary for this paper. Most of these facts are well-known for finite matroids. For a more detailed introduction to the theory of infinite matroids see [@nathanhabil].
Given a set $E$ and a set $\mathcal{I}$ of subsets of $E$ let us write $\mathcal{I}^{\text{max}}$ for the set of the $\subseteq$-maximal elements of $\mathcal{I}$. A pair ${(E,\mathcal{I})}$ where ${\mathcal{I} \subseteq \mathcal{P}(E)}$ is a *matroid* if
1. $\emptyset \in \mathcal{I}$;
2. If $I \in \mathcal{I}$ and $I' \subseteq I$, then $I' \in \mathcal{I}$;
3. For all $I \in \mathcal{I} \setminus \mathcal{I}^{\text{max}}$ and $I' \in \mathcal{I}^{\text{max}}$ there is an $x \in I' \setminus I$ such that $I+x \in \mathcal{I}$;
4. Whenever $I \subseteq X \subseteq E$ and $I \in \mathcal{I}$, the set $\{I' \in \mathcal{I} \colon I \subseteq I' \subseteq X \}$ has a $\subseteq$-maximal element.
We note that, when $E$ is finite, the above is equivalent with the usual axiomisation of a matroid in terms of its independent sets. We call the sets in $\mathcal{I}$ *independent* sets. All other subsets are *dependent*. The maximal independent sets are called *bases* and the minimal dependent sets are called *circuits*. Every dependent set contains a circuit (which fact in non-trivial for infinite matroids). A matroid is called *finitary* if all of its circuits are finite.
\[Fact-finitary\] A matroid is finitary if and only if for every $\subseteq$-increasing chain of independent sets the union of the chain is also independent.
One example of a finitary matroid is the finite cycle matroid of an infinite graph. That is, if $G=(V,E)$ is a connected infinite graph then we can define a matroid ${M(G)=(E,\mathcal{I})}$ on the edge set of $E$ by letting ${I \in \mathcal{I}}$ if and only if $I$ does not contain any finite cycle of $G$. It is straightforward to show that ${(E,\mathcal{I})}$ is an infinite matroid, and that the following facts are true:
- A set of edges $I \subseteq E$ is independent if and only if it is a forest;
- A set of edges $B$ is a base if and only if it is a spanning tree;
- A set of edges $C$ is a circuit if and only if it is a finite cycle.
As with finite matroids, there is a notion of duality for infinite matroids. The *dual* of a matroid ${M=(E,\mathcal{I})}$ is the matroid ${M^*=(E,\mathcal{I}^*)}$ such that ${I^* \in \mathcal{I}^*}$ if and only if there exists ${B \in \mathcal{I}^{\text{max}}}$ with ${I^* \cap B = \emptyset}$, and so the bases of $M^*$ are the complements of the bases of $M$. A matroid is called *cofinitary* if its dual is finitary.
Given a matroid ${M=(E,\mathcal{I})}$ and an $X\subseteq E$, the *restriction* of $M$ to $X$ is the matroid ${(X,\mathcal{I} \cap \mathcal{P}(X))}$ and it is denoted by ${M {\restriction}X}$. For the restriction of $M$ to $E \setminus X$ we also write ${M \setminus X}$ and call it the minor obtained by the *deletion* of $X$. We call the matroid ${(M^* {\restriction}X)^*}$ the minor of $M$ obtained by the *contraction* of $X$ and denote it by ${M \slash X}$. For the matroid obtained by the contraction of $E \setminus X$ we also write ${M{.}X}$ and call it the *contraction of $M$ onto $X$*. It is shown in [@BDKW13] that both of these structures are indeed matroids, moreover, for every disjoint $X,Y\subseteq E$, $(M \slash X)\setminus Y=(M \setminus Y) \slash X$. The *minors* of $M$ are the matroids of the form $(M \slash X)\setminus Y$ as above. The class of finitary (cofinitary) matroids is closed under taking minors.
We extends our notations for matroid families. For a family ${{\mathcal}{M} = ( M_i \colon i < \lambda )}$ of matroids on the same ground set $E$, we write ${{\mathcal}{M} \setminus X}$, ${{\mathcal}{M} {\restriction}X}$, ${{\mathcal}{M} \slash X}$ or ${{\mathcal}{M}{.}X}$, for the families ${( M_i \setminus X \colon i<\lambda)}$, ${( M_i {\restriction}X \colon i<\lambda)}$, ${( M_i \slash X \colon i<\lambda)}$ or ${( M_i {.} X \colon i<\lambda)}$, respectively.
We say ${X \subseteq E}$ *spans* ${e \in E}$ in matroid $M$ if either ${e \in X}$ or there exists a circuit ${C\ni e}$ with ${C-e \subseteq X }$. We denote the set of edges spanned $X$ in $M$ by $\mathsf{span}_{M}(X)$. The operator $\mathsf{span}_M$ is clearly extensive and increasing. It is less obvious but it is idempotent as well. An $S \subseteq E$ is spanning in $M$ if $\mathsf{span}_{M}(S)=E$.
\[Fact-cofinitary\] A matroid is cofinitary if and only if for every $\subseteq$-decreasing chain of spanning sets the intersection of the chain is also spanning.
For a $B\subseteq E$ the following are equivalent:
- $B$ is a maximal independent set,
- $B$ is a minimal spanning set,
- $B$ is an independent spanning set.
\[Fact-base\] If $ B_0, B_1 $ are bases of a matroid and $ \left|B_0\setminus B_1\right|<\aleph_0 $, then $ {\left|B_0\setminus B_1\right|=
\left|B_1\setminus B_0\right|} $.
\[Fact-unique-circ\] If $I$ is independent and spans $e\notin I$, then there is a unique circuit $C(e,I)$ contained in $I+e$. For every $f\in C(e,I) $, the set $I-f+e$ is independent.
Preparations {#sec_prep}
============
Let $\lambda$ be a finite or infinite cardinal and let ${{\mathcal}{M} = (M_i \colon i < \lambda)}$ be a family of matroids on the same ground set $E$ (we do not restrict our scope to finitary-cofinitary families until we say so explicitly). To allow ourselves some extra flexibility, we will broaden the concept for a covering and a packing for ${\mathcal}{M}$ slightly, and allows coverings for ${\mathcal}{M}$ to consist of independent subsets and packings to consist of spanning subsets. It is clear that there exist packing and coverings in the new sense if and only if there are packings and coverings in the more restricted sense as defined in the introduction of the paper. Let ${{\mathcal}{M} = (M_i \colon i < \lambda)}$ be a family of arbitrary matroids on the same ground set $E$. A *packing of $X$* (or *covering of $X$*, respectively) is a packing (or covering, respectively) for the family ${{\mathcal}{M} {\restriction}X}$.
Augmenting paths
----------------
The algorithmic proof of the rank function formula for sums of finite matroids by Edmonds and Fulkerson (see cite [@edmonds1965transversals]) turned out to be an efficient tool in the theory of infinite matroids as well. Lifting the proof to the infinite case does not require new ideas, however the “right” formulation slightly differs from the finite version.
Suppose that $ (I_i: i<\lambda) $ are pairwise disjoint sets such that each of them is independent in the corresponding matroid $ M_i $ and suppose $ e\in E
\setminus
\bigcup_{i<\lambda}I_i $. Roufly speaking, the next lemma tells that either there is another such a system $ (J_i: i<\lambda) $ of independent sets which is “very close” to the original system $ (I_i: i<\lambda) $ and covers $ \bigcup_{i<\lambda}I_i+e $ or there is a “witness” for the non-existence of such a sets $ J_i $. Note that on the contrast of the case of finite matroids, this witness does not rule out the existence of a cover for $ \bigcup_{i<\lambda}I_i+e $.
\[lem: augpath\] Let $I_i $ be independent in $ M_i $ for $ i<\lambda $ where $ I_i\cap I_j=\varnothing $ for $ i\neq j $ and suppose that $ e\in E \setminus
\bigcup_{i<\lambda}I_i $. Then exactly one of the following two statements holds.
1. \[item: primal\] There is a set system $ (J_i:i<\lambda) $ and a $ k<\lambda $ with the following properties:
1. $ J_i $ is independent in $ M_i $,
2. $ J_i\cap J_j=\varnothing $ for $ i<j<\lambda $,
3. $ \bigcup_{i<\lambda}J_i=\bigcup_{i<\lambda}I_i +e $,
4. $ \sum_{i<\lambda} \left|I_i \vartriangle J_i\right|<\aleph_0$,
5. $ \mathsf{span}_{M_i}(J_i)=\mathsf{span}_{M_i}(I_i) $ for $ i\neq k $ and $ {\mathsf{span}_{M_k}(J_k-f)= \mathsf{span}_{M_k}(I_k)} $ for some $ f\in J_k $.
2. \[item: dual\] There exists an $X\subseteq E $ containing $ e $ for which $ I_i\cap X $ spans $ X $ in $ M_i $ for $ i<\lambda $.
Let us start with an easy observation about exchanging multiple elements simultaneously in an independent set.
\[prop:simult change\] Let $ I $ be an independent and let $ e_0,\dots, e_{n-1}, f_0,\dots, f_{n-1}\in E $ where $ {f_m\in C(e_m, I) }
$ but $ f_m\notin C(e_\ell, I) $ for $ \ell>m $. Then $ {I\cup \{e_0,\dots, e_{n-1}\}\setminus \{f_0,\dots, f_{n-1}}\}$ is independent and spans the same set as $ I $.
We use induction on $ n $. The case $ n=0 $ is a tautology. Suppose that $ n>0 $. On the one hand, the set $ I-f_0+e_0 $ is independent and spans the same set as $ I $. On the other hand, ${C(e_m, I-f_0+e_0)= C(e_m, I)} $ for $ 1\leq
m <n $ because $ f_0\notin C(e_\ell, I) $ for $ \ell>0 $. Hence by using the induction hypothesis for $ I-f_0+e_0 $ and $ e_1,\dots e_{n-1}, f_1,\dots, f_{n-1} $ we are done.
To show that at least one of \[item: primal\] and \[item: dual\] holds, we build an auxiliary digraph $ D=(V,A) $ with ${V= E\cup \lambda} $. We have $ xi\in A $ whenever $ x\in E\setminus I_i $ and $ I_i+x $ is $ M_i$-independent. If $ I_i+x $ is dependent in $
M_i$, then there is an arc from $ x $ to each element of $ C_{M_i}(x, I_i) -x$. The definition of $ D $ is complete. We call a directed path from $ e $ to $ \lambda $ an *augmenting path*. Suppose first that there is no augmenting path. Let $ X\subseteq E $ be the set of vertices that are reachable from $ e $ in $
D $. Clearly $ e\in X $ and by the construction of $ D $ if $ g\in X $ and $i<\lambda $ then either $g\in I_i$ or $ C_{M_i}(g, I_i) $ is well-defined and a subset of $X$. It implies that $ I_i\cap X $ spans $ X $ in $ M_i $ for $ i<\lambda $.
Assume now that there is an augmenting path $ x_0,\dots, x_n $ in $ D $ where $ x_0=e $ and $ x_n = k<\lambda $. By trimming the path we may assume that there is no “jumping arc”, i.e., $ x_{\ell}x_m\notin A(D) $ for $ \ell+1<m $. We define $ J_i' $ as the symmetric difference of $ I_i $ and $\{ x_m, x_{m+1}: x_{m+1}\in I_i \} $. The non-existence of jumping edges ensures that Proposition \[prop:simult change\] is applicable for this simultaneous exchange at $ I_i $. Hence $ J_i' $ is independent in $ M_i $ and spans the same set as $ I_i $. For $ i\neq k $ we let $ J_i:=J_i' $. The arc $ x_nk\in A $ witnesses that $
I_k+x_n$ is independent in $ M_k$ and hence so is $ J_k'+x_n=:J_k$ because $ \mathsf{span}_{M_k}(I_k)= \mathsf{span}_{M_k}(J_k') $. Thus $ f:=x_n $ is suitable.
Suppose for a contradiction that both of \[item: primal\] and \[item: dual\] holds. Then $I_i^{ X }:= I_i\cap X $ is a base of $ M_i{\restriction}X
$. Edge $ e $ witnesses that the independent sets $ J_i^{ X }:=J_i\cap X$ cover strictly more elements of $ X $ than the bases $ I_i^{ X } $. By $ \sum_{i<\lambda} \left| I_i \vartriangle J_i\right|<\aleph_0$ and by the pigeonhole principle there is some $ j<\lambda $ with $ \left|J_j^{ X }\setminus I_j^{ X }\right|>\left|I_j^{ X }\setminus J_j^{ X }\right| $ which contradicts Fact \[Fact-base\] after extending $J_j^{ X }$ to a base of $ M_j{\restriction}X $ in an arbitrary way.
Tight sets
----------
Suppose that ${\mathcal}{M}$ admits a covering. A set ${X \subseteq E}$ is *tight* with respect to ${\mathcal}{M}$ (shortly ${\mathcal}{M}$-tight), if every covering ${(R_i \colon i < \lambda)}$ of $X$ is also a packing of $X$. A tight set is trivial if it is $\emptyset$ (which is always tight).
\[prop:tight-union\] Tight sets are closed under arbitrary large union.
Suppose that $X_\alpha$ is a tight set for $\alpha<\kappa$ and let $(R_i \colon i<\lambda)$ be an arbitrary covering of $\bigcup_{\alpha<\kappa}X_\alpha$. Since ${(X_\alpha\cap R_i \colon i<\lambda)}$ is a covering of the tight set $X_\alpha$, we know that $X_\alpha\cap R_i$ spans $X_\alpha$ in $M_i$ and ${\{ X_\alpha\cap R_i \colon i<\lambda \}}$ are pairwise disjoint. But then $R_i$ spans ${\bigcup_{\alpha<\kappa} X_\alpha}$ in $M_i$ and the sets $R_i$ are pairwise disjoint.
A set ${Y \subseteq E}$ is a *cowave* if ${{\mathcal}{M}{.}Y}$ admits a covering. It is easy to check that cowawes are closed under arbitrary large union (see [@bowler2015matroid]\*[Lemma 4.3 & last line, p. 179]{}). In the next lemma we characterise the situation when a given edge is never covered in a particular matroid whatever covering we consider.
\[lem:tight-char\] Let ${\mathcal}{M} = (M_i \colon i<\lambda) $ be a family of matroids on the same ground set $E$ admitting a covering and let ${e \in E}$. Then for all ${j<\lambda}$ the following statements are equivalent:
1. \[item:char1\] ${e \notin R_j}$ holds for every covering $ {(R_i \colon i<\lambda)}$;
2. \[item:char2\] there exists a tight set $X$ that does not contain $e$ such that ${e \in \mathsf{span}_{M_j}(X)}$.
Assume statement \[item:char2\] holds for some ${j < \lambda}$ and let ${(R_i \colon i<\lambda)}$ be a covering for ${\mathcal}{M}$. Then ${R_j \cap X}$ spans $X$ and hence $e$ in $M_j$. But since $R_j$ is $M_j$-independent and $e \notin X$ we get that $e \notin R_j$, as desired.
Now suppose that statement \[item:char1\] holds for some ${j < \lambda}$. We may assume that $e$ is not a loop in $M_j$, since otherwise $X := \emptyset$ is a witness for statement \[item:char2\]. Let ${M_j' := M_j/e}$ and ${M_i' := M_i-e}$ for ${i \neq j}$. Let us show first that we may assume without loss of generality that the maximal cowave $W$ with respect to ${{\mathcal}{M}' := (M_i' \colon i < \lambda)}$ is $\emptyset$. To do so, let ${ (R_i^{W} \colon i<\lambda)}$ be a covering for ${\mathcal}{M}'{.}W$. On the one hand, if ${\mathcal}{M}\setminus W$ admitted a covering ${(R_i \colon i<\lambda)}$ with $e\in R_j$ then ${ (R_i\cup R_i^{W} \colon i<\lambda)}$ would be a forbidden covering for ${\mathcal}{M}$. On the other hand, a desired tight set with respect to ${\mathcal}{M}\setminus W$ is a desired tight set for ${\mathcal}{M}$ as well.
We prove that $X:=E-e$ is as desired. To check tightness it is enough to consider coverings consisting of pairwise disjoint sets. Indeed, if there is a covering which is not a packing then “trimming” it to disjoint sets it cannot become a packing since the removed edges are no longer spanned. We take a covering ${(R_i^{X} \colon i<\lambda)}$ of $X$ for ${\mathcal}{M}$ where ${R_i^{X} \cap R_k^{X} = \emptyset}$ for ${i \neq k}$ but otherwise arbitrary. Then $R_j^{X}$ must span $e$ in $M_j$ otherwise adding $e$ to $R_j^{X}$ would lead a forbidden covering for ${\mathcal}{M}$. Since $e$ is not an $M_j$-loop, we can pick an ${f \in C_{M_j}(e,R_j^{X}) - e}$. Let ${Q_j^{X} := R_j^{X} - f}$ and ${Q_i^{X} := R_i^{X}}$ for ${i \neq j}$. Then the sets ${\{ Q_i^{X} \colon i<\lambda \}}$ are independent with respect to the corresponding elements of ${\mathcal}{M}'$ and cover ${X-f}$. By applying Lemma \[lem: augpath\] with ${\mathcal}{M}', {\{ Q_i^{X} \colon i<\lambda \}}$ and $f$, the augmentation cannot be possible since it would yield to a forbidden covering for ${\mathcal}{M}$. It follows that there is a $X'\subseteq X$ containing $f$ such that $Q_i^{X}\cap X'$ spans $X'$ in $M_i'$ for $i<\lambda$. Note that we must have $X'=X$ otherwise $X \setminus X'$ would be a non-trivial cowave with respect to ${\mathcal}{M}'$ witnessed by the sets $Q_i^{X}\cap X'$. But then by the definition of $Q_i^{X}$ and ${\mathcal}{M}'$ we can conclude that $R_i^{X} $ spans $X$ in $M_i$ for $i\neq j$ and $(R_j^{X} - f)$ spans $X$ in $M_j \slash e$. From the latter it follows that $R_j^{X}$ spans $X$ in $M_j$ since $ e\in \mathsf{span}_{M_j}(R^{X}_j) $.
Feasible extensions
-------------------
In the proof of the main result we will build the desired partitioning $(B_i: i<\lambda)$ by making decisions recursively about putting some edge to a $B_i$ or not. In the following approach the intended role of set $I_i$ is being the set of edges we decided already to belong to $B_i$ and set $S_i$ is the set of edges not ruled out to be in $B_i$.
A family of ordered pairs ${\mathcal{F} = (I_i, S_i \colon i<\lambda)}$ where ${I_i \subseteq S_i \subseteq E}$ for $i<\lambda$ and ${S_i \cap
I_j = \emptyset}$ for $ {i \neq j}$ is *covering-feasible* with respect to ${\mathcal}{M}$ if there exists a covering ${ (R_i \colon i<\lambda)}$ for ${\mathcal}{M}$ where ${I_i \subseteq R_i \subseteq S_i}$ for ${i < \lambda}$. We call such a covering ${\mathcal{F}}$-*compatible*. The definition of *packing-feasible* is analogous. Finally, $\mathcal{F}$ is *feasible* if it is covering-feasible and packing-feasible.
We say that ${\mathcal{F}' = (I_i', S_i' \colon i<\lambda)}$ is an *extension* of $\mathcal{F}$ if ${I_i \subseteq I_i'}$ and ${S_i \supseteq S_i'}$ holds for every ${i < \lambda}$. For a feasible $\mathcal{F}$, let ${{\mathcal}{M}(\mathcal{F})}$ denote the family ${(M_i(\mathcal{F}) \colon i<\lambda)}$ where we obtain ${M_i(\mathcal{F})}$ from $M_i$ by contracting $I_i$, deleting ${\bigcup_{j\neq i} I_j}$ and replacing every edge in ${(E \setminus
\bigcup_{i<\lambda} I_i) \setminus S_i}$ by a loop. Note that each ${M_i(\mathcal{F})}$ is a matroid on ${E \setminus \bigcup_{i<\lambda} I_i}$. Moreover, given a covering ${(R_i \colon i<\lambda)}$ for ${\mathcal}{M}(\mathcal{F})$ the family ${( R_i \cup I_i \colon i<\lambda)}$ is an $\mathcal{F}$-compatible covering for ${\mathcal}{M}$, and, vice versa, given an $\mathcal{F}$-compatible covering ${(R_i \colon i<\lambda)}$ for ${\mathcal}{M}$ the family ${( R_i \setminus \bigcup_{j<\lambda} I_j \colon i<\lambda)}$ is a covering for ${\mathcal}{M}(\mathcal{F})$. The analogous statement holds for packings.
\[cor:cov-feasible\] Let ${\mathcal{F} = (I_i,S_i \colon i<\lambda)}$ be ${\mathcal}{M}$-feasible and assume there is no non-trivial tight set with respect to ${{\mathcal}{M}(\mathcal{F})}$. If for a ${j < \lambda}$ and ${e \in S_j \setminus I_j}$ the set ${I_j + e}$ is independent in $M_j$, then the extension $\mathcal{F}'$ that we obtain by adding $e$ to $I_j$ and removing $e$ from $S_i$ for ${i\neq j}$ is covering-feasible with respect to ${\mathcal}{M}$.
Assume for a contradiction that the extension $\mathcal{F}'$ is not covering-feasible. It means that there is no covering ${(R_i \colon i<\lambda)}$ for ${{\mathcal}{M}(\mathcal{F})}$ with ${e \in R_j}$. Then Lemma \[lem:tight-char\] implies that there is a ${\mathcal}{M}(\mathcal{F})$-tight set $X$ spanning $e$ in $M_j(\mathcal{F})$. Now $X$ is empty since there is no non-trivial tight set by assumption. Thus $e$ is an $M_j(\mathcal{F})$-loop. But this contradicts the assumptions that ${I_j + e}$ is independent in $M_j$ and ${e \in S_j}$.
\[lem:elimin\] Let ${\mathcal{F} = (I_i,S_i \colon i<\lambda)}$ be feasible with respect to ${\mathcal}{M}$, let $X$ be the largest tight set in ${{\mathcal}{M}(\mathcal{F})}$ and let ${(R_i \colon i<\lambda)}$ be a covering of $X$ in ${{\mathcal}{M}(\mathcal{F})}$. Then the set system ${\mathcal{F}' = (I_i',S_i' \colon i<\lambda)}$ where ${I_i' = I_i \cup R_i}$ and ${S_i' = S_i \setminus \bigcup_{j\neq i} R_j}$ is a feasible extension of $\mathcal{F}$ with respect to ${\mathcal}{M}$. Furthermore, there is no non-trivial tight set with respect to ${{\mathcal}{M}(\mathcal{F}')}$ and $I_i'$ spans ${X \cap S_i}$ in $M_i$ for every ${i < \lambda}$.
Since $X$ is ${\mathcal}{M}(\mathcal{F})$-tight, it follows that $R_i$ spans $X$ in ${M_i(\mathcal{F})}$ and the $R_i$ are pairwise disjoint. Hence by the definition of ${M_i(\mathcal{F}})$ we obtain that $I_i'$ spans ${X \cap S_i}$ in $M_i$. Covering-feasibility follows from the fact that for any covering $(Q_i:i<\lambda)$ for ${\mathcal}{M}(\mathcal{F})$ the set $Q_i\cap X$ spans $X$ in $M_i(\mathcal{F})$ and therefore the sets $Q_i\setminus X $ form a covering for ${\mathcal}{M}(\mathcal{F}')$. Let $ {(P_i \colon i<\lambda)}$ be an $\mathcal{F}$-compatible packing and let ${P_i' := (P_i \setminus X) \cup I_i'}$ for ${i<\lambda}$. Then $P_i'$ is spanning in $M_i$ because $P_i$ was spanning and $I_i'$ spans ${S_i \cap X \supseteq P_i \cap X}$.
Suppose that $Y$ is tight with respect to ${{\mathcal}{M}(\mathcal{F}')}$. We show that ${X \cup Y}$ is tight with respect to ${{\mathcal}{M}(\mathcal{F})}$. Let us take a covering ${(Q_i \colon i<\lambda)}$ for $X\cup Y$ in ${\mathcal}{M}(\mathcal{F})$. As earlier, the set $Q_i\cap X$ spans $X$ in $M_i(\mathcal{F})$ and the $Q_i\cap X$ are pairwise disjoint because the ${\mathcal}{M}(\mathcal{F})$-tightness of $X$. Therefore the sets $Q_i\setminus X$ form a cover of $Y$ with respect to ${\mathcal}{M}(\mathcal{F}')$. By the ${\mathcal}{M}(\mathcal{F}')$-tightness of $Y$, the set $Q_i\setminus X$ spans $Y$ in $M_i(\mathcal{F}')$ and $Q_i\setminus X$ are pairwise disjoint. By combining these, we obtain that $Q_i$ spans $X\cup Y$ in $M_i(\mathcal{F})$. It means that $X\cup Y$ is ${\mathcal}{M}(\mathcal{F})$-tight. But then by the maximality of $ X $ we have ${Y = \emptyset}$.
The main lemmas
---------------
The core of the proof of Theorem \[thm:main-thm-intro\] is the following three lemmas.
\[lem:cover-e\] Let ${\mathcal{F} = (I_i, S_i \colon i<\lambda)}$ be feasible with respect to ${\mathcal}{M}$ and let ${e \in E}$. Then there exists a feasible extension ${\mathcal{F}' = (I_i', S_i' \colon i<\lambda)}$ of $\mathcal{F}$ such that ${e \in \bigcup_{i<\lambda} I_j'}$.
By Lemma \[lem:elimin\], we may assume that there is no non-trivial tight set with respect to $\mathcal{F}$. Let ${\mathcal{P} = (P_i \colon i<\lambda)}$ be an $\mathcal{F}$-compatible packing consisting of bases. We may assume that ${e \in \bigcup_{i<\lambda} P_i}$, as otherwise we take some ${j < \lambda}$ such that $e$ is not an $M_j(\mathcal{F})$-loop (which exists by covering-feasibility) and replace $P_j$ by ${P_j + e - f}$ for some ${f \in C_j(e,P_j) - e}$. Suppose that ${e \in P_j}$ and add $e$ to $I_j$ as in Corollary \[cor:cov-feasible\]. Packing-feasibility is preserved with $\mathcal{P}$ as a witness. We also preserve covering-feasibility by Corollary \[cor:cov-feasible\].
\[lem:span-e\] Let ${\mathcal{F} = (I_i, S_i \colon i<\lambda)}$ be feasible with respect to ${\mathcal}{M}$, let ${e \in E}$ and assume that $M_j$ is finitary for some ${j < \lambda}$. Then there exists a feasible extension ${\mathcal{F}'' = (I_i'', S_i'' \colon i<\lambda)}$ of $\mathcal{F}$ such that $e$ is spanned by $I_j''$ in $M_j$.
We take an $\mathcal{F}$-compatible packing ${\mathcal{P} = (P_i \colon i<\lambda)}$ and pick a minimum sized ${I \subseteq (P_j \setminus I_j)}$ such that ${I_j \cup I}$ spans $e$ in $M_j$. Since $M_j$ is finitary, ${\left|I\right| =: n \in \mathbb{N}}$. We apply induction on $n$. If ${n = 0}$, then $I_j$ spans $e$ in $M_j$ and we are done. Suppose ${n > 0}$ and let ${f \in I}$. Adding $f$ to $I_j$ results in a packing-feasible system since $\mathcal{P}$ is a witness for this. If it respects covering-feasibility as well, then we take this extension and use induction via ${\left| I\setminus \{ f \} \right| = n-1}$. Hence assume that adding $f$ to $I_j$ results in a system which is not covering-feasible. We may conclude by Corollary \[cor:cov-feasible\] and Proposition \[prop:tight-union\] that the largest ${{\mathcal}{M}(\mathcal{F})}$-tight set $X$ spans $f$ in ${M_j(\mathcal{F})}$. We take a feasible extension ${ \mathcal{F}' = (I_i',S_i': i<\lambda)}$ that “distributes” $X$ applying Lemma \[lem:elimin\]. Let $J$ be a maximal ${M_j/I_j'}$-independent subset of ${I \setminus X}$. Then ${I_j' \cup J}$ spans $e$ in $M_j$ because ${I_j \cup I}$ does and ${I_j' \supseteq I_j}$ spans the edges ${X \cap S_j \supseteq I \setminus X}$ in $M_j$. Furthermore, ${f \notin J}$ since the tight $X$ spanned $f$ in ${M_j/I_j}$ and therefore so does $I_j'$ in $M_j$ by Lemma \[lem:elimin\]. Hence ${\left| J \right| < \left| I \right|}$ and we are done by induction.
\[lem:cospan-e\] Let ${\mathcal{F} = (I_i, S_i \colon i<\lambda)}$ be feasible with respect to ${\mathcal}{M}$, let ${e \in E}$ and assume that $M_j$ is cofinitary for a $j<\lambda$. Then there exists a feasible extension ${\mathcal{F}'=(I_i', S_i' \colon i<\lambda)}$ of $\mathcal{F}$ such that $e$ is spanned by ${E \setminus S_j'}$ in $M^{*}_j$.
We reduce the statement to Lemma \[lem:span-e\] using a dualisation argument. Let ${{\mathcal}{M}^{*}=(M_i^{*} \colon i<\lambda)}$. Assume first that ${\lambda=2}$. Then $R_0,R_1$ is a covering for ${\mathcal}{M}$ if and only if $E\setminus R_0, E\setminus R_1 $ is a packing for ${\mathcal}{M}^{*}$ and the analogue statement holds for packings. Therefore ${\mathcal{F} = (I_i,S_i \colon i<\lambda)}$ is ${\mathcal}{M}$-feasible if and only if ${\mathcal{F}^{*} := (S_i,I_i \colon i<\lambda)}$ is ${\mathcal}{M}^{*}$-feasible. Thus we can simply apply Lemma \[lem:span-e\] with $\mathcal{F}^{*}$ and ${\mathcal}{M}^{*}$ and dualising back the resulting extension.
For a general $\lambda$ the argument is essentially the same except that we need to overcome some unpleasant technical difficulties. Namely if ${(R_i \colon i<\lambda)}$ is a covering for ${\mathcal}{M}$ then ${( E \setminus R_i \colon i<\lambda)}$ is usually not a packing for ${\mathcal}{M}^{*}$ when ${\lambda>2}$. Indeed, instead of being pairwise disjoint, the sets ${E \setminus R_i}$ satisfy the condition that each edge is missing from at least one ${E \setminus R_i}$. A similar problem occurs with the the dual object of packings. To fix this we define an auxiliary matroid family ${\widehat{{\mathcal}{M}} = (\widehat{M_i} \colon i\leq\lambda)}$ on the common ground set ${E \times \lambda}$. For ${e \in E}$ and ${i < \lambda}$ the edge ${(e,j)}$ is a loop in $\widehat{M_i}$ whenever ${i \neq j}$. A subset of ${E \times \{ i \}}$ is independent in $\widehat{M_i}$ if and only if its projection to the first coordinate is independent in $M_i^{*}$. Finally a set is defined to be independent in $M_\lambda$ if it meets ${\{ e \} \times \lambda }$ in at most one element for every ${e \in E}$.
It follows directly from the definitions that if for an ${\mathcal}{M}$-feasible system ${\mathcal{F} = (I_i, S_i \colon i<\lambda)}$ we take ${\widehat{I_i} := (E \setminus S_i) \times \{ i \}}$, ${\widehat{S_i} := (E\setminus I_i) \times \{ i \}}$, ${\widehat{I_\lambda} := \emptyset}$ and ${\widehat{S_\lambda} := E \times \lambda}$. Then ${\widehat{\mathcal{F}} := (\widehat{I_i},\widehat{S_i} \colon i\leq \lambda)}$ is $\widehat{{\mathcal}{M}}$-feasible. On the other hand, if some ${\widehat{\mathcal{F}} = (\widehat{I_i},\widehat{S_i} \colon i \leq \lambda)}$ is $\widehat{{\mathcal}{M}}$-feasible, then $${\left( \mathsf{proj_E}\left[ (E\times \{ i \}) \setminus \widehat{S_i}\right] , \mathsf{proj_E}\left[ (E\times \{ i \}) \setminus \widehat{I_i}\right] \colon i<\lambda \right)}$$ is ${\mathcal}{M}$-feasible. From this point the proof goes the same way as in the case ${\lambda = 2}$ using $\widehat{{\mathcal}{M}}$ instead of ${\mathcal}{M}^{*}$.
Proof of the main result {#sec_mainresult}
========================
\[Proof of Theorem \[thm:main-thm-intro\]\] Let us start by reducing the theorem to its special case where ${\lambda = 3}$. (An uncountable $\lambda$ would cause technical difficulties at some point.)
\[claim:wlog-lambda=3\] Theorem \[thm:main-thm-intro\] is implied by its special case where ${\lambda = 3}$.
We build a new matroid family ${\mathcal}{M}'$ on the ground set ${E \times \lambda}$ with at most three members (each of which is either finitary or cofinitary) such that it admits a packing/covering/partitioning if and only if ${\mathcal}{M}$ does. First we take a copy ${\tilde{M_i}}$ of $M_i$ on ${E \times\{ i \}}$ via the bijection ${e \mapsto (e,i)}$ for ${i < \lambda}$. Let ${F := \{ i<\lambda \colon M_i \text{ is finitary} \}}$. Let $M$ be the direct sum of the matroids $\tilde{M_i}$ for ${i \in F}$ extended by the elements ${E \times (\lambda \setminus F)}$ as loops. The construction of $N$ is analogous, we take the direct sum of the cofinitary copies and declare everything else to be a loop. Finally the circuits of the third matroid $L$ are ${\{ \{ e \} \times \lambda \colon e \in E \}}$. It is easy to check directly from the definitions that ${\mathcal}{M}'$ has the desired property. Indeed, suppose for example that ${(R_i \colon i<\lambda )}$ is a covering for ${\mathcal}{M}$. Then we choose for each ${e \in E}$ an ${i_e < \lambda}$ such that ${e \in R_{i_e}}$. Then ${I_0 := \bigcup_{i\in F} R_i \times \{ i \}}$ is $M$-independent, ${I_1 := \bigcup_{i \in \lambda \setminus F} R_i \times \{ i \}}$ is $N$-independent and ${I_2 := \bigcup_{e\in E} (\{ e \} \times \lambda) \setminus \{ (e,i_e) \}}$ is $L$-independent. Moreover, ${I_0 \cup I_1 \cup I_2 = E\times \lambda}$. The proof of the remaining implications are similarly easy and we leave them to the reader.
By Claim \[claim:wlog-lambda=3\], we can assume that ${\lambda = 3}$. As above, let ${F := \{ i < 3 \colon M_i \text{ is finitary} \}}$. We apply transfinite induction on ${\left|E\right|}$. If $E$ is finite then every packing is a covering and every covering is a packing (under the assumption that both exists). Indeed, let us denote the rank of $M_i$ by $r_i$. Then by the existence of a covering and a packing we obtain the inequalities ${\sum_{i<3}r_i\geq \left|E\right|\geq \sum_{i<3}r_i} $ respectively, thus ${\sum_{i<3}r_i= \left|E\right|}$. Then for any covering $(R_i: i<3)$ necessarily $\left|R_i\right|=r_i$ for $i<3$ and the $R_i$ must be pairwise disjoint. It means that $(R_i: i<3)$ is a packing as well. Let ${\left|E\right| = \aleph_0}$ and let us fix an enumeration ${\{ e_n \colon n\in \mathbb{N} \}}$ of it. We build by recursion an increasing sequence ${(\mathcal{F}_n \colon n\in \mathbb{N})}$ of feasible systems where ${\mathcal{F}_{n} = (I_i^{n}, S_i^{n} \colon i<3)}$. We demand that for every ${n \in \mathbb{N}}$
1. \[item:main-1-1\] ${e_n \in \bigcup_{i<3}I_i^{n+1}}$,
2. \[item:main-1-2\] $e_n$ is spanned by $I_i^{n+1}$ in $M_i$ for ${i \in F}$,
3. \[item:main-1-3\] $e_n$ is spanned by ${E \setminus S_i^{n+1}}$ in $M_i^{*}$ for ${i \notin F}$.
Each step of the recursion can be done by applying Lemmas \[lem:cover-e\], \[lem:span-e\] and \[lem:cospan-e\] consecutively. The sets ${B_i := \bigcup_{n=0}^{\infty} I_i^{n}}$ are pairwise disjoint because so are ${\{ I_i^{n} \colon i<3 \}}$ for every ${n \in \mathbb{N}}$. By property \[item:main-1-1\], $e_n$ is covered by some $I_i^{n+1}$ and hence by $B_i$ as well. Therefore the sets $B_i$ form a partition of $E$. It remains to show that $B_i$ is a base of $M_i$. For ${i \in F}$, $B_i$ is independent by Fact \[Fact-finitary\]. Note that ${B_i = \bigcap_{n=0}^{\infty} S_i^{n}}$ since ${S_i^{n} \supseteq I_i^{n}}$ and ${S_i^{n} \cap I_j = \emptyset}$ for ${j \neq i}$ hold for every ${n \in \mathbb{N}}$. Then $B_i$ is spanning for ${i \notin F}$ by Fact \[Fact-cofinitary\]. By property \[item:main-1-2\], $e_n$ is spanned by $I_i^{n+1}$ in $M_i$ for ${i \in F}$ and spanned by ${E \setminus S_i^{n+1}}$ in $M_i^{*}$ for ${i \notin F}$. It means that $B_i$ is also spanning for ${i \in F}$ and independent for ${i \notin F}$ which completes the proof of the countable case.
Suppose ${\kappa := \left|E\right| > \aleph_0}$. Let ${(R_i \colon i < 3 )}$ be a covering for ${\mathcal}{M}$ consisting of bases and let ${\mathcal{P} = (P_i \colon i < 3 )}$ be a packing consisting of bases. We construct an increasing continuous chain ${\left\langle E_\alpha \colon \alpha \leq \kappa \right\rangle}$ of subsets of $E$ where
1. \[item:main-2-1\] ${E_0 = \emptyset}$,
2. \[item:main-2-2\] ${E_\kappa = E}$,
3. \[item:main-2-3\] ${\left| E_\alpha \right| < \kappa}$ for ${\alpha < \kappa}$,
4. \[item:main-2-4\] ${E_\alpha \cap R_i}$ and ${E_\alpha \cap P_i}$ spans $E_\alpha$ in $M_i$ for ${i \in F}$,
5. \[item:main-2-5\] ${E_\alpha \cap (E \setminus R_i)}$ and ${E_\alpha \cap (E \setminus P_i)}$ spans ${E_\alpha}$ in $M_i^{*}$ for ${i \notin F}$.
Putting some $e$ to an $E_\alpha$ indicates that some finitely many finite fundamental circuits should be also contained in $E_\alpha$, thus the construction of the chain above can be done by a straightforward transfinite recursion. Note that the sets ${E^{\alpha} := E_{\alpha+1} \setminus E_\alpha\ (\alpha<\kappa)}$ form a partition of $E$. Let $$M_i^{\alpha} :=
\begin{cases}
(M_i{\restriction}E_{\alpha+1})/E_\alpha & \mbox{if } i\in F \\
((M_i^{*}{\restriction}E_{\alpha+1})/E_\alpha)^{*} & \mbox{if } i\notin F.
\end{cases}$$
\[lem:smaller-problem\] For every $ \alpha<\kappa $ the matroid family ${{\mathcal}{M}_\alpha := (M_i^{\alpha} \colon i<3)}$ on $E^{\alpha}$ admits a covering and a packing.
Let ${R_i^{\alpha} := R_i\cap E^{\alpha}}$. We show that ${(R_i^{\alpha} \colon i<3)}$ is a covering for ${\mathcal}{M}_{\alpha}$. The non-trivial part of the statement is that $R_i^{\alpha}$ is independent in $M_{i}^{\alpha}$. Property \[item:main-2-4\] ensures that for ${i \in F}$, the set ${R_i \cap E_\alpha}$ spans $E_\alpha$ in $M_i$, and hence ${R_i \cap E^{\alpha}}$ remains independent after the contraction of $E_\alpha$ in $M_i$. For ${i \notin F}$, we know by property \[item:main-2-5\] that ${E_{\alpha} \setminus R_i}$ spans $E_{\alpha}$ in $M_i^{*}$. Thus ${E^{\alpha} \setminus R_i}$ spans $E^{\alpha}$ in ${M_i^{*}/E_\alpha}$ and hence in ${M_i^{*}/E_\alpha {\restriction}E_{\alpha+1}}$ as well. By taking the dual it means that ${E^{\alpha} \cap R_i}$ is independent in ${M_i^{\alpha}}$. The proof of the fact that ${\mathcal}{M}_\alpha$ admits a packing is analogous.
By applying Lemma \[lem:smaller-problem\] and then the induction hypothesis for ${\mathcal}{M}_\alpha$ (see property \[item:main-2-3\]), we may fix a partition ${(B^{\alpha}_i \colon i<3)}$ of $E^{\alpha}$ where $B_i^{\alpha}$ is a base of $M_i^{\alpha}$. Note that the set family ${\{ B_i^{\alpha}
\colon \alpha<\kappa, i<3 \}}$ forms a partition of $E$ and hence so does the family ${(B_i \colon i<3)}$ where ${B_i :=\bigcup_{\alpha<\kappa} B_i^{\alpha}}$. It remains to show that $B_i$ is a base of $M_i$.
\[lem:build-base\] For every $ i<3 $ and $ \alpha\leq\kappa $, $ B_{i,\alpha}:= \bigcup_{\beta<\alpha}B_i^{\beta} $ is a base of $\begin{cases}
M_i{\restriction}E_\alpha &\mbox{if } i\in F \\
M_i. E_\alpha & \mbox{if }i\notin F.
\end{cases}$
Let $ i\in F $. By the construction, $ B_i^{\alpha} $ is a base of $ (M_i{\restriction}E_{\alpha+1})/E_{\alpha} $. The induction hypotheses guarantees that $ B_{i,\alpha} $ is a base of $
M_i{\restriction}E_{\alpha}
$. By combining these, we obtain that $ B_{i,\alpha+1} $ is a base of $ M_i{\restriction}E_{\alpha+1} $. At limit steps we obviously preserve the spanning property and $ i\in F $ ensures that the independence as well. For $ i\notin F$, we use the reformulation that $ E_\alpha \setminus B_{i,\alpha} $ is a base of $ M_i^{*}{\restriction}E_{\alpha} $ for every $ \alpha $ which can be proved the same way.
Lemma \[lem:build-base\] tells for ${\alpha = \kappa}$ that $B_i$ is a base of $M_i$ which completes the proof.
We point out a consequence of the special case ${\lambda = 2}$. By applying Theorem \[thm:main-thm-intro\] for two matroids and dualising one of them we obtain the following corollary.
Let $M_i$ be a finitary or cofinitary matroid on the ground set $E$ for ${i = 0,1}$. Suppose that there are ${J_{i} \subseteq E\ (i=0,1)}$ such that $J_i$ is independent in $M_i$ and spanning in $M_{1-i}$. Then there exists a common base of $M_0$ and $M_1$.
Let us also state the reformulation with bases.
Let $M_i$ be a finitary or cofinitary matroid on the ground set $E$ for ${i = 0,1}$. If there are bases $ B_i$, $B_i'$ of $M_i$ such that ${B_0 \subseteq B_1}$ and ${B_1' \subseteq B_0' }$, then $M_0$ and $M_1$ share some base.
A consistency result {#s:consistently-false}
====================
It is natural to ask if Theorem \[thm:main-thm-intro\] remains true for arbitrary matroids. We show that it is unprovable in ZFC (unless ZFC is inconsistent), i.e., there is consistently a matroid family admitting a packing and a covering but not a partitioning. We do not know yet if the statement is independent or it is possible to construct a counterexample in ZFC alone.
This section is devoted to the proof of Theorem \[thm:unprovable-intro\]. It is based on a construction of N. Bowler and S. Geschke in [@bowler2016self]. Among other results they showed that the existence of a matroid admitting two bases with different infinite sizes is consistent (and actually independent). Let us denote the set of the infinite and (simultaneously) co-infinite subsets of a set $X$ by $\mathcal{P}^{*}(X)$ and we write ${X \subseteq^{*} Y}$ if ${\left| X \setminus Y \right| < \aleph_0}$. We repeat the theorem for convenience. The reaping number $\mathfrak{r}$ and uniform matroids were introduced right before Theorem \[thm:unprovable-intro\].
\[thm:unprov\] If $\mathfrak{r}=2^{\aleph_0}$ then there is a uniform matroid $U$ on $\mathbb{N}$ such that the matroid family consisting of two copies of $U$ admits a packing and a covering, but not a partitioning.
We construct a ${\mathcal{B} \subseteq \mathcal{P}^{*}(\mathbb{N})}$ which is a set of the bases of a desired $U$. Conditions \[item:unprov-1\]–\[item:unprov-3\] in the following list are expressing that $ \mathcal{B} $ is the set of the bases of a uniform matroid on $\mathbb{N}$ (see [@bowler2016self]), whereas \[item:unprov-4\] and \[item:unprov-5\] guarantee that $U$ admits a packing and a covering but not a partitioning.
1. \[item:unprov-1\] The elements of $ \mathcal{B} $ are pairwise $ \subseteq $-incomparable,
2. \[item:unprov-2\] Whenever $ e\in B\in \mathcal{B} $ and $ f\in \mathbb{N} \setminus B $, then $ B-e+f\in \mathcal{B} $,
3. \[item:unprov-3\] For every $I\subseteq X\subseteq \mathbb{N} $, there is a $ B\in \mathcal{B} $ such that one of the following holds:
1. \[item:unprov-3a\] $ B\subseteq I $
2. \[item:unprov-3b\] $ I\subseteq B \subseteq X $
3. \[item:unprov-3c\] $ X\subseteq B $
4. \[item:unprov-4\] There are no disjoint $ B,B'\in \mathcal{B} $ with $ B\cup B'=\mathbb{N} $,
5. \[item:unprov-5\] There are $R_0, R_1, P_0, P_1\in \mathcal{B} $ with $ R_0\cup R_1=\mathbb{N} $ and $ P_0\cap P_1=\emptyset $.
Let us fix a well-order $ \prec $ of the set $ \mathcal{A}:=\{ (I,X): I\subseteq X\subseteq \mathbb{N} \} $ of type $ 2^{\aleph_0} $ in which $ (\varnothing, I)\preceq (I,X) $ for every $ I\subseteq X \subseteq \mathbb{N} $. Let $ \{ (I_\alpha,X_\alpha):
\alpha<2^{\aleph_0} \} $ be the enumeration of $ \mathcal{A} $ given by $ \prec $. We build the family $\mathcal{B}$ by transfinite recursion as the union of an increasing continuous chain ${\left\langle \mathcal{B}_\alpha \colon \alpha<2^{\aleph_0} \right\rangle}$ where $\mathcal{B}_\alpha$ satisfies \[item:unprov-1\], \[item:unprov-2\], \[item:unprov-4\], \[item:unprov-5\] and the restriction of \[item:unprov-3\] to the pairs ${\{ (I_\beta,X_\beta) \colon \beta<\alpha \}}$ that we call \[item:unprov-3\]($\alpha$). Let ${R_0, R_1 \subseteq \mathbb{N}}$ be such that they cover $\mathbb{N}$ and all the sets ${R_0 \setminus R_1}$, ${R_1\setminus R_0}$, ${R_0\cap R_1}$ are infinite. Pick also disjoint sets ${P_0, P_1 \subseteq \mathbb{N}}$ such that all of ${P_0}, {P_1}, {\mathbb{N} \setminus (P_0 \cup P_1)}$ have an infinite intersection with any of ${R_0 \setminus R_1}$, ${R_1\setminus R_0}$, ${R_0\cap R_1}$. By defining $\mathcal{B}_0$ as the closure of the set ${\{ R_0, R_1, P_0, P_1 \}}$ under \[item:unprov-2\], we get neither $\subseteq$-comparable sets nor two sets forming a partitioning. Clearly the conditions cannot be ruined at limit steps. Suppose that $\mathcal{B}_\alpha$ is defined. If \[item:unprov-3\]($\alpha+1$) is satisfied by $\mathcal{B}_\alpha$ then let ${\mathcal{B}_{\alpha+1} := \mathcal{B}_\alpha}$. Suppose that it does not.
$ \left|X_\alpha\setminus
I_\alpha \right|=\aleph_0$
By the choice of $ \prec $, we know that $(\varnothing, I_\alpha) \preceq(I_\alpha, X_\alpha) $. The inequality must be strict otherwise $ I_\alpha=X_\alpha=\varnothing $ and hence $ \mathcal{B}_\alpha $ satisfies \[item:unprov-3\]($\alpha+1$) which is a contradiction. By the induction hypothesis, the condition corresponding to $ (\varnothing, I_\alpha) $ is satisfied in $ \mathcal{B}_\alpha $. We cannot have a $
B\in
\mathcal{B}_\alpha $ with $ B\subseteq I_\alpha $ because then it would exemplify that $ \mathcal{B}_\alpha $ satisfies \[item:unprov-3\]($\alpha+1$). Therefore we must have a $ B\in
\mathcal{B}_\alpha $ with $ B\supseteq I_\alpha $. If $ X_\alpha\setminus
I_\alpha $ was finite, then property \[item:unprov-2\] and $ B $ would give a $ B'\in \mathcal{B}_\alpha $ with either $ B'\supseteq
X_\alpha $ or $I_\alpha\subseteq B'\subseteq X_\alpha $ both of which contradicts the assumption that $ \mathcal{B}_\alpha $ does not satisfy \[item:unprov-3\]($\alpha+1$).
Consider the family $$\mathcal{F} =
\{B\cap (X_\alpha\setminus I_\alpha),\ (X_\alpha\setminus I_\alpha)\setminus B \colon B\in \mathcal{B}_\alpha \}.$$ Since $\left|\mathcal{F}\right|<2^{\aleph_0}$ and $\mathfrak{r}=2^{\aleph_0}$, there is a ${G \in \mathcal{P}^{*}(X_\alpha \setminus I_\alpha)}$ such that both $G$ and ${G \setminus (X_\alpha \setminus I_\alpha)}$ have an infinite intersection with each infinite element of $\mathcal{F}$. Let ${B_\alpha := I_\alpha \cup G}$ and we extend $\mathcal{B}_\alpha$ with all the sets $$\{ Y \subseteq \mathbb{N} \colon \left| Y \setminus B_\alpha \right| = \left| B_\alpha \setminus Y \right| < \aleph_0 \}$$ to obtain $\mathcal{B}_{\alpha+1}$. Conditions \[item:unprov-2\] and \[item:unprov-3\]($\alpha+1$) obviously hold for $\mathcal{B}_{\alpha+1}$. To show \[item:unprov-1\], it is enough to prove that there is no ${B \in \mathcal{B}_\alpha}$ for which ${B \subseteq^{*} B_\alpha}$ or ${B_\alpha \subseteq^{*} B}$. Suppose for a contradiction that ${B \subseteq^{*} B_\alpha}$ for some ${B \in \mathcal{B}_\alpha}$. Then ${B \cap (X_\alpha \setminus I_\alpha)}$ must be finite since otherwise it contains infinitely many elements not contained by $G$. But then ${B \subseteq^{*} I_\alpha}$ and by applying \[item:unprov-2\] we obtain a ${B' \in \mathcal{B}_{\alpha}}$ for which either ${B' \subseteq I_\alpha}$ or ${I_\alpha \subseteq B' \subseteq X_\alpha}$, thus \[item:unprov-3\]($\alpha+1$) was satisfied in $\mathcal{B}_\alpha $, a contradiction. Ruling out the existence of a ${B \in \mathcal{B}_\alpha}$ with ${B_\alpha \subseteq^{*} B}$ is analogous.
To check \[item:unprov-4\], take an arbitrary ${B \in \mathcal{B}_\alpha}$. If ${B \cap (X_\alpha\setminus I_\alpha)}$ is infinite then the choice of $G$ guarantees that ${B \cap G}$ is infinite and therefore $B\cap B_\alpha$ as well. If ${B \cap (X_\alpha\setminus I_\alpha)}$ is finite then ${(X_\alpha \setminus I_\alpha})\setminus B$ is infinite and $G$ has an infinite co-intersection with it. Therefore ${\mathbb{N}\setminus (B \cup B_\alpha)}$ is infinite.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
author:
- 'Y. Apertet'
- 'H. Ouerdane'
- 'O. Glavatskaya'
- 'C. Goupil'
- 'P. Lecoeur'
title: 'Reply to the Comment by S. Su '
---
In our Letter [@Apertet1], we showed how optimal working conditions for thermoelectric generators (TEG) with realistic thermal coupling to heat baths may be achieved. In their comment [@CommentSu], Su and co-workers claim that our results and conclusions [@Apertet1] are flawed because:\
$i/$ the maximum output power of the TEG $P_{\rm max}$ is a monotonically increasing function of the ratio of the heat exchangers’ thermal conductances to the open-circuit thermal conductance of the TEG: $K_{\rm contact}/K_{_{I=0}}$, which is in contradiction with Fig. 2 of Ref. [@Apertet1]\
$ii/$ $K_{_{I=0}}$ could not be used as a variable.\
$iii/$ Eq. (21) in Ref. [@Apertet1] is erroneous.\
We believe that the first two observations stem from a misunderstanding of the condition for which power and efficiency optimizations are performed. The third remark raises a point that we must clarify here. The notations used in the following discussion are the same as those adopted in Ref. [@Apertet1].
On the choice of the variables
==============================
In Ref. [@Apertet1], we study the impact of non-ideal heat exchangers on the TEG performances. We show in particular that the thermal matching between the heat exchangers’ thermal conductances, $K_{\rm contact}$, and that of the TEG, $K_{\rm TEG}$, yields power maximization. We highlight however that this condition only applies to cases where $K_{\rm contact}$ is fixed: the variable used to perform the optimization is $K_{_{I=0}}$ only. Therefore, though the maximum power $P_{\rm max}$ in Fig. (2) is plotted against the normalized dimensionless ratio $K_{_{I=0}}/K_{\rm contact}$, this ratio cannot be considered as the optimization variable: only $K_{_{I=0}}$ is. We thus rebut point $i/$ of Su and co-worker’s criticism.
We analyze power maximization for TEGs with different figures of merit $Z\overline{T}$ using $K_{_{I=0}}$ as a variable. For a fixed value of $Z\overline{T}$, we adapt the value of the Seebeck coefficient $\alpha$ for each value of $K_{_{I=0}}$: $\alpha = \sqrt{ Z R K }$. This method is adapted from the article from Nemir and Beck [@nemirbeck], in which the authors discuss the significance of the figure of merit. Contrary to the statement of Su and co-workers in their comment [@CommentSu], one should not consider that $\alpha^2 / R$ is a fixed parameter; so using $K_{_{I=0}}$ as a variable is not a mistake. The variation of $\alpha$ and $K_{_{I=0}}$ for a constant $Z\overline{T}$ allows to put forth the fact that the *individual* value of the different parameters may be as important for a TEG’s performance as the value of the *global* figure of merit $Z\overline{T}$ when realistic thermal contacts are accounted for. We thus prove wrong point $ii/$ of Su and co-worker’s criticism.
On the definition and optimisation of the efficiency
====================================================
In Ref. [@Apertet1], we used approximations in order to make the derivations tractable. The main hypothesis is that the electrical power delivered by the TEG is very small compared to the average thermal current $I_{Q}$, i.e., the efficiency is low. This allows to use a thermal anologue of the potential divider formula to determine the temperature difference experienced by the TEG. We then get $I_{Q} \approx I_{Q_{\rm in}} \approx I_{Q_{\rm out}}$. Consequently, in order to keep the expression of the efficiency $\eta$ straightforward we defined it as $\eta = P/I_{Q}$ instead of $\eta = P/I_{Q_{\rm in}}$. Although this derivation is coherent with the hypothesis made before, a statement that this expression is only a reasonable approximation is missing. We thank the Authors of the Comment to point out this omission.
In order to evaluate the difference between this approximation and the correct definition for the efficiency, we rewrite Eq. (8) of Ref. [@CommentSu]:
$$\frac{P}{I_{Q}} = \frac{P}{ \left( \dot{Q}_{\rm in} + \dot{Q}_{\rm out} \right) / 2} = \frac{\eta}{ 1 - \eta / 2}$$
The correction factor is then $1/(1 - \eta/2)$, which reduces to $1$ when $\eta$ is small compared to $1$.
Finally to demonstrate that the use of this approximation does not change the results and conclusion of Ref. [@Apertet1], we check that the analytical expressions for $m_{_{\eta=\eta_{\rm max}}}$ and $m_{_{P=P_{\rm max}}}$ given by Eqs. (14) and (19) of Ref. [@Apertet1] are in excellent agreement with the exact numerical results obtained using Eq. (9) of Ref. [@Apertet1]. The two parameters are ahown as functions of $K_{\rm contact}$ scaled to $K_{_{I=0}}$ in Fig. \[fig1\]. The values of the thermoelectric parameters for this example are taken from Ref. [@nemirbeck]. We see that the analytical expressions match perfectly the values for both $m_{_{\eta=\eta_{\rm max}}}$ and $m_{_{P=P_{\rm max}}}$ obtained numerically: all the results and conclusions in Ref. [@Apertet1] remain valid.
Summary
=======
We showed that the criticisms made by Su and co-workers concerning the choice of the variables are unsubstantiated, and we provided additional justification for the use of the approximation made in Eq. (21) of Ref. [@Apertet1] to derive the electrical condition leading to performance optimization.
[10]{}
.
.
.
| {
"pile_set_name": "ArXiv"
} | ArXiv |
---
abstract: |
We analyze the contributions of the one–pion–pole (OPP) exchange, caused by strong low–energy interactions, and the pseudoscalar interaction beyond the Standard Model (BSM) to the correlation coefficients of the neutron $\beta^-$–decays for polarized neutrons, polarized electrons and unpolarized protons. The strength of contributions of pseudoscalar interactions is defined by the effective coupling constant $C_{ps} = C^{(\rm OPP)}_{ps} + C^{(\rm BSM)}_{ps}$. We show that the contribution of the OPP exchange is of order $C^{(\rm
OPP)}_{ps} \sim - 10^{-5}$. The effective coupling constant $C^{(\rm
BSM)}_{ps}$ of the pseudoscalar interaction BSM can be in principle complex. Using the results, obtained by Gonzaĺez-Alonso [*et al.*]{}( Prog. Part. Nucl. Phys. [**104**]{}, 165 (2019)) we find that the values of the real and imaginary parts of the effective coupling constant $ C^{(\rm BSM)}_{ps}$ are constrained by $ - 3.5 \times
10^{-5} < {\rm Re}\,C^{(\rm BSM)}_{ps} < 0$ and ${\rm Im}\,C^{(\rm
BSM)}_{ps} < - 2.3 \times 10^{-5}$, respectively. The obtained results can be used as a theoretical background for experimental searches of contributions of interactions BSM in asymmetries of the neutron $\beta^-$–decays with a polarized neutron, a polarized electron and an unpolarized proton at the level of accuracy of a few parts of $10^{-5}$ or even better (Abele, Hyperfine Interact. [ **237**]{}, 155 (2016)).
author:
- 'A. N. Ivanov'
- 'R. Höllwieser'
- 'N. I. Troitskaya'
- 'M. Wellenzohn'
- 'Ya. A. Berdnikov'
title: |
Precision analysis\
of pseudoscalar interactions in neutron beta decays
---
Introduction {#sec:introduction}
============
Nowadays the neutron lifetime and correlation coefficients of the neutron $\beta^-$-decays for polarized neutrons, polarized electrons and unpolarized protons are calculated within the Standard Model (SM) at the level of $10^{-3}$ including the radiative corrections of order $O(\alpha/\pi)$ of and corrections caused by the weak magnetism and proton recoil of order $O(E_e/M)$ [@Bilenky1959]–[@Ivanov2019a], where $\alpha$, $E_e$ and $M$ are the fine–structure constant [@PDG2018], an electron energy and the nucleon mass, respectively. Such a SM theoretical background has allowed to make steps forwards investigations of contributions of interactions beyond the SM (BSM) of order $10^{-4}$ or even smaller [@Abele2016]. The analysis of interactions beyond the $V - A$ effective theory of weak interactions [@Feynman1958; @Sudarshan1958; @Marshak1959; @Nambu1960; @Marshak1969] (see also [@Shekhter1959a; @Shekhter1959b]) in the neutron $\beta^-$–decays with different polarizations of massive fermions has a long history and started in 50th of the 20th century and is continuing at present time [@Lee1956]–[@Severijns2019] (see also [@Gudkov2006; @Ivanov2013; @Ivanov2017d]). The most general form of the Lagrangian of interactions BSM has been written in [@Lee1956]-[@Severijns2006], including non–derivative vector $\bar{\psi}_p\gamma_{\mu}\psi_n$, axial–vector $\bar{\psi}_p\gamma_{\mu}\gamma^5\psi_n$, scalar $\bar{\psi}_p\psi_n$, pseudoscalar $\bar{\psi}_p\gamma^5 \psi_n$ and tensor $\bar{\psi}_p\sigma_{\mu\nu} \psi_n$ nucleon currents coupled to corresponding lepton currents in the form of local nucleon–lepton current–current interactions, where $\{1, \gamma_{\mu},
\gamma_{\mu}\gamma^5, \gamma^5, \sigma_{\mu\nu}\}$ are the Dirac matrices [@Itzykson1980], With respect to $G$–parity transformations [@Lee1956a], i.e. $G = C\,e^{\,i \pi I_2}$, where $C$ and $I_2$ are the charge conjugation and isospin operators [@Itzykson1980], the vector, axial–vector, pseudoscalar and tensor nucleon currents are $G$–even and the scalar nucleon current is $G$–odd. According to the $G$–transformation properties of hadronic currents, Weinberg divided hadronic currents into two classes, which are $G$–even first class and $G$–odd second class currents [@Weinberg1958], respectively. Thus, following Weinberg’s classification the non–derivative vector, axial–vector, pseudoscalar and tensor nucleon currents in the interactions BSM, introduced in [@Lee1956]–[@Severijns2006], are the first class currents, whereas the non–derivative scalar nucleon current is the second class one (see also [@Ivanov2018]). The analysis of superallowed $0^+
\to 0^+$ nuclear beta transitions by Hardy and Towner [@Hardy2015] and González–Alonso [*et al.*]{} [@Severijns2019] has shown that the phenomenological coupling constants of non–derivative scalar current–current nucleon–lepton interaction is of order $10^{-5}$ or even smaller. This agrees well with estimates of contributions of the second class currents, caused by derivative scalar $\partial^{\mu}(\bar{\psi}_p\psi_n)$ and pseudotensor $\partial^{\nu}(\bar{\psi}_p\sigma_{\mu\nu}\gamma^5 \psi_n)$ nucleon currents proposed by Weinberg [@Weinberg1958], to the neutron lifetime and correlation coefficients of the neutron $\beta^-$–decays carried out by Gardner and Plaster [@Gardner2001; @Gardner2013] and Ivanov [*et al.*]{} [@Ivanov2017d; @Ivanov2019]. The contemporary experimental sensitivities $10^{-4}$ or even better [@Abele2016] of experimental analyses of parameters of neutron $\beta^-$–decays (see, for example, [@Abele2018; @Seng2018; @Seng2018a]) demand a theoretical background for the neutron lifetime and correlation coefficients of the neutron $\beta^-$–decays with different polarizations of massive fermions at the level of $10^{-5}$ [@Ivanov2017b; @Ivanov2017d; @Ivanov2019; @Ivanov2019a]. As has been shown in [@Cirigliano2010]–[@Cirigliano2013a] in the linear approximation the contributions of vector and axial–vector interactions BSM can be absorbed by the matrix element $V_{ud}$ of the Cabibbo–Kobayashi–Maskawa (CKM) mixing matrix and by the axial coupling constant $\lambda$ (see also [@Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019]). As a result, taking into account the constraints on the scalar interaction [@Hardy2015] and [@Severijns2019] the contributions of interactions BSM to the neutron $\beta^-$–decay can be induced only by a tensor nucleon current [@Pattie2013; @Ivanov2018c]. As we show below the contribution of the one–pion–pole (OPP) exchange to the correlation coefficients of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton is of order $10^{-5}$. This is commensurable with the contribution of the isospin breaking correction to the vector coupling constant of the neutron $\beta^-$–decay calculated by Kaiser [@Kaiser2001] within the heavy baryon chiral perturbation theory (HB$\chi$PT). However, unlike Kaiser’s correction the contribution of the OPP exchange can be screened by the contribution of the pseudoscalar interaction BSM.
This paper is addressed to the analysis of contributions of the OPP exchange, caused by strong low–energy interactions, and the pseudoscalar interaction BSM introduced in [@Lee1956]–[@Severijns2006] to the neutron lifetime and correlation coefficients of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and unpolarized proton. The analysis of contributions of pseudoscalar interactions to the electron–energy and angular distribution of the neutron $\beta^-$–decay for a polarized neutron and unpolarized electron and proton has a long history [@Harrington1960]–[@Hayen2018] (see also [@Wilkinson1982; @Severijns2019]). For example the Fierz–like interference term [@Fierz1937], induced by pseudoscalar interactions, can be recognized in the electron–energy and angular distributions calculated in [@Harrington1960]–[@Hayen2018] (see also [@Wilkinson1982; @Severijns2019]). The contributions of the pseudoscalar interactions to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decay for a polarized neutron and unpolarized electron and proton can be, in principle, extracted from the electron–energy and angular distributions obtained by Harrington [@Harrington1960] (see Eqs.(9) – (13) of Ref.[@Harrington1960]) and Holstein [@Holstein1974] (see Appendix B of Ref.[@Holstein1974]) (see also section \[sec:schluss\] of this paper). In our work in addition to the results obtained in [@Harrington1960]–[@Hayen2018] (see also [@Wilkinson1982; @Severijns2019]) we calculate the contributions of pseudoscalar interactions to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays, caused by correlations with the electron spin. The analyze of contributions of pseudoscalar interactions to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and unpolarized proton, carried out in this paper, completes the investigations of contributions of interactions BSM to the electron–energy and angular distributions, which we have performed in [@Ivanov2017b; @Ivanov2017d; @Ivanov2019], where we have calculated i) the complete set of corrections of order $10^{-3}$, caused by radiative corrections of order $O(\alpha/\pi)$ and the weak magnetism and proton recoil corrections of order $O(E_e/M)$, and ii) contributions of vector, axial–vector, scalar and tensor interactions BSM introduced in [@Lee1956]–[@Severijns2006].
The paper is organized as follows. In section \[sec:amplitude\] we write down the amplitude of the neutron $\beta^-$–decay by taking into account the contributions of the OPP exchange and the pseudoscalar interaction BSM only. We analyze the contributions of energy independent corrections to the pseudoscalar form factor of the nucleon defined by the Adler-Dothan-Wolfenstein (ADM) term [@Adler1966; @Wolfenstein1970] and chiral corrections calculated within the HB$\chi$PT [@Bernard1995; @Bernard1996; @Kaiser2003]. We show that the ADM–term and chiral corrections, calculated in the two–loop approximation within the HB$\chi$PT by Kaiser [@Kaiser2003], are able in principle to induce sufficiently small real contributions to phenomenological coupling constants of the pseudoscalar interaction BSM of a neutron–proton pseudoscalar density coupled to a left–handed leptonic current. In section \[sec:spectrum\] we discuss the contributions to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays caused by the OPP exchange and the pseudoscalar interaction BSM. The distribution is calculated for a polarized neutron, a polarized electron and an unpolarized proton. Using the results, obtained in [@Bhattacharya2012; @Severijns2019; @Gonzalez-Alonso2014; @Gonzalez-Alonso2016] we estimate the phenomenological coupling constants of the pseudoscalar interactions BSM. We adduce the results in Table I. In section \[sec:schluss\] we discuss the obtained results, which can be used for experimental analyses of the neutron $\beta^-$–decays with experimental accuracies of about a few parts of $10^{-5}$ [@Abele2016]. Since the complete set of contributions of order $10^{-3}$, including the radiative corrections of order $O(\alpha/\pi)$ and corrections of order $O(E_0/M)$, caused by the weak magnetism and proton recoil, are calculated at the neglect of contributions of order $O(\alpha E_0/\pi
M) \sim 10^{-6}$ and $O(E^2_0/M^2) \sim 10^{-6}$ [@Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019], the results obtained in this paper should be tangible and important for a correct analysis of experimental data on searches of contributions of interactions BSM with an accuracy of a few parts of $10^{-5}$. We give also a comparative analysis of the results obtained in this work with those in [@Wilkinson1982; @Harrington1960]–[@Hayen2018]. This allows us to argue that the corrections, caused by pseudoscalar interactions, calculated for the correlation coefficients of the neutron $\beta^-$–decays, induced by correlations of the electron spin with the neutron spin and 3-momenta of decay fermions with standard correlation structures introduced by Jackson [*et al.*]{} [@Jackson1957], are fully new. Moreover all terms in Eq.(\[eq:A.6\]) with correlation structures beyond the standard ones by Jackson [*et al.*]{} [@Jackson1957] and proportional to the effective coupling constants $C'_{ps}$ and $C''_{ps}$ were never calculated in literature. In the Appendix we give a detailed calculation of the contributions of pseudoscalar interactions caused by the OPP exchange and BSM to the correlation coefficients of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton, completing the analysis of contributions of interactions BSM to the correlation coefficients of the neutron $\beta^-$–decays carried out in [@Ivanov2017b; @Ivanov2017d; @Ivanov2019].
Amplitude of the neutron $\beta^-$–decay with contributions of OPP exchange and pseudoscalar interaction BSM {#sec:amplitude}
============================================================================================================
Since the expected order of contributions of pseudoscalar interactions of about $10^{-5}$, we take them into account in the linear approximation additively to the corrections of order $10^{-4} -
10^{-3}$ calculated in [@Bilenky1959]–[@Severijns2019]. In such an approximation and following [@Ivanov2013; @Ivanov2017d; @Ivanov2019] the amplitude of the neutron $\beta^-$–decay we take in the form $$\begin{aligned}
\label{eq:1}
M(n \to p e^- \bar{\nu}_e) &=& -
\frac{G_F}{\sqrt{2}}\,V_{ud}\,\Big\{\langle p(\vec{k}_p,
\sigma_p)|J^{(+)}_{\mu}(0)|n(\vec{k}_n,
\sigma_n)\rangle\,\big[\bar{u}_e(\vec{k}_e, \sigma_e)\gamma^{\mu}(1
- \gamma^5)v_{\bar{\nu}}(\vec{k}_{\bar{\nu}}, +
\frac{1}{2})\big]\nonumber\\ &&\hspace{0.50in}~ +
\bar{u}_p(\vec{k}_p, \sigma_p)\gamma^5 u_n(\vec{k}_n,
\sigma_n)\,\big[\bar{u}_e(\vec{k}_e, \sigma_e) (C_p + \bar{C}_P
\gamma^5)v_{\bar{\nu}}(\vec{k}_{\bar{\nu}}, +
\frac{1}{2})\big]\Big\},\end{aligned}$$ where $G_F$ and $V_{ud}$ are the Fermi couping constant and the Cabibbo–Kobayashi–Maskawa (CKM) matrix element [@PDG2018]. Then, $\langle p(\vec{k}_p, \sigma_p)|J^{(+)}_{\mu}(0)|n(\vec{k}_n,
\sigma_n)\rangle$ is the matrix element of the charged hadronic current $J^{(+)}_{\mu}(0) = V^{(+)}_{\mu}(0) - A^{(+)}_{\mu}(0)$, where $V^{(+)}_{\mu}(0)$ and $A^{(+)}_{\mu}(0)$ are the charged vector and axial–vector hadronic currents [@Feynman1958; @Nambu1960; @Marshak1969]. The fermions in the initial and final states are described by Dirac bispinor wave functions $u_n$, $u_p$, $u_e$ and $v_{\bar{\nu}}$ of free fermions [@Ivanov2013; @Ivanov2014]. In the second term of Eq.(\[eq:1\]) we take into account the contribution of the pseudoscalar interaction BSM [@Lee1956]–[@Severijns2006] with two complex phenomenological coupling constants $C_P$ and $\bar{C}_P$ in the notation of [@Ivanov2013; @Ivanov2017d; @Ivanov2019].
For the analysis of contributions of pseudoscalar interactions to the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton we define the matrix element $\langle p(\vec{k}_p, \sigma_p)|J^{(+)}_{\mu}(0)|n(\vec{k}_n,
\sigma_n)\rangle$ as follows $$\begin{aligned}
\label{eq:2}
\langle p(\vec{k}_p, \sigma_p)|J^{(+)}_{\mu}(0)|n(\vec{k}_n,
\sigma_n)\rangle = \bar{u}_p(\vec{k}_p,
\sigma_p)\Big(\gamma_{\mu}(1 + \lambda \gamma^5) + \frac{2 M \lambda
\, q_{\mu}}{m^2_{\pi} - q^2 - i0}\gamma^5\Big)u_n(\vec{k}_n,
\sigma_n),\end{aligned}$$ where $\lambda$ is the axial coupling constant with recent experimental value $\lambda = - 1.27641(45)_{\rm stat.}(33)_{\rm
syst.}$ [@Abele2018]. The first term in Eq.(\[eq:1\]) is written in agreement with the standard $V - A$ effective theory of weak interactions [@Feynman1958; @Nambu1960; @Marshak1969] (see also [@Shekhter1959a; @Shekhter1959b]). The term proportional to $q_{\mu} \gamma^5$ defines the contribution of the OPP exchange, caused by strong low–energy interactions (see also [@Nambu1960]) with the $\pi^-$–meson mass $m_{\pi} =
139.57061(24)\,{\rm MeV}$ [@PDG2018] and $q = k_p - k_n = - k_e -
k_{\bar{\nu}}$ is a 4–momentum transfer. The OPP contribution is required by conservation of the charged hadronic axial–vector current in the chiral limit $m_{\pi} \to 0$ [@Nambu1960].
In the more general form the matrix element of the hadronic axial–vector current can be taken in the form accepted in the HB$\chi$PT [@Bernard1995; @Bernard1996; @Kaiser2003]. This gives $$\begin{aligned}
\label{eq:3}
\langle p(\vec{k}_p, \sigma_p)|A^{(+)}_{\mu}(0)|n(\vec{k}_n,
\sigma_n)\rangle = \bar{u}_p(\vec{k}_p,
\sigma_p)\Big(\gamma_{\mu}\,G_A(q^2) +
\frac{q_{\mu}}{2M}\,G_P(q^2)\Big)\,\gamma^5 u_n(\vec{k}_n, \sigma_n),\end{aligned}$$ where $G_A(q^2)$ and $G_P(q^2)$ are the axial–vector form factor and the induced pseudoscalar form factor, respectively, at $0 \le q^2 \le
\Delta^2$ for the neutron $\beta^-$–decay with $\Delta = m_n -
m_p$. The invariant 4–momentum transfer squared $q^2$ vanishes, i.e. $q^2 = 0$, at the kinetic energy of the proton $T_p = E_p - m_p =
\Delta^2/2 m_n$. In the chiral limit $m_{\pi} \to 0$ because of conservation of the charged hadronic axial–vector current [@Nambu1960] the form factors $G_A(q^2)$ and $G_P(q^2)$ are related by $G_P(q^2) = - (4M^2/q^2) G_A(q^2)$. In turn, for a finite pion mass the pseudoscalar form factor $G_P(q^2)$ has been calculated in the two–loop approximation within HB$\chi$PT by Kaiser [@Kaiser2003]. A precision analysis of the induced pseudoscalar form factor in the proton weak interactions has been also carried out by Gorringe and Fearing [@Gorringe2004].
Pseudoscalar interaction BSM as induced by corrections to the pseudoscalar form factor, caused by strong low–energy interactions
--------------------------------------------------------------------------------------------------------------------------------
According to [@Bernard1995], the axial–vector form factor $G_A(q^2)$ can be rather good parameterized by a dipole form (see also [@Liesenfeld1999]) $$\begin{aligned}
\label{eq:4}
G_A(q^2) = \frac{g_A}{\big(1 + q^2/M^2_A\big)^2} = g_A\Big(1 -
\frac{1}{6}\,\langle r^2_A\rangle q^2 + \ldots\Big),\end{aligned}$$ where $g_A = - \lambda$ is the axial–coupling constant, and $M_A$ is the cut–off mass related to the mean square axial radius of the nucleon $\langle r^2_A\rangle$ as $\langle r^2_A\rangle = 12/M^2_A =
0.403(29)\,{\rm fm^2}$ with $M_A = 1.077(39)\,{\rm GeV}$ extracted from charged pion electroproduction experiments [@Liesenfeld1999]. In turn, the cut–off mass $M_A =
1.026(17)\,{\rm GeV}$ extracted from (quasi)elastic neutrino and antineutrino scattering experiments [@Liesenfeld1999] gives $\langle r^2_A\rangle = 12/M^2_A = 0.440(16)\,{\rm fm^2}$. In the approximation Eq.(\[eq:4\]) the pseudoscalar form factor $G_P(q^2)$ acquires the following form [@Bernard1995] (see also [@Gorringe2004]) $$\begin{aligned}
\label{eq:5}
\frac{1}{2M\,}G_P(q^2) = \frac{2M g_A}{m^2_{\pi} - q^2 - i0} -
\frac{1}{3}\,g_A M \langle r^2_A\rangle,\end{aligned}$$ where the correction to the OPP exchange is the Adler–Dothan–Wolfenstein (ADW) term [@Adler1966; @Wolfenstein1970]. The ADW–term induces the BSM–like pseudoscalar interaction with the coupling constants $$\begin{aligned}
\label{eq:6}
C^{(\rm ADW)}_P = - \bar{C}^{(\rm ADW)}_P = - \frac{1}{3}\,\lambda
\langle r^2_A\rangle\,m_e M = 2.1\times 10^{-3}.\end{aligned}$$ According to Eq.(\[eq:11\]), this gives the contribution to the correlation coefficients of the neutron $\beta^-$–decays equal to ${\rm Re}\,C^{(\rm BSM)}_{ps} = C^{(\rm ADW)}_{ps} = - 4.9 \times
10^{-7}$. Using the results, obtained by Kaiser [@Kaiser2003] (see Eq.(7) of Ref.[@Kaiser2003]) in the two–loop approximation in the HB$\chi$PT, the induced BSM–like pseudoscalar coupling constants are equal to $$\begin{aligned}
\label{eq:7}
C^{(\rm K)}_P = - \bar{C}^{(\rm K)}_P = \frac{m_e m^2_{\pi}M}{32
\pi^4 f^4_{\pi}}\,\zeta_0 = 4.1\times 10^{-5}\,\zeta_0,\end{aligned}$$ where $f_{\pi} = 92.4\,{\rm MeV}$ is the charged pion leptonic (or PCAC) constant [@Bernard1995; @Kaiser2003]. Since $|\zeta_0|\sim
1$ [@Kaiser2003], we get $|C_P| = |\bar{C}_P|\sim 4.1\times
10^{-5}$. The contribution of $C^{(\rm K)}_P = - \bar{C}^{(\rm K)}_P$ to the coupling constant ${\rm Re}\,C^{(\rm BSM)}_{ps}$ (see Eq.(\[eq:11\])) is of order $|{\rm Re}\,C^{(\rm BSM)}_{ps}| \sim
9.6\times 10^{-9}$. This means that the SM strong low–energy interactions are able to induce the BSM–like pseudoscalar interaction with real coupling constants, the contributions of which are much smaller than the current experimental sensitivity of the neutron $\beta^-$–decays [@Abele2016]. Below we consider a more general pseudoscalar interaction BSM with complex phenomenological coupling constants $C_P$ and $\bar{C}_P$ such as $C_P \neq - \bar{C}_P$.
Non–relativistic approximation for the amplitude of the neutron $\beta^-$–decay Eq.(\[eq:1\])
---------------------------------------------------------------------------------------------
In the non–relativistic approximation for the neutron and proton the amplitude of the neutron $\beta^-$–decay in Eq.(\[eq:1\]) takes the form $$\begin{aligned}
\label{eq:8}
M(n \to p e^- \bar{\nu}_e) &=& - \frac{G_F}{\sqrt{2}}\,V_{ud}2 M
\Big\{[\varphi^{\dagger}_p\varphi_n][\bar{u}_e\gamma^0(1 -
\gamma^5)v_{\bar{\nu}}] -
\lambda[\varphi^{\dagger}_p\vec{\sigma}\,\varphi_n]\cdot [\bar{u}_e
\vec{\gamma}\,(1 - \gamma^5)v_{\bar{\nu}}]\nonumber\\ && +
\lambda\,\frac{m_e}{m^2_{\pi}}[\varphi^{\dagger}_p(\vec{\sigma}\cdot
\vec{k}_p)\varphi_n][\bar{u}_e(1 - \gamma^5)v_{\bar{\nu}}] -
\frac{1}{2 M}[\varphi^{\dagger}_p(\vec{\sigma}\cdot
\vec{k}_p)\varphi_n][\bar{u}_e(C_p + \bar{C}_P
\gamma^5)v_{\bar{\nu}}]\Big\},\end{aligned}$$ where $\varphi_j$ for $j = p,n$ are the Pauli spinorial wave functions of non–relativistic neutron and proton, and $\vec{k}_p = - \vec{k}_e
- \vec{k}_{\bar{\nu}}$ is a 3–momentum of the proton.
Electron–energy and angular distribution of the neutron $\beta^-$–decay for polarized neutron, polarized electron, and unpolarized proton {#sec:spectrum}
=========================================================================================================================================
The electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton has been written by Jackson [*et al.*]{} [@Jackson1957]. It reads $$\begin{aligned}
\label{eq:9}
\hspace{-0.15in}&&\frac{d^5 \lambda_n(E_e, \vec{k}_e,
\vec{k}_{\bar{\nu}}, \vec{\xi}_n, \vec{\xi}_e)}{dE_e d\Omega_e
d\Omega_{\bar{\nu}}} = (1 + 3
\lambda^2)\,\frac{G^2_F|V_{ud}|^2}{32\pi^5}\,(E_0 - E_e)^2
\,\sqrt{E^2_e - m^2_e}\, E_e\,F(E_e, Z = 1)\,\zeta(E_e)\,\Big\{1 +
b\,\frac{m_e}{E_e}\nonumber\\
\hspace{-0.15in}&& + a(E_e)\,\frac{\vec{k}_e\cdot
\vec{k}_{\bar{\nu}}}{E_e E_{\bar{\nu}}} +
A(E_e)\,\frac{\vec{\xi}_n\cdot \vec{k}_e}{E_e} + B(E_e)\,
\frac{\vec{\xi}_n\cdot \vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} +
K_n(E_e)\,\frac{(\vec{\xi}_n\cdot \vec{k}_e)(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{E^2_e E_{\bar{\nu}}}+
Q_n(E_e)\,\frac{(\vec{\xi}_n\cdot \vec{k}_{\bar{\nu}})(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{ E_e E^2_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& + D(E_e)\,\frac{\vec{\xi}_n\cdot (\vec{k}_e\times
\vec{k}_{\bar{\nu}})}{E_e E_{\bar{\nu}}} + G(E_e)\,\frac{\vec{\xi}_e
\cdot \vec{k}_e}{E_e} + H(E_e)\,\frac{\vec{\xi}_e \cdot
\vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} + N(E_e)\,\vec{\xi}_n\cdot
\vec{\xi}_e + Q_e(E_e)\,\frac{(\vec{\xi}_n\cdot \vec{k}_e)(
\vec{k}_e\cdot \vec{\xi}_e)}{(E_e + m_e) E_e}\nonumber\\
\hspace{-0.15in}&& + K_e(E_e)\,\frac{(\vec{\xi}_e\cdot \vec{k}_e)(
\vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{(E_e + m_e)E_e E_{\bar{\nu}}} +
R(E_e)\,\frac{\vec{\xi}_n\cdot(\vec{k}_e \times \vec{\xi}_e)}{E_e} +
L(E_e)\,\frac{\vec{\xi}_e\cdot(\vec{k}_e \times
\vec{k}_{\bar{\nu}})}{E_eE_{\bar{\nu}}} - 3\,\frac{E_e}{M}\,\frac{1
- \lambda^2}{1 + 3 \lambda^2}\,\Big(\frac{(\vec{k}_e\cdot
\vec{k}\,)^2}{E^2_e E^2} - \frac{1}{3}\,\frac{k^2_e}{E^2_e}\Big)
\nonumber\\
\hspace{-0.3in}&& + 3\,\frac{1 - \lambda^2}{1 + 3 \lambda^2}\,
\frac{m_e}{M}\,\Big(\frac{(\vec{\xi}_e\cdot
\vec{k}_{\nu})(\vec{k}_e\cdot \vec{k}_{\nu})}{E_e E^2_{\nu}} -
\frac{1}{3}\,\frac{\vec{\xi}_e\cdot \vec{k}_e}{E_e}\,\Big) +
3\,\frac{1 - \lambda^2}{1 + 3 \lambda^2}\,
\frac{1}{M}\,\Big(\frac{(\vec{\xi}_e\cdot \vec{k}_e)(\vec{k}_e\cdot
\vec{k}_{\nu})^2}{(E_e + m_e)E_e E^2_{\nu}} - \frac{1}{3}\,(E_e -
m_e)\,\frac{\vec{\xi}_e\cdot \vec{k}_e}{E_e}\,\Big) \Big\},\end{aligned}$$ where we have followed the notation [@Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019]. The last three terms in Eq.(\[eq:9\]) are caused by the contributions of the proton recoil calculated to order $O(E_e/M)$ [@Gudkov2006; @Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019]. Then, $\vec{\xi}_n$ and $\vec{\xi}_e$ are unit polarization vectors of the neutron and electron, respectively, $d\Omega_e$ and $d\Omega_{\bar{\nu}}$ are infinitesimal solid angels in the directions of electron $\vec{k}_e$ and antineutrino $\vec{k}_{\bar{\nu}}$ 3–momenta, respectively, $E_0 = (m^2_n - m^2_p
+ m^2_e)/2m_n = 1.2926\,{\rm MeV}$ is the end–point energy of the electron spectrum, $F(E_e, Z = 1)$ is the relativistic Fermi function equal to [@Blatt1952]–[@Konopinski1966] (see also [@Wilkinson1982; @Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019]) $$\begin{aligned}
\label{eq:10}
\hspace{-0.3in}F(E_e, Z = 1 ) = \Big(1 +
\frac{1}{2}\gamma\Big)\,\frac{4(2 r_pm_e\beta)^{2\gamma}}{\Gamma^2(3 +
2\gamma)}\,\frac{\displaystyle e^{\,\pi
\alpha/\beta}}{(1 - \beta^2)^{\gamma}}\,\Big|\Gamma\Big(1 + \gamma +
i\,\frac{\alpha }{\beta}\Big)\Big|^2,\end{aligned}$$ where $\beta = k_e/E_e = \sqrt{E^2_e - m^2_e}/E_e$ is the electron velocity, $\gamma = \sqrt{1 - \alpha^2} - 1$, $r_p$ is the electric radius of the proton. In the numerical calculations we use $r_p =
0.841\,{\rm fm}$ [@Pohl2010]. The function $\zeta(E_e)$ contains the contributions of radiative corrections of order $O(\alpha/\pi)$ and corrections from the weak magnetism and proton recoil of order $O(E_e/M)$, taken in the form used in [@Gudkov2006; @Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019]. Then, $b$ is the Fierz interference term defined by the contributions of interactions beyond the SM [@Fierz1937]. The analytical expressions for the correlation coefficients $a(E_e)$, $A(E_e)$ and so on, calculated within the SM with the account for radiative corrections of order $O(\alpha/\pi)$ and corrections caused by the weak magnetism and proton recoil of order $O(E_e/M)$ together with the contributions of Wilkinson’s corrections [@Wilkinson1982], are given in [@Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019].
Corrections to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays caused by pseudoscalar interactions
---------------------------------------------------------------------------------------------------------------------------------------------------------------
In the Appendix we calculate the contributions of the OPP exchange and the pseudoscalar interaction BSM to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton. The corrections to the correlation coefficients and the correction to the electron–energy and angular distribution are given in the Appendix in Eqs.(\[eq:A.5\]) and (\[eq:A.6\]), respectively. The strength of these corrections (see Eq.(\[eq:A.5\])) is defined by the effective coupling constants $C'_{ps}$ and $C''_{ps}$, which are the real and imaginary parts of the effective coupling constant $C_{ps}$ given by $$\begin{aligned}
\label{eq:11}
C_{ps} &=& C^{(\rm OPP)}_{ps} + C^{(\rm BSM)}_{ps} = C'_{ps} +
i\,C''_{ps},\nonumber\\ C^{(\rm OPP)}_{ps} &=& \frac{2\lambda}{1 +
3\lambda^2}\,\frac{m_e }{m^2_{\pi}}\,E_0 = - 1.47\times
10^{-5},\nonumber\\ C^{(\rm BSM)}_{ps} &=& - \frac{1}{1 +
3\lambda^2}\, \frac{E_0}{2M}\,(C_P - \bar{C}_P) = - 1.17\times
10^{-4}\,(C_P - \bar{C}_P),\nonumber\\ C'_{ps} &=& {\rm Re}\,C_{ps} =
C^{(\rm OPP)}_{ps} + {\rm Re}\, C^{(\rm
BSM)}_{ps},\nonumber\\ C''_{ps} &=& {\rm Im}\,C_{ps} = {\rm
Im}\,C^{(\rm BSM)}_{ps},\end{aligned}$$ where $C^{(\rm OPP)}_{ps}$ and $C^{(\rm BSM)}_{ps}$ are the effective coupling constants caused by the OPP exchange and the pseudoscalar interaction BSM, respectively. The numerical values are calculated for $\lambda = - 1.27641$ [@Abele2018], $m_e = 0.5110\,{\rm MeV}$, $m_{\pi} = 139.5706\,{\rm MeV}$ [@PDG2018], $E_0 = (m^2_n - m^2_p
+ m^2_e)/2m_n = 1.2926\,{\rm MeV}$ and $M = (m_n + m_p)/2 =
938.9188\,{\rm MeV}$ [@PDG2018], respectively. According to our analysis (see Eqs.(\[eq:6\]) and (\[eq:7\])), a real part of the phenomenological coupling constant $C^{(\rm BSM)}_{ps}$ can be partly induced by the SM strong low–energy interactions through the ADM–term (see Eq.(\[eq:6\])) and Kaiser’s two–loop corrections, calculated within the HB$\chi$PT (see Eq.(\[eq:7\])).
The corrections, caused by pseudoscalar interactions (see Eq.(\[eq:A.5\]) and Eq.(\[eq:A.6\])), to the electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton, taken together with the electron–energy and angular distributions calculated in [@Gudkov2006; @Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019] can be used as a theoretical background for experimental searches of contributions of interactions BSM of order $10^{-4}$ or even smaller [@Abele2016].
Estimates of the real and imaginary parts of the phenomenological coupling constant $C_P - \bar{C}_P$
-----------------------------------------------------------------------------------------------------
According to [@Cirigliano2013], the phenomenological coupling constant $C_P - \bar{C}_P$ can be defined as follows $$\begin{aligned}
\label{eq:12}
C_P - \bar{C}_P = 2\, g_P\,\epsilon_P,\end{aligned}$$ where $\epsilon_P$ is a complex effective coupling constant of the four–fermion local weak interaction of the pseudoscalar quark current $\bar{u}\gamma^5 d$, where $u$ and $d$ are the [*up*]{} and [ *down*]{} quarks, with the left–handed leptonic current $ \bar{\ell}(1
- \gamma^5)\nu_{\ell}$ [@Cirigliano2010] – [@Cirigliano2013a] (see also [@Gonzalez-Alonso2014; @Severijns2019]). Then, $g_P$ is the matrix element $\langle p|\bar{u}\gamma^5 d|n\rangle =
g_P\bar{u}_p\gamma^5 u_n$ caused by strong low–energy interactions, where $\bar{u}_p$ and $u_n$ are the Dirac wave functions of a free proton and neutron, respectively. According to González-Alonso and Camalich [@Gonzalez-Alonso2014], one gets $g_P = 349(9)$ (see Eq.(13) of Ref.[@Gonzalez-Alonso2014]).
---------------------------------------------------------------------------------------------------------------------------------
$0 \lesssim {\rm Re}\,(C_p - \bar{C}_P) \lesssim 0.3 $ $ -
3.5 \times 10^{-5} \lesssim {\rm Re}\,C^{(\rm BSM)}_{ps} \lesssim 0$
-------------------------------------------------------- ------------------------------------------------------------------------
${\rm Im}\,(C_P - \bar{C}_P) < 0.2$ $ {\rm Im}\,C^{(\rm
BSM)}_{ps} < - 2.3 \times 10^{-5}$
---------------------------------------------------------------------------------------------------------------------------------
: Estimates of the phenomenological coupling constant $C_P -
\bar{C}_p = 2g_P\,\epsilon_P$ for $g_P = 349(9)$ [@Gonzalez-Alonso2014] and the constraints on the parameter $\epsilon_P$ [@Bhattacharya2012; @Severijns2019; @Gonzalez-Alonso2014; @Gonzalez-Alonso2016].
Following [@Gonzalez-Alonso2014] and using the constraint $|\epsilon_P| < 5.8\times 10^{-3}$, obtained at $90\,\%$ C.L. from the experimental data on the search for an excess of events with a charged lepton (an electron or muon) and a neutrino in the final state of the pp collision with the centre-of-mass energy of $\sqrt{s} = 8\,{\rm
TeV}$ with an integrated luminosity of $20\,{\rm fb^{-1}}$ at LHC [@CMS2013], we get $|{\rm Re}(C_P - \bar{C}_P)| < 4.1$. In this case the pseudoscalar interaction BSM can dominate in the effective coupling constant $C'_{ps}$ in comparison to the OPP exchange, which is of order $|C^{(\rm OPP)}_{ps}| \sim 10^{-5}$.
In turn, the analysis of the leptonic decays of charged pions, carried out in [@Severijns2019] (see Eq.(113) and a discussion on p.51 of Ref.[@Severijns2019]), taken together with the results, obtained in [@Gonzalez-Alonso2016], gives one ${\rm Re}\,\epsilon_P =
(0.4\pm 1.3)\times 10^{-4}$ and, correspondingly, ${\rm Re}(C_P -
\bar{C}_P) = 0.03 \pm 0.09$. Such an analysis implies that the phenomenological coupling constants ${\rm Re}(C_P - \bar{C}_P)$ and $C^{(\rm BSM)}_{ps}$ are commensurable with zero. This leads to a dominate role of the OPP exchange in the effective coupling constant $C'_{ps}$ equal to $C'_{ps} = - 1.47 \times 10^{-5}$.
Then, following the assumption $\epsilon_P = 2 m_e(m_u +
m_d)/m^2_{\pi} \sim 4\times 10^{-4}$ [@Severijns2019],which is also related to the analysis of the leptonic decays of charged pions (see a discussion below Eq.(112) of Ref.[@Severijns2019]), we get ${\rm Re}(C_P - \bar{C}_P) \sim 0.3$ and ${\rm Re}\,C^{(\rm BSM)}_{ps}
\sim - 3.5 \times 10^{-5}$. As a result, according to the assumption $\epsilon_P = 2 m_e(m_u + m_d)/m^2_{\pi} \sim 4\times 10^{-4}$, the contribution of the pseudoscalar interaction BSM to the effective coupling constant $C'_{ps}$ should be of order $10^{-5}$, that makes it commensurable with the contribution of the OPP exchange.
Since the constraint $|\epsilon_P| < 5.8 \times 10^{-3}$ [@Gonzalez-Alonso2014] disagrees with the constraints following from the analysis of the leptonic decays of charged pions [@Severijns2019; @Gonzalez-Alonso2016], one may conclude that the phenomenological coupling constant ${\rm Re}(C_P - \bar{C}_P)$ should be constrained by $ 0 \lesssim {\rm Re}(C_P - \bar{C}_P) \lesssim
0.3$. This leads to the effective coupling constant ${\rm
Re}\,C^{(\rm BSM)}_{ps}$ restricted by $ - 3.5 \times 10^{-5}
\lesssim {\rm Re}\,C^{(\rm BSM)}_{ps} \lesssim 0$. This shifts the contributions of the pseudoscalar interaction BSM to the region of values $|{\rm Re}\,C^{(\rm BSM)}_{ps}| \sim 10^{-5}$ or even smaller.
The imaginary part ${\rm Im}(C_P - \bar{C}_P) = 2 g_P \,{\rm
Im}\,\epsilon_P$ we estimate using the upper bound ${\rm
Im}\,\epsilon_P < 2.8 \times 10^{-4}$, obtained at $90\,\%$ C.L. in [@Bhattacharya2012] (see also Eq.(114) of Ref.[@Severijns2019]). We get ${\rm Im}(C_P - \bar{C}_P) <
0.3$. The effective coupling constant $C''_{ps} = {\rm Im}\,C^{(\rm
BSM)}_{ps}$ is restricted by $C''_{ps} = {\rm Im}\,C^{(\rm
BSM)}_{ps} < - 2.3 \times 10^{-5}$. Since the contribution of the OPP exchange is real, the effective coupling constant $C''_{ps}$, constrained by $C''_{ps} < - 2.3\times 10^{-5}$, is fully defined by the pseudoscalar interaction BSM.
In Table I we adduce the constraints on the real and imaginary parts of the phenomenological coupling constant $C_P - \bar{C}_P$ and on the effective coupling constant $C^{(\rm BSM)}_{ps}$, which may follow from the results obtained in [@Severijns2019; @Gonzalez-Alonso2014; @Gonzalez-Alonso2016].
Discussion {#sec:schluss}
==========
The corrections of order $10^{-5}$, calculated within the SM, are needed as a SM theoretical background for experimental searches of interactions beyond the SM in terms of asymmetries and correlation coefficients of the neutron $\beta^-$–decays [@Ivanov2017b; @Ivanov2017d; @Ivanov2019]. An experimental accuracy of about a few parts of $10^{-5}$ or even better, which is required for experimental analyses of interactions BSM of order $10^{-4}$, can be reachable at present time [@Abele2016]. In this paper we have continued the analysis of corrections of order $10^{-5}$ to the correlation coefficients of the neutron $\beta^-$–decays, which we have begun in [@Ivanov2017b; @Ivanov2017d; @Ivanov2019; @Ivanov2019a]. In this paper we have taken into account the contributions of strong low–energy interactions in terms of the OPP exchange and the contributions of the pseudoscalar interaction BSM [@Lee1956]–[@Severijns2006], and calculated corrections to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decay for a polarized neutron, a polarized electron and an unpolarized proton.
In addition to the results, concerning the corrections caused by pseudoscalar interactions to the electron–energy and angular distributions of the neutron $\beta^-$–decay for a polarized neutron and unpolarized electron and proton, obtained in [@Harrington1960]–[@Hayen2018] and especially by Harrington [@Harrington1960] and Holstein [@Holstein1974], we have calculated corrections to the correlation coefficients, caused by correlations with the electron spin, i.e. for a polarized neutron and a polarized electron with an unpolarized proton.
We have shown that the energy independent contributions to the pseudoscalar form factor [@Bernard1995; @Bernard1996; @Kaiser2003; @Adler1966; @Wolfenstein1970], related to the Adler-Dothan-Wolfenstein (ADM) term Eq.(\[eq:6\]) and to the chiral corrections Eq.(\[eq:7\]), calculated by Kaiser [@Kaiser2003] in a two–loop approximation within the HB$\chi$PT, are able in principle to be responsible for sufficiently small real parts of the phenomenological coupling constants $C_P$ and $\bar{C}_P$ and at the level of $10^{-6}
- 10^{-8}$ of the effective coupling constant $C^{(\rm BSM)}_{ps}$. In turn, the isospin breaking corrections of order $10^{-5}$, calculated by Kaiser within the HB$\chi$PT [@Kaiser2001] to the vector coupling constant of the neutron $\beta^-$–decay, should be taken into account for a correct description of the neutron lifetime at the level of $10^{-5}$.
As has been shown in [@Cirigliano2013] the phenomenological coupling constant $C_P - \bar{C}_P$, introduced at the hadronic level [@Lee1956]–[@Severijns2006], can be related to the effective coupling constant $\epsilon_P$ of the pseudoscalar interaction of the [*up*]{} and [*down*]{} quarks with left–handed leptonic current by $C_P - \bar{C}_P = 2 g_P \epsilon_P$, where $g_P = 349(9)$ [@Gonzalez-Alonso2014] is the matrix element of the pseudoscalar quark current caused by strong low–energy interactions. Using the relation $C_P - \bar{C}_P = 2 g_P \epsilon_P$ [@Cirigliano2013] we have estimated the real and imaginary parts of the phenomenological coupling constant $C_P - \bar{C}_P$. Having summarized the results, concerning the constraints on the parameter $\epsilon_P$, obtained in [@Bhattacharya2012; @Severijns2019; @Gonzalez-Alonso2014; @Gonzalez-Alonso2016], and taking into account that $g_P = 349(9)$ [@Gonzalez-Alonso2014], we have got $ 0
\lesssim {\rm Re}(C_P - \bar{C}_P) \lesssim 0.3$ and ${\rm Im}(C_P -
\bar{C}_P) < 0.2$. Such an estimate agrees well with the analysis of the contributions of the pseudoscalar interaction BSM to the lifetimes of charged pions [@Severijns2019].
For the effective coupling constants ${\rm Re}\,C^{(\rm BSM)}_{ps}$ and ${\rm Im}\,C^{(\rm BSM)}_{ps}$, defining the strength of the contributions of the pseudoscalar interaction BSM to the correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays, we get $ - 3.5 \times 10^{-5} \lesssim {\rm
Re}\,C^{(\rm BSM)}_{ps} \lesssim 0$ and ${\rm Im}\,C^{(\rm
BSM)}_{ps} < - 2.3 \times 10^{-5}$, respectively. This implies that the effective coupling constant $C^{(\rm BSM)}_{ps}$ is of order $|C^{(\rm BSM)}_{ps}| \sim 10^{-5}$.
The analysis of contributions of pseudoscalar interactions to the electron–energy and angular distributions of weak semileptonic decays of baryons has a long history [@Harrington1960]–[@Hayen2018] (see also [@Wilkinson1982; @Severijns2019]). That is why it is important to make a comparative analysis of the results obtained in our work with those in [@Wilkinson1982; @Severijns2019; @Harrington1960]–[@Hayen2018]. For the first time the contributions of pseudoscalar interactions to the correlation coefficients of electron–energy and angular distributions for weak semileptonic decays of baryons for polarized parent baryons and unpolarized decay electrons and baryons were calculated by Harrington [@Harrington1960]. In the notation of Jackson [*et al.*]{} [@Jackson1957] Harrington calculated the contributions of the induced pseudoscalar form factor to the Fierz interference term $b(E_e)$ [@Fierz1937] and to the correlation coefficients $a(E_e)$, $A(E_e)$, $B(E_e)$ and $D(E_e)$, caused by electron–antineutrino angular correlations and correlations of the neutron spin with electron and antineutrino 3–momenta, respectively. The corresponding contributions of pseudoscalar interactions can be obtained from Eqs.(9) – (13) of Ref.[@Harrington1960] keeping the leading terms in the large baryon mass expansion. They read $$\begin{aligned}
\label{eq:B.1}
\hspace{-0.15in}&&\frac{d^5 \delta \lambda_n(E_e, \vec{k}_e,
\vec{k}_{\bar{\nu}}, \vec{\xi}_n, \vec{\xi}_e)}{dE_e d\Omega_e
d\Omega_{\bar{\nu}}} \propto - \frac{{\rm
Re}(g_1g^*_3)}{|f_1|^2 +
3|g_2|^2}\,\frac{m^2_e}{M^2}\,\frac{E_{\bar{\nu}}}{E_e} -
\frac{{\rm Re}(g_1g^*_3)}{|f_1|^2 +
3|g_2|^2}\,\frac{m^2_e}{M^2}\,\frac{\vec{k}_e\cdot
\vec{k}_{\bar{\nu}}}{E_eE_{\bar{\nu}}} - \frac{{\rm
Re}(f_1g^*_3)}{|f_1|^2 +
3|g_2|^2}\,\frac{m^2_e}{M^2}\,\frac{\vec{\xi}_n\cdot \vec{k}_e}{E_e}
\nonumber\\
\hspace{-0.15in}&&\hspace{1.5in}~~~ - \frac{{\rm Re}(f_1g^*_3)}{|f_1|^2 +
3|g_2|^2}\,\frac{m^2_e}{M^2}\,\frac{E_{\bar{\nu}}}{E_e}\,\frac{\vec{\xi}_n
\cdot \vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} + \frac{{\rm Im}(g_1
g^*_3)}{|f_1|^2 + 3|g_2|^2}\,\frac{m^2_e}{M^2}\,
\frac{\vec{\xi}_n\cdot \big(\vec{k}_e \times
\vec{k}_{\bar{\nu}}\big)}{E_e E_{\bar{\nu}}},\end{aligned}$$ where the first term describes the contribution of pseudoscalar interactions to the Fierz–like interference term [@Fierz1937]. The analogous corrections can be extracted from the expressions, calculated by Holstein [@Holstein1974] (see Appendix B of Ref.[@Holstein1974]). The corrections of pseudoscalar interactions to the Fierz–like interference term $\delta b_{ps}(E_e)$ and correlation coefficients $\delta a_{ps}(E_e)$, $\delta
A_{ps}(E_e)$, $\delta B_{ps}(E_e)$ and $\delta D_{ps}(E_e)$, calculated in Eqs.(\[eq:A.5\]) and (\[eq:A.6\]), agree well with those calculated by Harrington [@Harrington1960] (see Eq.(\[eq:B.1\])). Since in [@Wilkinson1982; @Severijns2019; @Shekhter1960; @Bender1968; @Armstrong1972; @Holstein1974; @BB1982; @Gonzalez-Alonso2014; @Hayen2018] the electron–energy and angular distributions were analyzed for weak semileptonic decays either for polarized parent baryons and unpolarized decay electrons and baryons or for unpolarized parent baryons and unpolarized decay electrons and baryons the overlap of our results with those obtained in [@Wilkinson1982; @Shekhter1960; @Bender1968; @Armstrong1972; @Holstein1974; @BB1982; @Gonzalez-Alonso2014; @Hayen2018] is at the level of the corrections shown in Eq.(\[eq:B.1\]). Indeed, the contribution of the Fierz–like interference term $\delta b_{ps}(E_e)$ in Eq.(\[eq:A.4\]) agrees well with the result, obtained by Wilkinson [@Wilkinson1982] and by González-Alonso and Camalich [@Gonzalez-Alonso2014] $$\begin{aligned}
\label{eq:B.2}
\hspace{-0.3in}&&\frac{d^5 \delta \lambda_n(E_e, \vec{k}_e,
\vec{k}_{\bar{\nu}}, \vec{\xi}_n, \vec{\xi}_e)}{dE_e d\Omega_e
d\Omega_{\bar{\nu}}} \propto C'_{ps}\,\lambda\, \frac{E_0 -
E_e}{E_0} \frac{m_e}{E_e} + \ldots \to - \frac{g_A g_{\rm IP}}{g^2_V +
3 g^2_A}\,\frac{E_0 - E_e}{M}\,\frac{m_e}{E_e}\nonumber\\
\hspace{-0.3in}&&\hspace{1.5in} - \frac{\lambda}{1 +
3 \lambda^2}\,g_P{\rm Re}\epsilon_P\,\frac{E_0 -
E_e}{M}\,\frac{m_e}{E_e} + \ldots,\end{aligned}$$ where the term proportional to $g_Ag_{\rm I P}$, describing the contribution of the OPP exchange with $g_{\rm IP} = 2 g_A
M/m^2_{\pi}$, was calculated by Wilkinson (see Table 1 and a definition of $g_{\rm IP}$ on p.479 of Ref.[@Wilkinson1982]), whereas the second term, caused by the contribution of the pseudoscalar interaction BSM and where we have taken into account the relation $C_P - \bar{C}_P = 2 g_P \epsilon_P$ [@Cirigliano2013], was calculated by González-Alonso and Camalich [@Gonzalez-Alonso2014] (see Eqs.(16) and (17) of Ref.[@Gonzalez-Alonso2014])).
In turn, the contributions of pseudoscalar interactions to the correlation coefficients, induced by correlations with the electron spin, were not calculated in [@Wilkinson1982; @Severijns2019; @Harrington1960; @Shekhter1960; @Bender1968; @Armstrong1972; @Holstein1974; @BB1982; @Gonzalez-Alonso2014; @Hayen2018]. Thus, the calculation of contributions of pseudoscalar interactions to the correlation coefficients, induced by correlations with the electron spin, distinguishes our results from those obtained in [@Wilkinson1982; @Severijns2019; @Harrington1960; @Shekhter1960; @Bender1968; @Armstrong1972; @Holstein1974; @BB1982; @Gonzalez-Alonso2014; @Hayen2018]. However, we would like to notice that in the book by Behrens and Bühring [@BB1982] there is a capture entitled “Electron polarization”, concerning an analysis of a polarization of decay electrons in beta decays. In this capture the authors propose a most general density matrix, which can be applied to a description of energy and angular distributions for beta decays by taking into account a polarization of decay electrons (see Eq.(7.6) and Eq.(7.7) of Ref.[@BB1982]). Of course, by using such a general density matrix and the technique, developed by Biedenharn and Rose [@Biedenharn1953], one can, in principle, calculate contributions of pseudoscalar interactions to the correlation coefficients induced by correlations with the electron spin. Nevertheless, the calculation of these corrections were not performed in [@BB1982]. The authors applied such a general density matrix to a calculation of a general formula for a value of a longitudinal polarization of decay electrons in beta decays only (see Eq.(7.151) of Ref.[@BB1982]). Thus, we may assert that all corrections of pseudoscalar interactions to the correlation coefficients, induced by correlations with the electron spin (see Eq.(\[eq:A.5\])), and also other terms proportional to the coupling constants $C'_{ps}$ and $C''_{ps}$ in Eq.(\[eq:A.6\]) are new in comparison to the results, obtained in [@Wilkinson1982; @Severijns2019; @Harrington1960]–[@Hayen2018] and were never calculated in literature. Moreover, a theoretical accuracy $O(\alpha
E_0/\pi M) \sim 10^{-6}$ and $O(E^2_0/M^2) \sim 10^{-6}$ of the calculation of a complete set of corrections of order $10^{-3}$ [@Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019] including radiative corrections of order $O(\alpha/\pi)$ and corrections of order $O(E_0/M)$, caused by the weak magnetism and proton recoil, makes the contributions of corrections of order $10^{-5}$, induced by pseudoscalar interactions, observable in principle and important as a part of theoretical background for experimental searches of contributions of interactions BSM in asymmetries of the neutron $\beta^-$–decays with a polarized neutron, a polarized electron and an unpolarized proton [@Abele2016].
Thus, in this work we have calculated the contributions of pseudoscalar interactions, induced by the OPP exchange and BSM, to the complete set of correlation coefficients of the electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton. The corrections to the Fierz interference term $b(E_e)$, the correlation coefficients $a(E_e)$, $A(E_e)$, $B(E_e)$ and $D(E_e)$, caused by electron–antineutrino angular correlations and correlations of the neutron spin with electron and antineutrino 3–momenta, respectively, and as well as the correlation coefficients, induced by correlations with the electron spin such as $G(E_e)$, $N(E_e)$ and so on, and also corrections, given by the terms proportional to the effective coupling constants $C'_{ps}$ and $C''_{ps}$ in Eq.(\[eq:A.6\]), are calculated by using one of the same theoretical technique. The agreement of the corrections to the Fierz interference term $b(E_e)$ and the correlation coefficients $a(E_e)$, $A(E_e)$, $B(E_e)$ and $D(E_e)$ with the results obtained in [@Wilkinson1982; @Severijns2019; @Harrington1960]–[@Hayen2018] may only confirm a correctness of our results.
The obtained corrections (see Eq.(\[eq:A.5\]) and Eq.(\[eq:A.6\])), caused by the OPP exchange and the pseudoscalar interaction BSM, complete the analysis of contributions of interactions BSM to the correlation coefficients of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton carried out in [@Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019]. For experimental accuracies of about a few parts of $10^{-5}$ or even better [@Abele2016] the exact analytical expressions of these corrections can be practically distinguished from the contributions of order $10^{-5}$, caused by the second class hadronic currents or $G$–odd correlations, calculated by Gardner and Plaster [@Gardner2013] and Ivanov [*et al.*]{} [@Ivanov2017d; @Ivanov2019].
Acknowledgements
================
We thank Hartmut Abele for discussions stimulating the work under corrections of order $10^{-5}$ to the neutron lifetime and correlation coefficients of the neutron $\beta^-$–decays for different polarization states of the neutron and massive decay fermions. The work of A. N. Ivanov was supported by the Austrian “Fonds zur Förderung der Wissenschaftlichen Forschung” (FWF) under contracts P31702-N27 and P26636-N20 and “Deutsche Förderungsgemeinschaft” (DFG) AB 128/5-2. The work of R. Höllwieser was supported by the Deutsche Forschungsgemeinschaft in the SFB/TR 55. The work of M. Wellenzohn was supported by the MA 23 (FH-Call 16) under the project “Photonik - Stiftungsprofessur für Lehre”. The results obtained in this paper were reported at International Workshop on “Current and Future Status of the First-Row CKM Unitarity”, held on 16 - 18 of May 2019 at Amherst Center of Fundamental Interactions, University of Massachusetts Amherst, USA [@CKM2019].
Appendix A: Calculation of corrections caused by pseudoscalar interactions to the electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton {#appendix-a-calculation-of-corrections-caused-by-pseudoscalar-interactions-to-the-electronenergy-and-angular-distribution-of-the-neutron-beta-decays-for-a-polarized-neutron-a-polarized-electron-and-an-unpolarized-proton .unnumbered}
===================================================================================================================================================================================================================================
A direct calculation of the corrections, caused by the OPP exchange and the pseudoscalar interaction BSM [@Ivanov2013], to the electron–energy and angular distribution of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarized proton yields $$\begin{aligned}
\label{eq:A.1}
\hspace{-0.15in}&&\frac{d^5 \delta \lambda_n(E_e, \vec{k}_e,
\vec{k}_{\bar{\nu}}, \vec{\xi}_n, \vec{\xi}_e)}{dE_e d\Omega_e
d\Omega_{\bar{\nu}}} = (1 + 3
\lambda^2)\,\frac{G^2_F|V_{ud}|^2}{32\pi^5}\,(E_0 - E_e)^2
\,\sqrt{E^2_e - m^2_e}\, E_e\,F(E_e, Z = 1)\,\frac{1}{E_0
E_eE_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&&\times \,\Big\{C' _{ps}\Big[\lambda\, \Big(- m_e
(\vec{k}_p\cdot \vec{k}_{\bar{\nu}}) - (\vec{k}_p\cdot
\vec{k}_e)(\zeta_e\cdot k_{\bar{\nu}}) + (\vec{k}_p\cdot
\vec{\zeta}_e)(k_e\cdot k_{\bar{\nu}})\Big) + (\vec{\xi}_n \cdot
\vec{k}_p)\big(m_e E_{\bar{\nu}} + E_e (\zeta_e \cdot k_{\bar{\nu}})
- \zeta^0_e (k_e\cdot k_{\bar{\nu}})\big)\nonumber\\
\hspace{-0.15in}&& + \lambda\,(\vec{\xi}_n \times \vec{k}_p) \cdot
\Big(- E_e(\vec{\zeta}_e \times \vec{k}_{\bar{\nu}}) + \zeta^0_e
(\vec{k}_e\times \vec{k}_{\bar{\nu}}) + E_{\bar{\nu}}(\vec{\zeta}_e
\times \vec{k}_e)\Big) + C''_{ps}\Big[\vec{\zeta}_e\cdot (\vec{k}_e
\times \vec{k}_{\bar{\nu}})(\vec{\xi}_n\cdot \vec{k}_p) + \lambda
\Big(E_e\,
\vec{k}_p\cdot (\vec{\zeta}_e \times \vec{k}_{\bar{\nu}})\nonumber\\
\hspace{-0.15in}&& - E_{\bar{\nu}}\, \vec{k}_p\cdot (\vec{\zeta}_e
\times \vec{k}_e) - m_e \,\vec{k}_{\bar{\nu}}\cdot (\vec{\xi}_n \times
\vec{k}_p) - \vec{k}_e \cdot (\vec{\xi}_n \times \vec{k}_p)(\zeta_e
\cdot k_{\bar{\nu}}) + \vec{\zeta}_e\cdot (\vec{\xi}_n \times
\vec{k}_p)(k_e \cdot k_{\bar{\nu}}) \Big)\Big] \Big\}.\end{aligned}$$ The strength of the contributions of pseudoscalar interactions is defined by the effective coupling constants $C'_{ps}$ and $C''_{ps}$, which are the real and imaginary parts of the effective coupling constant $C_{ps}$ given by $$\begin{aligned}
\label{eq:A.2}
C_{ps} &=& C^{(\rm OPP)}_{ps} + C^{(\rm
BSM)}_{ps},\nonumber\\ C^{(\rm OPP)}_{ps} &=& \frac{2\lambda}{1 +
3\lambda^2}\,\frac{m_e }{m^2_{\pi}}\,E_0 = - 1.47\times
10^{-5},\nonumber\\
C^{(\rm BSM)}_{ps} &=& - \frac{1}{1 +
3\lambda^2}\, \frac{E_0}{2M}\,(C_P - \bar{C}_P) = - 1.17\times
10^{-4}\,(C_P - \bar{C}_P),\nonumber\\
C'_{ps} &=& {\rm Re}\,C_{ps}
= C^{(\rm OPP)}_{ps} + {\rm Re}\, C^{(\rm
BSM)}_{ps},\nonumber\\ C''_{ps} &=& {\rm Im}\,C_{ps} = {\rm
Im}\,C^{(\rm BSM)}_{ps}.\end{aligned}$$ The numerical values are obtained at $\lambda = - 1.27641$, $E_0 =
(m^2_n - m^2_p + m^2_e)/2m_n = 1.2926\,{\rm MeV}$, $m_e = 0.511\,{\rm
MeV}$ and $M = (m_n + m_p)/2 = 938.918\,{\rm MeV}$ [@PDG2018]. Then, $\zeta_e$ is a 4–polarization vector of the electron [@Itzykson1980] $$\begin{aligned}
\label{eq:A.3}
\zeta_e = (\zeta^0_e, \vec{\zeta}_e) = \Big(\frac{\vec{\xi}_e\cdot
\vec{k}_e}{m_e}, \vec{\xi}_e + \frac{\vec{k}_e(\vec{\xi}_e\cdot
\vec{k}_e)}{m_e (E_e + m_e)}\Big)\end{aligned}$$ obeying the constraints $\zeta^2_e = - \vec{\xi}^{\,2}_e = - 1$ and $k_e\cdot \zeta_e = 0$. The right-hand-side (r.h.s.) of Eq.(\[eq:A.1\]) can be transcribed into the form $$\begin{aligned}
\label{eq:A.4}
\hspace{-0.15in}&&\frac{d^5 \delta \lambda_n(E_e, \vec{k}_e,
\vec{k}_{\bar{\nu}}, \vec{\xi}_n, \vec{\xi}_e)}{dE_e d\Omega_e
d\Omega_{\bar{\nu}}} = (1 + 3
\lambda^2)\,\frac{G^2_F|V_{ud}|^2}{32\pi^5}\,(E_0 - E_e)^2
\,\sqrt{E^2_e - m^2_e}\, E_e\,F(E_e, Z =
1)\,\Big\{C'_{ps}\Big[\lambda\,\Big(
\frac{E_{\bar{\nu}}}{E_0}\,\frac{m_e}{E_e}+
\frac{m_e}{E_0}\,\frac{\vec{k}_e\cdot \vec{k}_{\bar{\nu}}}{E_e
E_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& - \frac{m_e}{E_0}\,\frac{\vec{\xi}_e\cdot
\vec{k}_e}{E_e} - \Big(1 -
\frac{m^2_e}{E_0E_e}\Big)\,\frac{\vec{\xi}_e\cdot
\vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} + \Big(1 +
\frac{m_e}{E_0}\Big)\,\frac{(\vec{\xi}_e\cdot
\vec{k}_e)(\vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{(E_e + m_e) E_e
E_{\bar{\nu}}}\Big) + \Big(- \frac{m_e}{E_0}\,\frac{\vec{\xi}_n\cdot
\vec{k}_e}{E_e} -
\frac{m_e}{E_0}\,\frac{E_{\bar{\nu}}}{E_e}\,\frac{\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} + \frac{E_e}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_e)}{E_e}\nonumber\\
\hspace{-0.15in}&&\times \frac{(\vec{\xi}_e\cdot
\vec{k}_{\bar{\nu}})}{ E_{\bar{\nu}}} + \frac{
E_{\bar{\nu}}}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e\cdot
\vec{k}_{\bar{\nu}})}{E^2_{\bar{\nu}}} -
\frac{E_e}{E_0}\,\frac{(\vec{\xi}_n\cdot \vec{k}_e)(\vec{\xi}_e\cdot
\vec{k}_e)(\vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{(E_e + m_e) E^2_e
E_{\bar{\nu}}} - \frac{E_{\bar{\nu}}}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e\cdot \vec{k}_e)(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{(E_e + m_e) E_e E^2_{\bar{\nu}}}\Big) +
\lambda \Big(\frac{E_e E_{\bar{\nu}} - k^2_e}{E_e
E_0}\nonumber\\
\hspace{-0.15in}&&\times\,(\vec{\xi}_n\cdot \vec{\xi}_e) + \frac{E_e -
E_{\bar{\nu}}}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{\xi}_e)(\vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{E_e E_{\bar{\nu}}}
+ \frac{E_e - E_{\bar{\nu}} + m_e}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_e)(\vec{\xi}_e\cdot \vec{k}_e)}{(E_e + m_e) E_e} -
\frac{m_e}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e \cdot \vec{k}_e)}{E_e
E_{\bar{\nu}} } + \frac{E_{\bar{\nu}}}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_e)(\vec{\xi}_e \cdot \vec{k}_{\bar{\nu}})}{E_e
E_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& -
\frac{E_{\bar{\nu}}}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e \cdot \vec{k}_{\bar{\nu}})}{
E^2_{\bar{\nu}}} - \frac{E_e}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_e)(\vec{\xi}_e \cdot \vec{k}_e)(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{(E_e + m_e)E^2_e E_{\bar{\nu}}} +
\frac{E_{\bar{\nu}}}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e \cdot \vec{k}_e)(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{(E_e + m_e)E_e E^2_{\bar{\nu}}}\Big)\Big] +
C''_{ps}\Big[- \frac{E_e}{E_0}\,\frac{\vec{\xi}_e\cdot (\vec{k}_e \times
\vec{k}_{\bar{\nu}})(\vec{\xi}_n\cdot \vec{k}_e)}{E^2_e
E_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& - \frac{E_{\bar{\nu}}}{E_0}\,\frac{\vec{\xi}_e\cdot
(\vec{k}_e \times \vec{k}_{\bar{\nu}})(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})}{E_e E^2_{\bar{\nu}}} + \lambda
\Big(\frac{\vec{\xi}_e\cdot(\vec{k}_e \times \vec{k}_{\bar{\nu}})}{E_e
E_{\bar{\nu}}} + \frac{m_e}{E_0}\,\frac{\vec{\xi}_n\cdot (\vec{k}_e
\times \vec{k}_{\bar{\nu}})}{E_e E_{\bar{\nu}}} +
\frac{E_e}{E_0}\,\frac{\vec{\xi}_n\cdot (\vec{\xi}_e \times
\vec{k}_e)}{E_e} + \frac{E_{\bar{\nu}}}{E_0}\,
\frac{\vec{\xi}_n\cdot (\vec{\xi}_e \times
\vec{k}_{\bar{\nu}})}{E_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& - \frac{E_e}{E_0}\,
\frac{\vec{\xi}_n\cdot(\vec{\xi}_e \times \vec{k}_e)(\vec{k}_e \cdot
\vec{k}_{\bar{\nu}})}{E^2_e E_{\bar{\nu}}} -
\frac{E_{\bar{\nu}}}{E_0}\, \frac{\vec{\xi}_n\cdot (\vec{\xi}_e \times
\vec{k}_{\bar{\nu}})(\vec{k}_e \cdot \vec{k}_{\bar{\nu}})}{ E_e
E^2_{\bar{\nu}}} + \frac{E_{\bar{\nu}}}{E_0}\,
\frac{\vec{\xi}_n\cdot (\vec{k}_e \times
\vec{k}_{\bar{\nu}})(\vec{\xi}_e \cdot \vec{k}_{\bar{\nu}})}{ E_e
E^2_{\bar{\nu}}}- \frac{E_{\bar{\nu}}}{E_0}\,\frac{\vec{\xi}_n\cdot
(\vec{k}_e \times \vec{k}_{\bar{\nu}})(\vec{\xi}_e \cdot
\vec{k}_e)}{(E_e + m_e) E_e E_{\bar{\nu}}}\Big) \Big] \Big\}.\end{aligned}$$ We obtain the following contributions to the correlation coefficients $$\begin{aligned}
\hspace{-0.15in}\delta
\zeta_{ps}(E_e) &=&0\;,\;\delta
b_{ps}(E_e) = C'_{ps}\,\lambda\,\frac{E_0 -
E_e}{E_0}\;,\;\delta a_{ps}(E_e) = C'_{ps}\,
\lambda\, \frac{m_e}{E_0}\;,\;\delta A_{ps}(E_e) =
-\,C'_{ps}\, \frac{m_e}{E_0},\nonumber\\
\end{aligned}$$ $$\begin{aligned}
\label{eq:A.5}
\hspace{-0.15in}\delta B_{ps}(E_e) &=& -\,C'_{ps}\,
\frac{m_e}{E_0}\,\frac{E_0 - E_e}{E_e}\;,\;\delta K_{n
ps}(E_e) =\delta Q_{n
ps}(E_e) = 0\;,\; \delta
G_{ps}(E_e) = - C'_{ps}\,\lambda\,\frac{m_e}{E_0},\nonumber\\
\hspace{-0.15in}\delta H_{ps}(E_e) &=& -\,C'_{ps}\,\lambda\, \Big(1
- \frac{m^2_e}{E_0 E_e}\Big)\;,\;\delta Q_{e ps}(E_e) =
C'_{ps}\,\Big(\lambda \,\frac{2 E_e - E_0 + m_e}{E_0} + (\lambda -
1)\,\frac{1}{3}\frac{E_0 - E_e}{E_0}\Big),\nonumber\\
\hspace{-0.15in}\delta K_{eps}(E_e)&=& C'_{ps}\,\lambda\,\Big(1 +
\frac{m_e}{E_0}\Big)\;,\; \delta N_{ps}(E_e) =
C'_{ps}\Big(\lambda\,\frac{- 2 E^2_e + E_0 E_e + m^2_e}{E_0 E_e} +
(1 - \lambda)\,\frac{1}{3}\,\frac{E_0 - E_e}{E_0}\Big), \nonumber\\
\hspace{-0.15in}\delta D_{ps}(E_e)&=&
C''_{ps}\,\lambda\,\frac{m_e}{E_0}\;,\; \delta R_{ps}(E_e) =
C''_{ps}\,\Big( - \lambda\,\frac{E_e}{E_0} + (1 + 2
\lambda)\,\frac{1}{3}\,\frac{E_0 - E_e}{E_0}\Big)\;,\; \delta
L_{ps}(E_e) = C''_{ps}\,\lambda.\end{aligned}$$ In terms of corrections to the correlation coefficients Eq.(\[eq:A.5\]) the correction to the electron–energy and angular distribution Eq.(\[eq:A.4\]) is given by $$\begin{aligned}
\label{eq:A.6}
\hspace{-0.15in}&&\frac{d^5 \delta \lambda_n(E_e, \vec{k}_e,
\vec{k}_{\bar{\nu}}, \vec{\xi}_n, \vec{\xi}_e)}{dE_e d\Omega_e
d\Omega_{\bar{\nu}}} = (1 + 3
\lambda^2)\,\frac{G^2_F|V_{ud}|^2}{32\pi^5}\,(E_0 - E_e)^2
\,\sqrt{E^2_e - m^2_e}\, E_e\,F(E_e, Z = 1)\,\Big\{ \delta
b_{ps}(E_e)\,\frac{m_e}{E_e} + \delta
a_{ps}(E_e)\,\frac{\vec{k}_e\cdot \vec{k}_{\bar{\nu}}}{E_e
E_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& + \delta
A_{ps}(E_e)\,\frac{\vec{\xi}_n\cdot \vec{k}_e}{E_e} + \delta
B_{ps}(E_e)\, \frac{\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} + \delta G_{ps}(E_e)\,\frac{\vec{\xi}_e
\cdot \vec{k}_e}{E_e} + \delta H_{ps}(E_e)\,\frac{\vec{\xi}_e \cdot
\vec{k}_{\bar{\nu}}}{E_{\bar{\nu}}} + \delta Q_{e
ps}(E_e)\,\frac{(\vec{\xi}_n\cdot \vec{k}_e)( \vec{k}_e\cdot
\vec{\xi}_e)}{(E_e + m_e) E_e} + \delta K_{e ps}(E_e)\nonumber\\
\hspace{-0.15in}&&\times\,\frac{(\vec{\xi}_e\cdot \vec{k}_e)(
\vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{(E_e + m_e)E_e E_{\bar{\nu}}} +
\delta N_{ps}(E_e) \,(\vec{\xi}_n\cdot \vec{\xi}_e) + \delta
R_{ps}\,\frac{\vec{\xi}_n\cdot (\vec{k}_e \times \vec{\xi}_e)}{E_e} +
\delta L_{ps}\,\frac{\vec{\xi}_n\cdot (\vec{k}_e \times
\vec{k}_{\bar{\nu}})}{E_e E_{\bar{\nu}}} +
C'_{ps}\Big[\lambda\,\frac{E_0 - E_e}{E_0}\frac{(\vec{\xi}_n\cdot
\vec{\xi}_e)( \vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{E_e
E_{\bar{\nu}}}\nonumber\\
\hspace{-0.15in}&& + \frac{(1 - \lambda) E_e + \lambda E_0
}{E_0}\frac{(\vec{\xi}_n\cdot \vec{k}_e)( \vec{\xi}_e\cdot
\vec{k}_{\bar{\nu}})}{E_e E_{\bar{\nu}}} -
\lambda\,\frac{m_e}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e\cdot \vec{k}_e)}{E_e E_{\bar{\nu}}}
- (1 + \lambda) \frac{E_e}{E_0}\,\frac{(\vec{\xi}_n\cdot
\vec{k}_e)(\vec{\xi}_e \cdot \vec{k}_e)(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{(E_e + m_e)E^2_e E_{\bar{\nu}}} + (1 -
\lambda)\,\frac{E_0 - E_e}{E_0}\nonumber\\
\hspace{-0.15in}&&\times \,\Big(\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{\xi}_e\cdot
\vec{k}_{\bar{\nu}})}{E^2_{\bar{\nu}}} -
\frac{1}{3}\,\vec{\xi}_n\cdot \vec{\xi}_e\Big) + (\lambda -
1)\,\frac{E_0 - E_e}{E_0}\,\Big(\frac{(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{E^2_{\bar{\nu}}} -
\frac{1}{3}\,\vec{\xi}_n\cdot \vec{k}_e \Big) \frac{(\vec{\xi}_e\cdot
\vec{k}_e)}{(E_e + m_e) E_e}\Big] + C''_{ps}\Big[- \frac{E_e}{E_0}
\nonumber\\
\hspace{-0.15in}&&\times \,\frac{\vec{\xi}_e\cdot (\vec{k}_e \times
\vec{k}_{\bar{\nu}})(\vec{\xi}_n\cdot \vec{k}_e)}{E^2_e
E_{\bar{\nu}}} + \lambda\,\frac{E_0 - E_e}{E_0}
\,\frac{\vec{\xi}_n\cdot (\vec{\xi}_e \times
\vec{k}_{\bar{\nu}})}{E_{\bar{\nu}}} -
\lambda\,\frac{E_e}{E_0}\,\frac{\vec{\xi}_n\cdot (\vec{\xi}_e \times
\vec{k}_e)(\vec{k}_e\cdot \vec{k}_{\bar{\nu}})}{E^2_e E_{\bar{\nu}}}
- \lambda\,\frac{E_0 - E_e}{E_0} \frac{\vec{\xi}_n\cdot (\vec{k}_e
\times \vec{k}_{\bar{\nu}})(\vec{\xi}_e\cdot \vec{k}_e)}{(E_e + m_e)
E_e E_{\bar{\nu}}} \nonumber\\
\hspace{-0.15in}&& - \frac{E_0 - E_e}{E_0} \Big(\frac{\vec{\xi}_e\cdot
(\vec{k}_e \times \vec{k}_{\bar{\nu}})(\vec{\xi}_n\cdot
\vec{k}_{\bar{\nu}})}{E_e E^2_{\bar{\nu}}} - \frac{1}{3}
\frac{\vec{\xi}_e\cdot (\vec{k}_e \times \vec{\xi}_n)}{E_e} \Big) -
\lambda\,\frac{E_0 - E_e}{E_0} \Big(\frac{\vec{\xi}_n\cdot
(\vec{\xi}_e \times \vec{k}_{\bar{\nu}})(\vec{k}_e\cdot
\vec{k}_{\bar{\nu}})}{E_e E^2_{\bar{\nu}}} - \frac{1}{3}
\frac{\vec{\xi}_n\cdot (\vec{\xi}_e \times \vec{k}_e)}{E_e} \Big)
\nonumber\\
\hspace{-0.15in}&&+ \lambda\,\frac{E_0 - E_e}{E_0}\Big(
\frac{\vec{\xi}_n\cdot (\vec{k}_e \times
\vec{k}_{\bar{\nu}})(\vec{\xi}_e \cdot \vec{k}_{\bar{\nu}})}{ E_e
E^2_{\bar{\nu}}} - \frac{1}{3}\,\frac{\vec{\xi}_n\cdot (\vec{k}_e
\times \vec{\xi}_e)}{E_e}\Big)\Big]\Big\}.\end{aligned}$$ This correction to the electron–energy and angular distribution together with the results obtained in [@Gudkov2006; @Ivanov2013; @Ivanov2017b; @Ivanov2017d; @Ivanov2019], can be used for experimental analyses of asymmetries and correlation coefficients of the neutron $\beta^-$–decays for a polarized neutron, a polarized electron and an unpolarised proton with experimental uncertainties of a few parts of $10^{-5}$ [@Abele2016].
[9]{}
S. M. Bilen’kii, R. M. Ryndin, Ya. A. Smorodinskii, and Ho Tso-Hsiu, [*On the theory of the neutron beta decay*]{}, JETP [**37**]{}, 1759 (1959) (in Russian); Sov. Phys. JETP, [**37**]{}(10), 1241 (1960).
A. Sirlin, [*General properties of the electromagnetic corrections to the beta decay of a physical nucleon*]{}, Phys. Rev. [**164**]{}, 1767 (1967).
R. T. Shann, [*Electromagnetic effects in the decay of polarized neutrons*]{}, Nuovo Cimento A [**5**]{}, 591 (1971).
D. H. Wilkinson, [*Analysis of neutron beta decay*]{}, Nucl. Phys. A [**377**]{}, 474 (1982).
W. J. Marciano and A. Sirlin, [*Radiative corrections to $\beta$ decay and the possibility of a fourth generation*]{}, Phys. Rev. Lett. [**56**]{}, 22 (1986).
A. Czarnecki, W. J. Marciano, and A. Sirlin, [ *Precision measurements and CKM unitarity*]{}, Phys. Rev. D [**70**]{}, 093006 (2004).
W. J. Marciano and A. Sirlin, [*Improved calculation of electroweak radiative corrections and the value of $V(ud)$*]{}, Phys. Rev. Lett. [**96**]{}, 032002 (2006).
V. Gudkov, G. I. Greene, and J. R. Calarco, [ *General classification and analysis of neutron beta-decay experiments*]{}, Phys. Rev. C [**73**]{}, 035501 (2006); V. Gudkov, [ *Asymmetry of recoil protons in neutron beta decay*]{}, Phys. Rev. C [**77**]{}, 045502 (2008).
A. N. Ivanov, M. Pitschmann, and N. I. Troitskaya, [*Neutron beta decay as a laboratory for testing the Standard Model*]{}, Phys. Rev. D [**88**]{}, 073002 (2013); arXiv:1212.0332 \[hep–ph\].
A. N. Ivanov, R. Höllwieser, N. I. Troitskaya, M. Wellenzohn, and Ya. A. Berdnikov, [*Precision analysis of electron energy spectrum and angular distribution of neutron beta decay with polarized neutron and electron*]{}, Phys. Rev. C [**95**]{}, 055502 (2017); arXiv:1705.07330 \[hep-ph\].
A. N. Ivanov, R. Höllwieser, N. I. Troitskaya, M. Wellenzohn, and Ya. A. Berdnikov, [*Tests of the Standard Model in neutron beta decay with polarized neutron and electron, and unpolarized proton*]{}, Phys. Rev. C [**98**]{}, 035503 (2018); arXiv:1805.03880 \[hep-ph\].
A. N. Ivanov, R. Höllwieser, N. I. Troitskaya, M. Wellenzohn, and Ya. A. Berdnikov, [*Tests of the standard model in neutron beta decay with polarized electrons and unpolarized neutrons and protons*]{}, Phys. Rev. D [**99**]{}, 053004 (2019); arXiv:1811.04853 \[hep-ph\].
A. N. Ivanov, R. Höllwieser, N. I. Troitskaya, M. Wellenzohn, and Ya. A. Berdnikov, [*Radiative corrections of order $O(\alpha E_e/m_N)$ to Sirlin’s radiative corrections of order $O(\alpha/\pi)$ to neutron lifetime*]{}, Phys. Rev. D [**99**]{}, 093006 (2019); arXiv:1905.01178 \[hep-ph\].
M. Tanabashi [*et al.*]{} (Particle Data Group), Phys. Rev. D [**98**]{}, 030001 (2018).
R. P. Feynman and M. Gell–Mann, [*Theory of Fermi interaction*]{}, Phys. Rev.[**109**]{}, 193 (1958).
E. C. G. Sudarshan and R. E. Marshak, [*Chirality invariance and the universal Fermi interaction*]{}, Phys. Rev. [**109**]{}, 1860 (1958).
R. E. Marshak, S. Okubo, and G. Sudarshan, [ *V-A theory and the decay of the hyperon*]{}, Phys. Rev. [**113**]{}, 944 (1959).
Y. Nambu, [*Axial vector current conservation in weak interactions*]{}, Phys. Rev. Lett. [**4**]{}, 380 (1960).
R. E. Marshak, Riazuddin, and C. P. Ryan, in [*Theory of weak interactions in particle physics*]{}, Wiley-Interscience, A Division of John Wiley $\&$ Sons, Inc. New York, p. 41 (1969).
V. M. Shekhter, [*Hyperon beta decay*]{}, JETP [**8**]{}, 316 (1959). V. M. Shekhter, [*Beta decay of strange particles*]{}, JETP [**9**]{}, 920 (1959).
T. D. Lee and C. N. Yang, [ *Question of parity conservation in weak interactions*]{}, Phys. Rev. [**104**]{}, 254 (1956).
T. D. Lee, R. Oehme, and C. N. Yang, [*Remarks on possible noninvariance under time reversal and charge conjugation*]{}, Phys. Rev. [**106**]{}, 340 (1957).
J. D. Jackson, S. B. Treiman, and H. W. Wyld Jr., [*Possible tests of time reversal invariance in beta decay*]{}, Phys. Rev. [**106**]{}, 517 (1957).
M. E. Ebel and G. Feldman, [*Further remarks on Coulomb corrections in allowed beta transitions*]{}, Nucl. Phys. [ **4**]{}, 213 (1957).
P. Herczeg, [*Beta decay beyond the standard model*]{}, Progr. Part. Nucl. Phys. [**46**]{}, 413 (2001).
N. Severijns, M. Beck, and O. Naviliat-Cuncic, [*Tests of the standard electroweak model in beta decay*]{}, Rev. Mod. Phys. [**78**]{}, 991 (2006).
V. Cirigliano, J. Jenkins, and M. González-Alonso, [*Semileptonic decays of light quarks beyond the Standard Model*]{}, Nucl.Phys. B [**830**]{}, 95 (2010).
T. Bhattacharya, V. Cirigliano, S. D. Cohen , A. Filipuzzi, M. González-Alonso, M. L. Graesser, R. Gupta, and Huey-Wen Lin, [*Probing novel scalar and tensor interactions from (ultra)cold neutrons to the LHC*]{}, Phys. Rev. D [**85**]{}, 054512 (2012).
V. Cirigliano, M. Gonzáles-Alonso, and M. L. Graesser, [*Non-standard charged current interactions: beta decays versus the LHC*]{}, J. High Energy Phys. [**02**]{}, 046 (2013).
V. Cirigliano, S. Gardner, and B. Holstein, [*Beta decays and non-standard interactions in the LHC era*]{}, Prog. Part. Nucl. Phys. [**71**]{}, 93 (2013).
S. Gardner and C. Zhang, [*Sharpening low-energy, Standard-Model tests via correlation coefficients in neutron beta decay*]{}, Phys. Rev. Lett. [**86**]{}, 5666 (2001).
S. Gardner and B. Plaster, [ *Framework for maximum likelihood analysis of neutron beta decay observables to resolve the limits of the V - A law*]{}, Phys. Rev. C [**87**]{}, 065504 (2013).
M. González–Alonso, O. Naviliat–Cuncic, and N. Severijns, [*New physics searches in nuclear and neutron beta decay*]{}, Prog. Part. Nucl. Phys. [**104**]{}, 165 (2019); arXiv: 1803.08732 \[hep-ph\].
C. Itzykson and J.–B. Zuber, in [*Quantum field theory*]{}, McGraw–Hill Inc., New York, 1980.
T. D. Lee and C. N. Yang, [ *Charge Conjugation, a New Quantum Number $G$ , and Selection Rules Concerning a Nucleon Anti-nucleon System*]{}, Nuovo Cimento [**10**]{}, 749 (1956).
S. Weinberg, [*Charge symmetry of weak interactions*]{}, Phys. Rev. [**112**]{}, 1375 (1958).
A. N. Ivanov, [*Lorentz structure of vector part of matrix elements of transitions $n
\longleftrightarrow p$, caused by strong low–energy interactions and hypothesis of conservation of charged vector current*]{}, J. of Phys. G: Nucl. Part. Phys. [**45**]{}, 025004 (2018).
J. C. Hardy and I. S. Towner, [*Superallowed $0^+ \to 0^+$ nuclear beta decays: 2014 critical survey, with precise results for $V_{ud}$ and CKM unitarity*]{}, Phys. Rev. C [ **91**]{}, 025501 (2015).
H. Abele, [*Precision experiments with cold and ultra-cold neutrons*]{}, Hyperfine Interact. [**237**]{}, 155 (2016).
B. Märkisch, H. Mest, H. Saul, X. Wang, H. Abele, D. Dubbers, M. Klopf, A. Petoukhov, C. Roick, T. Soldner, and D. Werder, [*Measurement of the weak axial-vector coupling constant in the decay of free neutrons using a pulsed cold neutron beam*]{}, Phys. Rev. Lett. [**122**]{}, 242501 (2019); arXiv:1812.04666 \[nucl-ex\].
Ch.-Y. Seng, M. Gorchtein, H. H. Patel, and M. J. Ramsey-Musolf, [*Reduced hadronic uncertainty in the determination of $V_{ ud}$*]{}, Phys. Rev. Lett. [**121**]{}, 241804 (2018); arXiv:1807.10197 \[hep-ph\].
Ch.-Y. Seng, M. Gorchtein, and M. J. Ramsey-Musolf, [*Dispersive evaluation of the inner radiative correction in neutron and nuclear beta decay*]{}, Phys. Rev. D [**100**]{}, 013001 (2019); arXiv:1812.03352 \[nucl-th\].
R. W. Pattie Jr., K. P. Hickerson, and A. R. Young, [*Limits on tensor coupling from neutron beta decay*]{} Phys. Rev. C [**88**]{}, 048501 (2013); arXiv:1309.2499 \[nucl-th\], Erratum: Phys. Rev. C [**92**]{}, 069902 (2015).
A. N. Ivanov, R. Höllwieser, N. I. Troitskaya, M. Wellenzohn, and Ya. A. Berdnikov, [*Neutron dark matter decays and correlation coefficients of neutron beta decays*]{}, Nucl. Phys. B [**938**]{},114 (2019); arXiv:1808.09805 \[hep-ph\].
N. Kaiser, [*Isospin breaking in neutron beta decay and $SU(3)$ violation in semileptonic hyperon decays*]{}, Phys. Rev. C [**64**]{}, 028201 (2001).
D. R. Harrington, [*Lepton decays of hyperons*]{}, Phys. Rev. [**120**]{}, 1482 (1960).
V. P. Belov, B. S. Mingalev, and V. M. Shekhter, [*Possibility of determining the form factors in leptonic decay of hyperons*]{}, JETP [**11**]{}, 392 (1960).
L. Bender, V. Linke, and H. J. Rothe, [ *Leptonic decays of baryons*]{}, Z. für Phys. [**212**]{}, 190 (1968).
Ll. Armstrong, Jr., and C. W. Kim, [*Coulomb corrections in nuclear $\beta$ decay: elementary-particles treatment*]{}, Phys. Rev. C [**5**]{}, 672 (1972).
B. R. Holstein, [*Recoil effects in allowed beta decay: the elementary particle approach*]{}, Rev. Mod. Phys. [ **46**]{}, 789 (1974).
H. Behrens and W. Bühring, in [*Electron radial wave functions and nuclear beta-decay*]{}, Oxford University Press, New York 1982.
M. González-Alonso and J. M. Camalich, [*Isospin breaking in the nucleon mass and the sensitivity of beta decays to new physics*]{}, Phys. Rev. Lett. [**112**]{}, 042501 (2014); arXiv: 1309.4434 \[hep-ph\].
L. Hayen, N. Severijns, K. Bodek, D. Rozpedzik, and X. Mougeot, [*High precision analytical description of the allowed $\beta$ spectrum shape*]{}, Rev. Mod. Phys. [**90**]{}, 015008 (2018).
M. Fierz, [*Zur Fermischen Theorie des $\beta$-Zerfalls*]{}, Z. Physik [**104**]{}, 553 (1937).
S. L. Adler and Y. Dothan, [*Low-energy theorem for the weak axial-vector vertex*]{}, Phys. Rev. [**151**]{}, 1267 (1966); Erratum, Phys. Rev. [**164**]{}, 2062 (1967). L. Wolfenstein, [*Weak interactions of pions and muons*]{}, in [*High Energy Physics and Nuclear Structure*]{}, edited by S. Devons (Plenum, New York, 1970), p.661; High–energy physics and nuclear structure, Proceedings of the Third International Conference on High Energy Physics and Nuclear Structure; held at Columbia University, New York City, September 8 - 12, 1969.
V. Bernard, N. Kaiser, and Ulf-G. Meißner, [*Chiral dynamics in nucleons and nuclei*]{}, Int. J. Mod. Phys. E [**4**]{}, 193 (1995). V. Bernard, N. Kaiser, and Ulf-G. Meißner, [*Nucleon electroweak form-factors: Analysis of their spectral functions*]{}, Nucl. Phys. A [**611**]{}, 429 (1996). N. Kaiser, [*Induced pseudoscalar form-factor of the nucleon at two loop order in chiral perturbation theory*]{}, Phys. Rev. C [**67**]{}, 027002 (2003).
T. Gorringe and H. W. Fearing, [*Induced pseudoscalar coupling of the proton weak interaction*]{}, Rev. Mod. Phys. [**76**]{}, 31 (2004).
A. N. Ivanov, M. Pitschmann, N. I. Troitskaya, and Ya. A. Berdnikov, [*Bound-state $\beta^-$-decay of the neutron re-examined*]{}, Phys. Rev. C [**89**]{}, 05550 (2014).
A. Liesenfeld [*et al.*]{}, [*A measurement of the axial form factor of the nucleon by the $p(e,e'\pi^+)n$ reaction at $W = 1125$MeV*]{}, Phys. Lett. B [ **468**]{},20 (1999).
J. M. Blatt and V. F. Weisskopf, [*Theoretical nuclear physics*]{}, John Wily $\&$ Sons, New York 1952.
J. D. Jackson, S. B. Treiman, and H. W. Wyld Jr., [*Coulomb corrections in allowed beta transitions*]{}, Nucl. Phys. [**4**]{}, 206 (1957); [*Note on relativistic coulomb wave functions*]{}, Z. Phys. [**150**]{}, 640 (1958).
E. K. Konopinski, in [*The theory of beta radioactivity*]{}, Oxford at the Clarendon Press, 1966.
R. Pohl [*et al.*]{}, [*The size of the proton*]{}, Nature [**466**]{}, 213 (2010).
V. Khachatryan [*et al.*]{}, (CMS Collaboration), [*Search for new physics in the final states with a lepton and missing transverse energy at$ \sqrt{s}= 8\,{\rm TeV}$*]{}, CERN Report No. CMS-PAS-EXO-12-060 (2013).
M. González-Alonso and J. M. Camalich, [*Global effective-field-theory analysis of new-physics effects in (semi)leptonic kaon decays*]{}, J. High Energy Phys. [ **12**]{}, 052 (2016).
L. C. Biedenharn and M. E. Rose, [*Theory of angular correlations of nuclear radiation*]{}, Rev. Mod. Phys. [ **25**]{}, 729 (1953).
A. N. Ivanov, [*Status of the Free Neutron Decay, Wilkinson’s corrections and role of pseudoscalar interactions in neutron beta decays*]{}, Invited talk at International Workshop on “Current and Future Status of the First-Row CKM Unitarity”, held on 16 - 18 of May 2019 at Amherst Center of Fundamental Interactions, University of Massachusetts Amherst, USA; http://www.physics.umass.edu/acfi/seminars-and-workshops/current-and-future- status-of-the-first-row-ckm-unitarity.
| {
"pile_set_name": "ArXiv"
} | ArXiv |